PDL1-inhi.predictor: a two-tiered machine learning framework for predicting and prioritizing PDL1 inhibitors

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning in oncology

Background:

• Small-molecule inhibition of PD-1/PDL1 interactions, specifically targeting PDL1 dimerization, is a promising strategy for cancer treatment.
• Developing effective PDL1 inhibitors remains challenging, necessitating advanced computational approaches.

Purpose of the Study:

• To accelerate the discovery of effective PDL1 inhibitors using machine learning (ML) models.
• To develop a two-tiered screening strategy combining ML-based activity prediction and affinity estimation.

Main Methods:

• Systematic construction and evaluation of 60 ML models (30 classifiers, 30 regressors) for PDL1-specific predictions.
• Screening of a 12-million-compound library, followed by molecular docking and HTRF-based assays.
• Molecular dynamics simulations and SwissADME profiling for lead compound validation.

Main Results:

• Identified Stacking_standard as the best classifier and SVM_standard as the optimal regressor.
• Screened library yielded 6704 potential inhibitors; 7 top candidates showed strong binding affinity (≤ -11.5 kcal/mol).
• Two compounds confirmed potent PDL1 inhibition (IC50: 17.49 nM and 63.17 nM) with stable binding conformations.

Conclusions:

• The developed ML pipeline effectively identifies potent PDL1 inhibitors.
• Discovered critical substructural motifs for rational inhibitor design and identified lead compounds with favorable drug-likeness.
• The PDL1-inhi.predictor web server provides a valuable resource for researchers in PDL1 inhibitor discovery.