AtropoBench: A Benchmark for Geometric Chirality in Molecular Machine Learning

Restricted rotation around a bond can give molecules left- and right-handed configurations, a form of axial chirality. Recognising their handedness and predicting resistance to interconversion are distinct tasks: reflection reverses handedness but preserves the rotational energy barrier. We introduce AtropoBench, a diagnostic benchmark with nearly one million exact mirror pairs and over 21,000 computational rotational-barrier annotations from ligand and catalyst candidates. Its protocols separately test whether models can recognise handedness, generate novel structures with requested handedness, and predict rotational barriers on unfamiliar chemistry. Experiments with eleven generators and eight stability-prediction configurations reveal that high generation quality can coexist with negligible graph novelty, while similar classification accuracy can conceal substantially different errors on readily rotating molecules. Controlled interventions further distinguish limitations in reflection sensitivity, bond localisation, and uncertainty calibration. AtropoBench supports systematic model comparison while identifying specific capabilities that require improvement.