We used AutoGluon to optimise a classical machine-learning ensemble with up to three stacking layers, using CatBoost, LightGBM, and ExtraTrees as core learners.
Molecular representations included:
- ADMET Predictor 13 descriptors
- bRo5-specific descriptors
- LLM-derived embeddings
- RDKit and Mordred descriptors
To reduce redundancy, we removed highly correlated features and retained a single representative from each correlated group. Model performance was assessed via repeated cross-validation using multiple random seeds.
Step-wise stacking produced modest gains in cross-validation but did not translate into improved leaderboard ranking. We suspect this reflects a challenging out-of-domain generalisation setting, where validation performance is not fully predictive of test-set behaviour.