7Function and Property Prediction

A model that reads one molecule and predicts a number you could otherwise get only from an experiment: how tightly a drug candidate binds, whether a point mutation cripples an enzyme, how well a protein survives being heated, how much of it a cell will actually express. That is property prediction. The molecule is given — you are not searching for a new one, as in design (Chapter 9), and you are not folding it, as in structure (Chapter 8). You want a property, and you want it cheaply, before the assay (Chapter 4) that would measure it for real. The central bet of this chapter is that the same protein language models that learned sequence statistics can score properties they were never explicitly trained on, and that when you do have labels, a small supervised head on top does even better.

7.1The problem: from sequence to phenotype

Property prediction takes a single molecule and returns a measurable quantity. For proteins the workhorse quantities are: variant effect or fitness (how much a mutation helps or hurts function), folding stability (the free-energy change on mutation, written ΔΔG, where a positive value means destabilizing), binding affinity (typically a dissociation constant Kd — smaller means tighter), and expression / solubility (whether the cell makes usable protein at all). For small molecules the analogous bundle is ADMET — absorption, distribution, metabolism, excretion, and toxicity — the pharmacological properties that decide whether a promising binder can ever become a drug.

Why bother predicting these instead of measuring them? Scale. Deep mutational scanning (DMS), an assay that measures the functional effect of thousands to millions of variants of one protein in a single experiment, is powerful but still bounded by what one lab can build. A model that generalizes from a few such experiments can score every possible variant of a protein in seconds, turning a wet-lab screen into a ranked shortlist. That shortlist feeds target discovery (Chapter 6), where you ask which variants of a gene are damaging, and it feeds engineering, where you ask which mutations to try next.

7.2Models that attempt it

Two families do the work, and they map onto the model families from the ML on-ramp (Chapter 3). The first is zero-shot scoring with a protein language model — zero-shot meaning no task-specific labels are used at all. A masked language model like ESM-2, or an autoregressive one like ProGen2, is trained only to predict amino acids from sequence context across UniRef-scale data, and in doing so it learns which sequences look "natural" (Rives et al., 2021). To score a mutation you compare how probable the model finds the mutant residue versus the wild-type residue at that position: the log-likelihood ratio (LLR), log p(mutant) − log p(wild type). A large negative score means the model thinks the substitution is unlike anything evolution kept. This is the ESM-1v recipe (Meier et al., 2021).

The idea that likelihood tracks fitness rests on a real mechanism: natural sequences are the survivors of selection, so residues that would break function were removed from the training data, and the model assigns them low probability. The community benchmark is ProteinGym, a suite of 217 DMS assays covering millions of variants against which methods are scored by Spearman rank correlation to measured fitness (Notin et al., 2023). On it, single-sequence PLMs land around 0.4–0.5 average Spearman (ESM-1v and ESM-2 650M are essentially tied there), and — worth internalizing — they do not dominate. Methods that exploit a multiple sequence alignment (MSA), the stack of evolutionarily related sequences for a protein, remain competitive or better: the alignment-based GEMME and hybrids like TranceptEVE cluster near the top, and AlphaMissense, an AlphaFold derivative fine-tuned on population variant frequencies, is state of the art for human missense pathogenicity (Cheng et al., 2023). Recent analyses even suggest single-sequence PLM likelihood is hitting a scaling plateau on this task.

The second family is supervised: when you have labeled measurements for the exact property you care about, take the PLM's per-residue embeddings and train a regression head on them. This wins whenever labels exist. The clearest case is stability: the Tsuboyama megascale dataset measured folding stability for ~776,000 domain variants by proteolysis (Tsuboyama et al., 2023), and models fine-tuned on it — ThermoMPNN — reach Spearman around 0.72; note that the strongest stability predictors read the 3D backbone (ThermoMPNN builds on ProteinMPNN) rather than sequence-PLM embeddings on held-out stability, far above any zero-shot likelihood (Dieckhaus et al., 2024). For small-molecule ADMET the same pattern holds without a language model at all: graph neural networks trained on the Therapeutics Data Commons benchmarks (Huang et al., 2021), as in ADMET-AI, are the practical default (Swanson et al., 2024).

7.3What they do well and what is still hard

These models are genuinely useful at what they are for: ranking single substitutions, flagging likely-deleterious mutations, and pre-screening a large variant library down to a testable set. Where they struggle traces to a few recurring failures.

Off-distribution generalization. PLMs are anchored to natural sequence space, so they extrapolate poorly to de novo or heavily engineered proteins that no longer resemble anything in UniRef — exactly the regime protein design (Chapter 9) operates in. The plateau in single-sequence performance on ProteinGym is a symptom: more parameters stop helping once the model has absorbed the evolutionary signal there is to absorb.

Epistasis. The effect of two mutations is often not the sum of their individual effects, and likelihood-based scores, computed largely per position, miss these higher-order interactions. In reciprocal sign epistasis, two individually beneficial mutations combine to break the protein — the case that most punishes an additive predictor.

Calibration. Scores rank well but are not calibrated probabilities, and the useful decision threshold shifts from protein to protein. An LLR of −5 may mean "dead" for one enzyme and "mildly worse" for another, so a fixed cutoff over-flags on some proteins and under-flags on others.

Fitness versus the activity you care about. This is the deepest gap. Likelihood and stability are correlated with function broadly, but the specific number you want — a catalytic turnover, selectivity between two targets, an off-target toxicity — is a narrow slice that generic evolutionary plausibility does not resolve. Closing that gap is what the design–build–test loop (Chapter 17) is for, and it is why property models augment rather than replace experiments.

For interpreting human variants specifically — pathogenic or benign rather than fitness in a dish — the framing shifts toward mechanism and population genetics, which is the subject of variant-to-mechanism (Chapter 13); the single-cell foundation models that predict expression phenotypes are covered lightly in the appendix.