Structure-based virtual screening for PDL1 dimerizers: Evaluating generic scoring functions.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning in bioinformatics

Background:

• The programmed cell death protein 1 (PD1) and its ligand PDL1 pathway promotes tumor progression and resistance to apoptosis.
• Inhibiting the PD1/PDL1 interaction via small-molecule binders that induce PDL1 dimerization is an innovative therapeutic strategy.
• Structure-based virtual screening (SBVS) is a key computational method for identifying potential small-molecule inhibitors.

Purpose of the Study:

• To evaluate the suitability of different generic scoring functions for structure-based virtual screening of small-molecule PD1/PDL1 inhibitors.
• To compare the performance of machine learning-based scoring functions against classical and fingerprint-based methods.

Main Methods:

• Utilized CNN-Score (a convolutional neural network ensemble), Smina (a classical scoring function), and IFP (a structural fingerprint similarity scoring function).
• Evaluated scoring functions on two test sets comprising known small-molecule PD1/PDL1 inhibitors and different types of inactive molecules (true vs. decoy).
• Assessed predictive performance using both true and property-matched decoy inactive sets to mitigate bias.

Main Results:

• CNN-Score significantly outperformed both Smina and IFP across both test sets.
• Smina demonstrated superior performance compared to IFP.
• The robust performance of CNN-Score, even without decoy bias, highlights its predictive power for PDL1 inhibitors.

Conclusions:

• CNN-Score is a highly effective scoring function for structure-based virtual screening of PD1/PDL1 inhibitors.
• Re-scoring docked molecules with CNN-Score represents a promising approach for discovering novel small-molecule inhibitors targeting this crucial cancer pathway.