Bioactivity Deep Learning for Complex Structure-Free Compound-Protein Interaction Prediction.

Area of Science:

• Computational Chemistry and Cheminformatics
• Bioinformatics and Computational Biology
• Drug Discovery and Development

Background:

• Accurate protein-ligand binding affinity prediction is crucial for virtual drug screening.
• Traditional methods depend on limited protein-ligand crystal structures.
• Structure-free compound-protein interaction (CPI) methods offer alternatives using bioactivity data but face challenges with data heterogeneity and activity cliffs (ACs).

Purpose of the Study:

• To address limitations in structure-free CPI prediction, particularly data heterogeneity and ACs.
• To introduce a large-scale benchmark dataset (CPI2M) with AC annotations.
• To develop and validate a novel deep learning model (GGAP-CPI) for robust CPI prediction.

Main Methods:

• Creation of CPI2M, a dataset with ~2 million bioactivity points across four activity types (Ki, Kd, EC50, IC50) and AC annotations.
• Development of GGAP-CPI, a structure-free deep learning model utilizing integrated bioactivity learning and advanced protein representation.
• Comprehensive evaluation of GGAP-CPI against 19 baseline methods across 4 prediction scenarios and 7 benchmark datasets.

Main Results:

• GGAP-CPI significantly outperforms 12 target-specific and 7 general CPI baselines.
• The model demonstrates superior performance in general CPI prediction, rare protein prediction, transfer learning, and virtual screening.
• GGAP-CPI provides stable bioactivity predictions, measures prediction uncertainty, and identifies binding pocket residues and interactions.

Conclusions:

• GGAP-CPI represents a significant advancement in structure-free CPI prediction, effectively addressing data heterogeneity and activity cliffs.
• The model's ability to predict bioactivity, quantify uncertainty, and enrich interaction data highlights its practical utility in drug discovery.
• The CPI2M dataset and GGAP-CPI model provide valuable resources for advancing computational drug screening and bioactivity assessment.