#3 · ML / AI methods

Protein language model embeddings enable sequence-based prediction of protein traits

The common thread is reuse of pretrained protein representations as input features for task-specific prediction. Evoformer latent representations from AlphaFold2 are evaluated for calpain cleavage and nucleic-acid-binding site prediction, while mean-pooled ProtT5-XL embeddings feed logistic regression, gradient boosting, and neural networks for CAFA-6 Gene Ontology prediction. PALM applies transfer learning from protein language model embeddings to peptide aggregation and aggregation-prone residues, and an open-source ProtTrans-based predictor addresses plant food allergenicity, outperforming similarity-based allergen approaches. The DeepDrug Protein Embeddings Bank extends this direction to multimodal human protein representations for protein-protein interaction prediction. What is new is the breadth of targets addressed through general-purpose embeddings, spanning cleavage, aggregation, allergenicity, Gene Ontology terms, and interactions.