Protein Family Classification Using Sequence Embeddings

Which Pfam family does a protein belong to - and is a protein language model better than aligning it to known proteins? This project classifies 3,600 reviewed UniProtKB/Swiss-Prot proteins of 24 Pfam families with ESM-2 embeddings (logistic regression, SVM, neural network), composition and 3-mer descriptors, a CNN trained from scratch and a Smith-Waterman nearest-neighbour baseline, using cross-validation that keeps identity clusters together. A random split flatters every method; with relatives above 40 % identity removed, ESM-2 keeps 0.98 accuracy, Smith-Waterman 0.90 and composition 0.74. On unreviewed TrEMBL proteins the simple alignment is as good as ESM-2 (0.978 each).

Category: Bioinformatics - Proteomics and Protein Bioinformatics

Proteomics & Protein Bioinformatics Project B21 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.