Task-Dependent Performance of GNNs and Pretrained Models in ADMET Prediction

Across six Therapeutics Data Commons ADMET endpoints under scaffold-split evaluation, no model family dominates; the GNN is clearly best for Caco-2 permeability with R$^2 \approx 0.48$.

• Why it matters: Published ADMET benchmarks can mislead due to unequal tuning budgets and leakage-prone evaluation, so apparent performance gaps may reflect protocol artifacts rather than true model quality.
• What they did: The study benchmarked GNNs, frozen pretrained models (ChemBERTa, MolCLR), and fingerprint baselines using a matched 50-trial validation-selected tuning budget, medianed over ten seeds, and kept the held-out test split for final scoring.
• The result: Morgan fingerprints are competitive on hERG (AUC $\approx 0.77$) and best on both clearance endpoints, MolCLR is best on Tox21, while hepatocyte clearance and half-life show near-zero or negative R$^2.

AI summary of the abstract · may contain errors · not medical advice · read the paper.