ADMET-Stack: quantitative guidance on how much representational complexity an ADMET endpoint actually requires

Accurate absorption, distribution, metabolism, excretion, and toxicity (ADMET) prediction determines whether a drug candidate survives early development, yet how well it can be predicted depends heavily on how molecules are represented. Previous studies have generally focused on new model architectures, with less attention to the effects of molecular representations and features on ADMET prediction. Here, we evaluated 54 combinations of structural fingerprints, physicochemical descriptors, and SELFIES-derived encodings across all 22 ADMET endpoints of the Therapeutics Data Commons (TDC) Leaderboard, under a single, fixed stacked-ensemble architecture. A simple baseline proved remarkably competitive: RDKit molecular descriptors alone recovered 94.3% of best observed performance on average (median 96.4%), outperforming 2,363 combined structural fingerprints (93.2%) at a fraction of the feature count. Adding six further representation classes lifted mean retention to only 96.2%, and that gain was concentrated almost entirely on three endpoints, including a 67% relative improvement on Half_Life_Obach, while leaving eight others essentially unchanged. Benchmarked against four distinct leaderboard models (MiniMol, MapLight + GNN, ADMET-AI, ADMETrix), the advantage was uneven: regression endpoints improved by 9.0%–23.7% on average under per-endpoint configuration selection, whereas for classification no difference was detected against the two strongest baselines (Wilcoxon p = 0.339 and p = 0.685). To make individual predictions actionable, we further translated them into percentile ranks against 1,088 approved DrugBank compounds. All 54 per-configuration results and the 440 underlying leaderboard records are released, giving practitioners both a defensible default representation and an auditable basis for deciding when greater complexity is worth its cost.