Descriptor-First Approach for ADMET Prediction in the PolarisHub Antiviral Challenge

The prediction of absorption, distribution, metabolism,
excretion,
and toxicity (ADMET) properties remains a central bottleneck in small-molecule
discovery. We present the third-place solution from the PolarisHub
Antiviral Competition, covering five end points broadly relevant to
small-molecule design: human and mouse liver microsomal stability
(HLM, MLM), MDR1-MDCKII permeability, kinetic solubility, and lipophilicity
(LogD). Rather than pursuing complex machine learning architectures,
we adopted a descriptor-first strategy. We systematically curated
descriptors and models from ADMET Predictor as meta-features and then
applied high-capacity tabular learners. A pretrained foundation model
for tabular data (TabPFN), used in single-task regression, consistently
outperformed or matched a strong gradient boosting baseline (CatBoost),
yielding up to 44% mean absolute error (MAE) reduction across end
points while simplifying deployment by eliminating an extensive hyperparameter
search and producing compact models. Additionally, we engineered two
feature sets that delivered modest gains in randomized cross-validation
runs: (i) tuned fragment representations and (ii) site-of-metabolism
pattern features. Overall, we used four groups of features: mechanistic,
physicochemical, fragment, and metabolic. These results indicate that
in practical ADMET modeling scenarios, where rich, validated descriptors
are available, the competitive advantages often arise from principled
feature engineering combined with robust, rather than overly complex,
modeling approaches.