pLM-SAV

pLM-SAV is a pathogenicity classifier for single amino acid variants (SAVs) that works entirely from protein language model representations. Its central idea is the Δ-embedding: run a pretrained protein language model over both the wild-type and the mutant sequence, then subtract the two representations residue by residue. The resulting difference vector is fed to a small convolutional neural network that labels the substitution as neutral or effect-bearing. No multiple sequence alignment, predicted structure, substitution matrix, or curated annotation enters the pipeline.

The model was developed by Orsolya Gereben, Hedvig Tordai, Lana Khamisi, Erda Qorri, and Tamás Hegedűs at Semmelweis University with collaborators at the Hungarian Centre of Excellence for Molecular Medicine and the University of Szeged. It was posted to bioRxiv in May 2025 and published in Bioinformatics Advances in 2026. The work sits between two established approaches: supervised, feature-rich predictors such as SNAP2 and VESPA, and alignment- or structure-derived scorers such as AlphaMissense and the zero-shot ESM-1v. pLM-SAV asks a narrower question — how much signal survives in the raw arithmetic difference between two embeddings?

The answer, on the authors' own benchmarks, is a candid one. pLM-SAV matches feature-engineered supervised predictors on the datasets they were built for, but does not beat AlphaMissense or the REVEL ensemble on broad ClinVar evaluation. Where it does pull ahead is on the subsets those methods score ambiguously, suggesting the Δ-embedding carries signal complementary to alignment- and structure-based features.

#Key Features

• Δ-embedding representation: Wild-type minus mutant per-residue embeddings expose mutation-specific signal that extends several residues beyond the substitution site, reflecting the contextual encoding of protein language models.
• No handcrafted features: Unlike VESPA, which supplements embeddings with BLOSUM scores and substitution probabilities, pLM-SAV uses only pLM-derived vectors.
• Compact and locally deployable: The classifier is a two-layer 1D CNN, so scoring runs on a single GPU and requires no proteome-scale precomputed database.
• Leakage-controlled evaluation: Training used the Rost lab's original Eff10k fold assignment, and ClinVar test sets were PSI-BLAST filtered to exclude any protein sequence-similar to training or validation folds.
• Complementary on ambiguous variants: On the subsets where AlphaMissense or REVEL return low-confidence scores, pLM-SAV improved AUC, MCC, and F1 over the method that flagged them.

#Technical Details

Embeddings come from either Ankh (1,536 dimensions) or ProtT5-XL-U50 (1,024 dimensions), applied to full-length sequences. A seven-residue window — three before and three after the mutation — yields a 7×1,536 or 7×1,024 input array. The network stacks two 1D convolutions with ReLU, 8 output channels and kernel size 3, followed by layer normalization, a residual connection, ELU activation, global pooling, dropout, and a sigmoid-activated fully connected layer, implemented in PyTorch with BCEWithLogitsLoss and early stopping; the binary threshold is chosen by maximizing MCC. Training used the Eff10k set (100,737 annotations over 9,594 proteins; 99,926 used, 39,299 neutral and 60,627 effect) under nested k-fold cross-validation, in a 10-fold variant and a 9-fold variant that drops the outsized Fold 0. Ankh embeddings with the 9-fold scheme performed best throughout.

On PMD4k test folds, pLM-SAV reached MCC 0.274 and 0.271, against 0.280 for SNAP2 and comparable values for VESPA and VESPAl; on the human subset it led with MCC 0.292 and F1 0.497 for the neutral class. Independent ClinVar sets contained 17,677 and 83,642 variants. On ProteinGym v0.1, pLM-SAV averaged a Spearman correlation of 0.382 across 72 proteins, below AlphaMissense (0.512) and in the range of ESM-1b (0.358).

#Applications

The most direct use is triaging variants of uncertain significance in clinical genomics, particularly as a second opinion on substitutions that established scorers leave in their ambiguous band. Because the pipeline needs only a sequence and a mutation list, it also applies to non-human proteins, which proteome-scale human resources do not cover, and to protein engineering campaigns where the question is whether a designed substitution is tolerated. The lightweight inference path suits groups that must run scoring locally on their own sequences rather than querying a precomputed database.

#Impact

pLM-SAV's contribution is less a new state of the art than a clean measurement of how far a single representational trick goes. The authors are explicit that AlphaMissense and REVEL retain higher overall performance on broad ClinVar benchmarks, and that the mismatch between Eff10k's "any measurable effect" labels and ClinVar's benign/pathogenic distinction caps achievable accuracy. What the study establishes is that Δ-embeddings are a reusable feature worth folding into larger mutation-effect predictors and integrative models. Code, the nine best Ankh checkpoints with their optimal thresholds, and usage scripts are archived on Zenodo under CC BY 4.0; there is no GitHub or HuggingFace distribution, and the released training scripts contain hard-coded internal paths and omit the Eff10k data, which the authors are not licensed to redistribute.