PolarFlow: Pretrained Model Weights and Evaluation Data for Joint Continuous-Discrete Flow Matching for Multi-Task 3D Drug Design

This deposit accompanies the PolarFlow paper. PolarFlow is a continuous–discrete flow matching model for pocket-conditioned 3D molecular generation. It jointly generates atomic coordinates (continuous) and atom types, formal charges, and bond orders (discrete CTMC), and supports de novo design, scaffold hopping, R-group replacement, linker design, and fragment growing in a single model via multi-level masking, followed by optional DiffusionNFT reinforcement learning and a lightweight property prediction head for in-generation screening. Contents • ckpt/ — model weights used in the paper: • polarflow.ckpt — Stage A (unconditional pretraining on GEOM-Drugs) • polarflow-sbdd.ckpt — Stage B (pocket-conditioned SBDD fine-tuning) • polarflow-rl.ckpt — Stage C (RL with AutoDock Vina + SA rewards) • polarflow-sbdd-nopretrained.ckpt — ablation without Stage A • property_heads.ckpt — frozen-backbone Vina/SA/QED prediction heads • denove_data/ — generated molecules for PolarFlow and baseline methods under the unified CrossDocked2020 protocol • ablation/ — Stage A ablation and RL evaluation outputs • leadopt/ — lead-optimization (masking) task outputs • vina_score/ — property-head docking/screening benchmark summaries (Figure 3b–c metrics) UsageCode and configs: https://github.com/imaxtric/PolarFlow