How Much Structure Does a Biological Model Need?

I’ve been fairly biologically illiterate for most of my life. My 9th grade biology teacher, who wore a full suit every day and seemed to dye his hair with shoe polish, would often tell us falsehoods in service of bio “jokes” like “Al G. is a fun guy.” Algae are, in fact, not fungi. He also conveyed biology as a tedious field of taxonomy memorization. But a close biophysicist friend from graduate school has convinced me otherwise. Statistical mechanics, high-throughput experiments, and machine learning are now colliding in genuinely interesting ways. After spending the past few months reading this work, I agree. However, a concern I have that keeps popping up is that the experiment that generates the labeled output, and therefore determines what that label means, is increasingly treated as an implementation detail.

Biological foundation models have become remarkably good at learning patterns in natural sequences. Protein language models can predict the effects of many mutations without task-specific training. DNA models learn representations that transfer to genomic prediction tasks, and models such as Evo can predict and generate plausible sequences across DNA, RNA, and proteins. Evolution has already conducted an experiment of absurd scale, and sequence models are learning from its surviving record (Meier et al., 2021; Dalla-Torre et al., 2025; Nguyen et al., 2024).

But plausible is not the same as controllable. The question we want answered is “If I build this sequence, in this format, under these conditions, what will happen?” A masked sequence model is trained to answer “Given the sequences evolution has preserved, what belongs here?”

The divergence comes partly from the size of sequence space, but more importantly from the labels themselves: biological labels are produced by experiments. Measurements like enrichment, fluorescence, and binding score mix the molecular property we care about with things like concentration, expression, and kinetics.

A sufficiently flexible model may learn around those effects from enough data. An explicit assay or physical model may extract more information from fewer measurements, and carry that information more reliably into new conditions. So the core question should be whether explicitly modeling the assay recovers information that the experiment actually identifies and that improves the next engineering decision.

1. Evolution is a prior, not an engineering specification

Natural sequence databases are vast, but they were not assembled to answer a particular engineering question. Evolution preserves sequences under many overlapping pressures: folding, expression, reproduction, environment, and historical accident. A conserved residue may improve binding. It may instead stabilize the fold, prevent aggregation, or compensate for another mutation.

Sequence statistics can reveal that a pattern matters without identifying why it matters. The sample is also one-sided. They contain molecules that arose, persisted, and were sequenced. They do not contain a controlled catalogue of every alternative that misfolded, aggregated, failed to bind, or never arose.

Presence in nature does not prove success, and absence does not prove failure. Evolution does not explore every possibility, and natural proteins are rarely optimized for the exact objective an engineer has chosen. Target-specific experiments provide a different kind of evidence. They deliberately perturb a local sequence landscape under defined conditions and measure a chosen outcome.

2. The assay writes the label

A sequencing count, fluorescence bin, or enrichment score is an output of an apparatus, not an intrinsic molecular property. A selection experiment may depend on target concentration, display level, saturation, nonspecific adhesion, wash time, and sequencing noise.

Consider saturation. Once binding approaches saturation, molecules with meaningfully different affinities can produce nearly identical signals. A model trained directly on those signals may mistake an assay plateau for a ceiling in the underlying biology.

Or consider expression. A weak binder displayed at ten times the level of a strong binder may produce the larger fluorescence signal. Without an expression measurement or an adequate model of the assay, the observed label combines two different effects.

Tite-Seq addressed this problem by measuring titration curves across multiple antigen concentrations. This allowed quantitative affinities to be estimated for thousands of antibody variants rather than treating a single enrichment value as affinity (Adams et al., 2016).

This motivates a decomposition graph:

The latent variable might be affinity, kinetics, stability, or expression. The measurement layer might include saturation, sorting, growth, fluorescence, and sequencing noise.

MAVE-NN jointly learns genotype–phenotype maps and measurement processes, supporting additive, pairwise, neural, and custom biophysical models (Tareen et al., 2022).

ProBound uses a more assay-specific architecture. It models molecular recognition, selection, and sequencing together, and can infer equilibrium constants or kinetic rates when the experiment contains the information needed to identify them (Rube et al., 2022).

MoCHI extends this line of work to multimodal deep-mutational-scanning data and fits interpretable energetic models with higher-order interaction terms where the data support them (Faure et al., 2024).

The experimental results look better, but now it must be shown that the structured decomposition produces better downstream decisions.

3. What can the experiment actually identify?

Initially I would contrast meaningful physical quantities with noisy assay scores. A dissociation constant or off-rate may be more interpretable and transportable than enrichment, but it is still an intermediate variable. The real objective is generally something like cellular potency, tissue penetration, or efficacy. And by transport, I mean that a relationship learned in one setting remains useful after the concentration, assay, or some other biological context changes.

This graph looks something like:

This creates two separate transport problems.

Observation to property

Can the assay readout be converted into an intrinsic quantity such as affinity or kinetics?

This is where a measurement model can help separate expression, saturation, sampling, and other confounders.

Property to objective

Does that affinity or kinetic rate predict the functional outcome we actually care about?

A model can predict affinity accurately and still fail to predict potency, pharmacokinetics, or safety. Abundant affinity labels license a model to predict affinity. They do not solve the separate transport from affinity to a therapeutic outcome.

A physical variable is worth inferring only if the experiment identifies it, it matters to the downstream objective, and it improves decisions over the raw score. If a binding experiment uses one concentration, does not independently measure expression, and returns only selected and unselected counts, absolute affinity may not be identifiable.

Tareen et al describe MAVE-NN and non-identifiable transformations in which changes to the latent genotype–phenotype map can be absorbed by the measurement process without changing the predicted observations. A plausible latent quantity is therefore not necessarily a uniquely determined physical parameter (Tareen et al., 2022).

In short, the design principle is to infer the most physical quantity the experiment identifies, not the most physical quantity the model can emit.

4. Three Routes to Better Biological Design

How expensive it is to be wrong sets which criterion matters: raw accuracy when you can afford to brute-force through errors, information-per-experiment when you can't. I see three ways a campaign can improve its decisions:

Scale the model

One strategy is to make a general molecular model capable enough to work before much target-specific data exist. Isomorphic Labs is testing this. The goal is to shift more candidate evaluation into computation before a target-specific experimental campaign is mature. Isomorphic says IsoDDE outperforms other machine-learning methods on the FEP+ 4, OpenFE, and CASP16 affinity benchmarks (Isomorphic Labs, 2026).

Scale the experimental loop

When a team can generate large numbers of standardized labels for the exact measured property and refresh them cheaply inside a rapid experimental loop, the model does not need to explain every error before the next decision. A-Alpha generates large, comparable protein-interaction datasets through AlphaSeq and uses those measurements to train and refine AlphaBind. Its bet is that enough high-quality local data can teach a flexible model the relevant binding landscape. BigHat takes a related but broader closed-loop approach. It designs, builds, and tests hundreds of antibodies in weekly cycles across half a dozen biophysical and functional properties. Its model does not need to be globally correct if the laboratory can discover and correct errors quickly.

Extract more information from each experiment

Structure becomes more valuable when labels are slow or expensive, or when a result must carry into a new concentration, format, assay, or sequence neighborhood.

A silent failure is expensive here. It consumes a synthesis, assay, or experimental round without necessarily revealing why the model failed or what should be measured next. A structured model can be worth more even when its first-round accuracy is similar, if its errors carry more information. A failed prediction may reveal things like affinity and expression were confounded, the assay saturated, the assumed kinetic regime was wrong, and so on. The relevant question is how much decision information did the campaign gain from the experiment?

Assay-explicit modeling takes a more grounded approach. It asks whether one rich experiment contains more decision information than a sequence score or black-box model usually extracts. These approaches are not mutually exclusive. A strong system may combine a broad learned prior, standardized measurements, and a fast experimental loop. The unresolved question is which layer adds the most value for a particular campaign. So I'll hold my own bet to the same bar: a matched prospective comparison, same campaign data, same tuning effort, same validation budget, or it doesn't count. I don't think anyone in this list has run that yet, including the approach I'm most sympathetic to. MAVE-NN, ProBound, and MoCHI represent the model-the-assay philosophy.

An objection I take most seriously is that none of this requires an explicit assay model. A flexible model with calibrated uncertainty, evaluated under the same conditions, can also flag unreliable predictions. You don't need a biophysical parameterization to flag an out-of-distribution candidate.

Benchmarking of uncertainty quantification methods on protein-fitness tasks found no single method reliably best across datasets and splits, and that uncertainty-guided sampling often failed to beat greedy selection (Greenman et al., 2025). A new concentration, format, or sequence neighborhood, is where transport matters most. Predictive uncertainty may flag where the model is likely to fail without distinguishing among expression confounding, assay saturation, and a broken kinetic assumption. That distinction can determine the next experiment.

A structured model can still hand you a confident, wrong diagnosis, and legibility is only worth something if the explanation makes a falsifiable prediction and survives the next round. But a well-calibrated black box doesn't even attempt the diagnosis. So even when a wrong answer is expensive, being told where and why beats being told only where.

Conclusion

Biological modeling has three levers: a broader learned prior, more standardized measurements, and a more explicit model of how measurements arise. Their value depends on the campaign. When another round is cheap, empirical correction can substitute for explanation. When another round is slow or the conditions will change, the observation model should earn its complexity by extracting an identifiable quantity that improves the next decision.

The Bitter Lesson still applies. Sutton argued that general methods which exploit increasing computation, especially learning and search, have repeatedly overtaken systems built around brittle, handcrafted knowledge (Sutton, 2019).

But he says less about whether the experiment that generated the data should be modeled explicitly.

First-round accuracy is therefore an incomplete metric. What matters is how much each round improves the campaign’s subsequent decisions - the decision value per experimental round. Many modern design campaigns can generate more candidates than they can test intelligently. The bottleneck is choosing which ones deserve the next experimental round.

References

Adams, R. M., Mora, T., Walczak, A. M., & Kinney, J. B. (2016). Measuring the sequence-affinity landscape of antibodies with massively parallel titration curves. eLife.

A-Alpha Bio. (2024–2026). Our Technology and Predicting protein binding and engineering antibodies with AlphaBind. Company materials.

BigHat Biosciences. (2026). Our Technology & Science. Company materials.

Dalla-Torre, H., Gonzalez, L., Mendoza-Revilla, J., et al. (2025). Nucleotide Transformer: building and evaluating robust foundation models for human genomics. Nature Methods.

Faure, A. J., et al. (2024). MoCHI: neural networks to fit interpretable models and quantify energies, energetic couplings, epistasis and allostery from deep mutational scanning data. Genome Biology.

Greenman, K. P., Amini, A. P., & Yang, K. K. (2025). Benchmarking uncertainty quantification for protein engineering. PLOS Computational Biology.

Isomorphic Labs. (2026). The Isomorphic Labs Drug Design Engine unlocks a new frontier. Company technical materials.

Meier, J., Rao, R., Verkuil, R., et al. (2021). Language models enable zero-shot prediction of the effects of mutations on protein function. NeurIPS.

Nguyen, E., Poli, M., Durrant, M. G., et al. (2024). Sequence modeling and design from molecular to genome scale with Evo. Science.

Rube, H. T., Rastogi, C., Kribelbauer, J. F., & Bussemaker, H. J. (2022). Prediction of protein–ligand binding affinity from sequencing data with interpretable machine learning. Nature Biotechnology.

Sutton, R. S. (2019). The Bitter Lesson.

Tareen, A., Kinney, J. B., et al. (2022). MAVE-NN: learning genotype–phenotype maps from multiplex assays of variant effect. Genome Biology.