TRACER: Molecular Optimization Using Conditional Transformer for Reaction-Aware Compound Exploration with Reinforcement Learning

Designing molecules with desirable properties is a critical endeavor in drug discovery. Because of recent advances in deep learning, molecular generative models have been developed. However, existing compound exploration models often disregard the important issue of ensuring the feasibility of organic synthesis. To address this issue, we propose TRACER (molecular optimization using a conditional Transformer for reaction-aware compound exploration with reinforcement learning); this is a framework that integrates the optimization of molecular properties with the generation of synthetic pathways. At the core of TRACER is a conditional Transformer model trained on a dataset of chemical reactions. The model can predict the product from a given reactant under the constraints of a reaction type specified by a graph convolutional network. The results of molecular optimization on an activity prediction model targeting the dopamine receptor D2 showed that TRACER effectively generated compounds exhibiting high scores. The Transformer model, which recognizes the entire structure, captures the complexity of the organic synthesis and enables its navigation in the vast chemical space, with consideration of the real-world reactivity constraints. The source code of TRACER, the activity prediction model, and the curated dataset are available in our public repository at https://github.com/sekijima-lab/TRACER.