CRAFT: A Chemical Reaction-Aware Transformer for Synthesizable Molecular Optimization via Multimodal Learning

Abstract

Ensuring synthetic accessibility remains a major challenge in de novo drug design, as generative models can produce high-scoring molecules that are difficult to synthesize. To address this issue, we propose CRAFT, a reaction-aware molecular optimization framework that integrates Monte Carlo Tree Search (MCTS) with a multimodal strategy network. Unlike sequence-only baselines, CRAFT explicitly encodes two-dimensional molecular topology and atom-level reaction roles, enabling chemically feasible transformations during multi-step optimization. Comprehensive evaluations on DRD2, AKT1, and CXCR4 demonstrate that CRAFT consistently outperforms state-of-the-art methods, generating more high-scoring candidates while improving reaction-constrained synthetic accessibility and molecular diversity. CRAFT also rediscovers scaffolds homologous to known active ligands and identifies novel chemical entities beyond the training distribution, as supported by molecular docking simulations. Moreover, the model provides transparent forward-synthesis pathways for generated candidates, helping bridge in silico molecular design and practical synthesis.