Innovative Virtual Screening of PD-L1 Inhibitors: The Synergy of Molecular Similarity, Neural Networks, and GNINA Docking

Immune checkpoint inhibitors have garnered significant attention in oncological research over recent years. A plethora of studies have elucidated that inhibitors targeting the Programmed Death-Ligand 1 (PD-L1) play a pivotal role in circumventing the evasion mechanisms of cancer cells against the immune system. This study aimed to develop an integrated screening model combining an Artificial Neural Network (ANN), Molecular Similarity (MS) assessments, and GNINA 1.0 molecular docking, targeting PD-L1 inhibitors. A database of 2044 substances with known PD-L1 inhibitory activity was compiled from Google Patents and used to enhance molecular similarity evaluations and train the machine learning model. For retrospective validation of the docking procedure, the human PD-L1 protein, with the Protein Data Bank (PDB) ID: 5N2F, was employed as a control. In this phase of the study, 15,235 compounds from the DrugBank database were subjected to a series of screening processes: initially through medicinal chemistry filters, followed by MS assessments, the ANN model, and culminating with molecular docking using GNINA 1.0. The decoy generation yielded promising outcomes, evidenced by an AUC-ROC 1NN value of 0.52 and Doppelganger scores with a mean of 0.24 and a maximum of 0.346, indicating a high resemblance of the decoys to the active set. For MS, the AVALON emerged as the most effective fingerprint for similarity searching, demonstrating an Enrichment Factor (EF) of 1% at 10.96%, an AUC-ROC of 0.963, and an optimal similarity threshold of 0.32. The ANN model demonstrated superior performance in cross-validation, achieving an average precision of 0.863±0.032 and an F1 score of 0.745±0.039, outperforming both the Support Vector Classifier (SVC) and Random Forest (RF) models, albeit not significantly. In external validation, the ANN model maintained its superiority with an average precision of 0.851 and an F1 score of 0.790. GNINA 1.0, employed for molecular docking, was validated through redocking and retrospective control, achieving an AUC of 0.975, with a critical cnn_pose_score threshold of 0.73. From the initial 15,235 compounds, 128 were shortlisted using the MS and ANN models. Further screening through GNINA 1.0 identified 22 potential candidates, among which (3S)-1-(4-acetylphenyl)-5-oxopyrrolidine-3-carboxylic acid emerged as the most promising, with a cnn_pose_score of 0.79, a PD-L1 inhibitory probability of 70.5%, and a Tanimoto coefficient of 0.35.