Supporting data for "Prediction and Classification of Bacterial Virulence Factors using Deep Learning"

This study investigates the use of protein language model embeddings along with evolutionary features extracted from PSSM as features for deep learning models to identify and classify VF sequences. The source files contain raw data (i.e., VF and non-VF sequences), processed data, and the model training and validation scripts. The trained model is provided in a compiled standalone python package called DeepVIC. Positive datasets were aggregated from 4 VF databases (i.e., VFDB, Victors, UniProtKB, and PATRIC (or BV-BRC)) while negative sequences were retrieved from UniProtKB using keyword filtering. The Classification of VFs is based on the VFDB2022 schema, where non-VFDB data were assigned VFDB classes through BLASTP hits.This study also examines the interpretability of deep learning models using SHAP values by perturbation of VF-related domains in VF sequences. The SHAP values provide a quantitative examination of how VF-related domains affect model prediction output. Furthermore, feature importance examination also provides insight into how different features improve model predictions.