LysinFusion: Integrating Multi-Feature Encoding and Hybrid CNN-Transformer Architecture for Phage Lysin Prediction
Journal:
bioRxiv
Published Date:
Mar 16, 2026
Abstract
The growing threat of antimicrobial resistance underscores the demand for new therapeutic modalities. Phage lysins are promising candidates, yet their large-scale discovery from expanding genomic resources remains limited. Existing computational methods often rely on sequence homology or lack reproducibility, restricting their utility for high-throughput screening. Here, we introduce LysinFusion, a reproducible deep learning framework that integrates heterogeneous sequence features with a hybrid CNN-Transformer architecture for accurate lysin identification. Trained on curated datasets, LysinFusion outperforms DeepMineLys on an independent benchmark (n=148), achieving an accuracy of 0.8108, AUC of 0.8921, and MCC of 0.6225, with substantially fewer false positives, thereby reducing downstream validation cost. Model interpretability analyses reveal consistent attention to functionally critical regions--the N-terminal catalytic domain and C-terminal membrane-penetrating domain--supporting the biological relevance of predictions. The framework is available at https://github.com/sinuo560/LysinFusion.
