The preprint presents GRASP, a graph Transformer model developed to learn molecular representations by combining large-scale molecular structure pretraining with intermediate supervision from experimental assays. The authors note that while molecular structures are abundant, experimental bioactivity measurements are comparatively sparse and distributed across many assays. GRASP is proposed as a staged training pipeline that exploits both types of data sequentially to improve molecular-property prediction.
GRASP is described as a graph Transformer with 93.5 million parameters. The design emphasizes learning graph-based molecular representations suitable for downstream prediction of molecular properties. The paper frames the parameter scale and graph-transformer architecture as enabling effective transfer from structural pretraining to assay-level adaptation, although architectural specifics beyond the stated parameter count are not detailed in the abstract.
The first training stage focuses on molecular structure. GRASP is pretrained using a replaced-token detection objective on 1.54 billion molecular presentations drawn from ZINC20. This pretraining stage aims to capture generalizable structural features from massive, unlabelled molecular collections. The authors report a baseline mean MAE after this stage (structure-only), which was subsequently improved by assay adaptation; details of the replaced-token detection procedure and training hyperparameters are included in the full text but are not reported in the abstract.
After structure pretraining, GRASP is adapted to experimental bioactivity using supervision from 642 assays in ChEMBL. This intermediate adaptation step uses available, but sparse, bioactivity measurements to refine the representations learned during pretraining. The authors emphasize that assay-level supervision sits between abundant structural data and task-specific finetuning, and that it provides complementary signal that can improve downstream predictive accuracy on ADMET endpoints.
Model performance is evaluated on two types of benchmarks reported in the abstract: 23 cluster-held-out OpenADMET endpoints and 22 TDC ADMET tasks. The primary metric reported in the abstract for OpenADMET is mean mean absolute error (MAE). For TDC ADMET tasks the abstract reports comparative point estimates versus another model (CheMeleon). The abstract does not provide full per-task metrics, confidence intervals, or statistical testing details; these would be found in the full text.
On the 23 cluster-held-out OpenADMET endpoints, GRASP achieved a mean MAE of 0.374. This result is compared with several baselines: CheMeleon (0.383), ECFP4-LightGBM (0.408), and Chemprop (0.440). The authors state that adaptation to ChEMBL assays reduced GRASP's mean MAE from 0.407 (structure-pretrained but not assay-adapted) to 0.374, and that assay adaptation improved performance on 21 of the 23 endpoints. On the 22 TDC ADMET tasks, GRASP provided a better point estimate than CheMeleon on 13 tasks, indicating competitive performance across that benchmark set.
The authors interpret these results as support for a staged molecular representation learning paradigm. In this view, abundant structural data and sparse experimental measurements act as complementary supervision sources at different stages of training: large-scale, unsupervised or self-supervised pretraining captures general structural features, and intermediate assay supervision injects domain-relevant bioactivity signal that improves downstream prediction on ADMET endpoints. The reported reductions in mean MAE and per-endpoint improvements are presented as empirical evidence for this approach.
This work is presented as a preprint on bioRxiv and has not been certified by peer review. The abstract reports high-level results and summary statistics; granular details such as training schedules, full per-task results, uncertainty estimates, and potential limitations beyond those implicit in the staged approach are not reported in the abstract. The authors explicitly declared no competing interests.
For researchers and practitioners in molecular machine learning and predictive ADMET modeling, the GRASP results suggest that an intermediate assay-supervision stage—between large-scale structural pretraining and task-specific finetuning—can yield measurable gains in predictive accuracy. The presented comparisons to established baselines (CheMeleon, ECFP4-LightGBM, Chemprop) indicate that staged training may be a useful strategy when abundant structural databases such as ZINC20 and assay repositories such as ChEMBL are available.
All numerical results and methodological claims above are taken from the preprint abstract. The full paper contains additional experimental details and results that are not included in the abstract and should be consulted for implementation and reproducibility specifics.