Molecular property models skip the sparse assay data: 1.54B structure presentations, then 642 ChEMBL assays as a middle stage
Molecular property prediction has mostly gone from structural self-supervision straight to endpoint fine-tuning, skipping the sparse experimental activity data in between. GRASP is a 93.5M-parameter graph Transformer trained in two stages: replaced-token detection on 1.54 billion ZINC20 molecule presentations, then adaptation on 642 ChEMBL assays. On 23 cluster-held-out OpenADMET endpoints, full fine-tuning reaches a mean MAE of 0.374, versus 0.383 for CheMeleon, 0.408 for ECFP4-LightGBM and 0.440 for Chemprop; adding ChEMBL adaptation lowers mean MAE from 0.407 to 0.374 and improves 21 of 23 endpoints.