DeepVIC: modular prediction and classification of bacterial virulence factors using protein language model embeddings.

MOTIVATION: Virulence factors (VFs) play critical roles in bacterial pathogenesis, and identifying these proteins and elucidating their mechanisms is essential for developing effective infection treatments. While protein language models (PLMs) have revolutionized the protein analysis, current in silico methods for classifying VFs primarily rely on fine-tuned PLMs ProtBert-BFD that perform identification and classification in a single step. This monolithic approach lacks flexibility and is vulnerable to error propagation, particularly from false negatives. Moreover, the direct use of PLM embeddings as features for VF prediction and classification remains underexplored. RESULTS: To overcome these limitations, we introduce the Deep learning VF Identifier and Classifier (DeepVIC), a modular framework that separates identification and classification into distinct states. The identification module employs a binary classifier using information-rich PLM embeddings. The classification module then integrates these embeddings with evolutionary features derived from position-specific scoring matrices (PSSMs) in a multiclass classifier, structured according to the virulence factor database (VFDB) schema. We benchmarked DeepVIC against six state-of-the-art VF classifiers and predictors using a large independent holdout dataset and two literature-curated datasets for Streptococcus pneumoniae and Lactococcus spp. DeepVIC demonstrated strong generalizability, achieving robust and balanced performance on the independent holdout dataset and competitive recall of putative VFs from existing literature. The modular, task-separation architecture adopted in DeepVIC offers exceptional flexibility, making it a uniquely adaptable and competitive framework. Model interpretability strategies reveal that DeepVIC leverages biologically relevant motifs and effectively integrates PLM embeddings with evolutionary features, offering insights into how deep learning models interpret VFs.