PyaiVS unifies AI workflows to accelerate ligand discovery and yields ABCG2 inhibitors

Area of Science:

• Computational chemistry and cheminformatics
• Artificial intelligence in drug discovery
• Pharmacology and medicinal chemistry

Background:

• Optimized AI models for virtual screening necessitate careful selection of algorithms, molecular representations, and data splitting techniques.
• Existing tools lack integration, hindering the coordinated optimization of these critical components.
• The development of efficient AI-driven virtual screening models is crucial for accelerating drug discovery.

Purpose of the Study:

• To introduce PyaiVS, a Python package designed to integrate various machine learning algorithms, molecular representations, and data splitting strategies for virtual screening.
• To demonstrate the impact of coordinated optimization of algorithm architectures, molecular representations, and data splitting strategies on AI-driven virtual screening performance.
• To showcase the utility of PyaiVS by applying it in conjunction with pharmacophore modeling and docking to identify potential ABCG2 inhibitors.

Main Methods:

• PyaiVS integrates nine machine learning algorithms, five molecular representations (including ECFP4/MACCS fingerprints and graph-based representations), and three data splitting strategies (including clustering-based splitting).
• The study evaluated the performance of different combinations of algorithms, representations, and splitting strategies, highlighting the effectiveness of deep learning models (GCN, GAT, Attentive FP) and clustering-based splitting for optimal AUC-ROC.
• A large-scale virtual screening of 4,188,623 compounds for ABCG2 inhibitors was performed using PyaiVS, pharmacophore modeling, and docking.

Main Results:

• Coordinated optimization of algorithm architectures, molecular representations, and data splitting strategies is essential for efficient AI-driven virtual screening.
• Clustering-based data splitting achieved optimal AUC-ROC performance of 68.5%.
• Experimental validation identified four novel compounds (C1/C6/C7/C9) that bind to ABCG2 with sub-100 μM dissociation constants (kd) and potentiate topotecan cytotoxicity.

Conclusions:

• PyaiVS provides a unified and accessible platform for streamlining the development of optimized AI models for virtual screening.
• The findings underscore the importance of selecting appropriate molecular representations and data splitting strategies tailored to dataset size and model architecture.
• The identified ABCG2 inhibitors demonstrate the practical utility of PyaiVS in discovering potential therapeutic agents.