Ligand Based Virtual screening using Reinforcement Learning

Abstract

Ligand-based virtual screening (LBVS) is a widely employed strategy in drug discovery to identify potential ligands for target proteins based on molecular structure similarity. However, conventional LBVS methods often rely on manually designed molecular descriptors or fingerprints that may not capture the essential characteristics necessary for binding affinity and specificity. This thesis proposes a novel LBVS methodology utilizing a computer-learned representation of molecules derived from features extracted through reinforcement learning (RL) with the Proximal Policy Optimization (PPO) algorithm. The RL agent is trained to perform multi-label classification on a diverse and extensive dataset of drug-like molecules (MDDR), employing a custom-designed environment that facilitates episodic data access. Moreover, a meticulously designed reward function is implemented to incentivize the agent for accurate classification. The acquired representation enables similarity computation and ranking of query and database molecules. Our method is evaluated on 10 reference structures from diverse protein families and compared against state-of-the-art LBVS techniques. The findings demonstrate significant enhancements in the top 1% and 5% cut-off recall values, with an average improvement of 50.6% at the 1% cut-off and 20.73% at the 5% cut-off, effectively showcasing the efficacy of our proposed approach