Ark-ADMET: endpoint-specific selection of molecular representations and hyperparameters for gradient-boosted ADMET models under the Therapeutics Data Commons benchmark protocol

Typical submissions to the Therapeutics Data Commons (TDC) absorption, distribution, metabolism, excretion, and toxicity (ADMET) benchmark leaderboard apply a single architecture and a single feature representation across all 22 endpoints, despite substantial heterogeneity in evaluation metric, training-set size, and class balance. Rather than proposing a new architecture, we show that endpoint-specific selection of molecular representations and CatBoost hyperparameters, carried out under the official TDC ADMET benchmark protocol, is a competitive alternative to that uniform recipe. For each endpoint, a fixed CatBoost backbone was paired with configurations chosen from a common candidate set comprising published fingerprints and physicochemical descriptors, truncated three-dimensional PaDEL descriptors, frozen graph isomorphism network (GIN) and MiniMol embeddings, and a convex combination with a shallow network trained on the MiniMol embedding alone. All feature and hyperparameter choices were made on an internal scaffold validation split under a prespecified strict-improvement rule. Against a frozen snapshot of the public leaderboard, the resulting Ark-ADMET family of 22 endpoint-specific models matched or exceeded the best previously listed score on 7 of 21 rankable endpoints and strictly exceeded it on 3. All selected configurations retain the baseline fingerprint-and-descriptor block; optional components differ across endpoints. Taken together, these results show that competitive performance on this benchmark can be achieved without a new architecture or a single optional representation shared across all endpoints.