Explainable Deep Relational Networks for Predicting Compound-Protein Affinities and Contacts.

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Predicting compound-protein affinity accelerates drug discovery but often requires 3D structure data.
• Existing structure-free machine learning methods prioritize accuracy over interpretability.
• Attention mechanisms in current models are insufficient for understanding underlying interactions.

Purpose of the Study:

• To develop interpretable, structure-free machine learning models for compound-protein affinity prediction.
• To identify and leverage intermolecular contacts for enhanced interpretability.
• To assess the interpretability and generalizability of novel deep learning approaches.

Main Methods:

• Formulated a hierarchical multiobjective learning problem for contact and affinity prediction.
• Employed hierarchical recurrent neural networks for protein sequences and graph neural networks for compound graphs.
• Introduced joint attention mechanisms between protein residues and compound atoms.
• Developed three interpretability-enhancing advances: structure-aware attention regularization, attention supervision using known contacts, and an intrinsically explainable architecture.

Main Results:

• Achieved generalizable affinity prediction for novel and dissimilar molecules.
• Significantly improved interpretability compared to state-of-the-art methods.
• Boosted contact prediction AUPRC by 33- to 35-fold for various test sets.
• Demonstrated utility in contact-assisted docking and structure-free binding site prediction.

Conclusions:

• Developed novel, interpretable deep learning models (DeepAffinity+ and DeepRelations) for structure-free compound-protein affinity prediction.
• These models provide superior interpretability and comparable/better accuracy than existing methods.
• The approach offers potential for advancing drug discovery through explainable AI.