Research

Signal peptides direct proteins to the secretory pathway, and their accurate identification is a long-standing sequence-function prediction problem in proteomics. This project built an end-to-end pipeline for eukaryotic signal-peptide prediction using high-confidence UniProtKB/Swiss-Prot annotations, with MMseqs2 clustering applied for redundancy reduction and biologically meaningful train/benchmark splits.

Four approaches were benchmarked head-to-head: the classical von Heijne position-specific weight matrix, a biochemical-feature SVM, a CNN-BiLSTM sequence model, and a model built on ESM-2 protein language model embeddings. The ESM-2-based model achieved the strongest performance (F1 = 0.962, MCC = 0.957) on a held-out benchmark set. The analysis also examined amino-acid composition, signal-peptide length distributions, kingdom-specific bias, and cleavage-site motifs to ground the results biologically.

Protein domain annotation connects classical homology detection to modern representation-learning approaches. This project built an automated workflow using profile Hidden Markov Models (HMMER) and sequence-similarity search (MMseqs2) to detect Kunitz/BPTI-type protease inhibitor domains, processing structure-derived alignments to support domain-level annotation.

Under two-fold cross-validation, the profile-HMM approach reached MCC = 0.997, slightly outperforming the sequence-similarity approach (MCC = 0.991). The profile-HMM approach slightly outperformed sequence-similarity search, illustrating its usefulness for detecting divergent members of a protein family — relevant to remote homology detection more broadly.