Molecular property models skip the sparse assay data: 1.54B structure presentations, then 642 ChEMBL assays as a middle stage

Molecular property pipelines typically move from structural self-supervision directly to endpoint fine-tuning, leaving the sparse but real experimental activity data unused as a training stage. The common prior path used structure databases such as ZINC for self-supervision and then fine-tuned on a specific endpoint, skipping the link between structure and measured chemical behaviour held in incomplete ChEMBL assay matrices. Across 23 cluster-held-out OpenADMET endpoints, full fine-tuning reaches a mean MAE of 0.374, below CheMeleon at 0.383, tuned ECFP4-LightGBM at 0.408 and Chemprop D-MPNN at 0.440, with lower endpoint MAE on 12, 16 and 22 of the 23 endpoints respectively.

The model learns in two stages, first molecular structure alone and then sparse experimental activity as a supervision signal. The first stage uses replaced-token detection: a three-layer generator replaces each atom token with MASK at probability 0.25 and predicts the original token, and a 12-layer discriminator judges whether each position was replaced, over 1.54 billion ZINC20 molecule presentations. The second stage fine-tunes on 511,898 ChEMBL 36 molecules and 642 binding or functional assays for 80,000 updates, with labels from activity comments or pChEMBL thresholds, and downstream work uses the fixed 50,000-step encoder snapshot.

Adding the ChEMBL adaptation stage lowers error on most endpoints relative to the structure-only checkpoint. The structure-only GRASP-S1 checkpoint has a mean MAE of 0.407, which falls to 0.374 after ChEMBL adaptation, an 8.0% relative reduction. On the same 23 OpenADMET endpoints and the same three cluster-held-out splits, the adapted model has lower MAE on 21 endpoints, with the largest relative gains on ASAP solubility at 24.2%, ExpansionRx plasma binding at 19.5%, ASAP LogD at 16.4% and ExpansionRx brain binding at 14.3%.

The reported scope is 23 cluster-held-out OpenADMET endpoints and 22 TDC ADMET tasks, covering clearance, solubility, permeability, plasma and tissue binding, efflux, PXR activity and antiviral Mpro potency. Adaptation modes include full fine-tuning, LoRA and frozen layer mixing, with full fine-tuning giving the lowest mean MAE. A structural-overlap analysis shows that removing test molecules identical to ChEMBL adaptation data or with ECFP4 similarity at least 0.70 changes mean MAE from 0.374 to 0.372 or 0.369, and the advantage over ECFP4-LightGBM remains.

The GRASP-S1 versus GRASP comparison measures the complete ChEMBL adaptation stage and does not separate assay supervision from additional optimization or exposure to the ChEMBL molecular distribution. The structural-overlap analysis tests molecular exposure but cannot establish whether semantically related assay supervision exists between ChEMBL and downstream endpoints. The RTD versus MLM comparison runs in the GRASP-Small configuration and compares implemented pretraining recipes rather than isolating pretraining scale. The principal downstream experiments use the 50,000-step Step-2 snapshot without a systematic downstream checkpoint-sensitivity study. The TDC comparison with released CheMeleon scores is descriptive because the two methods were not trained and selected within a single paired experimental pipeline.