Bin Cui

Large ecotoxicity databases contain extensive evidence, yet heterogeneous reporting and sparse coverage across chemicals, species, end points, and conditions limit predictive use. From relational U.S. EPA ECOTOX, we linked 1,234,077 traceable records and constructed model-ready aquatic and soil data sets of 219,818 and 11,532 records. We developed a context-aware multitask model that preserves end point-specific responses while sharing information across molecular, biological, and experimental contexts. On held-out condition keys and parent-species-task combinations, four-seed ensembles achieved pooled R2 of 0.758/0.751 for aquatic toxicity and 0.628/0.537 for soil toxicity. Compared with same-input single-task models, multitask fitting reduced soil task-macro mean absolute error (MAE) by 0.136/0.156 across 16 task heads; aquatic point estimates also favored sharing, although uncertainty across all aquatic task heads was not estimable. Under canonical-parent holdout, frozen shared representations reduced task-macro MAE by 0.170–0.180 in aquatic and 0.093–0.125 in soil tree models versus molecular-only inputs, indicating reusable predictive information. Multidimensional training support stratified parent-holdout errors, meeting prespecified criteria fully for soil and partially for aquatic predictions; benefits weakened under stronger structural separation. These results show model-ready ECOTOX data can support context-aware toxicity prediction across a sparse experimental matrix, helping prioritize informative measurements and providing a predictive evidence base for environmental hazard assessment.