Under fair tuning and leakage-free splits, no model family wins across ADMET endpoints
- Graph neural networks, frozen pretrained encoders and fingerprint baselines were compared on six Therapeutics Data Commons ADMET tasks, using scaffold splits, an identical tuning budget per family, validation-only selection and multi-seed medians.
- Differences between families mostly fell within seed variability, so no approach dominated. Fingerprints were competitive on several endpoints, and the GNN's lone clear edge traced to added descriptors, not the graph itself.
- Selecting optimizers on test data or reporting single best runs inflated apparent gains, supporting endpoint-specific model choice and leakage-controlled evaluation.
💡 Points Based on the abstract
Once descriptors are added, the GNN's one clear win looks like featurization rather than graph structure, which makes me doubt many reported GNN gains. Six public TDC tasks and nothing prospective say little about my own series. The real takeaway is the checklist: matched tuning budgets, validation-only selection, multi-seed medians. Handy for vetting vendor claims.
Task-Dependent Performance of GNNs and Pretrained Models in ADMET Prediction ↗