ProVenTL: a transfer-learning framework for predicting peptide-protein interactions derived from snake venom for cancer therapeutics.

Accurate prediction of peptide-protein interactions (PepPI) is crucial for advancing peptide-based anticancer drug design. In this study, we introduce ProVenTL, a computer-aided molecular design framework that leverages transfer learning and protein language model embeddings to enhance PepPI prediction accuracy and interpretability. Two complementary strategies were explored: (i) fine-tuning a CAMP model pretrained on large-scale PepPI data from the Protein Data Bank (PDB) using a curated dataset of Calloselasma rhodostoma venom peptides and cancer-related proteins, and (ii) integrating ProtT5 embeddings with stacked autoencoder-deep neural networks (SAE-DNN) and TabNet classifiers. Models were comprehensively benchmarked against baseline configurations and representative deep-learning approaches using standard classification metrics, while biological relevance was evaluated through functional enrichment and pathway analysis of top-ranked predictions. Compared with baseline configurations and conventional deep-learning approaches, the ProtT5-based SAE-DNN model achieved the best performance (accuracy = 0.78; ROC-AUC = 0.86), demonstrating improved generalization capability on a small, domain-specific venom peptide dataset. The model identified key targets such as TRBC2, CD274, HIF1AN, PCSK9, and PLAU, which are associated with pathways involved in immune suppression, hypoxia regulation, lipid metabolism, and metastasis. This study highlights the utility of transfer learning and protein language models for PepPI prediction in data-limited scenarios and establishes a computational framework for prioritizing snake-venom-derived peptides for anticancer drug discovery and future experimental validation.