AI-driven design of novel PARP inhibitors.

Abstract

3091 Background: Inhibitors of the Poly (ADP-ribose) polymerase (PARP) family play a role in treating HER2-negative locally advanced or metastatic breast cancer with germline BRCA1/2 (gBRCA) mutations, as well as in the maintenance treatment of gBRCA-associated metastatic pancreatic ductal adenocarcinoma. However, the design of de novo small molecules targeting proteins like PARP remains time consuming and resource intensive. It is hypothesized that generative models trained on molecular graph encodings could accelerate the design of novel PARP inhibitors. Objective:This study aims to develop a generative model capable of designing novel, orally bioavailable PARP inhibitors. Methods: A large language model was pre-trained on 1 million chemical structures sourced from the ChEMBL database. Each structure was represented as a Simplified Molecular Input Line Entry System (SMILES) string, which was tokenized into discrete atomic and functional group-level tokens. The model leverages an Average-Stochastic Gradient Descent Weight-Dropped Long Short-Term Memory (AWD-LSTM) architecture. Transfer learning was applied to adapt the pre-trained model to specific target chemical structures, enabling domain-specific fine-tuning for the de novo design of PARP inhibitors. Results: The model demonstrated robust performance in generating chemically valid, unique, novel, and diverse PARP inhibitors. It achieved a validity rate, uniqueness rate, and novelty rate of 100%, along with a diversity score of 81.53%. Furthermore, the generated molecules exhibited favorable physicochemical properties, including a molecular weight of 417.52 Da, a logarithm of the partition coefficient (LogP) of 2.58, a topological polar surface area (TPSA) of 93.21 Angstrom squared, an average of 4.05 rotatable bonds, 1.84 hydrogen bond donors, and 5.14 hydrogen bond acceptors. Conclusions: A generative model was developed to design novel, orally bioavailable PARP inhibitors. It provides an efficient and automated tool for de novo small molecule design with tailored molecular and pharmacological properties, potentially accelerating the development of PARP inhibitors.