Cross-Modal Multi-Source Transfer Learning for Predicting Host Species of Bacteriophage Receptor-Binding Proteins

Abstract:
Predicting host species of bacteriophages through receptor-binding protein (RBP) sequences remains critical for phage therapy and microbial ecology. However, limited labeled biological data significantly constrain traditional prediction models. We propose a cross-modal multi-source transfer learning framework, leveraging abundant chemical toxicity data (ClinTox dataset) to enhance host prediction accuracy from sparse bacteriophage sequence data. Our approach integrates a shared BiLSTM encoder and Maximum Mean Discrepancy (MMD) loss for effective domain alignment. Experimental results demonstrate that our method achieves rapid convergence and up to 30% accuracy improvement during early training epochs compared to baseline models. Robust performance of the chemical domain classifier further validates the feasibility of our transfer learning strategy, offering promising implications for targeted phage therapeutic interventions.