Predicting which protein mutations break function is one of the highest-value open problems in ML for biology, and nearly all current predictors are supervised: they train regressors or classifiers on scarce, expensive, assay-biased fitness labels. Yet an entire neighboring field solves structurally identical problems without labels. In industrial and medical imaging, anomaly detection works by learning a generative model of normal data only, projecting a suspect input back onto that "normal manifold," and reading the anomaly out of the difference between the input and its "healed" reconstruction — a pseudo-healthy counterfactual. This proposal asks whether the same maneuver works for proteins: train exclusively on wild-type evolutionary and structural ensembles, and detect deleterious variants as inputs that resist being projected back onto the wild-type manifold.
The proposal treats the four input directions asymmetrically. Reconstruction-based anomaly detection is the core mechanism; multimodal protein representation is the application regime; high-dimensional feature redundancy is the scoring constraint that decides which residual dimensions carry signal; and video transformation robustness supplies the invariance machinery that keeps the residual from firing on trivial transformations. The claim is concrete: a variant is deleterious when, after we remove every transformation a protein can undergo without consequence — rigid-body motion, loop flexibility, homologous substitution, insertion-induced misalignment — the reconstruction residual, measured in a small set of deliberately selected feature dimensions, remains large.
Core Claim
Three findings motivate the design, one from each supporting direction. First, redundancy is fatal to naive scoring: Gaussian Image Anomaly Detection with Greedy Eigencomponent Selection shows that in deep feature spaces the highest-variance principal components carry background, not anomaly, signal, and that a handful of non-consecutive, low-variance components can reach 100% AUROC where full-space scoring degrades. Second, naive residual subtraction fails: DRÆM - A discriminatively trained reconstruction embedding for surface anomaly detection shows pixel-wise differences between input and reconstruction reach only 55.0 average precision on MVTec-AD localization while a learned treatment of the residual pair reaches 68.4, because raw differences are dominated by misalignment and reconstruction artifacts. Third, when modalities are fused without de-redundancy, the extra modality adds almost nothing: TriFit: Trimodal Fusion with Protein Dynamics for Mutation Fitness Prediction reports that adding static 3D structure to sequence embeddings yields only +0.2 AUROC — the structural modality is almost fully redundant with what the protein language model already encodes — whereas adding dynamics yields +2.2.
Our claim is that deleteriousness can be scored without any fitness label by combining a mutation-trained denoising reconstruction over multimodal residue representations with a redundancy-aware, alignment-robust residual, and that this closes part of the gap between zero-shot sequence models (ESM3 at 0.769 AUROC on ProteinGym's 217 DMS assays) and supervised multimodal systems (TriFit at 0.897 AUROC) without touching a fitness label.
Research Program
The system, which we call pseudo-wild-type (PWT) scoring, has three components.
The first component is the normal manifold itself. For each protein family we assemble a multimodal residue-level representation per homolog and per wild type: ESM-2 sequence embeddings injected as node features into a geometric graph network over residues, following the integration recipe of Integration of pre-trained protein language models into geometric deep learning networks, which reports an average 20% gain across CASP, docking, and binding tasks; structure-constrained attention in the spirit of SSEmb: A joint embedding of protein sequence and structure enables robust variant effect predictions; and Gaussian Network Model dynamics features per TriFit. Unlike TriFit, which fuses these with a trained router for supervised classification, we fit an untrained density model — a per-family multivariate Gaussian over residue embeddings — and train a single denoising network on corrupted normal inputs, where the corruption distribution is mutations themselves: random substitutions, short insertions and deletions, and local structure perturbations. This is the protein analog of DRÆM's Perlin-noise corruption strategy, but the corrupting distribution is the actual test-time transformation family rather than arbitrary texture noise, which directly targets DRÆM's known weakness that synthetic outliers unlike real defects degrade detection.
The second component is the alignment-robust residual, imported from video copy detection. A variant and its reconstruction are misaligned by construction: an insertion shifts every downstream residue, exactly as frame drops or temporal re-sampling shift a copied video segment. Rather than subtracting embeddings position-by-position, we align the two residue sequences with a dynamic-programming match and score only over the aligned correspondence, in direct analogy to the temporal Hough and weak geometric consistency verification of INRIA-LEAR'S Video Copy Detection System, which rejects random matches by requiring consistent transformation parameters. Rigid-body pose is removed by using an equivariant graph network, so the residual is invariant to rotation and translation by construction. The conceptual transplant from video robustness is the principle that the residual must be computed after estimating and factoring out the nuisance transformation, never before.
The third component is redundancy-aware scoring. The residual vector at each residue is mapped into the eigenbasis of the family's normal-covariance, and the anomaly score is a Mahalanobis distance restricted to a small selected set of eigencomponents. The selection criterion is the key departure from prior work: Glossy greedy eigencomponent selection selects components by maximizing AUROC against labeled anomalies, which is unusable here; we replace the label with leave-one-out homolog reconstruction residuals, selecting the components on which held-out normal homologs have near-zero residual variance while variant residuals concentrate — a purely internal, label-free separation criterion. Across modalities, we additionally apply determinantal-point-process-style repulsion, as in Priors for Diversity in Generative Latent Variable Models, which forces redundant correlated components to collapse into shared "background" slots so the surviving slots carry distinct information; we mitigate its known $O(M^3)$ cost by applying it only to the small residue-level component pool. Formally, for variant $v$ with residue residuals $r_i = x_i(v) - \hat{x}_i(v)$ after alignment, the score is
$$s(v) = \max_i \; r_i^\top \Sigma_S^{-1} r_i,$$
where $\Sigma_S$ is the family covariance restricted to the selected component set $S$ and $\hat{x}_i(v)$ is the denoised, aligned reconstruction.
Proposed Experiments
The primary benchmark is ProteinGym, whose 250+ deep mutational scanning assays and 2.7 million variants are described in ProteinGym: Large-Scale Benchmarks for Protein Fitness Prediction and Design. We evaluate on the same 217-assay, ~696k substitution suite used by TriFit, with both Spearman correlation against continuous fitness and AUROC on the binarized top/bottom 30% split, so every published number is directly comparable. A second suite covers indels, where TranceptEVE reports Spearman $\rho_s = 0.516$ on 300k indels and where most multimodal models simply fail; our alignment-based residual is designed for exactly this regime, and this is where we most expect to beat structure-fixed competitors. A third evaluation uses ClinVar clinical annotations, where TranceptEVE reports 0.915 average AUC. Finally, we evaluate on the shallow-alignment regime ($N_{eff}/L < 1$), where SSEmb reports $\rho_s = 0.45$ against 0.39 for the MSA Transformer ensemble, because a method that never fits a per-family language model should be least damaged when homologs are scarce — though we flag the honest risk that our per-family covariance is itself unreliable with very few homologs, and we will report performance stratified by family depth.
Baselines
The naive reference point is zero-shot sequence log-likelihood scoring with a pretrained protein language model: on the identical 217-assay binarized suite, ESM3 zero-shot achieves 0.769 AUROC as reported by TriFit. This baseline needs no training and no structures, so any complexity we add must pay for itself above 0.769. The strongest existing competitors are the supervised multimodal systems on the same suite: TriFit at 0.897 AUROC, Kermut at 0.864, and ProteinNPT at 0.844. We will run all of them from their published configurations on identical splits and binarization, and we will be explicit that these methods consume fitness labels that ours does not — the fair claim is not "we beat TriFit" but "we close X points of the label-free gap." For compute fairness we match the frozen-encoder budget (ESM-2 650M) and report a matched-compute supervised control: a ridge regression on the same frozen multimodal features, since ProteinNPT's authors showed pooled-representation regressions are the honest low-compute supervised reference. The decisive ablation isolates each ingredient: full-space Mahalanobis (all eigencomponents, no selection) against selected-component scoring; per-position residual subtraction without alignment against the aligned residual; and monolithic fused features against DPP-repulsed components. Beating the naive reference means exceeding 0.769 AUROC zero-shot; beating the strongest fair competitor means closing more than half the 0.128 gap between ESM3 and TriFit without labels.
Why not just…?
Why not just use sequence log-likelihoods from a big protein language model? This is the most serious threat, and we do not argue it away: ESM3 reaches 0.769 AUROC with no training at all, and TranceptEVE reaches 0.915 AUC on ClinVar. But the corpus gives two concrete blind spots our design targets. Pure sequence models degrade badly on shallow alignments — the MSA Transformer ensemble falls to $\rho_s = 0.39$ where SSEmb holds 0.45 — and they fail or degrade on indels, the regime TranceptEVE itself flags. Our experiments stratify by $N_{eff}/L$ and by substitution-versus-indel so this is settled empirically rather than by assertion; if PWT does not beat likelihood scoring on indels and low-depth families, the proposal's central claim fails.
Why not just concatenate ESM-2 and structure embeddings and train an MLP or graph network? Because the corpus quantifies how little this buys: TriFit found adding static structure to sequence yields only +0.2 AUROC, evidence that the modalities are nearly redundant as raw concatenations; and Video Infringement Detection via Feature Disentanglement and Mutual Information Maximization shows monolithic concatenated vectors overfit redundant noise, losing 2.8 mAP when disentanglement is removed. Moreover this alternative requires fitness labels, whose scarcity and assay bias are precisely the problem ProteinGym documents, and ProteinNPT's authors showed pooled-representation regressions collapse under extrapolation splits. Our matched-compute ridge control in the baselines makes the comparison direct: if the supervised concatenation wins on every split, the label-free setting was not worth the machinery.
Why not just apply PCA to the fused features? PCA is cheaper than our selection procedure and needs no homolog held-out scheme. But the greedy-eigencomponent paper directly tested this in anomaly detection and found maximal-variance components capture background rather than anomaly signal, with performance dropping as more high-variance components are added, while fewer than 10 non-consecutive chosen components reached 100% AUROC. Our design keeps their finding and removes their label dependence; the PCA-versus-selection ablation replicates this contrast in protein space, and if unsupervised selection does not beat PCA there, the label-free selection criterion is the component to abandon, not the whole proposal.
Decisive Checks
Three results would decide the proposal. First, on the 217-assay suite, selected-component scoring must beat both full-space Mahalanobis and PCA by a margin mirroring the vision finding, confirming that variant signal lives in low-variance, non-consecutive eigencomponents of protein feature spaces — untested until now, since all protein-dynamics work (TriFit included) uses maximum-variance lowest-frequency modes. Second, on the indel and low-depth strata, the aligned residual must beat raw log-likelihood scoring, vindicating the transplant of video temporal-alignment machinery. Third, the DRÆM-style mutation-as-corruption training must beat an uncorrupted autoencoder, testing whether the identical-reconstruction shortcut — a generative model healing the anomaly it should flag — is avoided by construction. If all three hold even partially, the payoff extends past this benchmark: the same label-free manifold machinery applies to any scientific domain where normal data is abundant and anomalies are open-set, expensive, and poorly simulated.
References
- Gaussian Image Anomaly Detection with Greedy Eigencomponent Selection
- DRÆM - A discriminatively trained reconstruction embedding for surface anomaly detection
- TriFit: Trimodal Fusion with Protein Dynamics for Mutation Fitness Prediction
- Integration of pre-trained protein language models into geometric deep learning networks
- SSEmb: A joint embedding of protein sequence and structure enables robust variant effect predictions
- INRIA-LEAR'S Video Copy Detection System
- Priors for Diversity in Generative Latent Variable Models
- ProteinGym: Large-Scale Benchmarks for Protein Fitness Prediction and Design
- TranceptEVE: Combining Family-specific and Family-agnostic Models of Protein Sequences for Improved Fitness Prediction
- Video Infringement Detection via Feature Disentanglement and Mutual Information Maximization