GeminiMol

PD-1/PD-L1 are key immune checkpoint targets that suppress T-cell-mediated anti-tumor immunity, representing a major focus in cancer immunotherapy. While antibody drugs dominate the clinic, they are limited by administration challenges and immune-related side effects. Small-molecule PD-1/PD-L1 inhibitors, with oral availability, good tissue penetration and low cost, have emerged as a promising next-generation strategy.

A PD-1/PD-L1 lead-like library was built via a five-step virtual screening process. After collecting 8,947 inhibitors from BindingDB and PubChem and filtering by activity and duplicates, AI similarity screening was performed using GeminiMol. Key pharmacophores were extracted from the PPI interface of co-crystal structures, and molecular was screened via a pharmacophore model, effectively enhancing target activity.

Containing 10,000 structurally diverse and drug-like molecules well-matched to the PD-L1 pocket, the library supports virtual docking, high-throughput screening and hit discovery, enabling efficient and rapid development of small-molecule immunotherapies.

RNA is crucial for the regulation of numerous cellular processes and functions. With the in-depth study of disease mechanisms, processes such as RNA expression, splicing, translation, and stability regulation have become new targets for disease intervention. RNA has provided new therapeutic modalities for metabolic diseases, genetic disorders, and cancer patients, resulting in several innovative drugs.

MCE R&D team collected small molecules targeting RNA from the PDB, R-BIND, ROBIN, and internal database as the positive dataset, and non-targeting RNA small molecules from ROBIN as the negative dataset. Based on the GeminiMol pre-trained model, we encoded the molecules and calculated over 1700 molecular descriptors using Mordred as inputs for the model. Subsequently, we employed 13 deep learning models to learn from the data. All of which yielded good training results, with AUROCs greater than 0.75. Ultimately, we selected the Finetune model to screen HY-L901P, which exhibited the best classification performance, achieving an AUROC of 0.82 and a prediction accuracy of 0.76. We then applied filtering based on StaR rules (with at least two of the following properties: cLogP ≥ 1.5, Molar Refractivity ≥ 4, Relative Polar Surface Area ≤ 0.3) to obtain a library containing approximately 5,000 small molecule compounds targeting RNA. This library serves as a valuable tool for screening small molecules that interact with RNA.