Models that predict how a protein mutation changes the protein's behavior almost all make two simplifications at once. First, they keep the protein's 3D backbone rigid: every mutant is scored against a single frozen snapshot of the wild-type structure, even though a mutation that swaps a small amino acid for a bulky one physically shoves neighboring atoms aside. Second, they treat the measurement being predicted — binding energy, fluorescence, thermal stability, viral escape — as an anonymous number, one scalar output head per assay, even though "does this mutation destabilize the fold" and "does this mutation let the virus escape an antibody" are causally related questions that share biophysical vocabulary. This proposal builds a single model that relaxes both simplifications together: it lets the backbone move, and it lets the prediction target be described in natural language.
The two research communities involved have approached these problems from opposite sides and stopped exactly at the boundary. Structural-plasticity models — EBM-DDG, which lets mutant coordinates physically relax via Langevin sampling, TriFit, which adds harmonic vibration modes to sequence and structure features, and S3F, which adds molecular surface geometry — all end in bare scalar or binary heads. Semantic-grounding models go the other way: MutaPLM maps mutations to scientific-text explanations via an explicit vector difference between wild-type and mutant embeddings, and TALE embeds Gene Ontology label definitions jointly with sequence representations — but both operate on 1D sequences only; MutaPLM's own authors name the absence of 3D structure as their main limitation. Nothing in the indexed literature conditions structural relaxation on what the assay actually measures, and nothing grounds language in a moving structure.
Core Claim
A variant-effect predictor that (i) represents each mutant by a relaxed structural ensemble rather than a fixed wild-type backbone, and (ii) represents each assay by a semantic embedding of its natural-language description rather than a private scalar head, will match the best fixed-backbone supervised models on standard substitution benchmarks while doing three things none of them can: predict on assay types it never saw during training (zero-shot assay transfer), score insertion/deletion variants whose length changes break every fixed-backbone pipeline, and remain accurate precisely where rigidity is known to fail — glycine/proline mutations and flexible loops.
The mechanism is a bilinear match between two embeddings:
$$\hat{y}_{v,a} = h_a^{\top} W\, z_v$$
where $z_v$ encodes variant $v$ (a concatenation of the sequence-level wild-type–mutant difference, in the spirit of MutaPLM's delta network, and a summary of the mutant's relaxed structural ensemble), $h_a$ is the embedding of assay $a$'s textual description ("change in fluorescence intensity of GFP after mutagenesis", "binding of antibody to SARS-CoV-2 spike"), and $W$ is a learned coupling matrix. One model, one set of weights, all assays simultaneously.
Research Program
Three components get built. The relaxer is the physically substantive piece. Following EBM-DDG's decomposition of $\Delta\Delta G$ into a sequence-based inverse-folding term plus a structural corrective term, we train a lightweight equivariant network (the $E(3)$-equivariant design of ProtEGNN shows this need not be large — 400K parameters sufficed there) that, conditioned on the mutant sequence, outputs a small ensemble of backbone conformations rather than perturbing a fixed template. Crucially, the relaxer is generative in length: because conformations are produced from the mutant sequence (inverse-folding style, pretrained on the ~31,000 CATH domains as in S3F), insertions and deletions yield backbones of the correct new length instead of crashing the kernel, which is exactly where Kermut's fixed-coordinate structural kernel and all static-PDB pipelines fail. The relaxer is pretrained on Megascale's ~776,000 folding-stability measurements, the same corpus STAB-DDG used before fine-tuning on binding data.
The assay encoder maps each ProteinGym assay description into a shared space using a frozen text encoder, augmented — following TALE's demonstration that hierarchical label structure carries real signal (it maintained accuracy in the "midnight zone" below 30% sequence identity) — with structured annotation: the assay's protein's Gene Ontology terms, the measurement modality, and the organism. This addresses the known semantic-drift problem for fine-grained labels by anchoring descriptions in an ontology rather than free text alone. The matcher is the bilinear head above, trained multi-task across all assays with a ranking loss (Spearman-consistent), with an auxiliary generative-verbalization objective on MutaDescribe's 171K literature-annotated single-site mutations so the model can also explain predictions in words — the capability MutaPLM established but never connected to structure.
The scientific analysis asks two questions beyond leaderboard numbers. First, where does relaxation pay: we expect gains concentrated on glycine/proline mutations and flexible-loop regions, the documented failure modes of EBM-DDG's rigid competitors and of STAB-DDG on TCR binding loops. Second, what does the assay embedding space learn: we will test whether assays cluster by mechanism (stability assays near each other, binding assays near each other) and whether removing the ontology annotation degrades zero-shot transfer — separating genuine semantic transfer from mere multi-task regularization.
Proposed Experiments
The primary benchmark is the ProteinGym substitution suite of 217 assays (ProteinGym), evaluated by Spearman correlation, matching the supervised cross-validated protocol under which Kermut reports 0.662 average Spearman (vs. 0.613 for ProteinNPT), and by AUROC, where TriFit reports 0.897. Three evaluation regimes matter more than the aggregate number. Zero-shot assay transfer: train on a fixed subset of assay descriptions, hold out entire assays, and predict them from text alone — scalar-head models cannot enter this regime without retraining, so this is where semantic labels must earn their place or the idea fails. Indel benchmark: ProteinGym's 66 indel assays (~300K variants), on which fixed-backbone models cannot run at all; no reported result found on Lacuna for any baseline on this benchmark, so we will establish the first reference numbers, with sequence-only ESM likelihoods as the only runnable comparison. Flexible-region subsets: stratify SKEMPI v2.0 by glycine/proline involvement and by loop versus core position, comparing against EBM-DDG's reported per-structure Pearson correlation of 0.5681 on SKEMPI v2.0. Language quality is evaluated on MutaDescribe, where MutaPLM reports a 6.5% ROUGE-L improvement over baselines and top-50 recall of 0.409 for text-guided mutation proposal.
Baselines
The naive reference point is zero-shot ESM likelihood scoring — no structure, no semantics, no training; on the ProteinGym AUROC protocol, TriFit reports ESM3 at 0.769 AUROC, which any trained model must clear by a wide margin to justify its existence. The strongest existing competitor on the substitution benchmark is Kermut (Kermut), a Gaussian-process regressor combining an ESM-2 sequence kernel with a ProteinMPNN structural kernel over fixed wild-type coordinates, reporting 0.662 average Spearman across the 217 assays; we will rerun it with its public protocol on identical splits and identical assay partitions (including our held-out-assay splits, where it must be retrained per assay), giving it the same per-assay tuning budget our single model gets in total — a deliberately conservative comparison, since Kermut's 10-minute cross-validation already makes generous reruns cheap. For AUROC comparisons on the binarized protocol, TriFit (TriFit), reporting 0.897 AUROC with dynamics features, is the competitor; we reproduce its thresholding exactly. On SKEMPI v2.0, EBM-DDG's 0.5681 per-structure Pearson is the number to beat. Two ablations isolate the proposal's ingredients: a rigid-backbone version of our own model (identical architecture, but $z_v$ built from the static wild-type coordinates instead of the relaxed ensemble — this isolates relaxation), and a sequence-plus-text version (identical matcher and assay encoder, but $z_v$ from MutaPLM-style sequence deltas only — this isolates structure). Beating Kermut on the aggregate substitution Spearman is not the success criterion; matching it within noise while winning decisively on held-out assays, indels, and flexible-region strata is. If the rigid ablation matches the full model everywhere, the relaxation component is dead weight and the proposal's central claim is falsified.
Why not just…?
Why not just use a large sequence-only protein language model with assay text appended? This is the most serious alternative, and we run it as the sequence-plus-text ablation rather than argue around it. The literature gives it a real chance on substitutions: TriFit found that adding static structure to sequence bought only +0.2 AUROC while dynamics bought +2.2, and ProtEGNN showed small models with explicit geometry can match 5.7B-parameter sequence models — evidence cuts both ways. The decisive tests are the regimes where sequence alone is documented to fail: indels (where structure-free models are the only current option, so this is where sequence-text may genuinely win and we will report it honestly if so), low-homology assays where S3F showed sequence features collapse, and glycine/proline and loop strata where EBM-DDG showed rigid geometry fails. If the sequence-plus-text ablation matches the full model on all of these, the relaxer is unnecessary and the paper becomes a (still useful) negative result about semantic assay conditioning.
Why not just keep the backbone rigid and add semantic labels? That is precisely the rigid ablation, and it is the honest null hypothesis for the structural half of the proposal. The literature predicts it fails in specific, measurable places: EBM-DDG's rigid baselines collapsed on structure-breaking residues (its Langevin relaxation lifted Pearson correlation from FoldX's 0.3789 to 0.5681), BA-DDG showed ignoring unbound flexible states costs correlation on SKEMPI (0.4324 to 0.5134), and STAB-DDG recorded outright failure on TCR loop targets. The stratified evaluations are designed so this ablation's predicted failure mode is either confirmed or refuted, not assumed.
Why not just run real physics — Rosetta Flex ddG or molecular dynamics? Because the cost is prohibitive at any scale that matters: STAB-DDG documents Flex ddG at 15 CPU-hours per mutation against its own 0.2 GPU-seconds at comparable accuracy (0.42 vs. 0.45 Spearman on SKEMPI), and the benchmarks here involve hundreds of thousands to millions of variants. Physics simulation also produces a number, not a transferable assay representation, so it cannot address zero-shot assay transfer at any compute budget. We use Flex ddG's reported figures as the physical reference point rather than rerunning it.
Why not just verbalize everything with an LLM, MutaPLM-style? MutaPLM already grounds mutations in literature text, but by its own stated limitation it is 1D only, restricted to single-site substitutions, and it produces explanations rather than calibrated fitness scores conditioned on an assay. Our proposal reuses its delta-network idea and its MutaDescribe corpus for the auxiliary explanation objective, but the predictive head is the structure-aware bilinear matcher — the two are complementary, and the human-evaluation bar it set (nearly 40% of explanations rated accurate) is the number our verbalization component should be compared against.
Decisive Checks
The proposal stands or falls on three falsifiable outcomes. Zero-shot assay transfer must beat the best transfer a per-assay scalar model can manage after retraining on a small sample of the held-out assay's data (a few-shot adaptation curve, not just a single point). The indel benchmark must show the length-generative relaxer producing scores that correlate with measured indel fitness — no baseline number exists on Lacuna for this, so even modest correlation is a first, but the sequence-only ablation is the real test of whether structure helps there. And the flexible-region strata must show a gap between the full model and its rigid ablation on SKEMPI v2.0 in the direction EBM-DDG's results predict. If all three hold, semantic assay conditioning and conformational relaxation are mutually reinforcing, not just compatible; if only the semantic half holds, the relaxer gets cut and the finding is reported as such.
References
- Energy-Based Models for Predicting Mutational Effects on Proteins
- TriFit: Trimodal Fusion with Protein Dynamics for Mutation Fitness Prediction
- Multi-Scale Representation Learning for Protein Fitness Prediction
- MutaPLM: Protein Language Modeling for Mutation Explanation and Engineering
- TALE: Transformer-based protein function Annotation with joint sequence-Label Embedding
- Just Add Structure: Protein Language Models Combined with Structural Equivariance Excel at Protein Tasks
- Kermut: Composite kernel regression for protein variant effects
- Predicting mutational effects on protein binding from folding energy
- ProteinGym: Large-Scale Benchmarks for Protein Fitness Prediction and Design