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

Ark-ADMET builds endpoint-specific CatBoost ADMET models that matched or exceeded the public TDC best score on 7 of 21 endpoints and strictly exceeded it on 3.

• Why it matters: The TDC ADMET benchmark uses 22 heterogeneous endpoints, yet prior leaderboard submissions often reuse one representation and architecture, leaving performance limited by mismatched feature and hyperparameter needs.
• What they did: Under the official TDC protocol, each endpoint used a fixed CatBoost backbone with configurations selected from a shared candidate set of fingerprints/physicochemical descriptors, truncated 3D PaDEL descriptors, frozen GIN and MiniMol embeddings, and a MiniMol-only shallow-network convex combination, chosen via internal scaffold validation with a strict-improvement rule.
• The result: Compared to a frozen public leaderboard snapshot, the 22-model Ark-ADMET family achieved competitive scores without introducing a new architecture, with optional representation components varying by endpoint while retaining the baseline fingerprint-and-descriptor block.

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