Introduction Essential proteins are key to cellular viability, yet their experimental identification is costly and time-consuming. Methods In this study, DLAM is introduced as a deep learning framework that integrates four complementary biological cues, namely, domain composition, subcellular localization, orthology, and gene expression, together with a weighted protein–protein interaction network. The heterogeneous signals are encoded into compact representations and learned by an attention-enhanced network to score protein essentiality. Results On the DIP dataset, DLAM achieves consistently better performance than representative centrality measures and conventional machine-learning classifiers. In a further expanded baseline study based on the larger BioGRID dataset containing more proteins, we conducted comparative experiments between DLAM and four recently proposed deep learning methods (TCBB2021, EPGAT, BMC2022, and ACDMBI). On the BioGRID dataset, we evaluated DLAM using stratified five-fold cross-validation. Across folds, DLAM achieves consistently strong discrimination and ranking performance (reported as the mean ± std. for ROC-AUC and AP) and maintains a stable F1-score under a validation-selected decision threshold. This suggests that, under the same evaluation protocol, DLAM has strong ranking and discrimination capability. Moreover, it also exhibits good and stable performance on other metrics such as accuracy, precision, recall, and F-measure. Discussion These results indicate that jointly modeling multi-source biological information with interaction topology yields more reliable essential-protein prediction under class imbalance.
Zhou et al. (Tue,) studied this question.