BERT-T6: Toward High-Accuracy T6SS Bacterial Toxin Identification Using a Protein Language Model

Type VI secretion system effectors (T6SEs) are key virulence factors that disrupt critical cellular components in target cells, facilitating bacterial competition or host immune evasion. Accurate identification of T6SEs is therefore crucial for understanding bacterial pathogenesis. Although several computational predictors based on traditional machine learning and deep learning have been developed, their performance still requires improvement. In this study, we systematically evaluated a range of sequence-based features and embeddings from pretrained protein language models for T6SE prediction. Among these, ProtBert embeddings were identified as the most effective representation. Building on this finding, we present BERT-T6, a predictor fine-tuned from ProtBert via transfer learning for T6SE classification with imbalance awareness. It demonstrated good generalizability in independent test and achieved state-of-the-art performance, with a mean accuracy of 0.959, a sensitivity of 0.909, a specificity of 0.973, a precision of 0.905, an F1-score of 0.907, and an MCC of 0.881 in five repeated experiments. This work highlights the effectiveness of using a protein language model with transfer learning and imbalance-aware training for T6SE prediction. BERT-T6 provides a valuable tool for identifying T6SEs, supporting further investigation of bacterial virulence mechanisms.