Towards novel small-molecule inhibitors blocking PD-1/PD-L1 pathway: From explainable machine learning models to...

Area of Science:

• Computational chemistry
• Drug discovery
• Immunotherapy

Background:

• The programmed cell death-1 (PD-1)/programmed cell death ligand-1 (PD-L1) pathway is a key target in cancer immunotherapy.
• Machine learning (ML) accelerates drug design but often lacks interpretability, hindering optimization.
• Existing ML models for drug design are often

Purpose of the Study:

• To develop explainable ML models for rational drug design targeting the PD-1/PD-L1 pathway.
• To identify key molecular fragments influencing bioactivity for inhibitor optimization.
• To discover novel small-molecule inhibitors with improved efficacy and safety profiles.

Main Methods:

• Constructed five explainable ML models by integrating ML algorithms with the SHAP method.
• Pretrained models on over 4000 molecules, achieving R² values from 0.835 to 0.86.
• Modified BMS-1166 based on model insights, generating 60 novel compounds, followed by docking and ADMET prediction.

Main Results:

• Identified structure-activity relationships for BMS-1166 using explainable ML.
• Screened three novel compounds (C27, C52, C54) with superior docking scores and lower toxicity than BMS-1166.
• Validated enhanced binding affinity of C27 and C52 to the PD-L1 dimer via molecular dynamics simulations.

Conclusions:

• Developed an efficient, explainable ML-driven protocol for rational design of PD-1/PD-L1 inhibitors.
• Demonstrated the utility of explainable AI in overcoming ML model interpretability challenges in drug discovery.
• Identified promising lead compounds for further development in cancer immunotherapy.