Bracha Leah Shapira

Motivation: Designing novel molecules that simultaneously satisfy several, often conflicting, properties is a long-standing challenge in drug design. While recent deep learning advances optimize single targets effectively, navigating the tradeoffs required for multiple objectives remains a difficult problem. Results: We introduce the Multi-Objective-Conditioned-Return-Blocks Decision Transformer (M-GoRT), a novel architecture for multi-target de-novo drug design incorporating return-to-go blocks conditioned on goal embeddings. Trained without exploration, reward weighting or online fine-tuning, M-GoRT efficiently optimizes for multiple targets. M-GoRT achieves competitive multi-objective success rates against state-of-the-art baselines while maintaining superior structural diversity and novelty, even in high-dimensional tasks incorporating biological targets (DRD2, JNK3). In these high-dimensional tasks, M-GoRT maintains superior structural diversity and novelty establishing it as a practical tool for multi-property molecular generation. Availability and implementation: M-GoRT and the data used are freely available as an open-source software under the MIT license at GitHub https://github.com/eyalmazuz/MolGen/tree/GCRB-DT.