GRASP: Graph Representation Learning with Assay Supervision for Molecular Properties

Abstract

Molecular structures are abundant, but measured properties are scarce and scattered across assays that each cover different compounds. Most molecular pretraining pipelines use only the structures, moving directly from self-supervision to endpoint-specific fine-tuning and leaving large, sparsely labelled bioactivity databases unused. We present GRASP, a 93.5M-parameter graph Transformer trained in two stages: replaced-token detection over atom tokens with shortest-path relative attention using 1.54B molecule presentations from ZINC20, followed by adaptation on sparse labels from 642 ChEMBL assays. On 23 cluster-held-out OpenADMET endpoints, fine-tuned GRASP reaches a mean MAE of 0.374, compared with 0.383 for CheMeleon and 0.408 for ECFP4-LightGBM, and has lower error than Chemprop on 22 of 23 endpoints. Introducing the intermediate ChEMBL stage reduces mean MAE from 0.407 to 0.374 and improves 21 of 23 endpoints. In a controlled smaller-scale recipe comparison, replaced-token detection yields lower MAE than a masked-token recipe on 16 of 23 OpenADMET endpoints and favorable point estimates on 17 of 22 TDC ADMET tasks. Sparse multi-assay data, used as an intermediate training stage, is an effective complement to structural pretraining.