Activity Landscape Roughness Anticipates Machine-Learning Reliability Across Environmental Chemistry Endpoints

Machine learning increasingly supports chemical risk assessment under REACH, TSCA, and OECD QSAR frameworks, but random-split performance can be optimistic for structurally novel chemicals. We ask whether task difficulty can be anticipated before model training. EnvMolBench spans 45 datasets, 114,500+ end point-specific records (55,984 unique structures), and 6500+ model–dataset combinations from 13 algorithms and 6 representation categories. Training-set activity landscape roughness, quantified by nearest-neighbor disagreement rate (classification) or the Structure–Activity Landscape Index (regression), was associated with best-observed performance (Spearman ρ = −0.71, 95% CI [−0.94, −0.26], p = 0.0009 for classification; ρ = −0.47, 95% CI [−0.81, −0.04], p = 0.019 for regression), surviving family-wide FDR correction for classification (q = 0.0085) but not regression (q = 0.063). On high-roughness end points, excluding activity-cliff compounds post hoc raised AUC by a mean 0.134 and lowered standardized RMSE by 0.067, whereas removing cliffs from training did not help. The three evaluated pretrained models did not outperform well-tuned baselines on the ten roughest end points. Roughness therefore provides an empirical, representation-relative diagnostic of achievable performance rather than a fundamental ceiling, motivating a workflow linking it to method selection, applicability-domain assessment, and conformal calibration. EnvMolBench, its datasets, splits, and baselines are released openly.