Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
The paper introduces GRXForm, a Graph Transformer model for amortized molecular optimization that sequentially adds atoms and bonds to a molecule. To improve generalization, the authors identify and address the high variance in rewards caused by heterogeneous starting structures by using Group Relative Policy Optimization (GRPO). GRXForm demonstrates strong generalization to out-of-distribution molecular scaffolds, achieving competitive performance with instance optimizers in multi-objective optimization without requiring inference-time oracle calls or refinement.
Forget instance optimizers: GRXForm generalizes molecular optimization to unseen scaffolds by normalizing rewards relative to the starting structure, achieving state-of-the-art results without expensive per-molecule restarts.
Molecular design encompasses tasks ranging from de-novo design to structural alteration of given molecules or fragments. For the latter, state-of-the-art methods predominantly function as"Instance Optimizers'', expending significant compute restarting the search for every input structure. While model-based approaches theoretically offer amortized efficiency by learning a policy transferable to unseen structures, existing methods struggle to generalize. We identify a key failure mode: the high variance arising from the heterogeneous difficulty of distinct starting structures. To address this, we introduce GRXForm, adapting a pre-trained Graph Transformer model that optimizes molecules via sequential atom-and-bond additions. We employ Group Relative Policy Optimization (GRPO) for goal-directed fine-tuning to mitigate variance by normalizing rewards relative to the starting structure. Empirically, GRXForm generalizes to out-of-distribution molecular scaffolds without inference-time oracle calls or refinement, achieving scores in multi-objective optimization competitive with leading instance optimizers.