Residue-Level Multiview Deep Learning for ATP Binding Site Prediction and Applications in Kinase Inhibitors.

Area of Science:

• Computational biology
• Drug discovery
• Structural bioinformatics

Background:

• Accurate identification of adenosine triphosphate (ATP) binding sites is critical for understanding cellular mechanisms and developing targeted therapies, especially for kinase inhibitors in cancer treatment.
• Current methods for ATP binding site identification are often hindered by lengthy precomputed feature requirements and significant data imbalance, limiting their direct applicability in drug discovery.
• The utility of existing prediction models in practical drug discovery scenarios, particularly for enhancing inhibitor design, remains underexplored.

Purpose of the Study:

• To introduce novel computational methods, Multiview-ATPBind and ResiBoost, for accurate and efficient prediction of ATP binding sites.
• To address the limitations of existing methods by integrating diverse data types and mitigating data imbalance issues.
• To demonstrate the practical utility of the proposed methods in drug discovery, specifically for kinase inhibitors.

Main Methods:

• Developed Multiview-ATPBind, an end-to-end deep learning model integrating 1D sequence and 3D structural data for residue-level prediction of pocket-ligand interactions.
• Introduced ResiBoost, a residue-level boosting algorithm to counteract data imbalance and improve the prediction of rare positive binding residues.
• Validated the model's performance on benchmark datasets using both experimental and AI-predicted protein structures.

Main Results:

• The proposed approach, Multiview-ATPBind combined with ResiBoost, significantly outperformed state-of-the-art models on benchmark datasets.
• Achieved substantial improvements in balanced performance metrics, indicating robust prediction capabilities across diverse datasets.
• Demonstrated successful application in predicting binding sites for kinase inhibitors like imatinib and dasatinib, and enhancing molecular docking simulations.

Conclusions:

• Multiview-ATPBind and ResiBoost offer a rapid, precise, and robust method for identifying ATP binding sites, overcoming key limitations of previous approaches.
• The integrated deep learning and boosting strategy effectively handles complex binding site data and improves prediction accuracy.
• The demonstrated success in kinase inhibitor drug discovery highlights the significant potential of this method for advancing therapeutic development.