BioMolFormer: A Novel Approach to Molecular Generation with SELFIES and Monte Carlo Tree Search

In this study, we present BioMolFormer, a novel model to enhance the generation and optimization of drug candidates, especially for their binding affinity to target proteins. BioMolFormer integrates Transformer networks and MCTS to improve the power of exploring chemical space by deploying robust molecular representation based on using SELFIES. Results: Experiments show that BioMolFormer excels traditional methods in both molecular validity and diversity, with 100% chemical validities and an average docking score of 9.9; they highlighted the promise of dramatically improving drug discovery with this model.