facebookresearch/esm README.md

What this source contributed

Records in this release that cite this source. A record can cite several sources.

• 6 configurations
• 6 pipelines
• 6 evidence claims

Claims

Statements this source supports, each with where in the source it was checked.

• CLAPE-SMB with ESM-2: description: Description Predicts small-molecule-binding residues from sequence alone, using frozen esm2_t33_650M_UR50D residue embeddings with a contrastive-learning binding-site predictor. (Reviewed profile (2026-09-16): summary at Conclusion (paragraph 1); Results/Comparison of CLAPE-SMB with DeepProSite and heuristic analyses in three protein case studies (paragraph 3); architecture fact at Methods/Sequence embedding (paragraph 2); Methods/Loss function (paragraph 5); description text also read at Abstract)
• ESM-2 650M embeddings + classifier: description: Description Chimeric antibody-deamidation predictor: a global module of ESM-2 650M embeddings of the full antibody sequence is concatenated with a local windowed-sequence module and classified by a dense head. (Reviewed profile (2026-09-16): summary at 3. Results/3.4. Performance Evaluation of Models (paragraph 4); Abstract (paragraph 1); architecture fact at Abstract (paragraph 1); 4. Discussion and Conclusions (paragraph 2); description text also read at 3. Results/3.3 Model architecture (chimeric model figure); 2. Methods (esm2_t33_650M_UR50D selection))
• ESM-2 embedding + paper classifier: description: Description ESM-2-only feature configuration of a clathrin-identification study: 1,280-dimensional ESM-2 embeddings feed an LSTM classifier, separately from the fused multi-model PLM-CLA predictor. (Reviewed profile (2026-09-16): summary at Materials and methods/Feature selection method (paragraph 1); Introduction (paragraph 3); architecture fact at Materials and methods/Overall framework of PLM-CLA (paragraph 1); Results and discussions/The effect of feature selection methods on the predictive performance (paragraph 1); description text also read at Results and discussions/Performance evaluation of individual feature embeddings (embedding dimensions and per-PLM LSTM models))
• ESM2 OFS pseudo-perplexity: description: Description Estimates protein fitness from the negative pseudo-perplexity of a sequence, approximated in a single unmasked forward pass by an ensemble of eight multilayer perceptrons that predict the ESM-2 650M masked residue profiles. (Reviewed profile (2026-09-16): summary at Section III, One Fell Swoop; Figure 1; Section IV fitness evaluation; architecture fact at Section III, One Fell Swoop; Figure 1; Section IV fitness evaluation; description text also read at Abstract; Section III Pseudo-perplexity in one fell swoop)
• ESM2_AMPS: description: Description Predicts binary protein-protein interactions from ESM-2 representations of sequence segments fused by a transformer encoder; AMPS is the segment-feature variant of the framework. (Reviewed profile (2026-09-16): summary at Methods/Construction of dual models based on feature integration (paragraph 3); Results/Attention mechanism reveal protein–protein interaction mechanism (paragraph 1); architecture fact at Methods/Interpretability analysis utilizing attention mechanisms (paragraph 1); Methods/Construction of dual models based on feature integration (paragraph 3); description text also read at Abstract)
• SPIN + ESM2-35M: description: Description Classifies proteins into family-specific functional classes with an ESM2-35M backbone whose first six encoder layers are frozen, adapting a question-answering architecture to locate a domain span and classify the selected region. (Reviewed profile (2026-09-16): summary at 3 Results/3.4 Analysis of the SH3 protein family and representational augmentation across clades (paragraph 4); 3 Results/3.3 SPIN on thioredoxin, a widespread protein family (paragraph 1); architecture fact at 2 Materials and methods/2.1 SPIN architecture (paragraph 1); 2 Materials and methods/2.1 SPIN architecture (paragraph 7); description text also read at Abstract; 2 Materials and methods/2.1 SPIN architecture (layer freezing))

Models, benchmarks and other records

Evidence

Source checking verifies the cited claim or transcription. It does not establish independent reproduction.