DATE CHANGED, GRXForm: Graph Transformers and Group Relative Policy Optimization for Molecular Design

Berke Kisin will discuss his work in progress. Molecular design includes tasks such as de novo generation and structural modification of existing molecules. Current state-of-the-art methods for modification typically act as instance optimizers, restarting the search for each input molecule and incurring high computational cost. Although model-based approaches promise amortized efficiency by learning policies transferable to unseen structures, they often fail to generalize due to high variance caused by varying difficulty across different starting scaffolds. 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.