Issue № 76 // What LLMs Get Wrong About Multiple Testing Correction

In The Orchestrator’s Edge I wrote:

When you’re no longer capable of tracing every line of AI-generated code back to your own reasoning, your ability to read, interrogate, and sanity-check what a machine produces becomes the last line of defense against silent compounding errors, which ultimately determines the validity of downstream scientific results.

The word downstream is key here. The most consequential errors in a computational biologist’s analysis pipeline aren’t the ones that crash their script. They’re the ones that embed incorrect methodologies into an analysis without raising red flags.

Consider something as routine as differential expression analysis on proteomic data. If you ask an LLM to write the code for you, it will. But what decisions did the model make along the way? Did it opt for a t-test when the data’s non-normal distribution called for a Wilcoxon rank-sum test? Did it apply a Perseus-style filter you didn’t want? And when it corrected for multiple comparisons, did it choose Bonferroni or Benjamini-Hochberg?

The computational biologist who can spot wrong turns—or better yet, who has enough methodological intuition to write a prompt that forecloses bad choices before they happen—is the one whose science will hold up best in the long run. There’s no shortcut to building this skill; you really do need to learn the ins and outs of your methods and when certain contexts elevate one over another. In our world that demands instant results, there’s the temptation to try and hack this process by asking AI to bang out some code using “gold-standard” analysis methods, bypassing the need to write detailed and informed prompts. Unfortunately, this isn’t a winning strategy.

In today’s piece, I’m going to focus on one consequential choice that comes up constantly when analyzing multiomic data: whether to use the Benjamini-Hochberg (BH) or Bonferroni procedure for multiple testing correction. Ask an LLM to write a differential expression analysis pipeline and it will almost always default to Benjamini-Hochberg. This almost certainly reflects the composition of its training data rather than a principled methodological choice—BH dominates the published bioinformatics literature, so models trained on that literature will reproduce the pattern. Nine times out of ten, that default is actually correct. But when it isn’t, the mistake can carry serious clinical repercussions. After reading this piece you’ll understand why, and how you can make more informed decisions in your own work.

Both the Benjamini-Hochberg method and Bonferroni correction are multiple testing procedures used to reduce the likelihood of type I errors (i.e., false positives), which is the risk of saying a finding significant when it is not. However, each method asks a slightly different question, and therefore involves a different tradeoff. Benjamini-Hochberg asks “what fraction of my hits are false positives, and is that fraction acceptable?” Bonferroni by contrast asks “am I certain that none of these results are false positives?” Keeping this distinction in mind will help you understand when one method is more appropriate than the other. But before we get into those differences, it’s worth discussing how running many statistical tests increases the risk of type I errors in the first place.

Suppose we want to run a Wilcoxon rank-sum test to determine whether the expression of AKT T308 differs between treatment responders and non-responders in triple negative breast cancer. Before running the analysis, we set an alpha value (α), which is the threshold used to determine whether our result is statistically significant. A common alpha value is 0.05, meaning that if our p-value is less than 0.05, we consider the result significant.

Any time we run a statistical test, the risk of a type I error equals alpha. So, in the example above, the probability that we incorrectly conclude AKT T308 expression differs between groups is 5%. Now consider that we’re rarely testing just one protein. Often we’ll compare the abundances of hundreds of proteins, if not more. When we run that many tests, the probability that at least one result is a false positive grows quickly, and can be quantified with the family-wise error rate formula 1−(1−α)ⁿ. Using this formula, we can see that with 1, 5, and 20 tests, the risk of at least one false positive is 5%, 23%, and 64%, respectively. By around 100 tests, the probability of at least one false positive approaches 100%, as you can see in the figure below. If we’re comparing hundreds of proteins or thousands of genes, false positives are essentially guaranteed.

The purpose of multiple testing correction is to reduce this risk of false positives as the number of tests increases. In practice, both the Benjamini-Hochberg and Bonferroni procedures accomplish this, but through different mechanisms and with different tradeoffs, as we’ll expand on below.

The Benjamini-Hochberg method controls the false discovery rate (FDR)—the expected proportion of false positives among all results we call significant. The procedure works as follows:

1. Rank all individual p-values in ascending order p(1) ≤ p(2) ≤ … ≤ p(m), where m in the number of statistical tests/comparisons we run.

2. Assign a rank i to to each ordered p-value, where i=1 is the smallest p-value and i=m is the largest.

3. Choose your acceptable FDR level (Q). A standard value is 0.05, meaning we tolerate up to 5% false positives among significant results.

4. Calculate the BH critical value corresponding to each p-value using the following formula: critical value = i/m * Q.

5. Find the largest p-value that is less than or equal to its corresponding critical value. That p-value and all smaller ones are declared significant.

As an example, suppose we’ve run 7 statistical tests and have the following ordered p-values: 0.01 ≤ 0.02 ≤ 0.04 ≤ 0.048 ≤ 0.05 ≤ 0.057 ≤ 0.06. Setting Q = 0.05, the BH critical values for ranks 1 through 7 are: 0.007, 0.014, 0.021, 0.029, 0.036, 0.043, 0.050. Comparing each p-value to its critical value, we find that none of the p-values are less than or equal to their corresponding critical value, so none of the results survive BH correction, despite several being below the naive p<0.05 threshold.

Using this method, you can see how the more tests you do, the harder it is for any given p-value to be significant after multiple testing correction. The chart below shows p-value rank plotted against critical thresholds. When m=2, the first (smallest) ordered p-value being significant at p<0.025, but as m increases we see this value getting smaller and smaller. By the time we reach m=6 we’re already at the point where p must be <0.01 to be significant.

The algorithm procedure above gives us a binary yes/no output for each tests—is it significant or not? The reason I introduced this first is that it helps build intuition for what BH correction is doing, but there is actually a simpler way to arrive at the same conclusion. One that produces a reportable adjusted p-value. To perform this method, we first rank order our p-values as we did previously, where p(1) ≤ p(2) ≤ … ≤ p(m) and m is the total number of tests. Then, for each of your p-values, use the following formula:

where i is the ordered ran and i=1 is the smallest p-value. However, after computing the adjusted p-value for all original p’s, we then apply a step-down forcing rule from the largest to smallest value to ensure the adjusted sequence never decreases as rank decreases. To definition for this rule as follows:

where you begin by applying the method to the largest p-value (called the base case) and work backwards to the smallest value. Using the same raw p-values as above, we now get the following results:

When the adjusted p-value (aka q-value) is smaller than our original α, then the adjusted p-value is significant.

The Bonferroni procedure controls the family-wise error rate (FWER)—the probability of making even a single false positive across all tests performed. Rather than tolerating a controlled fraction of false positives as BH does, Bonferroni aims to ensure that the entire set of conclusions contains no false positives at all1. There are two ways to do this procedure, the first being the decision gate method, which scaled down the target significance (α) value based on the number of tests. To do this, we use the formula:

We then reject the null hypothesis for any test whose raw p-value is ≤ α_new. For example, with 20 tests and α = 0.05, the corrected threshold drops to 0.0025. With 100 tests, it drops to 0.0005, as you can see demonstrated in the chart below. As the number of comparisons grows into the hundreds or thousands—as is routine in proteomics and genomics—only the very strongest signals will clear this threshold. This makes Bonferroni highly conservative: it substantially reduces the risk of any false positive, but at the cost of a higher type II error rate (i.e., missing real findings).

Bonferroni correction can also be expressed as an adjusted p-value, which allows direct comparison to your original alpha. For each raw p-value, the Bonferroni-adjusted p-value is:

So, if your original p-value is 0.02 and m=10, our new p-value is 0.2. We can then compare this to alpha, where we get 0.2 > 0.05, therefore the result is not significant. Working backwards, you’d need a raw p-value of 0.005 or smaller to achieve significance at α = 0.05 with 10 tests—a demanding standard that grows more demanding with every additional comparison.

Whether you choose BH or Bonferroni comes down to what kind of error you’re more willing to tolerate. BH maximizes discovery—it finds more true positives, but accepts that some fraction of your significant results will be false. Bonferroni by contrast minimizes the chance of any false positive, but increases the risk of missing real effects.

For most exploratory bioinformatics analyses, BH is the appropriate choice. Proteomics and transcriptomics studies are typically hypothesis-generating, meaning their findings inform downstream validation efforts, not immediate clinical action. In this context, a false positive that gets filtered out in the next experiment is a manageable cost, and the greater statistical power of BH ensures you’re not missing real biology. There’s also an interpretive buffer built into pathway-level analyses: if AKT, mTOR, 4EBP1, and eIF4E all appear significantly elevated in TNBC non-responders, a single false positive among them doesn’t materially change the conclusion that the pathway is activated.

Our reasoning changes though when results directly inform clinical decision making. Imagine a clinical proteomics study identifying biomarkers to stratify patients onto different treatment arms in a cancer trial. A false positive in that context—a protein that appears to predict response but doesn’t—could result in patients receiving an ineffective or harmful treatment. Here, the acceptable false positive rate is not 5% of your hits; ideally it’s zero. Bonferroni correction, with its strict family-wise error rate control, is the appropriate tool when the downstream consequence of a single false positive is clinical harm rather than a failed replication.

Practically, we can use a simple heuristic to determine which test is appropriate: use BH for discovery-stage analyses where findings will be subject to further validation, and default to Bonferroni when results will directly inform clinical or regulatory decisions. This is the type of methodological judgment you don’t want an LLM to make for you. It doesn’t know whether your analysis is exploratory or consequential, or whether your hits will be validated in a follow-up experiment or used to stratify patients in a trial. When you prompt an AI to write your analysis pipeline, you need to supply that context and to do that, you need to understand the choice well enough to know it matters. That’s the irreducible core of what it means to be a computational biologist in the age of AI-assisted science.

Principle component analysis is a notoriously confusing topic for beginners in computational biology and applied ML. While the purpose of PCA—summarizing and compressing high-dimensional data—is relatively easy to understand, technical explanations are often overloaded with linear algebra, making them unwieldy and overwhelming. Personally, even after I had a firm grip on the math underlying PCA, I had trouble understanding what PCA was really doing to my data from a biological perspective. For example, after compressing 1,000 protein’s abundances into 30 principle components, what was I left with? More importantly, how should I reason about these abstract features when my ultimate goal is to cluster or classify patients? That’s what this article is about—providing an intuitive take on principle component analysis for biologists.

Issue № 75 // An Intuitive Take on PCA For Biologists

This article grew out of a simple question: practically speaking, what’s the purpose of principle component analysis (PCA) in computational biology and bioinformatics? The academic, and wholly unsatisfying, answer to this question is that PCA finds the directions of maximum variance in your data, transforming a set of high-dimensional correlated variables—for example, a bunch of protein’s abundances—into a smaller set of uncorrelated features, or principle components. For this reason, PCA is considered a form of dimensionality reduction, which is a smart sounding way of saying it reduces the number of input variables (features) in a dataset while retaining as much relevant information as possible. However, because lower-variance features are discarded, PCA is lossy, meaning you can never perfectly reconstruct your original dataset after it’s performed. Instead, what you’re left with is a series of orthogonal principle components starting with PC1, which captures the highest variance, then PC2, PC3, and so forth (we’ll come back to this shortly).

As a practical example, suppose we have a dataset with 50 protein’s values, each measured in an arbitrary number of breast cancer patients. This feature set corresponds to a 50-dimensional space, which while small in the grand scheme of bioinformatics analyses is impossible to directly visualize or reason about nonetheless. To make this concrete, imagine you wanted to visualize your heart rate data (a 1-dimensional data point) following a trail run. This could easily be viewed with a line graph. Now, let’s say you want to see two data points in relation to one another—heart rate vs. pace—this can be visualized on an x-y axis (i.e., a plane). Three? An x-y-z coordinate system. Anything beyond three dimensions though and we’re lost—our brains lack the ability to perceive a fourth spatial dimension and beyond, making it impossible to form a mental image of it.

Now, what principle component analysis allows us to do is visualize high-dimensional data in a low dimensional space. For example, we can compress our 50-dimensional proteomic data into just a few ordered components that can be viewed in two dimensions (PC1 on the x-axis, PC2 on the y-axis). Additionally, after running this analysis we can plot the cumulative explained variance per principle component, which tells us how many PCs we need to retain the majority of information contained in the original dataset. Suppose we find that 20 components explains ~90% of the variance—this would allow us to work with just 20 features per subject instead of 50, which is helpful for a few reasons. First, many analyses downstream from PCA, such as k-means clustering, work better in lower dimensional space (see footnote #1 for a detailed explanation for when we should feed principle components into clustering algorithms and when we should use raw features instead)1. Second, PCA can tell you whether our data actually has a low-dimensional structure, despite existing in a high-dimensional space, or whether it’s variance is diffusely spread across features, suggesting a more dilute signal (as is often the case in biology)2.

To this point, I’ve primarily been speaking about PCA from an information-centric perspective. Understanding that is step 1. But, what really confused me about principle component analyses even after I understood the above is what the principle components really represent. When we compress 50 proteins abundaces’s into a small handful of uncorrelated components, how should we think about these things? Additionally, if a given subject in our study has high level of PC1 compared to other subjects, what does that actually mean?

Assuming you know some linear algebra, the easiest way to think about a principle component is as a specific weighted linear combination of all your original features. For easy math, let’s say we have 5 proteins in a dataset (labeled proteins a-e). Each principle component is a weighted linear combination of these features, which can be understood with the formula below where Zᵢ is the 𝑖-th principal component, wᵢⱼare the weights (aka loadings) for that component, and Xₐ..Xₑ are the original feature’s values.

When we run PCA, every protein in the dataset gets a unique loading value, or weight, for each principle component. This loading value indicates the degree to which that protein contributes to given component and in which direction (positive or negative). If a protein has a high positive loading for PC1, it means that protein’s expression level is positively correlated with the PC1 score. Therefore, if a patient has high expression (high abundance) of that protein, the PCA algorithm calculates a positive score for them, moving them to the positive end of the PC1 axis. On the flip side, a protein with a high negative loading for PC1 is negatively correlated with the PC1 score. As a result, is a patient high abundance of that protein, they will be pushed towards the negative (opposite) end of the PC1 axis.

Now, if we take a specific linear combination of proteins (i.e., a principle component) we can determine whether it’s biologically meaningful by extracting its loadings. In simple terms, this is the weight of each protein in our original dataset for that component. We can then rank order proteins by their absolute loading values (while still being mindful of their direction) and ask “are the top contributes enriched in a particular pathway?” For example, we may see that PI3K/AKT/mTOR pathway proteins have high positive loadings for PC1, and immune checkpoint proteins have high positive loadings for PC2. This would indicate that the activation of the PI3K/AKT/mTOR signaling axis and the strength of immune checkpoint signaling are the two largest sources of variation in our data. From this, we might hypothesize that patients in the bottom right hand corner of a PC1 vs. PC2 x-y plot have the worst survival outcomes (high proliferation, low immune engagement) whereas those in the top left corner have the best survival (low proliferation, high immune engagement).

Alternatively, we may see that PI3K/AKT/mTOR pathway proteins have high positive loadings for PC1, and immune checkpoint proteins have high negative loadings for PC2, which would suggest that high PI3K/AKT/mTOR signaling and immune engagement are inversely associated in our data. Interrogating these types of outputs—rather than treating PCA as a push button analysis to be performed before clustering—is what sets a great computational biologist apart from the rest3.

Issue № 74 // A Cluster Is Not A Category

The distinction between classification and clustering is typically introduced early in computational biology and applied ML courses to help differentiate supervised from unsupervised learning. As a general heuristic, classification predicts outcomes, such a therapeutic response, while clustering finds patterns in data. Understanding this distinction is important because these two approaches both require different inputs and answer different biological questions.

Classification requires training data with labeled examples—such as patient’s protein abundances and treatment outcomes—then learns a decision boundary which is applied to new samples (unseen during training), thus answering the question: “given what we know about these labeled examples, which category does this new sample belong to?” From this problem statement its clear that classification models are appropriate when categories are defined in advance and your goal is to assign new instances to these categories (for example, predicting whether a tumor will response to a drug or calling a variant pathogenic or benign).

Clustering by contrast asks, “what structure exists in our data that we didn’t specify in advance?” When a k-means clustering algorithm groups patients by their gene expression or protein abundance it’s finding underlying patterns without being told what to look for. As a result, clustering is most useful for discover—for example, identifying new candidate genomic subtypes in breast cancer, which can latter be clinically validated. However, we need to be careful with how we interpret clustering results, remembering that a cluster is not a category; it’s a region of high density a particular feature space, defined by a particular distance metric, found by a particular algorithm initialized with particular parameters. Changing any of those will change the clusters.

This distinction is particularly important for computational biologists, as bioinformatics analyses often blur the lines between clustering and classification. Take single-cell RNA-sequencing for example, where unsupervised clustering is used to define cell types and identified clusters are subsequently treated as labels for classification models, resulting in a circular argument. This isn’t always a problem—such as when cell-type clusters are validated by experts or orthogonal measurements prior to being used as labels—but it’s easy to make mistakes here, assuming clusters represent something “real” and concrete when they may just be statistical artifacts (I wrote about this previously in the post below).

Now, that we’ve covered the delineation between clustering and classification, I want to pivot to discussing how these methods can be used synergistically.

Suppose we want to build a classifier to predict the following categorical labels assigned to patients by a molecular oncologist: treatment responder, treatment non-responser, partial responser (stable cases). In this case, let’s assume we have a labeled dataset where index rows are samples (i.e., patients), columns 0-99 are z-score normalized protein abundance values, and the final column contains one of the three aforementioned category labels for each sample. Before we even consider training a classification model, it’s worth asking a prior question: do the labels assigned by the molecular oncologist actually capture real biological variation, or are they arbitrary clinical distinctions imposed on a continuous underlying distribution?

Note: This is exactly the type of question that differentiates someone who is good at ML in computational biology from someone who can run push-button analyses. The separator from good and great practitioners is rarely the code they write—it’s how they interrogate outputs and what they bring to that interrogation that no automated pipeline can replicate.

In order to answer this question, you can run k-means clustering (with the “optimal” number of clusters) on your z-score normalized proteomic features along without touching the category labels. Then, you can use Pandas’ cross-tab function to compare the resulting cluster assignments against patients known response categories. If responders predominantly group in one cluster with non-responders and partial-responders in another, the unsupervised learning model and clinical annotations are well aligned suggesting that the oncologist’s response categories are biologically “real” (i.e., they don’t represent arbitrary boundaries imposed by the human mind).

If on the other hand, responders, non-responders, and partial-responders are distributed across clusters (co-mingling with one another), we have to entertain one of two possibilities: the proteomic features you’ve measured don’t capture the relevant biology or the response categories don’t reflect clean molecular distinctions (or both). In either case, training a classifier would be a waste of time. As a result, we can consider clustering and cross-tabulating to be a sanity check on whether supervised learning (classification) is even warranted.

Assuming our cross-tabulation results look good (i.e., clusters recapitulate classifications), we’re in a good position to build a classifier that can generalize to new, unseen, samples (if you’re interested in learning the details about this process check out How to Develop Predictive Biomarkers). To pull this off we’d take our labeled data—protein features as input, response category as the target—then train a model to learn the mapping that represents a linear or non-linear combination of proteins that best predict treatment response from our labeled examples (training data).

The ability to generalize to new unseen data is one of the qualities that sets classification apart from clustering. A trained classifier learns a decision boundary that can be applied to patients it has never encountered. When we add a new patient’s baseline protein abundance data through a k-means model on the other hand, we’ll get a cluster assignment—importantly, one based on proximity to centroids that shifts when we add new data or alter the distance metric. Clinically, this is the difference between a usable tool and a novel research finding.

In summary, clustering can be used to validate whether label assignments in your data are consistent with biological reality. Classification, by contrast, can be used to build a model that generalizes to new unseen data. The two methods reinforce each other and work synergistically. However, used interchangeably or without understanding what question each one answers, they produce findings that are nonsensical, or misleading.

Thanks for reading! If you found this post useful, please consider subscribing. I share hands-on computational biology techniques, fresh ways to think about tough problems, and perspectives on a range of related topics. All free, straight to your inbox:

Subscribe now

Issue № 73 // The Orchestrator’s Edge

As a dry lab scientist, I occasionally worry about AI’s impact on job availability or what it will mean to be a researcher in a world where machines begin to make scientific discoveries that humans can’t understand.1 Despite being skeptical of AI companies’ strategic messaging—which I previously wrote about in how to win markets and influence policy—I still find myself thinking "what if the pool of well paying computational biology jobs dries up?" or "what if I spent years earning a PhD in a field that’s soon to be consumed by agentic AI?"

Yet, ironically, on my best days—the ones where I’m most engaged with my work—I use AI to debug complex programming scripts, help refactor spaghetti code, and summarize documentation for unfamiliar Python libraries. Inevitably this leads to cognitive dissonance over using the very system whose negative externalities I’m worried about to augment my own capabilities. It’s not dissimilar to professional cyclists who rail against PEDs, then use them anyway because everyone else does—reinforcing their necessity and entrenching their influence.

I suspect I’m not the only one with mixed feelings about AI’s role in science and society, yet I’ve seldom seen this perspective reflected in a way that resonates for me. This is my motivation for writing this piece—not to shape the narrative around AI in computational science or influence your opinion, but to clarify my own thinking while grappling with these ideas. As a result, you can expect this piece to move between the personal and the practical—from my current perspective to where I see things heading, and from my thoughts on future-proofing my role as a scientist to the skills I’m actively trying to develop.

When Perfect Code Produces Imperfect Science Redux

In when perfect code produces imperfect science I hypothesized that the most successful bioinformaticians in the next few years will be the ones who understand computation deeply enough to decompose complex biological questions into analytical components, orchestrate AI tools to implement their vision, and maintain the skepticism to question whether their code truly answers their biological questions.2 What I was trying to articulate in this piece was that rather than AI replacing computational biologists wholesale, I believe it will instead push us to orchestrate analyses at higher levels of abstraction. Eleven months later, the New York Times Magazine published an article that came to the same conclusion—Coding After Coders: The End of Computer Programming as We Know It—where the author Clive Thompson writes:

"With A.I., though, programmers ascend to an even higher level of abstraction. They describe, in regular language, what the program should do, and the agents translate that idea — that human intent — into code. Writing software no longer means mentally juggling the nuances of a language like Python, say, or JavaScript or Rust. Coding no longer involves messing up an algorithm and then trying to figure out where your error lies. That part, too, has been abstracted away."

Taking it a step further, AI isn’t just pushing programmers to work at higher levels of abstraction—it’s changing the selection criteria for what makes a good programmer in general, especially in fields like computational biology that require significant domain expertise in addition to scientific computing skills. In a recent piece titled what actually makes someone good at ML in computational biology?, I expanded on these ideas, discussing how the separator between good and great ML practitioners in comp bio isn’t the code they write, but how they interrogate its output and what they bring to that interrogation that no automated pipeline can replicate. Then, in why interpretable models outperform black boxes in biology I wrote:

"Accuracy, in a sense, is a commodity that can be bought. Given enough money, you can buy your way to a better model—more high quality training data, more compute, larger teams, and greater access to proprietary datasets. This gives early movers a huge competitive advantage, but given enough resources a well-funded competitor can close almost any accuracy gap. Interpretability, on the other hand, doesn’t work that way—it can’t be purchased at scale. Instead, it requires deep domain expertise, biological intuition, and the kind of hard-won insight that you can’t list on a procurement order—it’s a bit of a dark art, and that’s what makes it defensible."

These pieces both point to the same idea—that the value of computational biologists isn’t shrinking, but shifting. The skills that once set a computational biologist apart—like writing pipelines, automating workflows, keeping a lab’s analysis infrastructure running—are no longer sufficient on their own. Instead, we’re living in a world that places high value on combining computational thinking in a broad sense with biological knowledge and mechanistic understanding.3

This shift from implementation to orchestration may suggest to some that learning to code is a waste of time in 2026, and in years to come. I’d argue the opposite: coding literacy has never mattered more, even when AI is capable of writing most code for major projects. When you’re no longer capable of tracing every line of AI-generated code back to your own reasoning, your ability to read, interrogate, and sanity-check what a machine produces becomes the last line of defense against silent compounding errors, which ultimately determines the validity of downstream scientific results.

The case for learning to code isn’t, then, to compete with generative AI. It’s to stay fluent enough to know when you can trust it. And yet, for a growing number of students and early-career researchers that case is increasingly difficult to make.

The Psychology of Learning to Code In the Age of AI

Learning to code has never been more accessible than it is right now—the resources are better, more abundant, and more personalized than ever. Yet, for many, the motivation to learn gets harder to muster by the month. This must be how medical residents training to become radiologists felt ten years ago, after Geoffrey Hinton stated the following:

"If you work as a radiologist, you’re like the coyote that’s already over the edge of the cliff but hasn’t yet looked down, so he doesn’t realize there’s no ground underneath him. I think we should stop training radiologists now. It’s just completely obvious that within five years, deep learning is going to do better than radiologists… we’ve got plenty of radiologists already"

For residents in the middle of their training, that must have been a destabilizing thing to hear: years of investment, a clear career trajectory, suddenly reframed as a sunk cost. As it turned out, Hinton was wrong. Radiologists are in higher demand than ever, with average US salaries reaching $673k as of 2026.4 But knowing that now doesn’t tell you what it felt like to be a radiology trainee then—absorbing that verdict mid-training, with no way to know whether the expert making it was right.

That feeling has a contemporary equivalent. I recently spoke to a fellow PhD student who described a feeling of nihilism about their technical skill development—they want to improve, but easily become demotivated when they imagine AI always outpacing them. What troubled them wasn’t just the capability gap, but something harder to articulate: that orchestrating AI to execute tasks made their own involvement feel interchangeable. "If I can prompt Claude Code to do this," they said, "so can anyone else. What am I actually contributing?"

I’ve grappled with the same thought. Like the cyclist using PEDs, I simultaneously feel it’s necessary to use AI to keep pace with colleagues and their expectations, and that doing so diminishes the value of my own contributions. Yet, this obscures what’s really happening. I still need to make the key decisions during a project, whether using an AI co-pilot or not: what questions to ask, how to frame the analyses, what the results mean biologically, and how to situate findings within the broader literature. These are the intellectual contributions that matter. Code generation is a tool, the same way statistical software is a tool. What you cannot outsource is the scientific reasoning.

Knowing this is one thing. Believing it is another. It’s easy to feel effective, but fraudulent, when using AI to augment your workflow. I’d argue that feeling has two distinct sources. First, we sometimes use AI to rapidly prototype code, and end up using tools or outputs we don’t understand well enough to audit—this is knowledge debt, and it’s a real problem that needs to be actively managed. The second source is a psychological artifact of the AI era—the sense that because AI is doing the heavy lifting, you’ve become a passenger rather than a contributor, that your skills no longer count. This is a valid feeling, but it’s worth reframing.

Consider a film director who relies on a cinematographer to shoot the footage for their movie. Nobody questions whose vision the film represents, or whether the director’s creative contribution is diminished by not having operated the camera themselves. The judgment about what story to tell, how to frame it, and what it means is still entirely the director’s. Similarly, the fact that AI can execute code faster doesn’t change who is responsible for the question and the interpretation.

The knowledge debt problem and the psychological artifact are easy to conflate because they produce similar surface symptoms—both make you feel like you don’t fully own your work. But their origins are different. Knowledge debt is epistemic: there are things you genuinely don’t understand about what your pipeline is doing, and that gap has real consequences for scientific validity. The psychological artifact is emotional: you understand the work, you made the key decisions, but the feeling of authorship hasn’t caught up with the reality of how the work was produced. Treating one as the other makes both worse—auditing code you already understand doesn’t resolve the feeling of being a passenger, and reframing your mindset doesn’t close a genuine gap in your understanding of what your model did.

Underneath all of this sits a harder question though. If AI gets better not just at writing code but at reasoning through problems, what exactly is left for the scientist who is still early in their career, still building the intuition that’s supposed to be irreplaceable?

The Role of the Junior Scientist

In the previous section I argued that code generation is a tool in the same way that statistical software is a tool, and that scientific reasoning is both what matters most and the skill that cannot be outsourced. But this raises a harder question: what happens if AI advances to the point where it can genuinely reason through problems—not pattern-match at scale, but formulate hypotheses, design experiments, and interpret results the way a scientist does? The idea that AI is a bicycle for the mind is a compelling one. But if the bicycle eventually thinks for itself, what remains for the human rider?

The question of whether reasoning can be outsourced to machines is a debate. It’s also the thesis behind multiple companies developing "AI scientists". I don’t know if this will ever be possible in the way it’s advertised. I’m also not sure it matters. Take the following example: you bring Einstein into your company and ask him to make it better. In order to direct him usefully, you need to understand the problem well enough to know what you’re asking him to solve. The same logic applies to AI. Telling it to "do research" produces nothing of value without specificity. You have to know what problem is worth solving, roughly how it should be approached, and how the work makes contact with the real world. AI may find a better implementation than what you had in mind, but you provide the scaffold that allows it to realize its potential. This points to the following conclusion: the orchestrator role isn’t a consolation prize—it’s where the real intellectual work lives.

If the orchestrator model is where things are headed, a reasonable concern follows: what happens to junior scientists? If a senior computational biologist can just orchestrate AIs and architect a plan, what need is there for a junior scientist? This is a legitimate problem, though I also think it overlooks the complexity of actually running companies and all of the different odds and ends you can have people work on.

While there may be a dearth of jobs for junior scientists in established companies in the short term, I see their career prospects increasing with time. Here’s why: historically, a single computational biologist or data scientist at a small company was rarely cost-effective. The value these roles generate typically required teams, which confined them to large pharma or well-funded biotechs. But if one scientist with AI tools can produce what previously required three to five, the economics change for a much broader set of employers. Small companies that could never justify building out a computational team may now be able to justify hiring one strong person. The total number of companies that can deploy this kind of expertise has expanded—and with it, potentially, the total number of roles available, even if the headcount per company remains low.

I’ve seen this play out myself. Over the past decade, I’ve had contact points in many companies and have seen ways that just a little bit of data science expertise can take them from good to great, but there was never an ROI to building out the team to make that doable—conversations with founders have told me as such. For a concrete example, take a hardware product like an EMS device and imagine how integrating biosensor feedback for closed-loop modulation can make it way better. For example, utilizing real-time muscle oxygenation data to adjust the stimulus magnitude and frequency. The value proposition is there, but before now making that happen was very expensive. I’ve had some of these same companies reach out to me in recent months asking if I can help them with these same types of problems. Before, some of these projects would have required at least a team of three people and a year of dedicated effort, often at a minimum cost of exceeding $225k. Now, I can run one of these projects end to end by myself in a few months, drastically reducing costs and making it a real option.

A lot of companies can benefit from this type of work. A junior computational biologist making $75k per year can bring in more value than they cost, especially for smaller companies. Additionally, the number of more established companies that can justify 2-3 senior scientists have multiplied, creating more jobs in total. That said, there’s a devil’s advocate case worth taking seriously. The idea that small-to-mid sized companies will absorb displaced talent assumes those companies can identify the need, afford even one person, and that the value is legible to non-technical founders. In practice, demand doesn’t materialize automatically—a lot of small companies don’t know what they don’t know, and they won’t go looking for a computational biologist or data scientist if they’ve never had one. The demand isn’t automatic; it has to be created or surfaced.

This argument is fair, and it points to something important: as the comp biologist, you need to be diligent about networking, writing, and finding these opportunities. Eventually, the market will sort this out once the concept is proven, but in the short term you need to make these opportunities for yourself. I’ve given this advice to junior scientists before—they may be the best in the world, but no one is going to know that if they don’t prove it. Writing increases your surface area for luck. This also ties into the idea of making sure you know where your paycheck comes from, which Sean Goedecke articled well in an article titled Knowing where your engineer salary comes from (i.e., if you cannot clearly articulate your value, no-one else will understand it either).

How to Uniquely Add Value in the Age of AI

Given everything above, what does a concrete training regimen actually look like? The following is what I’d focus on—not as an exhaustive list, but as the highest-leverage investments for a computational biologist trying to build durable value in an AI-augmented field.

Practice the scaffold, not the implementation. The step AI is worst at is also the most important one: taking a vague biological question and decomposing it into a tractable computational problem. This is where projects are won or lost before a single line of code is written. Spend deliberate time here—not just asking "how do I analyze this data" but "what would a convincing answer to this question actually look like, and what would falsify it?" That kind of problem formulation is a skill, and like most skills it improves with intentional practice.

Build cross-domain pattern recognition. Reading broadly isn’t just intellectually satisfying—it’s also strategically useful. Complexity science, ecology, economics, and network theory all share structural similarities with problems in computational biology. The scientist who recognizes that a gene regulatory network behaves like a market will generate hypotheses that an AI trained on biology literature alone will not. This has become something I’d consider an advantage in my own work. Operating across precision oncology, biodefense, and human performance can look like a liability from the outside—the jack of all trades, master of none. In practice, I’m constantly importing ideas from one discipline into another, and the cross-pollination produces things that a narrower focus wouldn’t.

Get good at scoping and saying no. In the embedded scientist model, a significant portion of your value comes from knowing which projects will hit a dead end before the team invests six months finding out. This is one of the most valuable skills I developed in my time at NNOXX. There were often app features or use cases that seemed interesting, and even viable, superficially, but upon working on them we quickly realized they couldn’t be reduced to practice. Over time, it became easier and easier to spot these things before wasting days or weeks working on them. This skill is learnable, but mostly through deliberate reflection on past projects rather than doing more projects faster, which is ironically what AI encourages. Build the habit of asking, after every project, where the dead end was and how early the signal was visible.

Develop a point of view. The scientists who will matter most in an AI-augmented world are the ones with strong, defensible opinions about what’s biologically true and why. The person who can say "this result is almost certainly an artifact because of X biological reason" and be right is providing a something scare. Cultivate opinions. Be willing to be wrong in specific, correctable ways rather than vague, unfalsifiable ones.

Learn enough about how the models work to know when not to trust them. Not AI-research-level depth, but enough to understand where LLMs hallucinate confidently, where they fail on compositional reasoning, and where their training data is thin. This is what separates someone who uses AI well from someone who inadvertently introduces AI errors into published science.

A concrete example from my own work illustrates why this matters. AI is consistently poor at interpreting phosphoprotein data. If your data shows EGFR T654 is upregulated in a breast cancer sample, a model will generate a perfectly coherent explanation about EGFR hyperactivation driving downstream signaling—except EGFR T654 is an inhibitory phospho site. The same failure mode appears with PTEN S380. When a model encounters PTEN, it defaults to PI3K inhibition and reduced AKT signaling. It doesn’t recognize that elevated PTEN S380 may indicate the opposite. The reason is straightforward—most papers discuss proteins without phosphorylation-site context, so the model has no basis for the distinction. The output sounds like expertise. It isn’t. This is the kind of failure that domain knowledge catches and automated pipelines don’t, and it’s a good illustration of why fluency in the underlying biology remains non-negotiable even as the code becomes increasingly optional.

Closing Thoughts

I started this piece admitting to cognitive dissonance—using the very system I’m conflicted about to do my best work, feeling simultaneously that it’s necessary and that it diminishes something. I haven’t resolved this tension. What I’ve come to instead is something more like a working position. AI is a force multiplier, which means it amplifies what you bring to it. If you bring shallow domain knowledge and poor problem formulation, it will help you produce shallow work faster. If you bring genuine biological intuition, hard-won judgment about what questions are worth asking, and the ability to recognize when an output is wrong it will extend your reach in ways that weren’t possible before.

The scientists who flourish in this environment won’t be the ones who resisted AI or the ones who outsourced their thinking to it. They’ll be the ones who stayed fluent enough to direct it, skeptical enough to audit it, and grounded enough in the real world to know what it still can’t do. That’s a narrower target than it sounds, and it requires deliberate effort to hit. But it’s also, I think, a genuinely exciting place to be building from.

**A Practical Addendum: On Becoming the Go-To Person for Companies That Don’t Know They Need You

One practical addendum worth making explicit because it’s the piece most scientists skip entirely. Earlier I argued that demand for computational talent won’t materialize on its own—it has to be created. Here’s what that actually looks like in practice.

The first thing to recognize is that this is a visibility and translation problem, not a credentials problem. The companies you’re targeting can’t search for what they don’t know exists. A founder running a hardware company with a biosensor problem isn’t browsing LinkedIn for computational biologists—they don’t have the vocabulary for what they need. Your job is to show up in their world, speaking their language, before they know to look for you.

That starts with developing a sharp, non-technical articulation of what you do. Not "computational biology and machine learning" but something closer to “I help companies that have biological data or biological products figure out what that data is actually telling them, and build systems around it.” The more precisely you can describe the business problem you solve rather than the technical method you use to solve it the more recognizable you become to founders who have the problem but not the vocabulary. This is harder than it sounds. Most scientists are trained to lead with methods. Inverting that instinct takes practice.

The next step is creating artifacts that demonstrate value before any conversation happens. Writing publicly—a newsletter, a LinkedIn post, a project summary with confidential details obscured—does more than any credential. A founder who reads "I helped a hardware company integrate biosensor feedback for closed-loop modulation, cutting what would have been a $225K team effort down to a few weeks of one person’s time" will share that with every founder they know with a similar problem. Concrete, specific, outcome-oriented writing is the highest-leverage marketing available to someone in this position, and few people do it well without practice.

Then get into the rooms where founders talk to each other. Accelerators, industry-specific conferences, founder communities—not as a vendor, but as a participant with something to contribute to the conversation. The moment you solve one problem for one company inside a network, word travels in ways that no amount of cold outreach replicates. Most of the relationships that actually lead somewhere don’t start from a pitch—they start from a warm referral from someone who saw the value firsthand.

The timeline here is long, which makes it easy to deprioritize. It shouldn’t be. I can trace my role as co-founder at NNOXX in 2020 directly to an event at Metier coffee in Seattle in 2015. Several of the consulting projects I’m working on now are direct offshoots of conversations I had at in-person events two or three years ago. The payoff isn’t linear and it isn’t fast, but it compounds—and it’s extraordinarily difficult to replicate through any shortcut.

Finally, make it easy for non-technical people to refer you. Most of your best leads will come from people who don’t fully understand what you do but know you’re useful. Give them a sentence they can repeat without distorting it. “She’s the person you call when you have data and don’t know what it’s telling you” travels further than an accurate technical description, because it’s repeatable.

As always, I’d like to hear what other scientists in this space are thinking. Where does this resonate, and where does it miss? Feel free to reach out.

Thanks for reading! If you found this post useful, please consider subscribing. I share hands-on computational biology techniques, fresh ways to think about tough problems, and perspectives on a range of related topics. All free, straight to your inbox.

Subscribe now

Issue № 72 // Building A Statistical Pipeline For Group Comparisons

I recently received a question about whether ANOVA is the best catch-all test for detecting differences between groups. The student asking this question wanted to build a pipeline with simple logic—the idea being that if ANOVA is the most versatile option, you could reduce code complexity and avoid needing fallbacks for situations like two-group comparisons or non-parametric data. The question also touched on whether you could bypass ANOVA entirely and build a pipeline around post hoc testing to simplify things further. This is the second time I’ve gotten essentially the same question, so I’m addressing it here for future reference.

Importantly, this is less of a pure statistical question and more of a practical one—it’s asking how firm the rules underlying these statistical tests are, and how far we can bend them before they break. My response is as follows:

An ANOVA compares the means of three or more groups to determine if they are significantly different from one another, with the assumption that the data is normally distributed and has an ~equal variance. Practically speaking though, ANOVAs are mildly robust to variations in normality, especially with large sample sizes which is often the case in bioinformatics analyses. However, when data really isn’t normally distributed and you want to compare three or more groups, a Kruskal-Wallis H-test can be used in place of ANOVA. Notably, this test analyzes ranks instead of means, and is therefore robust to non-normal distributions. This allows us to start with some simple logic when building an analysis pipeline:

Comparing 3+ groups?
├── Normally distributed → ANOVA
└── Not normally distributed → Kruskal-Wallis H-test

Suppose we’re working in a scenario where the data is always normally distributed can we then use ANOVA as a catch-all? After all, an ANOVA could work with just two groups, which would make the code much simpler, so what’s the problem?

Technically an ANOVA can be used to compare the means of two groups, but it’s not ideal. Assuming data is normally distributed, a t-test would be the best bet in this scenario as the calculation is more straightforward (which itself isn’t a great reason, imo) and its statistical power—the ability to detect true differences—is higher1. Alternatively, given two groups and non-normal data, a Wilcoxon rank sum test is a good choice (this is often an ideal test when comparing protein abundance between two groups, as this data is seldom normally distributed in my experience). Now, we have even further branching logic:

Comparing 2 groups?
├── Normally distributed → t-test
└── Not normally distributed → Wilcoxon rank-sum test

Comparing 3+ groups?
├── Normally distributed → ANOVA
└── Not normally distributed → Kruskal-Wallis H-test

Going back to ANOVA, this time addressing the portion of the question about post-hoc testing and whether we can bypass ANOVA entirely. If an ANOVA detects a significant difference between group means, post-hoc testing should be performed—the reason for this is that an ANOVA is an omnibus test, meaning it confirms that significant differences exist somewhere (this is also true of the Kruskal-Wallis H-test). But, it doesn’t give you specifics. For example, it may say there is a difference in fragment length ratios between groups A, B, C, and D, but we don’t know which of those group means actually differ. Figuring this out is the purpose of post-hoc testing. For simplicity, there are two easy options for post-hoc testing: Tukey’s Honest Significant Difference (HSD) test, and pairwise t-tests with Benjamini-Hochberg (BH) correction to control for the false discovery rate.

This raises the question though, why bother doing the ANOVA to begin with if we can just jump straight to post-hoc testing? The primary reason for this is that we want to control the type I error rate, which is the chance of making a false positive (ie, saying a difference exists between group means when none exists). If you just run a bunch of pairwise t-tests, for example, you increase your chances of falsely finding a significant difference between groups2. Aside from that, there are also cases where an ANOVA can detect significant differences that a post hoc test, like pairwise t-tests of Tukey’s HSD cannot. In these cases, the overall group variation may be statistically significant, but specific pairwise differences are too small or noisy to meet the strict thresholds that post-hoc tests apply to control type I errors (thus making it harder to achieve significance). Finally, if you’re using Tukey’s HSD for post hoc testing specifically, ANOVA is a prerequisite assumption. In other words, Tukey’s HSD assumes the omnibus test’s results were significant and running it after a non-significant ANOVA (or without an ANOVA at all) violates the procedure. This is both considered methodologically unsound and is considered “fishing” for results, which is ill advised.

So, from a pipeline standpoint, our logic is now increasingly complex:

Comparing 2 groups?
├── Normally distributed → t-test
└── Not normally distributed → Wilcoxon rank-sum test

Comparing 3+ groups?
├── Normally distributed → ANOVA
│   └── Significant → post hoc test (Tukey's HSD or pairwise t-test w/ BH correction)
└── Not normally distributed → Kruskal-Wallis H-test
    └── Significant → Dunn's test

Whether you choose to use a pairwise t-test with BH correction or Tukey’s HSD after ANOVA comes down to a number of factors, part of which is personal preference. Personally, I find Tukey’s HSD more useful and easier to implement as it compares all possible pairs of group means and controls for the family-wise error rate (the probability of committing one or more Type I errors when performing multiple statistical tests within a single study) making it an “exact” test for determining which specific groups are different. At the same time Tukey’s HSD also limits type II errors (false negatives) by maintaining high statistical power.

However, there may be cases where you don’t care to compare all groups means, and may only want to perform a small number of predetermined tests, especially if you have a lot of groups. This scenario tends to be less common in my own work, but in these cases a paired t-test with BH correction is favorable, as it’s statistical power is higher, and there is a low risk of type I errors due to multiple comparisons (again, because only a small number of planned comparisons are being performed).

Now, because we’re biologists, we don’t just want to know if the means between groups are statistically significant. We want to know if those differences are meaningful. For example, we run an ANOVA and find a significant difference between the mean three groups, A, B, and C. We then use Tukey’s HSD and find significant differences between the means of each pairwise comparison (A/B, A/C, B/C). To figure out if these differences and practically meaningful, and not just statistically significant, we want to perform an additional test such as Cohen’s d, which measures the size of the difference between two means in units of standard deviation (ie, the effect size). For example, imagine a scenario where p<0.05, but Cohen’s d is small— this would indicate that the result is statistically significant, but practically insignificant or weak. On the flip side, a p>0.05, but large Cohen’s d could mean that your experiment found a practically meaningful difference between groups, but it was not statistically significant, potentially indicating that the study was underpowered.

Interpreting Cohen’s d can be a little tricky, but a good rule of thumb is as follows:

• Small effect = ~|0.2|,

• Small-to-medium effect = ~|>0.2 - <0.5|,

• Medium effect = ~|0.5|,

• Medium to large effect = ~|>0.5 - ,0.8|,

• Large effect is >~|0.8|, and

• Very large effect >~|1.5|.

Note that we use the |absolute| values to determine the effect size. The specific sign of the value just tells us which group is larger. For example, a positive Cohen’s d indicates the control group’s mean is lower, and a negative Cohen’s d tells us the mean of the experimental group is lower. So, for the pipeline this now gives us:

Comparing 2 groups?
├── Normally distributed → t-test
└── Not normally distributed → Wilcoxon rank-sum test

Comparing 3+ groups?
├── Normally distributed → ANOVA
│   └── Significant → post hoc (Tukey's HSD or pairwise t-test w/ BH)
│       └── Effective size → Cohen's d
└── Not normally distributed → Kruskal-Wallis H-test
    └── Significant → Dunn's test
        └── Effect dize → Cohen's d

Importantly, Cohen’s benchmarks were based on behavioral sciences and do not transfer to all fields; what is “large” in psychology may be considered small in genomics, or vice versa3. As a result, we can’t assume that a small p-value and large Cohen’s d means that something is biologically meaningful. To do that, we’d need to understand what amount of change matters in our system being studied (ideally, we should also pre-specify this ahead of time, even if it’s informally in our own lab notebook—this practice is underused in company biology and more valuable than most people appreciate). For example, even moderate differences in phospho-protein abundance have large biological repercussions, but the same can’t be said of all molecules or biological systems. Thus, this isn’t something we can bake into our code and pipeline. It’s something we need as end-users of the tools we build, which speaks to the same issues I discussed in Issue #69 // What Actually Makes Someone Good at ML in Computational Biology? where I wrote:

"In my opinion, the most important skill for a bioinformatics scientist or computational biologist isn’t their ability to code or do advanced math and statistics. It’s their ability to contextualize the outputs of their code and extract meaningful biological insights."

This also relates to something I discussed in Issue #40: When Perfect Code Produces Imperfect Science, which is that coding is perhaps the simplest part of a bioinformatics scientist’s job. The real challenge is understanding the vast networks of connections—how data flows through analytical pipelines, why certain analyses or have to precede others, and which statistical assumptions underpin different methods (as we covered in this piece). All of these interdependencies make bioinformatics projects less like building a house, and more like baking a cake.

When you build a house specifications translate directly to outcomes. If you add an extra wall the finished product is just the house with an extra wall. When you bake a cake, on the other hand, small variations in technique or timing can produce dramatically different results. Add a little extra butter and the result may not just be an extra buttery cake—instead, it could be an oily blob. Similarly, when performing bioinformatics analyses small details make all the difference—choose Bonferroni correction over Benjamin Hochberg, use a t-test over a Wilcoxon sum rank test, or select a different normalization approach and your biological conclusions can transform entirely, often for the worst. These aren’t just technical choices but scientific judgments requiring deep understanding of both computational methods and biological systems.

Thanks for reading! If you found this post useful, please consider subscribing. I share hands-on computational biology techniques, fresh ways to think about tough problems, and perspectives on a range of related topics. All free, straight to your inbox.

Subscribe now

Issue № 71 // Deeper Signal

For the better part of the last five years, I’ve been predicting that the wearable technology industry would eventually abandon its fixation with single-point sensing—the use of a single localized device to monitor physical, physiological, or environmental data from one specific location on the body.

The smart watch, wristband, and smart ring-based form factors aren’t ubiquitous in continuous health monitoring because the wrist and finger are the best places to measure human physiology. They’re ubiquitous because they’re convenient and fashion forward—the epitome of form-first wearable design. Convenience has a way of calcifying into convention though, and for years the dominant players in the space—Whoop, Oura, Apple—showed little appetite for disrupting their own product lines.

Then, just a few days ago, I saw a patent by Whoop and a trademark filing by Garmin—both pointing in the same direction—that changed my read on where the industry may be heading. To understand why, it helps to have context about a biometric that most people in the consumer health space have never heard of: muscle oxygenation (SmO2). As I’ve written previously in Dampening the Noise: Making Sense of Variability In Biometric Measurements, SmO2 is best understood as the balance of oxygen supply and demand in skeletal muscle.

The basic principle behind measuring muscle oxygenation is fairly simple. Near-infrared light penetrates biological tissue with relative ease, and when directed at muscle, some of those photons are absorbed by hemoglobin and myoglobin—the oxygen-carrying proteins in blood and muscle cells respectively—while the rest scatter back to a detector on the skin surface. Oxygenated and deoxygenated forms of these proteins absorb light differently across wavelengths, so by measuring how much light returns at each wavelength, you can infer the ratio of oxygenated to deoxygenated hemoglobin in the tissue below the sensor.

This is different from what your Apple Watch, Whoop, or Oura ring is doing when it estimates blood oxygen (SpO2). Those devices use pulse oximetry—a technique that isolates the pulsatile arterial signal in the capillary beds near the skin’s surface (1-2mm deep), giving you a read on how well oxygenated your blood is. Muscle oxygenation, by contrast, is measured much deeper (~20mm), past the skin and subcutaneous fat, reaching the muscle itself. Where pulse oximetry tells you about oxygen in arterial blood, mNIRS tells you about oxygen in the tissue’s microvasculature—a meaningfully different thing. When you do a heavy set of squats, the SpO2 reading on your wrist barely moves. The SmO2 signal in your quadriceps on the other hand drops rapidly (as much as a ~75% change in oxygen saturation) as demand outpaces supply, then rebounds during rest—telling you something concrete about how your muscles are responding to load, recovering between sets, and adapting over time.

I first became interested in muscle oxygenation monitoring and mNIRS about twelve years ago, well before it had a meaningful foothold in consumer health. I later developed the first NSCA-certified course on training with muscle oxygenation for Moxy, and eventually co-founded NNOXX, one of the companies bringing this technology to a broader market. So when I say I’ve been watching this space closely, I mean it.

Over the past year, I’ve been trying to understand the trajectory of consumer awareness of this technology and how it compares to other measurements. The figure below shows Google Trends data for muscle oxygenation and heart rate variability—HRV, the wearable industry’s favorite recovery metric over the past decade—over the previous 25 years. Looking at the chart, two very different trajectories emerge. HRV’s rise—starting in ~2014—was steep and relatively sustained, carried upward by a succession of product launches from Whoop and Oura, each new device generation producing another wave of consumer interest. But look more carefully at what drove those waves, and something becomes apparent: with each successive product release, both companies seem to be extracting diminishing returns in terms of search popularity. Their hardware iterations still move the needle, but by less each time. The single biggest driver of consumer interest in wearable health monitoring over the past decade wasn’t a product launch at all—it was COVID-19, which sent search interest in biometric tracking sharply upward across the board as people became acutely aware of their own physiological states in a way they hadn’t been before.

Against that backdrop, the muscle oxygenation trend line looks different—not a smooth climb, but a series of spikes and retreats, each one triggered by a new company entering the space. Moxy, the first company to release a consumer-focused muscle oximeter, generated an initial spike of interest that then subsided. The dynamics of emerging technology markets are counterintuitive in this way: a lone pioneer often fails to create a category, because without competitors, there’s no broader narrative for consumers or media to latch onto. Whoop and Oura grew together precisely because they were foils for one another—each company’s marketing reinforcing the other’s, collectively expanding the conversation about wearable health monitoring. Moxy had no such foil. Lacking real competitors, it also lacked the network effects, market validation, and sustained consumer awareness that competition, paradoxically, tends to generate.

It wasn’t until Humon—a VC-backed spinout from MIT’s Venture Development Center—commercialized its flagship device in 2018 that muscle oxygenation seemed poised to achieve the kind of mainstream traction that eluded Moxy. Then, two years later, Humon shut down. In their closing letter, they cited the core challenge: muscle oxygen was a new and somewhat misunderstood metric that required sustained market education to exist. Building a user base around a concept most consumers have never encountered is expensive and slow. Without the runway to outlast that educational curve, even a good product can fail. Notably though, Whoop saw the potential in what Humon had built, acquiring their assets and team in a move that looked less like a corporate merger than a targeted acquihire—tellingly, Humon’s CTO Daniel Wiese became Whoop’s Director of R&D shortly after.

Following Humon’s closure, search interest in muscle oxygenation drifted back toward pre-Humon levels, reinforcing the impression that the category might simply not be ready. Then my co-founders and I announced the launch of NNOXX in 2021, after which Google Trends for muscle oxygenation reached an all-time high. Two years later when NNOXX and Train.Red—a European mNIRS company that launched around the same time—both brought their products to market in 2023, something new happened: for the first time, two well-funded, consumer-focused muscle oximeter companies were operating simultaneously, creating the competitive dynamic the category had always needed. From that point, interest in muscle oxygenation didn’t just recover—it climbed steadily and has continued to do so, now on a trajectory to overtake heart rate variability in search popularity within the next few years. Given how thoroughly HRV came to define the wearable technology conversation over the past decade, that’s an incredible thing to say about a metric that was essentially invisible to the general public not long ago.

Partly based on the above, my hypothesis was that Whoop and Oura would be forced toward distributed sensor architectures—multi-point sensing systems that capture physiological data from multiple locations on the body, rather than a single wrist or finger—both in order to stay competitive with one another and to fight off up and comers, which I previously wrote about in Distributed by Design. My logic was as follows: there’s only so many signals you can measure at any one anatomical site, and the first company to move beyond that constraint would gain a durable competitive advantage. But I also wrote, perhaps too confidently, that Whoop appeared to be doubling down on single-point sensing while Oura was already gesturing toward distributed architecture:

"Companies like Oura have already started moving in this direction by itegrating data from constant glucose monitors into their platform. This transition from single to multi-point sensing represents a schism opening up between companies like Oura, which are embracing distributed network sensing, and companies like Whoop who are doubling down on their single-point sensing technologies, making progressively bolder claims about its capabilities."

It was against this backdrop that I came across Whoop’s latest patent filing last week—listing, among its inventors, the founding team from Humon. On one hand, I’ve suspected Whoop would eventually deploy the assets they acquired from Humon in a new product rollout. On the other, I’d grown skeptical of their direction, assuming they’d continue abstracting away from their sensors’ underlying capabilities in favor of surface-level differentiation as I wrote about in Castes on Quicksand: A Consumer Health Tracking Story.

That said, the filing itself raises more questions than it answers. The sensor architecture it describes appears to rely on placing the existing Whoop device against different muscle groups to capture mNIRS data—rather than introducing a new, purpose-built sensor. Recent speculation about a mystery device spotted on Tadej Pogačar during the 2026 Strade Bianche seems to point in a related direction—though what exactly was being measured remains unclear.

Why does this mater? In February 2026, Whoop launched Whoop Advanced Labs in the UAE. Two months later, Team UAE Emirates—the professional cycling team Pogačar rides for and which he led to back-to-back Tour de France victories in 2024 and 2025—announced a partnership with Whoop, naming it their "Official Health and Performance Wearable." Pogačar is widely considered the GOAT of cycling, and equipment choices during races attract the kind of scrutiny in cycling circles that an Apple launch might generate in tech. A mystery sensor on his arm during Strade Bianche is not an accidental sighting. Whether it represents a Whoop 5.0 variant with enhanced optical capabilities (cleaner HRV, heart rate, and blood oxygenation readings during exercise), or something further along a development roadmap toward muscle oximetry, is the question.

If that reading is correct, the patent may be deliberately vague about a problem they haven't solved yet. That problem is source-detector geometry. Near-infrared light needs sufficient separation between the emitter and the photodetector—typically around 30mm—to ensure that photons penetrate deep enough into tissue to interact with muscle hemoglobin rather than simply scattering off the skin surface or being absorbed by subcutaneous fat. This isn’t a software problem or something a better algorithm can compensate for. It’s also the reason that Moxy, NNOXX, Train.Red, and BSX—companies that arrived at their designs independently, in different countries, over different time periods—all ended up with eerily similar form factors. The geometry is a constraint, not an aesthetic choice. A device optimized for wrist-based photoplethysmography isn’t going to become a muscle oximeter by virtue of being strapped to a bicep or quadricep.

If Whoop’s plan is to enable meaningful mNIRS readings from their existing hardware, they will almost certainly run into this wall. The alternative—developing a genuinely new sensor form factor, purpose-built for deep tissue optics—would represent a more significant strategic departure than anything they’ve announced publicly. What I suspect is that the patent represents an early-stage disclosure and that the actual product roadmap is still in flux. Whoop is reportedly eyeing an IPO, which creates real pressure to demonstrate a meaningful hardware differentiation story to investors. A credible move into muscle oxygenation, done properly and with the right sensor geometry, would be precisely that story. It would also be their first genuine step toward a distributed sensor architecture.

The question is what they’d actually do with a new purpose-built muscle oximeter. I have two predictions, and both stem from a gap that every serious athlete who has tried to train by wearable data has eventually run into. The first is readiness. Whoop already offers a readiness score, as does Oura and virtually every other consumer health wearable. But, these scores share a common limitation: they’re systemic. They measure how recovered your cardiovascular and autonomic nervous systems are and infer from that whether your body is prepared to perform. The problem is that systemic recovery and local muscle recovery don’t always move together—I first learned this when beta-testing an "AI coach" feature for OmegaWave in ~2013. Your HRV can look excellent the morning after a hard lower-body session while the muscles you trained are still compromised and a wrist-worn device has no way to see that discrepancy. What NNOXX built, and what Whoop could build with the Humon’s technology, is a readiness score that distinguishes between these two things—aerobic readiness on one hand, local muscle readiness on the other. The practical implication is significant: rather than asking "is my body ready?" athletes could ask "is this specific muscle group ready?"—and get an answer grounded in the tissue’s actual physiological state rather than a systemic proxy for it.

The second application is resistance training, and it’s where I think the opportunity is most underappreciated. The fundamental problem with quantifying strength training using wrist-worn wearables is that the two metrics most devices rely on—heart rate and accelerometry—are poor proxies for what actually determines training quality: metabolic demand and mechanical tension at the local muscle level. Heart rate during a set of heavy front squats tells you something about cardiovascular response, but it lags behind the actual effort, saturates quickly, and doesn’t distinguish between a set that was truly taxing on the working muscles and one where large increases in intra-abdominal pressure were driving increases in heart rate, without a proportional drop in muscle oxygenation. Additionally, accelerometry can count reps, but it says nothing about the physiological cost of producing them.

SmO2 changes this picture. During a set, you can watch muscle oxygenation drop in real time as demand outpaces supply and the steepness of that drop is a direct readout of exercise intensity at the tissue level, not a cardiovascular surrogate for it. Between sets, the reoxygenation curve tells you how quickly the muscle is recovering: a fast, complete rebound suggests the muscle is ready to work again; a slow or incomplete recovery suggests it isn’t, regardless of how long the clock says you’ve been resting. The combination of these two signals—depletion rate during effort, replenishment rate during recovery—gives you something that doesn’t currently exist in any mainstream wearable: an objective, real-time measure of strain from resistance training.

Note: On a further reading, a lot of this patent hinges around the pressure sensitive strap. To me, there could be two uses for this. One is ensuring that pressure from the strap securing the sensor doesn’t impact muscle oxygenation readings. The other is developing an open or closed-loop feedback system for muscle oxygenation-guided BFR training. My sense is that the latter is the target, given that the Humon sensor’s readings didn’t appear to be impacted to a large degree by the pressure exerted on the leg from it’s strap (~30 mmHG).

Consider the total addressable market for this type of technology. Endurance athletes have had access to objective intensity metrics for decades, both external and internal—power meters on bikes, pace data from GPS, HR readings, blood lactate, VO2. Strength athletes have had RPE and rep counts. The gap between those two worlds is not just a data gap; it’s a coaching and training design gap. Closing it with a wearable that can actually see into the muscle would be, for strength training, what the power meter and heart rate were for cycling: a shift from subjective feel to ground truth.

Whether Whoop gets there, or whether they will be the first major player to enter this space, is uncertain. The constraints of adding an SmO2 reading to a Whoop band are real, and if that is their plan there are reasons to be skeptical. But the pieces are in place in a way they haven’t been before—the acquired technology from Humon, the recent patent filing, the IPO pressure, and a market that has spent the last two years discovering that muscle oxygenation is a metric worth caring about. What happens next will be one of the more consequential product stories in wearables in some time.

About the Author

I’m a computational biologist and bioengineer whose work spans precision oncology, biodefense, and human performance. As a founder at NNOXX, I helped build the first platform to measure muscle oxygenation and nitric oxide bioactivity non-invasively and in real time, bringing a first-to-market technology from concept to commercialization. Since then, my work has focused on developing computational models to understand cancer evolution and mechanisms of treatment resistance. You can read about what I’m currently learning, working on, and thinking about on Sequence & Destroy, or reach me at evanpeikon@gmail.com.

I recently had an interesting conversation on the tension between accuracy and interpretability with the head of ML at a small biotech working on liquid biopsies, which grew out of two earlier pieces: How to Develop Predictive Biomarkers and What Actually Makes Someone Good at ML in Computational Biology. One of the more interesting facets of the discussion was on how accuracy and interpretability can act as competitive moats, but through totally different mechanisms.

Accuracy, in a sense, is a commodity that can be bought. Given enough money, you can buy your way to a better model—more high quality training data, more compute, larger teams, and greater access to proprietary datasets. This gives early movers a huge competitive advantage, but given enough resources a well-funded competitor can close almost any accuracy gap. Interpretability, on the other hand, doesn’t work that way—it can’t be purchased at scale. Instead, it requires deep domain expertise, biological intuition, and the kind of hard-won insight that you can’t list on a procurement order—it’s a bit of a dark art, and that’s what makes it defensible.

But, the more interesting question isn’t which moat is harder to breach. It’s which produces better science. In this piece, we’ll explore how interpretable models don’t just offer a strategic advantage; in biology specifically, they outperform black boxes where it matters most (i.e., clinical translation, generalization to new patient cohorts). Additionally, unlike accuracy, that advantage isn’t one that more spending can replicate.

Issue № 70 // Why Interpretable Models Outperform Black Boxes in Biology

Accuracy is a property of a model’s performance on a test set—it tells you how often the model is right, but nothing about why. Interpretability, on the other hand, describes a model’s relationship to the system its modeling, meaning the degree to which the model’s internal logic mirrors the underlying biology. The distinction between accuracy and interpretability can be made more clear with an example.

After building a biological knowledge graph, you can predict whether an edge exists between two nodes using a graph neural network (GNN). Alternatively, you can predict that same edge with a graph traversal, following chains of known biological relationships to their logical conclusion. The GNN might outperform the traversal on a test set, but the traversal tells you how those nodes are connected and why that connection is plausible. One approach gives you an answer while the other gives you a hypothesis. This same tension pops up when trying to predict treatment response. For example, a deep learning model trained on reverse phase protein array data may predict an HR-/HER2- breast cancer patient’s response to immunotherapy with high accuracy. A 5-gene signature, by contrast, will likely perform worse on held-out data, but is has the advantage of telling you which biological processes are doing the work and it gives a clinician tangible information to act on.

In most domains, we accept the trade-off between accuracy and interpretability without much deliberation. A bit of interpretability is worth giving up if the model generalizes reasonably well and ships on time. But, in biology there’s a third consideration that changes the calculus entirely: the model needs to suggest the next experiment.

This is why the standard machine learning framework—train, validate, test, deploy—doesn’t map cleanly onto biological research. In a recommendation system, deployment is the goal. In biology, deployment is the beginning of the interesting part. A classifier that predicts tumor response with 95% accuracy but offers no mechanistic hypothesis about why certain tumors respond is a dead end from a scientific standpoint. An interpretable model that achieves 88% accuracy but identifies a specific regulatory relationship—say, that tumors with elevated hypoxia signatures and low immune infiltration are the ones failing to respond—gives you a lever to pull. It tells you what to measure in the next cohort, what to interrogate in the next cell line experiment, and what pathway to target in the next therapeutic design.

George Church touched on this in a recent interview with the Lifespan Research Institute, which I tend to agree with:

Would you prefer a weaker but more interpretable AI or a stronger but less interpretable one?

GC: I lean on the interpretability side. It’s not an either-or, but… we’re in science. Few engineers are willing to just pull a rabbit out of a hat, just a black box. Scientists and engineers, by and large, want to know the mechanism. The FDA likes to know mechanisms. Typically, the autocatalytic loop where you learn something and then you invent something is better if it’s mechanistically grounded. So, I lean pretty heavily in the direction of interpretability, explainability, transparency, et cetera, and also it’s safer.

I just honestly think that we will soon be faced with this dilemma, where we will have to choose between the power of the model to do things and its actual interpretability, but maybe we’re not there yet.

GC: If you look at the human scientist experience, the most powerful sciences are the ones that are better articulated mechanistically on a solid foundation rather than black boxes. The black boxes tend to include artifacts, dead ends. Most of the progress in science and engineering has been part of community efforts with strong mechanistic underpinnings.

There’s another, subtler, version of this argument that goes beyond experimental utility. Interpretable models that align with biological mechanisms are more likely to generalize across contexts, because they’ve learned something about the causal structure of the system rather than correlations in a particular dataset. A gene expression signature that works as a classifier in one cohort may fail in another if the cohort-specific technical and demographic variables are confounded with the biology. For example, if your training cohort happens to be predominantly collected at a single institution, on a particular sequencing platform, with a skewed demographic composition, the model will quietly bake all of that in. Then, when it’s applied to a new cohort with different technical and demographic variables, it often fail (and more frustratingly, it fails without telling you why).

A model built around a mechanistically interpretable pathway is more robust to this kind of distribution shift because the mechanism travels with the biology, not with the dataset. For example, a hypoxia signature exists in a tumor regardless of where the sample was collected or how it was sequenced. Anchoring your model to that kind of structure gives it something real to hold onto across contexts. This isn’t always true, of course. Mechanisms can be context dependent as previously discussed in Issue #55 // Molecular Moonlighting. But, even when interpretable models fail, they tend to fail informatively. If your 5-gene signature stops being predictive in a new cohort, that’s a hypothesis: perhaps hypoxia isn’t the rate-limiting factor in this population, or perhaps there’s a moderating variable you haven’t accounted for. A black box achieving 0.51 AUC on held-out data gives you nothing to work with. The failure mode of an interpretable model is, in a sense, still science.

Thanks for reading! If you found this post useful, please consider subscribing. I share hands-on computational biology techniques, fresh ways to think about tough problems, and perspectives on a range of related topics. All free, straight to your inbox.

Subscribe now

Earlier this week I shared a new tutorial, How to Develop Predictive Biomarkers, which is designed to fill a conceptual gap—something I’ve identified when searching for resources that help computational biologists with intermediate machine learning experience leverage their unique skillset to create biomarker panels for predicting clinical outcomes (like treatment response in triple-negative breast cancer, for example).

Most existing resources cover isolated parts of ML model development, but none that I’ve come across explain the big picture—how the different parts of the pipeline fit together. This tutorial aims to do that, guiding you through the thinking process at each step from data acquisition to model deployment. Check it out, and please pass it along if you find it useful!

Read more

Thanks for reading! If you found this post useful, please consider subscribing. I share hands-on computational biology techniques, fresh ways to think about tough problems, and perspectives on a range of related topics. All free, straight to your inbox:

Subscribe now

Issue № 68 // Connectivity Is the Target

In the early 2000s, Barabási and colleagues introduced the idea of using network science to identify essential, highly connected protein "hubs" in molecular interaction networks as prime therapeutic targets. The core idea here is that by targeting, and disrupting, these central nodes (which often reside within disease-specific network modules) drugs can collapse pathological signaling while minimizing off target/adverse effects elsewhere.

Conceptually, idea makes a lot of sense to me. But, in practice it’s surprisingly difficult to prove that a given hub is actually a good drug target rather than just a topologically interesting node. This piece is about a framework I’ve been playing around with to do exactly that. I’m writing this partly to clarify my own thinking, and partly to invite feedback from others working on similar problems.

What Makes a Good Drug Target?

In Issue #58: How to Kill a Tumor, I wrote about different schools of thought on drug target identification in cancer. These include finding proteins that are differentially expressed on tumor cell surfaces relative to healthy tissue (the ADC approach), targeting molecular drivers of disease (HER2-directed therapy, AKT inhibitors, and so on), and disrupting network topology, which is the angle Barabási’s work opens up. In an ideal scenario, all three converge on the same protein, resulting in total pathway collapse.

For the first two approaches, the evidentiary chain is relatively clean and easy to understand. Let’s take the case of identifying an ADC (antibody drug conjugate) target, for example. To do this you’d first want to find proteins that are up-regulated in tumors relative to healthy tissue, ensuring that (a) the cytotoxic payload hits the right target and (b) that off-target effects are limited. Next, you’d want to confirm your candidate targets are associated with treatment non-response by seeing if they are further elevated in patients who fail standard-of-care treatment relative to those who response favorable (here, we can define "respond" as achieving pathological complete response, or pCR, for example). Finally, we’d want to show that high expression of the candidate protein predicts disease recurrence in the treatment-refractory population via Kaplan-Meier and Cox proportional hazards analysis. There’s a legible chain of cause and effect here, and the logic maps directly onto clinical intuition—the protein is elevated in tumors, associated with non-response, and patients with high expression of it have worse survival. No one will batt an eye if you use this logic to propose a new drug target.

The network hub case is harder. Many of the standard intuitions about what makes a good drug target don’t cleanly transfer, and even defining what it means for a protein to be a hub is non-trivial. Degree centrality, betweenness centrality, and closeness centrality all capture different aspects of topological importance, and there’s no consensus on which metric is most clinically relevant.

Defining "Core" Hubs

Recently, I’ve been developing the concept of "core hubs," which I define as proteins satisfying two criteria simultaneously1…

1. First, core hubs must be rich-club members, which are proteins whose degree exceeds the threshold k where the normalized rich-club coefficient (φ_norm) reaches its maximum value in the protein co-expression network. The rich-club phenomenon describes the tendency of nodes of similar degrees to be more densely interconnected with each other than would be expected by chance, forming a kind of elite backbone of the network.

2. Second, core hubs must also rank in the top 10% of all proteins by degree centrality.

An example: The figure below shows the same protein co-expression network visualized with two complimentary color schemes. The left version is color-coded by pathway membership with nodes sized by their degree centrality. The right version is color-coded by each node’s normalized rich-club coefficient (φ_norm) at their respective degree, with color intensity scaled to the maximum φ_norm.

The node delineated as A is categorized as a core hub, whereas nodes B and C are not. From the visual, you can se that node B has a very high degree centrality (it’s the largest node in the network), but it is not a rich club member (hence, the white color in the rightmost visual). Node C on the other hand is a rich-club member, but it is not highly connected. Only node A, in this comparison, fits both criteria being both highly connected (top 10% of nodes by degree centrally) and a rich-club member.

Conceptually, core hub proteins represent the network's most influential nodes, forming a highly connected, mutually reinforcing, central component that acts as a backbone for efficient global communication. Based on these properties, i’ve hypothesized that disrupting these nodes should cause disproportionate, cascading network failures relative to equivalent disruption of non-hub proteins, making them prime candidates for targeted intervention.

One early observation that gave me some confidence in this framework is how it maps onto established biology. For example, I’ve seen that HER2+ breast cancer cohorts have HER2 family proteins appearing as core hubs in their co-expression networks, which is consistent with the known efficacy of HER2-targeted therapies in these populations and suggests that part of what those therapies are doing is producing network-level collapse rather than simply reducing the abundance of a single target. The framework has also surfaced less expected results: HER2 family core hubs appearing in a subset of TNBC patient networks, and certain ADC target candidates appearing as core hubs in molecular subtypes where you might not anticipate it.

The Validation Problem

The challenge is proving these targets out. Kaplan-Meier curves are less informative when a candidate protein shows limited expression variance across a cohort—which is often the case for network hubs, since their importance is topological rather than purely expression-level. What does become compelling is when a given protein appears as a core hub in a specific molecular subtype (say, TNBC), and then when that population is stratified by treatment response, the hub is specific to the non-responder network. That’s a meaningful signal. But it still doesn’t tell you whether disrupting that hub would actually collapse the network in a therapeutically relevant way.

To address this, I’ve been developing a targeted network attack simulation that works as follows…

• Given a candidate protein A, I simulate a targeted attack by removing that protein and all proteins with co-expression relationships exceeding a defined threshold (|r| > 0.7, for example) from the network.

• I then quantify network disruption across four metrics: change in network density, change in the size of the largest connected component, change in average local clustering coefficient, and change in the number of connected components. These metrics collectively capture both global connectivity loss and the degree to which the network fragments into isolated subgraphs.

• To contextualize the disruption caused by targeting protein A, I run upwards of 1,000 random attack simulations on the same network, each time selecting proteins at random and removing them along with their high-correlation neighbors. The disruption scores from the targeted attack on protein A are then expressed as a percentile relative to this empirical random attack distribution.

Using the above framework, a core hub should score near the top of this distribution, indicating that its removal causes significantly more network disruption than would be expected from removing a randomly selected protein of equivalent size—evidence that its topological position, not just its connectivity, is what makes it influential.

To further validate the target we can then perform a parallel comparison. For example, if the targeted attack is performed on the TNBC non-responder network where the core hub is present, I perform the same attack on the TNBC responder network where the core hub is absent. The prediction is that attacking the hub in non-responders produces near-maximal network disruption, while the same attack in responders produces disruption near the 50th percentile—indistinguishable from random. That divergence would constitute meaningful evidence that the hub’s topological importance is specific to the biological context in which it’s relevant.

One of the most obvious limitations of this approach that I see is that it currently operates on coexpression networks, which capture statistical co-variation between proteins but are agnostic to mechanism. In practice, this could limit some of the causal interpretability of the attack simulations, though I do still think this framework can still meaningfully complement more traditional target validation analyses2.

Additionally, over the next few weeks/months, the extension I’m most interested in is replacing the co-expression networks with causal protein networks, like those generated by CausalEdge. A causal network would allow the attack simulations to operate on directional, mechanistically grounded relationships rather than symmetric correlations, which would substantially strengthen the interpretability of the results. The constraint for now is data: inferring reliable causal networks requires longitudinal proteomic measurements that I don’t currently have at the necessary scale. But it’s a natural next step, and one worth keeping in mind as the framework develops.

Liked this piece? If so, tap the 🖤 in the header above. It’s a small gesture that goes a long way in helping me understand what you value and in growing this newsletter.

This issue of Sequence & Destroy will be a bit different from the norm. Rather than writing about proteomics, molecular medicine, or any number of other topics I’ve been exploring as of late, this piece is about a new project I’ve been working on. One that lives at the intersections of human performance, wearable technology, and machine learning. If you’re interested in learning more, shoot me an email at evanpeikon@gmail.com.

Issue № 67 // Doping with Data

Every endurance athlete, coach, and human performance scientist is familiar with training zones—they’re ubiquitous across all work capacity-based sports, whether running, cycling, skiing, or rowing. In practice, these zones serve as a common language for coaches and athletes to prescribe, execute, and communicate training‡.

‡Conceptually, training zones are intensity classifications that divide exercise into discrete bands, usually based on heart rate, power, or blood lactate. Additionally, each zone is presumed to elicit a distinct physiological response and training adaptation, and as a result training zones are commonly used to prescribe training (i.e., run for 60-minutes in a zone-2 heart rate between X and Y bpm).

Zone-based training models have become so embedded in coaching culture that questioning them feels almost heretical. But, it’s my opinion that the standard 5-zone training model is based on two flawed assumptions:

1. The first is that human physiological can be adequately described by a single biomarker (for example, heart rate).

2. The second is that exercise intensity can be uniformly characterized by five (or any hard-coded number, really) distinct physiologic states.

Neither of these assumptions holds up to serious examination and the compounding effect of these two over simplifications—zones based on a single biomarker combined with arbitrary zone counts—leads to a conceptual framework that often obscures more than it reveals. I think we can do much better. Let’s dive in!

Problem 1—The Single Biomarker Fallacy

Zone-based training typically relies on one biomarker—most commonly heart rate. When you prescribe training based on a single biomarker, you’re by grouping data points alone a line, which forces you to make arbitrary cutoffs. For example, between 109 and 127 beats per minute (bpm) is zone 2, above 165 bpm is zone 5, and so forth. The idea is simple and easily communicated (and hence widely adopted), yet incomplete.

Physiology doesn’t respect these boundaries because physiology isn’t one-dimensional1. To better understand what I mean, let’s take the following scenario: a competitive Nordic skier completes the same 15-minute performance assessment (5-minute ramp, 10-minutes slightly above critical speed) on a ski-treadmill twice, with sessions separated by ~10 weeks. The first time they complete the test, their heart rate climbs from ~140 to ~175 over the first 5-minutes, then they maintain that heart rate (roughly) for the remaining 10-minutes. However, during those 10-minutes their respiratory rate and VO2 steadily climb, reaching near peak values by the end, as depicted below.

Now, let’s look at the second assessment after ~10 weeks of limitation-based training. We can see that the athlete’s heart rate kinetics are roughly the same through the duration of the test, but their respiratory rate and VO2 follow a very different trend. Whereas they progressively increased throughout the first assessment, they initially increase, then flatline in the second. Two tests with the same heart rate, but profoundly different overall responses—yet, we only see this discontinuity when moving from 1→3 dimensions.

The above case highlights how defining zones based on a single biomarker can result in two different physiologic states being lumped together as one (false positive). However, defining zones with discrete heart rate thresholds is just as likely to produce false negatives. For example, according to the Polar app, cycling at 133 bpm puts me in zone 2, but at 134 bpm I’m suddenly in zone 3, despite there being negligible physiological differences between the two measurements (and, as mentioned previously, this approach treats all exercise at 134 bpm as being equivalent, regardless of what’s happening to my ventilatory drive, substrate utilization, or muscle oxygenation)‡.

‡This problem actually compounds during training. You might complete a hard twenty-minute threshold effort where your heart rate stays locked between 168-172 bpm. A conventional five-zone model may classify the entire block as Zone 4. Perfect execution, according to your training app. But your subjective experience may tell a different story. There’s the moment when the effort first becomes hard, the period where it’s sustainably hard, and the final minutes where everything starts to unravel. Most athletes can feel this gradation. Traditional single-metric zoning flattens these experiences into a single category.

Multi-dimensional zoning allows us to circumvent these issues. With two biomarkers, you can place datapoints on an x-y plane, allowing you to identify regions characterized by high value in one dimension and low in another, for example.

With three biomarkers, data points can be spaced along x-y-z axes, providing even more granularity (computationally, we can work in much higher dimensional spaces, though it’s not possible to visualize anything greater than three).

The model my colleagues Frank Kolar, Daniel Crumback, and I have been developing uses a minimum of 3 and maximum of 18 dimensions based on the data a user provides us with us. In the best case scenario, the end-user of our platform uploads heart rate, muscle oxygenation, VO2, respiratory frequency, fat oxidation, and carbohydrate oxidation data. From those six features, we then calculate first and second derivatives, allowing us to capture both absolute rates and their rates of change (we also perform additional feature engineering to reduce measurement noise and prevent any single metric from dominating the analysis due to scale differences)‡.

‡Calculating these derivates allows us to answer questions like "How quickly is heart rate climbing (dHR/dt)? And is muscle oxygen saturation declining, and if so, is that decline accelerating (d²SmO2/dt²)?" During a ramp test or progressive effort, these temporal dynamics reveal metabolic transitions that static measurements miss. For example, an athlete whose heart rate is rising by 2 bpm/minute while SmO2 remains stable is in a different physiological state than one whose heart rate is rising at the same rate while SmO2 is plummeting.

This approach reveals patterns invisible to traditional zoning. For example, let’s say we have two cyclists with identical max power and max heart rate values who are both completing a ramp test to exhaustion. At 250 watts and 158 bpm, traditional single-metric zoning would classify them identically—same power, same heart rate, same zone. Multi-dimensional analysis might reveal that Athlete A has SmO2 of 58%, respiratory rate of 28 breaths/min, and RER of 0.89, while Athlete B has SmO2 of 44%, respiratory rate of 38 breaths/min, and RER of 0.97. Despite having the same heart rate and power values, they are not in equivalent physiological states. As a result, they will not receive equivalent training adaptations, and therefore should not be assigned to the same zone.

Problem 2—Arbitrary Zone Counts

For arguments sake, let’s say we accepted single biomarker zoning as adequate. Even then, the second problem still remains: why five zones? The answer, historically, is "because five seems like a reasonable number." Not too few to be useless and not too many to be unwieldy.

But, as many experienced coaches have learned, different athletes have different numbers of distinct physiological states, and that number often correlates with both training status and metabolic flexibility ( has a nice post that touches on this topic, which you can find here). For example, a recreational runner three months into their first structured training program may only have two or three "gears". Forcing this athlete’s physiology into five zones means at least two of them are arbitrary subdivisions with little underlying physiological justification. Conversely, a grand tour caliber cyclist with extraordinary metabolic flexibility developed over years of systematic training might have eight or night district physiological states. Compressing this into five zones obliterates meaningful distinctions—lumping together intensity bands that produce different training adaptations and have different sustainability characteristics.

As you can see from these examples, the arbitrary five-zone model creates two problems simultaneously: it over-partitions beginners (assigning five zones where fewer exist) and under-partitions elites (assigning five zones where many more may exist). The result is a system that’s simultaneously too coarse for advanced athletes and too granular for beginners, while being perfectly adequate for few. It’s a one-size-fits-all solution to what should be individualized.

Revealing Hidden Structure—What Optimal Multi-Dimensional Zoning Shows

The interaction between these two problems—single biomarker measurements and arbitrary zone counts—compounds their individual limitations. Consider a forced five-zone model applied to heart rate data alone. We’re taking a one-dimensional signal and arbitrarily dividing it into five categories. The zones might have some relationship to underlying physiology (higher heart rates generally indicate higher metabolic demand), but the boundaries are arbitrary percentages of maximum or threshold heart rate.

This is the problem Frank, Daniel Crumback, and I have been working on solving with our new project Assess.Works. Rather than abandoning training zones as a concept, we want the framework to evolve—from discrete to continuous, from single to multi-dimensional, and most importantly, from arbitrary to data-driven. The core principle is as follows: let unsupervised machine learning identify the natural number of distinct physiological states in an athlete’s data, rather than imposing a predetermined structure, and then characterize each state using whatever multi-dimensional physiological data is available. In practice, here’s what the process looks like: first, an athlete uploads their assessment data (typically from a ramp test or progressive effort going from easy to maximal intensity).

Then, our system searches for all available physiologic metrics, then loads them for easy charting and analysis.

Following that, the system selects certain biomarkers compatible with our zone prediction (whenever are available) model such as heart rate, muscle oxygenation, VO2, and respiratory frequency, substrate oxidation rates. It then engineers features, applies smoothing and standardization, then runs two parallel clustering algorithms:

• Forced five-cluster model: Constrains the data into exactly five zones for comparison with traditional frameworks and communication with coaches trained in conventional systems

• Optimal clustering model: Uses unsupervised machine learning to determine the natural number of distinct physiological states present in the data, typically ranging from two to nine clusters depending on athlete development.

Clusters are then ordinally ranked, using an in-house developed algorithm, to ensure logical progression from low to high intensity. Below is an example session where our model identified 10 distinct states during a ramp incremental exercise test including a recovery state (I often use the terms "state", "intensity cluster", and "zone" interchangeably. For all intents and purposes, I mean the same thing in each of these cases).

Looking at the chart above, you’ll quickly notice that adjacent intensity clusters overlapping ranges for individual metrics (ie., SmO2 data points with the same value can be color-coded by many different states depending on where in the workout they occur). This effect is even more pronounced in the figure below, which shows individual biomarkers data distribution per intensity cluster/zone (in this visual, I’ve forced the data back into a five-cluster model as described previously).

This is precisely where multi-dimensional analysis proves essential—and where single-biomarker zoning fails. For example, you can have the same heart rate (say, 145 bpm) assigned to multiple different clusters (recovery, and clusters 1-4 in this case). This overlap isn’t a flaw. It’s the system correctly recognizing that physiology is multivariate and that no single metric fully captures metabolic state. While individual metrics may overlap between clusters, their multi-dimensional combinations remain distinct‡.

‡When you examine mean values across clusters, you see the expected progression: mean heart rate and VO2 increase from one zone to the next, while SmO2 typically decreases. But the distributional overlap in any single metric reflects the reality that physiology is complex and multi-factorial.

This is why both problems—single-metric measurement and arbitrary zone counts—need to be solved simultaneously. Multi-dimensional analysis without optimal zone counting still imposes artificial structure. Optimal zone counting based on a single metric still misses critical physiological distinctions. The solution requires addressing both.

Interpretable Machine Learning—Combining Data-Driven Patterns with Objective Physiologic Metrics

One of the more interesting implications of data-driven intensity clustering is how it can be combined with objective physiologic thresholds, such as FATmax, critical power, or the metabolic crossover point, among others. After performing multi-dimensional zoning, you can overlay these boundaries where they exist. You might find, for instance, that FATmax corresponds with a cluster transition. Or you might discover it doesn’t—that FATmax occurs in the middle of a stable cluster, but a zone transition happens when FATmax is crossed and muscle oxygen saturation simultaneously drops below a threshold. This reveals that FATmax alone doesn’t drive a broader multi-dimensional state change in this particular athlete’s physiology.

This is interpretable machine learning. Rather than creating a black box that pronounces mysterious judgments about training zones, we’re building a system that shows you where objective physiological events occur and lets you identify which combinations of metrics signal meaningful transitions.

Point-by-point classification provides another means by which you can probe the underling physiology behind predictions. In contrast to interval-based zoning, Assess.Works uses a continuous zoning model with high resolution. Rather than assigning minutes 10-20 were Zone 4, for example each data point receives its own cluster assignment based on the instantaneous multi-dimensional physiological state. This reveals gradual drift during nominally constant-intensity efforts—the progressive metabolic stress that athletes subjectively experience but traditional zoning fails to capture. Each cluster can be interrogated: Why is this data point assigned to Cluster 3? Because it exhibits high heart rate combined with suppressed SmO2 and elevated respiratory frequency acceleration—a specific, interpretable multi-dimensional signature.

A New Direction for Human Performance

Traditional zoning models have guided effective training for decades, and they’ll continue to do so. The five-zone model is a useful heuristic, especially for athletes and coaches who don’t have access to sophisticated physiological monitoring or who prefer simplicity over precision. But, as wearable sensors become more capable and less expensive, as athletes routinely collect data on muscle oxygen saturation, respiratory metrics, and metabolic markers during training, we have an opportunity to move beyond both arbitrary assumptions that limit traditional zoning:

• From single-biomarker to multi-dimensional: Recognize that physiological state can’t be adequately captured by heart rate or power alone, but requires consideration of multiple interacting variables2.

• From arbitrary five to optimal N: Acknowledge that different athletes operate across different numbers of distinct metabolic states, and let individual physiology dictate zone structure rather than imposing predetermined categories

The shift is subtle but important. Instead of asking “what zone am I in?” based on a single metric and a lookup table, we can ask “what is my multi-dimensional physiological state?” and receive an answer grounded in the patterns present in our own data, informed by objective metabolic markers, and interpretable in terms of the specific combination of factors that characterize each state.

What emerges is a view of training intensity as a continuous, high-dimensional landscape rather than a series of discrete rooms. You don’t jump from Zone 3 to Zone 4 at some magic heart rate threshold. You transition gradually through a space defined by multiple interacting variables, and the “zones” are simply regions of that space where certain combinations of variables tend to cluster together naturally. Sometimes those regions align with traditional five-zone boundaries. Often they don’t. And that’s not a failure of the model—it’s the model working as intended, revealing structure in the data that simpler frameworks necessarily miss.

For now, most athletes and coaches will continue using traditional zones because they’re familiar, because they work well enough, and because the infrastructure of training plans and coaching wisdom is built around them. But as we instrument athletes more completely, as we collect richer data and develop more sophisticated analytical tools, the five-zone single-metric model will increasingly look like what it is: a useful approximation of a phenomenon too complex to fit neatly into five categories defined by one number. The question isn’t whether we should abandon zones entirely. It’s whether we’re ready to let them evolve.

The team at Assess.Works just finished a major milestone, and are now moving onto the next one: auto-identifying physiological limitations. If you’re interested in learning more or know organizations that might benefit from a conversation about our technology and strategic partnership opportunities, we’d welcome the connection. Feel free to respond to this email, or to reach out directly (evanpeikon@gmail.com).

Enjoy this piece? You can subscribe to have the next post delivered to your inbox:

Subscribe now

Issue № 66 // Crossing the Proteomics Chasm

I recently read an article titled "Why Proteomics is Still Stuck in the Basement" on the Talus Bio blog, and in it and said something that stuck with me:

"My grad school cohort mostly consisted of genomicists and computational biologists, with us proteomicists frequently the outliers, so despite DIA-MS being a Big Deal in my mass spec circles, my departmental research reports mostly got glazed-over eyes and scattered polite claps. Amongst my fellow trainees, the joke was that proteins weren’t real – they’re "gene products" and therefore just a consequence of genomics – and to an extent, it’s true. Proteomics has definitely lagged behind genomics, for a variety of reasons." [source]

Prior to starting lab rotations early in my PhD, almost all of my prior bioinformatics experience was working with bulk, single-cell, and spatial RNA-sequencing data. But, having now spent what—with some back of the napkin math—must have amounted to >1,000 hours doing applied proteomics research, I’ve noticed a subtle shift in my thinking. When I was primarily working with RNA/transcriptional data, proteins were relegated to "gene products", as Alex and Lindsay alluded to, and were almost an afterthought1. Now, I find myself thinking more along the lines of "a gene is a protein’s way of making another protein."

The biggest challenge going from working with transcriptomic to proteomic data for me wasn’t this mental frameshift though. It was learning the nuances of working with a new type of omics data and understanding that the simplest conventions—from how to normalize data to specific thresholds to use during differential expression analysis—can be totally different2. This is particularly challenging in proteomics, where so much knowledge is platform-specific; as a result there is less standardization of protocols and best practices aren’t as well codified. Because of this lack of centralized information, becoming proficient in working with proteomics data is a bit like a dark art, handed down from mentor to trainee through spoken word. This can be particularly true, when starting to work with more niche types of proteomic data, like that from reverse protein arrays.

Most people reading this are probably familiar with RNA-sequencing, which measures mRNA transcript abundance. Reverse phase protein array (RPPA) data on the other hand is less common, but at the simplest level it measures proteins and post-translational modifications, allowing you to assess function and active signaling status (for this reason, RPPA proteomics is often called functional proteomics). Aside from measuring a different layer of biological abstraction, one of the biggest differences between RNA-seq and RPPA data is that in the former case you’re measuring thousands (even tens of thousands) of genes at a time, whereas in the case of RPPA you’re typically measuring a few hundred proteins at most (often as little at ~150). Additionally, whereas RNA-seq provides a broad, genome-wide, "snapshot" of transcription, RPPA offers a targeted, high-sensitivity, measurement of specific pre-specified proteins.

Why use RPPA over mass spectrometry-based proteomics? For my work studying tumor phenotypes and drug vulnerabilities, RPPA offers several advantages: it provides highly quantitative measurements with excellent reproducibility across large sample cohorts (often hundreds or thousands of samples), has superior sensitivity for detecting phosphorylation events (which are critical for understanding active signaling states), and most importantly, allows for hypothesis-driven interrogation of specific signaling pathways I'm most interested in. While mass spec offers unbiased, proteome-wide coverage, RPPA's targeted approach means I can deeply profile the signaling networks most relevant to cancer biology across many more samples than would be practical with mass spec (in some cases we do have access to both RPPA and mass-spec proteomics data, or mRNA data, but these tend to be less common).

In addition to RPPA measuring a more focused set of proteins, we also need to consider how protein kinetics differ from mRNA. For example, when performing differential expression analysis with RNA-seq data we may be used to seeing a much higher dynamic range of values, and as a result it’s common to use stricter log2FC filters. As compared to genes, protein levels are often more stable, and have smaller, more consistent, changes in the range of ~20-50%, and as a result lax filters like |log2FC|≈0.26-0.58 can provide meaningful results.

Where things became more of a challenge for me is when I wanted to start working with RPPA protein co-expression networks, both because there was no existing software available for this, and because so many assumptions from working with gene expression data don’t hold. For example, when working with RNA-seq data, there are conventional guidelines for performing weighted gene coexpression network analysis, such as selecting an appropriate soft thresholding power to achieve scale-free topology. In WGCNA, you typically test a range of soft threshold powers (often 1-20) and select the lowest power where the scale-free topology fit index (R²) exceeds 0.8 or 0.9, indicating the network approximates a scale-free structure. This soft thresholding approach preserves the continuous nature of gene-gene relationships and emphasizes strong connections while maintaining weaker ones.

RPPA networks require a different approach. Because RPPA measures a pre-selected panel of functionally related proteins—typically key signaling nodes and their phosphorylation states—the network has inherently different topological properties than genome-wide transcriptomic networks. Rather than aiming for scale-free topology, I often use hard thresholding based on correlation strength to identify the most robust, coordinated signaling relationships. This approach is well-suited for RPPA data because…

1. With only ~175-250 proteins, we’re specifically looking for strong, direct signaling relationships rather than the broader, more diffuse co-expression modules typical of transcriptomics;

2. These pre-selected proteins are chosen because they’re known signaling hubs, so we expect—and want to capture—tight coordination between functionally related proteins; and

3. The resulting networks are more interpretable for identifying coordinated signaling architecture, such as kinase cascades or feedback loops, rather than loosely connected modules. The density of these networks reflects the biological reality: these proteins were selected precisely because they work together in known pathways.

Below you’ll find a sample output from a co-expression analysis software I built: here, each node is a phosphoprotein and edges between them are coordinated relationships. Each node is sized based on it’s degree centrality, and proteins are color-coded based on their pathway.

Because I’m typically working with the same ~175-250 phospho-proteins, it’s also allowed me to create some cool tools like a Figma pathway map, part of which is depicted below, with all of the proteins and their interactions included. This resource has been extremely valuable when trying to interpret analysis results, as I can cross reference to see if any specific "pathway stories" emerge. For example, I may see that AKT T308, 4E-BP1 S65, and eIF4E S209 are up-regulated. At a quick glance this may not seem as meaningful as seeing PI3K, AKT, and mTOR all up-regulated, but using the pathway map it quickly becomes clear that there is a signal here: phosphorylation at T308 activates AKT, and 4E-BP1 is a downstream effector whose phosphorylation leads to the release, and activation, of eIF4E.

Speaking of pathways, this raises another point. When working with RPPA data, traditional pathway enrichment tools fail to perform. Not only do we not have enough proteins for significant enrichment, but because of the specific antibody names and phospho-sites, they are also unrecognized in most databases. To solve this, I’ve had to create a massive dictionary with each protein, the pathway(s) it belongs to, and whether its phosphorylation is functionally activating or inhibiting. Using this, I’ve created my own custom enrichment tool that looks at both the protein’s log2FC and regulatory relationship to interpret the data (for example, a down-regulated inhibitor is functionally activating). This allows for some neat visualizations, like the one below, which combines differential expression and network topology data to create a composite score (y-axis) for a proteins influence in a given condition, and under the x-axis we have color indicators for each proteins pathway membership and functional effect.

I plan to share some of these tools publicly after finishing my PhD as part of an open-source software package for working with RPPA data. In the meantime, I’m curious to hear from others making similar transitions between omics platforms: What conventions have you established? What unexpected challenges did you encounter? The proteomics community is smaller and more fragmented than genomics, but perhaps that’s precisely why sharing our solutions and learning from each other’s experiences is so valuable.

Liked this piece? If so, you’ll want to check out other articles in Sequence & Destroy’s Wearable Technology & Biometrics collection. Additionally, you can show your support by tapping the 🖤 in the header above. It’s a small gesture that goes a long way in helping me understand what you value and in growing this newsletter. Thanks!

Issue № 65 // Multiplex Molecular Biosensing

I’ve been involved in the wearable technology industry for more than a decade now, first as a technical research consultant and scientific advisor to early-stage companies in this space and later as a co-founder at NNOXX. During this time I’ve seen two parallel trends unfold: the first is a massive cost-reduction for existing sensing modalities, and the second is a significant increase in the number of biomarker measurements that can be obtained with a wearable device.

I still remember working with the original Omegawave system—a cumbersome, expensive contraption that felt like a relic from the Soviet cosmonaut training program (in a way, it was). Omegawave was the first commercially available device to measure HRV and DC potential with the explicit goal of predicting "readiness" and was a precursor to Whoop and Oura. Back then, wearable devices were limited to a small number of measurements like heart rate, HRV, blood oxygenation, and skin temperature—vital signs that any 20th century physician would recognize. Today, we live in a different world entirely with muscle oximeters, portable metabolic analyzers, in ear EEGs, hydration sensors, and more being made available all the time.

After all of this progress, we’re approaching a conceptual precipice. The industry is now experiencing what looks to be a natural continuation of these trends with molecular biosensors capable of measuring specific proteins and metabolites in blood and interstitial fluid continuously. Companies are already developing devices that can measure real-time blood lactate, monitor ketone levels throughout the day, and track inflammatory cytokines like IL-6 at regular intervals. Unsurprisingly, VCs have taken notice, funding startups that promise to monitor various proteins with minimally invasive sensors continuously and in real-time. The implicit assumption seems to be that if measuring one molecule—glucose—revolutionized diabetes management, then surely measuring other new molecules will unlock similar insights for a wide range of applications.

But what if this assumption is wrong? What if molecular biosensors aren’t a natural progression from our current wearable technologies—what if they represent a category shift that the industry as a whole doesn’t seem to recognize? Continuous protein measurements and other molecular biomarkers aren’t just higher-resolution versions of gross physiological metrics like heart rate or blood oxygenation (collectively termed 'digital biomarkers'). They are fundamentally a different kind of measurement that requires a different interpretive framework. What follows are three connected arguments:

• First, molecular biomarkers operate according to different principles than the digital biomarkers we’re accustomed to tracking;

• Second, current business incentives are driving development toward single-analyte sensors that may not deliver their promised value; and

• Third, alternative paths—as opposed to those currently being pursued—might better match the technology to contexts where its capabilities are most needed.

Digital and molecular biomarkers are complementary, but they’re fundamentally different types of measurements that each require distinct interpretive frameworks. The jump from measuring heart rate to measuring individual proteins isn’t incremental progress—it’s moving from integrated physiological outputs to individual molecular components. More precisely, it’s crossing layers of biological abstraction—from emergent system-level phenomena down to the molecular entities that underlie them. From my experience working with both digital and molecular biomarkers independently, as well as developing methods to integrate them via knowledge-graph-based systems, it’s clear that the industry is attempting to apply a digital biomarker framework to molecular measurements—and that approach is built on a flawed assumption about how proteins work.

Consider what happens when you measure Interleukin-6 (IL-6) continuously. IL-6 is commonly described as a pro-inflammatory cytokine, one of the key mediators of fever and the acute phase response. It would seem reasonable to interpret elevated IL-6 as indicating more inflammation. But the biology is more complex than this. IL-6 can signal through two different pathways: classical signaling via membrane-bound receptors tends toward regenerative and anti-inflammatory effects, while trans-signaling through soluble receptors drives chronic inflammation. The same molecule at the same concentration can have entirely different biological meanings depending on context, which I previously wrote about in Issue #55: Molecular Moonlighting.

Interleukin-6 isn’t unique in this regard. All proteins exist in dynamic networks where function emerges from context, relationships, and timing. A kinase that phosphorylates one substrate during cell cycle progression might phosphorylate a different substrate during stress response, creating entirely different downstream effects. The protein hasn’t changed—its relational context has. This has implications for how we should think about molecular biosensors. The quantified self movement has largely embraced single-molecule monitoring: track your cortisol for stress, measure your C-reactive protein for inflammation, monitor your glucose for metabolic health. But, proteins are components within networks, not standalone indicators. Measuring a single protein is like measuring the voltage across one resistor in a complex circuit—the measurement may be accurate, but its meaning depends on the state of the rest of the system.

This is where molecular biosensors differ from the gross physiological metrics we’re used to tracking. Heart rate and blood oxygenation are already emergent phenomena that integrate signals from multiple systems. They represent the output of countless molecular interactions, abstracted up to a level where single measurements carry interpretable meaning. A single protein measurement, by contrast, is looking at one node in a network where meaning comes from relationships rather than absolute values.

Glucose monitoring deserves special consideration here, since it’s the standard reference point for continuous molecular measurement. Glucose occupies an unusual middle ground between gross physiological metrics and true molecular biomarkers. While glucose is technically a molecule, its behavior in the bloodstream functions more like a systems-level measurement. Glucose concentration reflects the integrated output of multiple regulatory processes—insulin signaling, hepatic glucose production, peripheral uptake, hormonal counter-regulation—operating within relatively stable kinetic parameters. The interpretation of a glucose value doesn’t depend heavily on knowing the concentration of specific enzymes or transporters at that moment; the meaning is largely contained in the glucose value itself, contextualized by simple factors like food intake and activity. This is precisely what makes it tractable as a single-analyte measurement.

Most proteins don’t work this way. Their function depends on post-translational modifications, binding partners, subcellular localization, and the expression levels of their targets and regulators—information that isn’t captured in a concentration measurement. Glucose monitoring succeeded not because all single-molecule measurements are inherently useful, but because glucose happens to be one of the few molecules whose concentration alone carries robust meaning across most contexts. The practical implication here is that single-analyte molecular biosensors will have limited utility until we can measure enough proteins simultaneously to capture meaningful biological context. The technical achievement of measuring proteins like IL-6 in interstitial fluid is genuinely impressive, but the value proposition may be more limited than current market enthusiasm suggests. At the same time, the longer-term potential once multiplexing becomes feasible may be under-appreciated.

Building a multiplex molecular biosensor requires substantial capital, sophisticated engineering, and regulatory pathways that are still being established (and rapidly changing at the time of this writing), creating an interesting challenge for technology development. For a small biotech company, developing a single-analyte sensor is the more tractable path. You can focus on one molecule, establish a clear use case, validate the technology, navigate regulatory approval, and potentially become an acquisition target for established players like Dexcom or Abbott, among others. This pathway is already clear, and it makes business sense. Major continuous monitoring companies have strong incentives to expand into molecular sensing. But there’s a potential issue with this trajectory. When these technologies get integrated into existing platforms, they’re likely to be presented as independent metrics—just another number on the dashboard—rather than as nodes in a network that require contextual interpretation. This reflects the current business model of these companies, which centers on providing users with actionable numbers they can understand and respond to.

The alternative would be to develop multiplex sensing from the outset, but this requires investors who understand why single-analyte approaches might not deliver the anticipated value and who are willing to fund a longer, more capital-intensive development path toward a market that doesn’t yet exist. This is a harder pitch in the current early 2026 funding environment.

That said, there may be alternative entry points worth considering. Critical care medicine already involves multiplex measurements through continuous monitoring combined with frequent blood tests. Clinicians in intensive care settings are accustomed to thinking about complex physiological states rather than individual metrics. A multiplex interstitial fluid sensor that helps predict complications like sepsis before vital signs change would have clear clinical utility without requiring a shift in how practitioners interpret biological data. Once proven in this context, the technology could potentially scale to consumer applications. This inverts the typical medical device development trajectory, but it might better match the technology to contexts where its value is most apparent.

One concern about the incremental single-analyte path is that design decisions made early on can constrain future development—a topic I discussed in Issue #48: The Garden of Technological Possibilities. Sensor architecture, regulatory strategy, manufacturing processes, and user interfaces optimized for single measurements may not easily extend to multiplexing. Adding measurements later isn’t just an engineering challenge; it may require rebuilding substantial portions of the technology stack and rethinking how information is presented to users.

When multiplex molecular biosensors do become practical, they’ll enable some genuinely new capabilities. Rather than reporting that a single biomarker is elevated, they could describe an individual’s network state and how that relates to physiological outcomes for that specific person. We could generate personalized protein expression networks, identify which regulatory hubs are most important in an individual’s biology, and potentially detect transitions between physiological states before conventional metrics change. This represents a form of precision medicine that’s complementary to genomics—these networks are dynamic and change with disease progression, treatment, and recovery.

We’re not at that point yet. The current development trajectory shows many parallels to early genomics, where there was an assumption that identifying components would quickly translate to understanding systems. The Human Genome Project, for all its triumph, didn’t immediately yield the therapeutic revolution that many expected—because knowing the parts list doesn’t tell you how the machine works. We’re building increasingly sophisticated tools to measure individual proteins while grappling with the question of how much information a single measurement actually provides when biological function emerges from molecular relationships.

The question facing the field is not whether we’ll eventually need multiplex approaches—the biology makes that clear. The question is whether we’ll recognize this early enough to avoid spending years and considerable resources developing single-analyte sensors that provide limited actionable information. From where I sit, the current trajectory looks misguided rather than suboptimal. The industry is building toward a capability that glucose monitoring seems to validate, but glucose is the exception that proves the rule. Most proteins require network context for meaningful interpretation, and no amount of sophisticated algorithm development or machine learning will extract signal that isn’t there in the first place.

Investors and companies have a choice: continue down the incremental path of single-analyte sensors because it fits established business models and funding timelines, or acknowledge that we’re working at a different layer of biological abstraction that demands different technological and interpretive approaches from the outset. The technical achievements in molecular sensing are real and impressive. The question is whether we’re building the right things, or just the things that are easiest to build and sell with our current conceptual frameworks.

Thanks for reading! If you found this post useful, please consider subscribing. I share hands-on computational biology techniques, fresh ways to think about tough problems, and perspectives on a range of related topics. All free, straight to your inbox.

Liked this piece? If so, tap the 🖤 in the header above. It’s a small gesture that goes a long way in helping me understand what you value and in growing this newsletter. Thanks so much!

Issue № 64 // Lossy Compression: What RNA-seq Doesn’t Tell You About Cellular State

In Manolis Kellis’ course Computational Biology: Genomes, Networks, Evolution, he states, "The most intuitive way to investigate a certain phenotype is to measure the expression levels of functional proteins present at a given time in the cell. However, measuring the concentration of proteins can be difficult, due to their varying locations, modifications, and contexts in which they are found, as well as due to the incompleteness of the proteome. mRNA expression levels, however, are easier to measure, and are often a good approximation. By measuring the mRNA, we analyze regulation at the transcription level, without the added complications of translational regulation and active protein degradation, which simplifies the analysis at the cost of losing information." This notion is easy to lose sight of when doing RNA-seq analysis. But consider what it means to lose information in the Shannon information theoretic sense.

In Claude Shannon’s framework, information is about uncertainty reduction. Each measurement collapses the space of possible states a system could occupy. By stopping at the transcriptional level, we deliberately remain uncertain about an entire regulatory layer. The post-transcriptional processes like alternative splicing, protein modifications, localization, and degradation represent additional sources of information that specify the actual functional state of the cell. The mutual information between mRNA and protein levels is often low to moderate, with correlations around 0.4-0.7, as depicted in the images below, meaning substantial uncertainty remains about the actual effector molecules (as a reference, a correlation of 0.5 between mRNA and protein means ~75% of the variance in protein levels remains unexplained by transcriptomics)1. Some proteins persist for days while their mRNAs turn over in minutes; others are rapidly degraded despite abundant transcripts. Additionally, post-translational modifications can completely alter function without any transcriptional change.

This is where functional proteomics comes in, both as a means of getting closer to a tissue’s phenotype and as a systematic uncertainty reduction tool. Consider HER2-amplified breast cancer. Fluorescence In Situ Hybridization (FISH) can identify gene amplification, giving us a binary answer that reduces uncertainty about genomic state but leaves open whether HER2 drives the tumor’s biology or predicts therapeutic response. Transcriptomics then answers whether HER2 mRNA is elevated, collapsing another degree of freedom, but leaves open the question of whether HER2 protein is abundant. Mass spectrometry proteomics then confirms protein abundance, but functional activity requires something like a reverse phase protein array (RPPA): a high-throughput antibody-based technique that measures protein expression and post-translational modifications, measuring phosphorylated HER2 states. Each layer answers a yes/no question, progressively narrowing the space of possible cellular states until we understand why, within a HER2-amplified population, only a subset responds to targeted therapy.

A simpler way to think about this is as follows: before data collection, we don’t know if a tumor has HER2 gene amplification, whether HER2 gene expression is elevated, or HER2 proteins are abundant and if they are, which sites are phosphorylated, reflecting their activity. Each measurement we take answers one yes or no question. FISH shows HER2 amplification; that’s one point of uncertainty reduced. RNA-seq shows high HER2 expression—we have even less uncertainty about the tissues phenotype. But, it’s not until we move further down the hierarchy that uncertainty is reduced to such a degree that we can make a firm conclusion one way or another2.

The information flow also runs in reverse. Starting from phospho-HER2 elevation and working back through transcriptomics to genomics reveals mechanism—and mechanism determines treatment strategy. High HER2 protein resulting from gene amplification signals oncogene addiction, making pathway inhibition viable. High HER2 protein from impaired degradation, without amplification, suggests the protein may not be a driver, favoring HER2 as a target for antibody-drug conjugates rather than kinase inhibition. The phenotype is identical, but the causal path matters.

Note: Absence of gene amplification doesn’t definitively rule out oncogene addiction—elevated HER2 could still drive proliferation through other mechanisms (e.g., autocrine signaling, receptor heterodimerization). However, it shifts the prior probability: without amplification, HER2 is more likely a passenger or consequence of upstream dysregulation rather than the initiating driver. This probabilistic distinction matters for treatment selection—oncogene addiction predicts durable responses to pathway inhibition, while passenger alterations favor using the protein as a delivery target rather than inhibiting its function. For more on this topic check out Issue #58 // How to Kill a Tumor: Three perspectives on rationale drug selection in personalized cancer therapy.

Understanding this through information theory makes explicit what we’re assuming in our experimental designs. RNA-seq gives us a compressed representation of cellular state—one that is remarkably productive for discovery, suggesting transcriptional regulation captures much of the cell’s control logic. But, the practical convenience of RNA-seq shouldn’t obscure the fundamental gap between what we measure and what actually drives biology. We should recognize when this lossy compression is acceptable and when we need measurements closer to the phenotype itself. In data-driven research, moving down the biological hierarchy reduces uncertainty about what the data mean. In clinical practice, moving up the hierarchy reduces uncertainty about how to intervene. These directions are complementary faces of the same information-theoretic process.

Data-driven research follows the path from genotype to phenotype, asking “what is the cellular state?” to interpret measurements correctly. Clinical practice reverses this, starting from phenotype and tracing back to mechanism, asking “why is this the state?” to select interventions. Both navigate the same multilevel system, but the direction of inquiry determines what uncertainty gets resolved—interpretive uncertainty versus mechanistic uncertainty. Precision medicine requires both: characterizing the actual state (down the hierarchy) and understanding its causal origin (up the hierarchy).

Thanks for reading! If you found this post useful, please consider subscribing. I share hands-on computational biology techniques, fresh ways to think about tough problems, and perspectives on a range of related topics. All free, straight to your inbox.

Subscribe now

Enjoy this piece? Whow your support by tapping the 🖤 in the header above. It’s a small gesture that goes a long way in helping me understand what you value and in growing this newsletter. Thanks!

Issue № 63 // When Absence Is Information—Missing Data in Proteomics

I first started coding seriously during COVID-19 lockdowns, shortly after co-founding a wearable technology company, NNOXX. At the time my primary interest was analyzing time-series physiological data, which quickly evolved into building ML-based models to predict fatigue status, MSK injury risk, and other factors using distributed sensor networks. Early into my PhD though, my interests shifted from applied machine learning—building and deploying models for tasks like prediction, decision making, and pattern recognition—to data-driven research, specifically focusing on computational cancer biology. With this transition came unanticipated second-order effects.

When I was focusing on ML, I spent a lot of time thinking about how to prepare and preprocess data in such a way as to best exposure the underlying structure of a given prediction problem to an algorithm to achieve the best performance. Then, when my attention shift to research, I found myself thinking more about the nuances of how to best clean data to preserve biological signal without introducing noise or spurious association in down stream analysis. In theory, these two things aren’t so different—after all, data cleaning is data cleaning—but, as I’ve recently come full circle something clicked for me. Perhaps this is obvious to some of you reading, but there are different constraints and tradeoffs when cleaning or preparing a given dataset for applied ML and data-driven research. In this piece I’m going to discuss some of those differences as it applies to working with proteomic data, specifically in how we handle missing data. But, before getting into the weeds I want to back up and clarify what I mean when I say data cleaning.

At the simplest level, data cleaning is the process of identifying and correcting systematic errors in your data. For example, data cleaning could include correcting mistyped data, removing corrupted, missing, or duplicate data points, and sometimes adding missing data values back in. Data cleaning may also include tasks such as encoding categorical variables into binary variables (i.e., turning no/yes into 0/1). For what may seem like a rote task, data cleaning actually requires a fairly high degree of domain expertise. For example, let’s say you’re working with a dataset where vO2 (volume of oxygen consumption) is one of your features and you notice a row where the value is 92 ml/kg/min. While not impossible, as a subject matter expert you may realize how improbably this value is, whereas someone without that expertise may overlook it.

Similarly, when looking at missing values in proteomic data, you may spot a pattern to the missingness, which can give clues as to whether these values should be dropped, whether you can impute values in, or whether you should leave them be. For example, you may notice that in a dataset of 500 breast cancer patients there are NaN values for the protein ERBB2 across ~15% of the samples. Upon closer inspection, you may see that those samples correspond to patients classified as HR+HER2- and HR-HER2-, suggesting that missingess is due to protein expression being below the detection threshold. Alternately, suppose you saw these same ERBB2 values missing, but with no clear separation by patient classification. In this case, maybe those NaN samples happened to correspond to patient samples (stored in rows) where many of the corresponding protein columns are blank. This might suggest that there was not enough tumor sample to process, and as a result this type of "missingeness" is technical, not biological.

Notice that identifying the cause of missing data isn’t just about looking at the data and finding patterns. You also have sufficiently understand the experimental context, the underlying biology, and potential sources of technical failure, whether during biological sample acquisition or sequencing. It’s for these reasons that I don’t advise fully automating data cleaning or outsourcing it to coding agents. While they may better understanding the statistical nature of your dataset, they often lack the context dependence to (a) understand the cause of missingness and (b) how to address it in a specific problem scenario. Importantly, these are two different skills; understanding the biological of your experimental condition and the sequencing technology generating your data allows you to solve the first problem. But, in order to figure out how to proceed after figuring out why your data is missing you need to understand the assumptions of the analyses you plan to perform.

Unlike missing data in many other contexts, missingness in proteomics is seldom random, and when it is it’s often hard to identify. Additionally, the appropriate strategy for handling different types of missingness differs between machine learning applications (where the goal is prediction) and data-driven research (where the goal is biological interpretation). This creates a tension between best practices for avoiding data leakage during ML applications (which typically means handling missing data after splitting the dataset) and the practical realities of working with proteomics data. In proteomics experiments, missing values typically result from one of two sources:

• Missing completely at random (MCAR): Some samples may have a high proportion of missing protein measurements (e.g., >50% of proteins unmeasured) due to insufficient tumor material, poor sample quality, or technical failures during sample processing. These samples represent failed experiments rather than a biological signal and should be removed from the dataset entirely before model development if the goal is prediction (this removal is a quality control step and does not constitute data leakage).

• Missing not at random (MNAR): Individual proteins may be missing in specific samples because their expression levels fall below the instrument’s detection limit. This is a common source of missingness in proteomics data and is informationally rich—the absence of a measurement often indicates very low or absent protein expression, which may be biologically and clinically relevant. For example, the absence of a measurable signal for an oncogene in a tumor sample might indicate good prognosis.

For machine learning and predictive modeling applications a standard approach is to dealing with for MNAR-based missingness is to impute missing values with a small fraction of the protein-specific minimum detected value (typically 10% of the minimum for linear-scale data, or minimum minus 1-2 units for log-transformed data). This approach is biologically justified because it represents a plausible lower bound for expression (the protein is present but below detection) and preserves the biological information that "this protein is very low in this sample". Additionally, this approach is protein-specific, respecting that different proteins have different detection ranges and avoids the assumption that missing = zero, which would often be an incorrect interpretation. For log-transformed proteomics data (such as log2-scaled measurements), imputing with the protein-specific minimum minus 1.0 in log space corresponds to expression approximately half as abundant as the minimum detected level in linear space, which represents a biologically reasonable lower bound.

Did you find this piece helpful? Subscribe for free to have the next post delivered to your inbox:

Subscribe now

The challenge for machine learning workflows is that this imputation ideally should occur before train-test splitting, which appears to violate the principle of preventing information leakage from test to training data. However, this pre-split imputation is widely implemented in the proteomics biomarker literature for several reasons. For starters, the imputation is based on biological knowledge (below detection = very low) rather than statistical properties of the full dataset. Additionally, the calculation uses only protein-specific minima, not cross-sample statistics that would truly constitute leakage. Most importantly though, attempting to impute separately within train/validation/test splits can lead to different imputation schemes for the same protein across splits, which is biologically inconsistent.

Of course, there is an alternative to imputation. We could also just remove all proteins with any missing values. The problem with this approach is that in most cases it would eliminate nearly the entire dataset as some degree of below-detection missingness is nearly universal in proteomics. As a result, for machine learning applications, I often use the following three stage approach:

• Stage 1: Remove samples with >20% missing proteins (quality control for MCAR) before splitting. This stage removes samples with insufficient tumor material or severe technical failures, excluding poor-quality samples that don’t represent successful experiments.

• Stage 2: Remove proteins with >20% missing values before splitting. After removing poor-quality samples, this stage identifies and removes proteins that are frequently unmeasured across the remaining high-quality samples. Proteins missing in more than 20% of samples likely reflect unreliable assays or technical measurement issues (MCAR) rather than biological signals. This ensures that the feature set consists only of proteins that can be reliably measured (after removing bad samples, high missingness in a protein indicates a fundamental measurement problem rather than sample-specific issues).

• Stage 3: Impute remaining sparse missing values (representing MNAR below-detection measurements) before splitting. At this point, remaining missing values are sparse (typically <10% per protein) and likely represent true biological phenomena where specific samples have expression below the detection limit. For log-transformed data, impute with the protein-specific minimum log value minus 1.0 (corresponding to half the minimum expression in linear space). For linear-scale data, use 10% of the protein-specific minimum. While this technically uses information from the full dataset, it is justified by the biological nature of the imputation and the practical impossibility of the alternative (we can’t train models with missing values).

For hypothesis-driven research and differential expression analysis (DEA), the approach to handling missingness is fundamentally different. DEA methods are specifically designed to handle missing values appropriately, and many can accommodate proteins with missingness in a subset of samples. Importantly, because the goal is to explain biology rather than build a predictive model, imputation is often undesirable—introducing imputed values can obscure true biological patterns and create artificial signals. Instead, a Perseus-style filtering approach is typically preferred: remove proteins with excessive missingness (e.g., missing in >70% of samples within a condition or across all samples, for example) while retaining proteins with moderate missingness that still provide biological information. This preserves the interpretability of the results while ensuring adequate data quality.

For normal data-driven analyses with continuous expression data, I handle zeros (representing samples with expression too low to detect) differently depending on context. When I have untransformed data and I’m confident that zeros represent true below-detection expression rather than technical artifacts, I typically impute these with 10% of the protein-specific minimum value. However, if the source of missingness is ambiguous—I cannot distinguish whether it results from biological low expression or technical failure—I do not impute. Similarly, if the data has already been transformed before I receive it, I avoid imputation since I lack the information needed to apply biologically appropriate imputation strategies.

Regardless of how you handle this approach you have to be clear in documenting your rationale, explaining all three types of missingness (sample-level MCAR, protein-level MCAR, and MNAR) and justifying your choice of imputation versus filtering based on your analytical goals. For machine learning applications, You should also justify the pre-split imputation based on the MNAR nature of below-detection measurements and acknowledge that this represents a compromise between statistical ideals and biological realities . When possible, it’s also helpful perform sensitivity analyses showing that your results are robust to alternative imputation strategies. For other omics data types (transcriptomics, metabolomics), similar principles apply, though the specific imputation strategies may differ based on the measurement technology and the meaning of missingness in that context. The key is always to understand the biological and technical sources of missingness in your specific dataset and to choose imputation strategies that respect the underlying biology while minimizing the risk of introducing spurious patterns that could compromise model generalization.

Did you find this piece helpful? If so, tap the 🖤 in the header above. It’s a small gesture that goes a long way in helping me understand what you value and in growing this newsletter. You can also subscribe for free to have the next post delivered to your inbox:

Subscribe now

Issue № 62 // Collider Biases In Proteomics Research

This past week my friend Tommy Tang shared a post on collider biases in genomics research. Collider bias is a statistical distortion that occurs when a variable that is a common effect (a "collider") of two other variables is inadvertently controlled for, creating a spurious association between those two variables. Tommy’s post got me thinking, as it seemed like a potential explanation for something I recently observed in my own research.

First, to give some background, androgen receptor S650, estrogen receptor alpha, and cyclin D1 expression have previously been identified as biomarkers of global treatment resistance among high risk breast cancer patients as reported by Gallagher et al., in Protein signaling and drug target activation signatures to guide therapy prioritization: Therapeutic resistance and sensitivity in the I-SPY 2 Trial. Specifically, when comparing responders versus non-responders in the ISPY-2 trial participant population—where responder is defined by achieving pathologic complete response (pCR) following treatment—these proteins were both up-regulated in non-responders and associated with an increased risk of recurrence (shorted distance recurrence free survival).

An important note is that despite the ISPY-2 trial being limited to high risk breast cancer, the patient population in the study is fairly heterogeneous, including a mix of patients with HR+/HER2-, HR-/HER2-, HR+/HER2+, and HR-/HER2+ tumors1. Among these HR/HER2 status-based cohorts in the ISPY-2 trial, the HR+/HER2- group had the lowest pCR rate2, and as a result a significant fraction of non-responders in the overall population (approximately ~45%) came from this group. So, it was to my surprise when I was working with reverse phase protein array data from this study, available on the Gene Expression Omnibus via accession number GSE196093, that I found that androgen receptor S650, androgen receptor total, and estrogen receptor alpha S118 were positive prognostic markers in the treatment-refractory HR+/HER2- patient population, with increased expression of all three proteins being associated with a lower risk of recurrence, as demonstrated by the survival curves below.

To make it more explicit, here we have a case where AR and ER are up-regulated in non-responders versus responders and are demonstrated to be negative prognostic markers across then entire patient population (i.e., associated with worse survival outcomes). However, within the non-responder patient population, these same proteins are positive prognostic markers and are associated with more favorable survival outcomes.

Now, because cancer is a highly heterogeneous disease, we do also need to consider the fact that the prognostic value of androgen and estrogen receptor expression depends on disease context and molecular subtype. Previous survival analyses have shown that AR acts as a tumor suppressor in ER-positive breast cancers, making it a favorable prognostic marker, and a tumor promoter in ER-negative breast cancers, including both ER-/HER2+ and triple negative breast cancer, making it a poor prognostic factor in these contexts (You et al., 2022). This biological heterogeneity could explain part of what we’re observing: the HR+/HER2- non-responders who express higher AR/ER may represent a biologically distinct subgroup where these receptors retain their tumor-suppressive functions despite treatment resistance in this population. This raises an important question: is the observed association in HR+/HER2- non-responders spurious (collider bias) or real, but context-dependent (effect modification)? Can it be both?

The answer to the above question is yes. Collider bias and real effect modification are not mutually exclusive. Collider bias is a statistical structure issue—conditioning on treatment response can induce spurious associations or amplify existing associations. Effect modification, in contrast, reveals a true biological signal. In this case, the causal effect of androgen and estrogen receptor expression genuinely differs by context (breast cancer subtype, tumor stage, treatment arm, etc.), suggesting effect-modification is in play. However, this doesn’t mean that a collider bias isn’t also present. We know that AR/ER have different prognostic effects in hormone receptor positive (HR+) versus hormone receptor negative (HR-) breast cancer, but by conditioning on treatment response (pCR status in this case), we’re selecting a non-random subset of patients. This selection process can still induce or amplify associations beyond what the true biological effect would predict. As a result, even where there’s a real context-dependent biological effect, collider bias can distort the magnitude of effect we observe in the conditioned subset. Thus, the association we observe might be partly real (effect modification) and partly artifactual (collider bias).

This raises the question, how do we determine the degree to which what we’ve observing is collider bias versus real effect? One way to do this is to analyze the protein-outcome association separately within each HR/HER2 subtype in the overall population, then compare these associations with what you observe when you condition on non-responders within each HR/HER2 subtype individually. In my case, I first looked at whether AR/ER are positive or negative prognostic markers within each HR/HER2 subtype population overall. This yielded the following results:

• Overall HR+/HER2- population → AR/ER = positive prognostic

• Overall HR-/HER2- population → AR/ER = negative prognostic

• Overall HR-/HER2+ population → AR/ER = negative prognostic

• Overall HR+/HER2+ population → AR/ER = no significant associations

Next, I did a non-responder only analysis within each of these subtypes:

• HR+/HER2- Non-responders only → AR/ER = positive prognostic

• HR-/HER2- Non-responders only → AR/ER expression is still unfavorable

• HR-/HER2+ Non-responders only → AR/ER expression is still unfavorable

• HR+/HER2+ Non-responders only → still no significant associations

Notice, the directional associations did not change after conditioning, which suggests that the biological effect we’re observing is real. In other words, androgen and estrogen receptor expression are actually associated with better survival outcomes in the treatment-refractory HR+/HER2- patient population.

To further parse this effect we can also look to see if there is a negative correlation between androgen and estrogen receptor expression and other predictors of non-response, even if these predictors are positively correlated with androgen and estrogen receptor expression in the overall study population. The logic here is that when you condition on a collider (treatment non-response), you create ratification negative correlation between the causes of that collider, even if they are positive or independent in the overall population. Specifically, when conditioning on non-response, we should expect to see that non-responders (patients with high AR/ER expression in this case) will tend to have lower levels of other non-response predictors such as Ki67 levels3. When I tested this with the aforementioned dataset, this does not appear to be the case. Instead, there is no significant association between Ki67 and treatment response in the HR+/HER2- population.

So, back to Tommy’s original post, understanding collider biases and how to disentangle statistical artifacts from true biological signals in an important skill for computational biologists. This is even more so in computational cancer biology, where context-dependence is the norm. We’re constantly faced with scenarios where a given protein acts as a tumor promoter in one context but a suppressor in another, or where a protein signature predicts treatment response in one molecular subtype but non-response in another. As a result, we need to pay careful attention to how studies are designed, what variables are conditioned on (or not), and how our analytical choices may distort true biological signals or introduce spurious associations. The I-SPY 2 dataset is a great testbed for these types of analyses because of its rich omics and clinical data, but we’re rarely so fortunate in practice. In most cases, we lack comprehensive clinical outcome data, detailed patient stratification, or sufficient demographic information. As a result, we need to tread carefully when interpreting our data, assess whether our findings align with the broader literature, and temper our confidence when we lack the data needed to properly disentangle collider bias from effect modification.

Thanks for reading! If you found this post useful, please consider subscribing. I share hands-on computational biology techniques, fresh ways to think about tough problems, and perspectives on a range of related topics. All free, straight to your inbox:

Subscribe now

Issue № 61 // Orthogonal Validation Over Replication

I recently shared the following note, which was a reflection on the fact that molecular biology research typically relies on indirect measurement techniques to infer the presence, abundance, and activity of genes, proteins, and metabolites (as well as the downstream consequences).1

Anyone who works with biological data knows this intuitively, yet it’s easy to forget that our measurement technologies themselves are not perfectly transparent and can distort our observations in subtle ways. In computational biology we address this issue through quality control, normalization, and batch correct. But, the unsettling reality is that a portion of what we consider to be biological signal might reflect variations in how different samples interact with our measurement technologies2.

For example, RNA-sequencing is prone to GC content bias and amplification preferences, stemming from the physicochemical properties of DNA/RNA, which affects which transcripts we detect and how we quantify them. Additionally, mass spectrometry-based proteomics favors certain peptides based on ionization efficiency and often misses low-abundance proteins entirely, while antibody-based methods depend on epitope accessibility that may not reflect true protein abundance or activity. These types of errors aren’t random and as a result they don’t just average out given enough data. They’re systematic biases inherent to how each measurement technology interrogates biological systems.

This brings us to something I’ve been thinking a lot about recently, which is that when different sequencing runs, technicians, or reagent lots produce different results, we often attribute it to batch effects that need correction. But, batch correction, by definition, can only address variations between uses of the same measurement technology. It doesn’t, and cannot, address the inherent way that technology affects what we ultimately measure. We’re always measuring biology filtered through the lens of our chosen instrumentation, limitations and biases included, and no amount of normalization or batch correction can remove these types of systematic distortions. If this is true—and I’m open to the possibility that my logic is flawed—then it leads to the realization that we often can’t distinguish between measurement technology-intrinsic bias and true biological signals when working within a single measurement modality. Using the RNA-sequencing example again, consider a given reproducible finding… can we be certain we’re identifying true transcriptional regulation, or is it possible we’re capturing a systematic bias in how RNA-seq captures certain transcript features?

When we apply batch correction, we make certain assumptions about what constitutes technical versus biological variation. However, these assumptions can’t be validated within a given measurement technology. Additionally, if biological and technical variation correlate (for example, diseases tissue being processed differently than controls, samples being collected at different time points, population structure confounding with processing batch, etc), our corrections may inadvertently remove real biological signals or introduce spurious patterns.

This is where an indiscriminate fixation on scientific reproducibility, which while foundational to science, can introduce a blind spot. We treat replication across labs, datasets, and study cohorts as the gold standard for biological truth, but reproducible doesn’t always mean correct—it could mean that our systematic errors are consistent. For example, if a given set of measurement protocols introduce similar biases and downstream analysis pipelines make the same normalizations assumptions we’ll get perfectly reproducible results that all reflect the same technical artifacts. As a result, some of our most reproducible findings—the ones that replicate across dozens of studies—might be measuring properties of our measurement technologies rather than properties of biology.

Of course, not all reproducibility is created equal. Replication using identical protocols in different labs is weaker evidence than replication using varied protocols within the same measurement technology. But both are fundamentally limited by the fact that they're looking at biology through the same technological lens. The solution to this problem isn’t more reproducibility through the same analytics lenses—It’s orthogonal validation using fundamentally different measurement approaches. When RNA-sequencing, high-throughout proteomics, and functional assays all independently point to the same conclusion despite having different technical biases you’re likely to have triangulated in on a true biological signal3.

This is also why the reflexive integration of multi-omics data can be counterproductive. Sometimes we don’t want to merge transcriptomics and proteomics into a unified model—we want each to speak independently, then ask whether they’re telling the same story. If they converge despite looking at biology through entirely different technical lenses, the signal is real. If they diverge, we’ve learned something about either the biology or our measurement limitations. The goal isn’t consensus through integration; it’s triangulation through independence4.

Liked this piece? If so, tap the 🖤 in the header above. It’s a small gesture that goes a long way in helping me understand what you value and in growing this newsletter.

Liked this piece? If so, tap the 🖤 in the header above. It’s a small gesture that goes a long way in helping me understand what you value and in growing this newsletter. You can also subscribe for free to have the next post delivered to your inbox:

Subscribe now

Issue № 60 // Compression As Understanding

The most useful computational models in biology aren’t the ones with the most parameters—they’re the ones that achieve the most compression. When you can predict gene expression across dozens of conditions using only a handful of latent factors or explain variations in protein abundance with a small set of regulatory relationships you’ve found something closer to biological "ground truth" than any super high-dimensional model that just memorizes your training data and spits out a prediction, even if it’s more accurate (as defined in a specific testing context).

Beyond their use as data visualization tools, dimensionality reduction techniques—like PCA, t-SNE, and UMAP—can be used to parse the underlying structure of biological systems, which I wrote about previously in Issue #36 // Machine Learning: The Native Language of Biology, stating:

"In a neural network, high-dimensional input data is compressed into latent representations— abstract patterns that capture essential features while discarding noise. Similarly, a transcription factor distills complex environmental information into a binary state (active or inactive) that the cell can use to make decisions. This concept of latent spaces offers another window into why machine learning aligns so well with biology. In cell biology, high-dimensional data like gene expression profiles or microscopy images can be projected into lower-dimensional spaces that capture meaningful biological variation. Each dimension in this latent space ideally corresponds to some biological process or state—cell cycle phase, differentiation stage, or stress response."

If 20,000 genes can be projected into 50 principal components that capture most of the variance, this tells us the genome isn’t actually operating in 20,000-dimensional space. It’s operating in something far lower-dimensional, with the vast majority of genes acting as downstream readouts of a smaller set of regulatory "programs".

The same idea can be applied to patient stratification. Take breast cancer, for example. We know every patient’s cancer is unique with its own gene mutations and amplifications, alterations to transcriptional programs, behaviors, and responses to treatment. This idea is encapsulated by the Anna Karenina principle in Oncology, which states something to the effect that all healthy tissues are alike, whereas all diseased tissues are diseased in their own way. Yet, if we can compress thousands of clinical and molecular features into a small number of subtypes that meaningfully predict treatment response (such as ER/HER status, PAM50 subtypes, MammaPrint risks profiles, etc), there is an argument to be made that despite enormous molecular complexity, the relevant biological variation between tumors has low intrinsic dimensionality1.

However, there is a strong argument to be made that compression is most useful when it leads to increased understanding of the underlying biological systems you’re working on. An auto-encoder can compress gene expression data perfectly, while learning latent representations that cannot be biologically, or mechanistically, interpreted. The representations that matter are the ones that align with biological processes we can observe, measure, or manipulate because those are the dimensions along which we can actually intervene.

This is why interpretable models often outperform black boxes in biological and clinical applications/use cases, even when they’re less accurate. A decision tree that says "if local tissue hypoxia and immune infiltration, then poor prognosis" is testable biology. A neural network that achieves 2% better accuracy but can’t explain its reasoning in biological terms may just be performing complex pattern matching. In this way, the goal of computational modeling isn’t prediction for its own sake. It’s to build models that can compress biological complexity and heterogeneity in ways that suggest new experiments, reveal potential mechanisms or action, and ultimately show us levers that we can pull to intervene in disease.

George Church recently gave a great interview on this topic for the Lifespan Research Institute which you can find here. I’ve reposted a short except below:

About that: it’s always interpretability versus the model’s power. Where are you in this debate? Would you prefer a weaker but more interpretable AI or a stronger but less interpretable one?

GC: I lean on the interpretability side. It’s not an either-or, but… we’re in science. Few engineers are willing to just pull a rabbit out of a hat, just a black box. Scientists and engineers, by and large, want to know the mechanism. The FDA likes to know mechanisms. Typically, the autocatalytic loop where you learn something and then you invent something is better if it’s mechanistically grounded. So, I lean pretty heavily in the direction of interpretability, explainability, transparency, et cetera, and also it’s safer.

I just honestly think that we will soon be faced with this dilemma, where we will have to choose between the power of the model to do things and its actual interpretability, but maybe we’re not there yet.

GC: If you look at the human scientist experience, the most powerful sciences are the ones that are better articulated mechanistically on a solid foundation rather than black boxes. The black boxes tend to include artifacts, dead ends. Most of the progress in science and engineering has been part of community efforts with strong mechanistic underpinnings.

These ideas stand opposed to the core thesis of predictive biology, which argues that predicting the outcome of an unknown experiment is equivalent to understanding a system. While beyond the scope of this article, the tension between "mechanism as understanding" and "prediction as understanding" is interesting to reflect on. As a biologist, I tend to lean towards the former.

This tension has played out in my own research a bit, where I’ve worked with both Cypher query-based graph traversal methods and graph neutral network-based approaches to predict phenotypes from molecular perturbation data. The former is more interpretable, allows for novel pathway discovery, and provides more causal information, which both makes it more trustworthy from a clinical standpoint. Yet, the GNN based approaches often make predictions that the traversal based ones miss, which while hard to validate the utility of are interesting nonetheless. Of course, there are ways to take the best of both of these worlds and create a hybrid system, but at the end of the day we still need to decide how much we’re willing to trust predictions we don’t fully understand, and the degree to which we need our models to be mechanistic and interpretable for them to be useful.

Enjoy this issue? If so, you can show your support by tapping the 🖤 in the header above. It’s a small gesture that goes a long way in helping me understand what you value and in growing this newsletter. Thanks!

Issue № 59 // Castles on Quicksand: A Consumer Health Tracking Story

Last month, a friend showed me the Whoop app’s latest feature with a level of enthusiasm typically reserved for winning the lottery. "Look!," he said, pointing to a number on his iphone screen, "My biological age is three years younger than my chronological age!"

Hearing my friend wax poetic about his Whoop Age left me in an uncomfortable position. As someone who founded a wearable biosensor company and understands the limits of current sensing modalities—and actively works on developing new protein biomarkers—I had a strong inkling that this measurement was less than accurate, to put it kindly. After all, his Whoop wasn’t measuring his telomere length, DNA methylation patterns, or protein aggregation markers. It was measuring his heart rate, HRV, blood oxygenation, skin temperature, and movement, then applying algorithms trained on population-level data to generate what appeared to be a legitimate biological assessment. This disconnect between the information Whoop claimed to provide and what their devices actually measures was surprisingly to say the least1.

Despite my strong opinions about the validity of some of Whoop’s digital biomarkers, I don’t believe the company is malicious, nor do I think they intend to deceive consumers. What I do believe is that they, and other consumer health companies, are operating in a hyper-competitive environment where the reward for sensational and easily marketable features is increased consumer adoption and the penalty for conservative claims is irrelevance. In this environment the question isn’t whether any individual company will make unsubstantiated claims or market a feature their technology can’t support—instead, it’s who will be the first to do it, and how far they can stretch these limits until consumers loose faith in their technology.

To fully appreciate why the current generation of consumer health devices represents such a profound case of measurement mismatch, we need to start with what we know about human biology and health from a bottom-up perspective. Take anti-aging—the poster child for speculative consumer health claims. The past two decades of research, propelled by next generation sequencing, have shown aging to be a complex multi-system phenomenon involving cellular senescence, mitochondrial dysfunction, protein aggregation, genomic instability, and epigenetic drift. All of which are processes several layers of biological abstraction removed from anything a wrist-worn optical sensor can detect2.

The same principle applies to brain health and immune function among other areas of health. The most meaningful biomarkers exist at the levels of gene expression, protein abundance, metabolite concentrations, and methylation patterns. Measuring these things accurately requires high-throughput RNA sequencing, mass spectrometry, immunoassays, imaging techniques, and other technologies. A device that measures heart rate, skin conductance, and movement patterns is an entirely different universe of biological information.

This isn't to say that the physiological parameters measured by consumer devices are meaningless or lack utility. Heart rate variability does correlate with autonomic nervous system function, sleep architecture does relate to metabolic health, and activity patterns do influence cardiovascular risk. But correlation is not causation, and correlation is certainly not equivalent to direct measurement. When devices measure heart rate variability and claim to be assessing "recovery" or measure blood glucose and claim to measure "metabolic health" their providers are making conceptual leaps over gaps that cannot be spanned by objective fact.

Whoop’s "Pace of Aging" score provides a perfect case study into how this measurement gap manifests itself in practice. Whoop, as company, has positioned its product as a leading form-first wearable3 and its devices genuinely excel at what their hardware is designed to measure — namely heart rate and heart rate variability.

The problems begin when accuracy with their primary sensing modality—photoplethysmography—becomes the basis for claims that venture far beyond what the underlying hardware and sensors can support. The biological age and pace of aging measurements are the most egregious examples, but they are hardly unique.

Consider what would actually be required to estimate biological age accurately. Morgan Levine and colleague’s DNAm PhenoAge, one of the more sophisticated approaches to biological age prediction, incorporates chronological age, albumin, creatinine, glucose, C-reactive protein, lymphocyte percentage, mean corpuscular volume, red cell distribution width, alkaline phosphatase, and white blood cell count in it’s model. Yet, even this approach has its limitations in that biological aging is organ-specific. Your liver, brain, and cardiovascular system don't age in lockstep, and as a result any single metric inevitably obscures important biological heterogeneity. Capturing this heterogeneity is possible, but it requires high-throughout proteomics, measuring thousands of plasma proteins simultaneously, as demonstrated in this seminal paper from of the Wyss-Coray lab at Stanford.

Whoop’s approach, by contrast, relies on algorithms trained on population-level data to identify correlations between lifestyle factors and age-related outcomes. In their own words, "Whoop Age and Pace of Aging are designed to reflect well-established links between behavior, physiology, and long-term health. While WHOOP Age correlates with perceived health and general wellness at a population level, there is no clinical benchmark for validating these metrics​."4 This approach might have some predictive value at the population level, but presenting it as a precise measurement of biological age feels a bit suspect.

The tragedy is that Whoop’s underlying, unadulterated, measurements are valuable. They don’t even require inflated claims to justify their use. So, why then do they take this approach? The answer lies in what I call the single-point sensing trap. When you’ve built your entire brand around the premise that a wrist-worn band can provide comprehensive health insights— as Whoop claims with their promise to "improve performance, build healthier habits, and extend health span with continuous health monitoring"— you’ve painted yourself into a technological corner. To admit that holistic health monitoring might require multiple distributed sensing modalities would undermine what makes your product marketable in the first place.

This constraint creates a perverse innovation cycle. Instead of expand their sensing capabilities to match their ambitious claims, companies like Whoop are forced to stretch their existing sensor data even thinner, creating increasing elaborate algorithmic castles built on quicksand. Each new feature, from sleep coaching to strain optimization to biological age estimation, represents another layer of abstraction designed to extract insights that their hardware just can’t capture.

Meanwhile, a schism is starting to open in the consumer health space between companies doubling down on single-point sensing, like Whoop, and those that are beginning to embrace a distributed sensor architecture, which I’ve previously written about in Issue #48: Distributed by Design. Oura’s recent integration with Stelo speaks to this. When Oura acknowledges that ring-based measurements need to be complemented by continuous glucose data to provide meaningful metabolic insights, they're admitting something that should be obvious but remains unspoken: human biology is too complex for any single sensing modality to capture.

The transition from single-point to multi-point sensing enables something more profound that just measurement diversity; it opens the door to discovering network biomarkers, a conceptual framework that treats biological relationships rather than isolated metrics as the fundamental unit of health assessment.

Consider what this might mean for the biological age measurement that so captivated my friend. Instead of trying to compress the complexity of multi-system aging into a single score derived from heart rate and sleep patterns, imagine a distributed sensor network that treats aging as what it actually is: a heterogeneous, organ-specific, process with complex interdependencies. Such a system might integrate data from a smartwatch or ring, continuous glucose monitor, muscle oximeter, smartphone sensors (including the camera and microphone), and periodic laboratory biomarker panels. Rather than generating disparate data streams, these inputs would feed into models designed to map causal relationships and treat the network itself as the biomarker.

Additionally, by collecting continuous digital biomarkers alongside periodic molecular measurements, we can map the causal relationships between different layers of biological organization, creating a fundamentally different paradigm for personalized health monitoring than what exists today5. For example, we could build causal models that map how changes in your diet, exercise, and daily behaviors impact sleep quality, specific inflammatory markers, and digital biomarkers like HRV (and how these things in turn impact each other, mapping the various circular dependencies and chains of cause-and-effect).

The practical applications of this approach are many. Imagine you notice your heart rate variability declining over several weeks coincident with changes in sleep architecture, more volatile blood glucose measurements following meals, and mild muscle tension dysphonia sensed via your iPhone’s microphone. Traditional approaches might flag this as "low readiness", "high strain", or some other vague wellness claim. A network biomarkers approach, on the other hand, might recognize this pattern as a potential signal of infection or systemic inflammation and recommend targeted laboratory testing for specific inflammatory cytokines, such as TNF-α, IL-1, and IL-6, not only allowing for it’s predictions to be validated, but further feeding data into the causal models. Perhaps more importantly, by mapping causal chains between all of these variables, and the relative strength of those causal relationships, we can determine the most impactful interventions, which is something I’ve written about in Breaking Biometric Babel.

This also seems to align with Oura's thinking, as indicated by Maz Brumand, their VP of product, who told CNET: "By combining Stelo data with Oura's existing insights, we're empowering members to better understand the cause-and-effect relationships between eating patterns, energy, mood and recovery and ultimately make sustainable, science-backed lifestyle changes."

Seeing one of the biggest players in the industry embrace the aforementioned approach validates what should be clear: that the measurement gap that characterizes today's consumer health landscape isn't immutable. It's the predictable result of trying to extract complex biological insights from individual sensors measuring isolated parameters. However, as the industry shifts from single-point sensing to distributed sensing networks and network biomarkers, the door to genuine biological insights opens wide.

This transition will likely create a bifurcation in the market. Companies that continue building increasingly elaborate castles on quicksand—layering algorithmic complexity onto fundamentally inadequate sensing—will find their structures collapsing under their own weight as consumers become more sophisticated about what these devices can and cannot measure. Meanwhile, those embracing distributed sensing will capture the growing segment of users who want real, actionable, health insights.

The question isn't whether this shift will happen, but how quickly the industry will recognize that solid ground lies not in more clever algorithms, but in acknowledging the inherent complexity of human biology and building sensing architectures that match that complexity.

Did you enjoy this issue? If so, you can show your support by tapping the 🖤 in the header above. It’s a small gesture that goes a long way in helping me understand what you value and in growing this newsletter. Thanks!

Issue № 58 // Høw to Kill a Tumør

I was staring at a series of protein heat maps when I realized I had a problem. Not the usual SyntaxError problem. My code ran to completion and the visualizations it generated looked fine. The problem was philosophical. Depending on which analytical lens I applied to the same clinical trial dataset different proteins appeared to be the "best" therapeutic target. The first protein that caught my attention was EGFR T1068, showing a clean eight-fold up-regulation in tumor samples compared to adjacent healthy tissue, with pathway enrichment analysis showing the activation of downstream growth signaling cascades. Kaplan-Meier and Cox hazard analysis even showed EGFR T1068 to be a strong negative prognostic marker in the trial’s patient population. This looked like textbook oncogene addiction, where cancer cells had become dependent on hyperactive EGFR signaling for their survival and proliferation. The verdict seemed clear: inhibit EGFR, break the addiction, eliminate the cancer.

But, when I shifted my focus to surface proteomics data, searching for potential antibody-drug conjugate (ADC) targets, a different protein stood out. TROP2, also known as TACSTD2, showed high tumor-specific expression despite having no obvious role in driving malignant transformation in this specific context. Yet, it was abundantly present on tumor cell surfaces, and had excellent internalization kinetics, making it an ideal postal address to deliver cytotoxic payloads to. The logic for drugging TROP2 was fundamentally different from EGFR. TROP2 didn’t need to be a driver of cancer; it just had to be present, accessible, and able to carry a lethal package across the cell membrane.

Network topology analysis revealed yet another drug target candidate entirely. A protein called NEDD9 consistently showed up as a high degree and betweenness centrality hub in tumor co-expression networks while maintaining low centrality in normal tissue networks. NEDD9 wasn't dramatically over-expressed, showing only a modest ~1.3-fold elevation compared to normal samples. But its position at the intersection of multiple signaling pathways made it a critical coordinator of cellular communication within the malignant "ecosystem". Disrupt NEDD9, the network analysis suggested, and you could trigger cascading failures throughout the tumor's organizational structure—like removing a critical router from a communication network.

Three different analytical approaches; three different answers for how to kill a tumor. Each one backed by strong statistical evidence and reasonable biological rationale, and each one representing a distinct philosophy about what makes cancer vulnerable to therapeutic intervention. Are these competing truths—or just different projections of, and perspectives on—the same underlying biology?

The multiplicity of tumor killing philosophies didn't emerge in a vacuum; it reflects the evolution of cancer biology itself. The driver-focused approach stems from the work of Bert Vogelstein and others, in the 1980’s and 1990’s, who showed that cancer progression follows predictable patterns of genetic alteration. Their discovery, that specific oncogenes like RAS and tumor suppressors like p53 are mutated across many cancer types, provided the conceptual framework for targeting the molecular engines of malignancy. This was the era when cancer research began to feel less like taxonomy and more like engineering. If we could identify the "broken parts" in a cancer cell, we could fix them, or at least jam them up enough to stop the disease.

This mechanistic worldview dominated cancer research for decades, leading to some amazing successes. When Brian Druker pioneered the research and clinical development of Imatinib (developed by Novartis) to target the BCR-ABL fusion protein in chronic myeloid leukemia, he wasn’t just advocating for a drug—he was validating an entire philosophical approach to cancer treatment1. Here was proof that if you understood the molecular driver of cancer and could drug it effectively, you could stop cancer in its tracks. The results were dramatic enough that some oncologists started talking about "functional cures" for a disease that had been a death sentence just a few years earlier. Yet, even as targeted therapies revolutionized treatment for specific cancer subtypes, researchers were discovering the limitations of pure driver-focused approaches. Cancer cells, frustratingly, have the ability to develop resistance through secondary mutations and pathway rewiring. The very genetic instabilities that create driver genes and proteins also enable escape mechanisms that render them undruggable.

The use of antibody-drug conjugates (ADCs) emerged partly as a response to these limitations. If cancer cells can adapt and evade our mechanistic understanding, perhaps it warrants a shift from disrupting their biology to simply delivering cytotoxic payloads with high precision. The philosophy here is both cynical and pragmatic—stop trying to understand what makes a cancer cell tick, and just find a way to selectively poison it. This concept culminated in the development of compounds like Trastuzumab-Emtansine, combining the anti-HER2 antibody Trastuzumab with the cytotoxic agent Emtansine, where HER2’s role as a cancer driver becomes secondary to its utility as a giant bullseye on a tumor cell’s surface.

Meanwhile, the explosion of high-throughput sequencing in the early 2000s enabled systems biologists to construct comprehensive molecular interaction networks and identify emergent properties that weren't apparent from studying individual proteins in isolation. These researchers were asking fundamentally different questions. Instead of "which protein is dysfunctional?" or "which protein is abundant?", they asked "which protein, if removed, would cause the most widespread systems failure?" Network-based approaches promised to reveal organizational vulnerabilities that traditional reductionist methods might miss—nodes whose disruption could cascade through interconnected pathways to achieve therapeutic effects that exceeded the sum of their molecular parts2.

The tension between these different tumor killing approaches became increasingly apparent as I dug deeper into the oncology literature. EGFR T1068, the clear winner from my driver analysis, has been targeted clinically for years with mixed results. While EGFR inhibitors like Erlotinib have proven to be efficacious in treating lung cancer, their performance has been less stellar in other forms of cancer, like breast cancer. Unfortunately, this happened to be the experimental condition I was working on. So, despite having strong statistical evidence suggesting a certain therapy, the underlying experimental system had too many feedback loops, compensatory pathways, or microenvironmental factors that could potentially blunt it’s impact.

TROP2, my ADC target candidate, represented a different set of trade-offs. Recent clinical trials with TROP2-targeting ADCs had shown promising early signals, but the approach carried its own risks. Any surface protein abundant enough to serve as an effective delivery address might also be expressed at low levels in healthy tissues, creating potential toxicity concerns. The elegance of identifying tumor-specific surface markers has to contend with the messy reality of off-target effects and normal tissue cross-reactivity. The pharmaceutical literature was full of ADCs that appeared promising in preclinical models but caused dose-limiting toxicities in patients because the "tumor-specific" marker turned out to be expressed in the liver, lung, or other tissue at levels just high enough to matter.

This made NEDD9, with its high network centrality, conceptually appealing, but making the jump from network topology to a rational therapeutic target required navigating somewhat uncharted territory. After all, despite there being a number of proteins that have been shown to be network hubs across various cancer types, few have been successfully targeted in the clinic. The problems with network-based drug targeting is partly technical, partly conceptual. First off, high centrality-proteins are often structurally challenging to drug (large, flat surfaces). Additionally, while we’ve perfected the mathematical frameworks for identifying important nodes in a biological network, we’re still learning how to translate graph theory-based metrics into therapeutic strategies. For example, a protein may have a high degree and betweenness centrality on paper, while simultaneously being functionally redundant in practice, with parallel pathways compensating for its loss.

Yet, rather than viewing the aforementioned challenges as failures of individual targeting philosophies, I’ve begun to see them as complimentary perspectives on the same problem, or rather, orthogonal measurement of tumor vulnerability. A protein can simultaneously serve as a mechanistic driver, a payload delivery target, and a network hub, with each property lending itself to different therapeutic opportunities. Additionally, proteins that score highly across multiple analytical frameworks may represent particularly attractive drug targets precisely because they offer multiple routes to cell killing. For example, if EGFR monotherapy could be circumvented through pathway rewiring—with a notable example being the T790M mutation in the EGFR receptor that leads Gefitinib resistance—what about a protein that was simultaneously a driver, abundant on the surface, and a network bottleneck? Resistance would require the tumor to solve multiple problems at once.

With this in mind, I’ve been thinking about the idea of a multi-dimensional vulnerability assessment for potential drug targets. Instead of choosing between driver analysis, surface targeting, and network disruption, what if we scored proteins across all three dimensions simultaneously? For example, we can identify candidate protein targets based on their differential expression, pathway involvement, or network centrality, then, for each candidate in the list, we can generate a normalized score for each of these criteria (which accounts for different scaling), then create a composite score that combines the three individual scores (where each component is weighted). The weighting itself becomes a tunable parameter—you might weight surface expression more heavily if you’re specifically interested in ADC development, or prioritize network centrality if you’re looking for non-obvious targets that conventional approaches might miss. While most proteins show strong signals in only one analytical framework, a small subset emerge as multi-dimensional targets with high scores across multiple approaches3.

When I implemented this scoring system on my clinical trial dataset, something interesting happened. The top-ranked proteins weren’t necessarily the ones with the most dramatic signals in any single dimension. Instead, they were proteins that performed consistently well across all three frameworks—showing moderate over-expression, reasonable surface accessibility, and meaningful network positions. These were the proteins that traditional single-lens analyses might overlook in favor of more exceptional, but one-dimensional, candidates.

Consider the example of HER2 (aka, ERBB2) in breast cancer—a protein that initially gained attention as a driver oncoprotein but proved equally valuable as an ADC target and network hub. HER2 overexpression clearly drives tumor growth through activation of downstream signaling cascades, but its abundant surface expression also makes it a viable target for ADCs like T-DM1. Additionally, network analyses consistently identify HER2 as a high-centrality node that coordinates multiple aspects of tumor biology as it sits at the intersection of the PI3K/AKT/mTOR pathway, the MAPK cascade, and multiple other signaling networks that cancer cells depend on for survival and proliferation.

Thus, the clinical success of HER2-targeting agents may derive not just from inhibiting a key driver pathway, but from simultaneously disrupting a critical network node and providing a reliable delivery address for cytotoxic payloads. This multi-dimensional vulnerability may explain why HER2-positive breast cancers, despite their aggressive biology and fast growth, often show excellent responses to targeted therapy combinations. We’ve been attributing HER2’s success primarily to its role as a driver oncogene, but perhaps we’ve been underselling the contribution of its other properties. The fact that you can both inhibit HER2 signaling with small molecules and deliver toxic payloads via HER2-targeting antibodies gives oncologists multiple tools to attack the same target, making resistance that much harder to evolve.

Instead of asking "which protein is the best drug target?" the question now becomes, "which combinations of vulnerabilities can we most effectively exploit, while simultaneously minimizing toxicity?" If tumors present multiple, orthogonal vulnerabilities, then therapeutic strategies that address these vulnerabilities simultaneously might achieve synergistic effects while reducing the likelihood of developing resistance. This reframing also suggests why some clinical trials fail despite strong preclinical rationale. A protein might score perfectly on one dimension—say, as a clear driver—but if it lacks the other properties that make it druggable or deliverable, the therapeutic window might be too narrow for clinical success.

By recognizing that different analytical approaches reveal different but equally valid aspects of tumor biology, we can move beyond the philosophical debates that have sometimes divided the cancer research community. The question isn't whether we should target drivers, exploit highly abundant surface proteins with cytotoxic payloads that can be internalized by the cell, or disrupt networks—it's how we can most effectively integrate these approaches to achieve better patient outcomes.

If you liked this post, you may enjoy the following pieces from Sequence & Destroy’s Backlog:

• Issue #36 // Machine Learning: The Native Language of Biology

• Issue #55 // Molecular Moonlighting

• Issue #57 // Graphs As Metaphor In Biological Systems

Liked this piece? If so, tap the 🖤 in the header above. It’s a small gesture that goes a long way in helping me understand what you value and in growing this newsletter. Thanks so much!

Issue #№ 57 // Graphs As Metaphør In Biøløgical Systems

There’s a particular kind of frustration that comes with studying living systems, one that doesn’t quite exist in physics or chemistry. You can spend years characterizing how a protein behaves in a carefully controlled experiment, publish your findings, and then watch as someone else discovers that same protein doing something completely different in another context. This anecdote isn’t personal; I've haven’t spent years studying a single protein, learning everything there is to know about it. But I have experienced a similar widening of my aperture as I went from undergrad where I learned that a protein’s structure determines its function to grad school where I learned about specific functions for a number of different oncoproteins to my doctoral research where I saw how protein function is highly context dependent, inherently relational, and not always entirely clear.

Along the way, I also learned to think about biological systems as graphs—networks of nodes connected by edges, which I wrote about at length in Issue #51 // Mapping Biology’s Dark Matter. At the molecular level, the nodes might be genes, proteins, or metabolites. The edges represent their relationships: how they interact, regulate each other, compete for resources, are co-expressed, and more. Zoom out, and those molecular nodes aggregate into functional modules such as signaling pathways, molecular functions, and biological processes. Zoom out further still, and you’re looking at cells, tissues, entire organisms, and populations. It’s graphs all the way up.

Most molecular biology research, when you strip away the particulars, does one of two things: it identifies a new node, or it identifies a new edge. A novel protein gets characterized—that’s a node. Someone discovers that this protein phosphorylates another protein under certain conditions—that’s an edge. The problem, and the thing that makes biology uniquely maddening, is that we don’t know what all the nodes are, even for relatively well-studied processes. And even when we do know the nodes, mapping the edges between them is extraordinarily slow work.

I’ve written before about molecular moonlighting, which is the phenomenon where a single protein performs multiple, often unrelated functions depending on context. It’s a good example of why the graph metaphor matters. If you study a protein in isolation, or even in a small set of pairwise interactions, you get an incomplete picture. You might tag it with a fluorophore, watch how it moves when you perturb the system, and conclude that it plays one specific role. But that role might be contingent on a dozen other factors you weren’t measuring: the presence of particular cofactors, the cell’s metabolic state, or what other proteins are competing for the same binding sites. Biology is full of higher-order feedback loops. A protein’s function isn’t intrinsic—it’s emergent from its position in the network.

The traditional experimental toolkit reflects these limitations1. If you wanted to understand a biological process twenty years ago, you could really only look at a few nodes at a time. Fluorescent tagging let you watch a handful of proteins and knockout models let you remove a node and observe the ripple effects. These approaches worked, in the sense that they produced real knowledge, but they were constrained by throughput. Designing an experiment that could simultaneously account for the behavior of dozens or hundreds of interacting components was essentially impossible.

The last fifteen to twenty years have been different. Next-generation sequencing technologies didn’t just make it cheaper to sequence DNA—they enabled a fundamentally different mode of inquiry. Suddenly you could measure the expression levels of thousands of genes in a cell at once. You could characterize tens of thousands of protein-protein interactions in a single experiment. RNA-seq, ChIP-seq, ATAC-seq, single-cell sequencing, spatial transcriptomics—each of these represents a way to observe many more nodes and edges simultaneously. The graph becomes denser, more complete.

What becomes visible at this scale are the things that pairwise studies miss. Redundancy, for example. Biological systems are full of backup mechanisms, parallel pathways that only become apparent when you knock out one component and discover that the system compensates. Or context-dependency; a protein that’s essential in one cell type might be practically irrelevant in another, not because the protein itself has changed, but because the surrounding network is different. These aren’t edge (pardon the pun) cases. They’re the norm.

The knowledge graph framework also clarifies what kinds of questions remain hard. Even with high-throughput tools, we’re still working with incomplete graphs. Some nodes are easier to observe than others—genes are straightforward, but post-translational modifications or transient protein complexes are slippery. Some edges are easier to infer: co-expression is a weak signal, but it’s measurable at scale. Physical protein-protein interactions are stronger evidence, but harder to capture comprehensively. Causal relationships—this protein activates that pathway, under these conditions—are the hardest of all, because biology resists clean causality2. The graph is dynamic. The edges change depending on the state of the system.

There’s an intellectual style I admire, the kind George Church exemplifies, where technical rigor coexists with a willingness to make unexpected connections across disciplines. The graph metaphor invites this. Network theory was developed to study social systems, electrical grids, and the internet3. The mathematics of centrality and modularity don’t care whether the nodes are people or proteins. This isn’t just analogy; there are real conceptual tools from network science that translate directly. The idea of network motifs, recurring patterns of connections that show up across different systems, has been productively applied to gene regulatory networks. The concept of scale-free networks, where a few highly connected hubs dominate, describes both the World Wide Web and protein interaction networks.

But the metaphor also has limits, and it’s worth being clear about them. Real biological systems aren’t static graphs. They’re dynamic, probabilistic, and context-dependent in ways that are difficult to capture in a single representation. An edge between two proteins might exist in one cellular state and disappear in another. The graph itself is a moving target. Still, as an organizing framework, it works. It lets us think clearly about what we know and what we don’t and it suggests where the next productive questions might be. The project of biology from this perspective is one of gradual graph completion. We’re filling in missing nodes, adding edges, and refining our understanding of which connections matter under which conditions. The pace has accelerated, but the work remains enormous. Even for the relatively well-characterized model organisms—yeast, worms, flies—the graphs are incomplete. For human biology, where the complexity is orders of magnitude higher, we’re still sketching the outline.

What feels different now isn’t just the speed of data generation, but the possibility of asking questions that were previously out of reach. If you can measure thousands of variables simultaneously, you can start to model how networks respond to perturbations. You can identify vulnerabilities—nodes whose removal would collapse an entire pathway. You can predict how a system might compensate for damage. This is the promise of systems biology, computational biology, network medicine: not just cataloging parts, but understanding how they fit together.

The frustration I mentioned earlier doesn’t go away. Biological systems remain irreducibly complex. But the graph gives us a way to think about that complexity that’s both rigorous and generative. It lets us map what we know, acknowledge what we don’t, and chart a course forward. The edges are still hard to draw, the nodes still incomplete, but at least we can see the shape of the problem.