Molecular Optimization with Mamba-Based GFlowNet

Molecular discovery involves the identification and design of novel chemical compounds with desired properties. Recently, this field has gained significant attention across various scientific domains for its potential to drive advancements in drug development and materials science. We propose an improved Generative Flow Network (GFlowNet) architecture for molecular optimization by replacing the traditional Gated Recurrent Unit (GRU) encoder with Mamba, which employs State Space Models (SSMs) to capture long-term dependencies more effectively. Experiments on single and multi-objective molecular optimization tasks demonstrate that our Mamba-based GFlowNet achieves superior sample efficiency and higher performance compared to existing methods, effectively generating high-quality molecular candidates within a limited number of trials.