CausAL: Causal Active Learning for Data-Efficient Molecular Property Prediction and Lead Optimization

Molecular optimization is often limited by scarce labeled data and costly experimental assays, while conventional active learning may prioritize uncertain yet unproductive candidates. We present CausAL, an active learning framework driven by structural causal discovery for bioactivity optimization, integrating molecular property prediction with uncertainty-aware candidate selection. CausAL learns directional feature--property relationships to provide interpretable guidance for structural optimization, while Gaussian-process uncertainty and predicted activity jointly prioritize candidates. Across three MoleculeNet regression tasks and seven TDC ADMET tasks, CausAL achieves the best performance among evaluated baselines on 7 of 10 tasks. In an anti-Klebsiella pneumoniae LpxH inhibitor study, CausAL shows strong few-shot ranking, enriches highly active compounds, and supports iterative Design--Synthesis--Assay--Update optimization, with one-round wet-lab validation supporting causally guided structural modification and uncertainty--activity-based candidate selection. To the best of our knowledge, CausAL provides the first demonstration of integrating causal structure discovery with uncertainty-aware active learning for molecular property modeling and causally guided structure optimization.