Compound-Conditioned Bioactivity Graphs: Structuring Intermediate Supervision for Low-Data Toxicity Prediction

Developmental toxicity is difficult to predict from molecular structure alone because adverse phenotypes emerge through intermediate biological responses, while organism-level labels remain sparse, imbalanced, and incomplete. Large-scale in vitro bioactivity offers transferable supervision, but predicted assay profiles are typically treated as flat vectors that discard relationships among assays. We introduce compound-conditioned bioactivity graphs, in which predictions from a frozen multitask model initialize a shared, mechanistically constrained assay topology and learned node and edge gates adapt both assays and assay–assay interactions to each chemical. Tier I pretrains a molecular graph neural network across 457 ToxCast assays. Tier II integrates the conditioned assay graph with molecular and physicochemical representations using cross-modal contrastive alignment, masked multi-label representation learning, and uncertainty-weighted multitask prediction. Across five matched held-out splits on 1,415 compounds and six zebrafish developmental phenotypes, the full model achieves 0.786 ± 0.020 AUROC, compared with 0.723±0.034 for flat-logit transfer, 0.718±0.040 for molecular structure alone, 0.730±0.016 for a feature-matched random forest, and 0.751±0.024 without physicochemical descriptors. A static assay graph reaches 0.740 ± 0.025, while node-only and edge-only gating reach 0.729 ± 0.057 and 0.722 ± 0.036, respectively, indicating that both adaptive mechanisms are jointly required. Gains occur across all six phenotypes, and qualitative analyses of training, validation, and held out compounds reveal heterogeneous assay-utilization patterns over the shared topology. These results show that the organization and compound-specific adaptation of intermediate supervision determine how effectively bioactivity transfers to low-data developmental-toxicity prediction.