A deep learning model for predicting essential proteins based on an attention mechanism.

Area of Science:

• Bioinformatics
• Computational Biology
• Systems Biology

Background:

• Identifying essential proteins is critical for understanding cellular viability.
• Experimental methods for essential protein identification are resource-intensive and time-consuming.

Purpose of the Study:

• To introduce DLAM, a novel deep learning framework for predicting essential proteins.
• To evaluate DLAM's performance against existing methods using benchmark datasets.

Main Methods:

• DLAM integrates four biological cues: domain composition, subcellular localization, orthology, and gene expression.
• A weighted protein-protein interaction network is incorporated into the framework.
• An attention-enhanced network encodes heterogeneous signals for essentiality scoring.

Main Results:

• DLAM demonstrates superior performance compared to centrality measures and conventional machine learning classifiers on the DIP dataset.
• On the larger BioGRID dataset, DLAM outperforms recent deep learning methods (TCBB2021, EPGAT, BMC2022, ACDMBI).
• DLAM achieves strong and stable discrimination and ranking performance across multiple metrics (ROC-AUC, AP, F1-score, accuracy, precision, recall).

Conclusions:

• Jointly modeling multi-source biological information with interaction topology enhances essential-protein prediction reliability.
• DLAM provides a robust and accurate computational approach for identifying essential proteins.
• The framework effectively addresses the challenge of class imbalance in essential protein prediction.