Accurately identifying essential genes in bacteria is critical for understanding microbial biology and developing novel antibiotics. However, the heterogeneity of biological data poses a challenge for reliable prediction. This study aims to enhance prediction accuracy by integrating diverse biological features through a multi-feature fusion framework. This study combined sequence data, gene annotations, protein–protein interaction networks, and subcellular localization information to construct a convolutional neural network (CNN)-based model, CNN4Essential. Feature importance was assessed using a random forest algorithm, and dimensionality reduction was performed with truncated singular value decomposition. The model was evaluated through intra-species prediction (INSP) and leave-a-species-out prediction (LASP) across 22 prokaryotic species. CNN4Essential achieved an average AUC of 0.884 in INSP, 0.726 in LASP, and 0.851 across all species, outperforming existing methods. Furthermore, predictions for Haemophilus influenzae were compared with known drug-target genes from DrugBank. A positive correlation between prediction scores and target gene matching rates was observed. The integration of multi-source features with a deep learning model significantly improves bacterial essential gene prediction. CNN4Essential not only surpasses single-feature and shallow models in performance but also holds promise for identifying potential drug targets.
Ye et al. (Mon,) studied this question.