Enhancing kinase-inhibitor activity and selectivity prediction through contrastive learning.

Area of Science:

• Biochemistry
• Computational Biology
• Drug Discovery

Background:

• Developing selective kinase inhibitors is difficult due to conserved protein structures and expensive screening.
• Accurate prediction of kinase-inhibitor affinity and specificity is crucial for efficient drug development.

Purpose of the Study:

• To present MMCLKin, a deep learning framework for predicting kinase-inhibitor activity and selectivity.
• To demonstrate MMCLKin's superior performance and generalizability compared to existing methods.

Main Methods:

• Developed MMCLKin, an attention consistency-guided contrastive learning framework.
• Integrated geometric graph and sequence networks with multi-head attention and multimodal, multiscale contrastive learning.
• Validated MMCLKin on multiple 3D kinase-drug, protein-drug, and mutation-aware datasets.

Main Results:

• MMCLKin outperformed existing methods across diverse datasets.
• The framework demonstrated strong generalizability on known and unknown kinase structures.
• Attention analysis identified key residues and functional groups for kinase-inhibitor binding.
• Experimental validation confirmed MMCLKin's ability to identify potent inhibitors, including against the LRRK2 G2019S mutant.

Conclusions:

• MMCLKin provides accurate and interpretable predictions of kinase-inhibitor interactions.
• The framework effectively screens for potent and selective kinase inhibitors.
• MMCLKin is a valuable tool for advancing kinase inhibitor drug discovery.