Mechanism-aware multimodal learning for organ-specific drug toxicity prediction through adaptive modality gating and text-neutralization stress testing

Early prediction of drug-induced organ toxicity remains a major bottleneck in drug discovery and clinical pharmacotherapy. Most data-driven toxicity models behave as endpoint predictors: they output a label but provide limited transparency about why a compound is risky or which evidence channel dominated the decision. Here, I present a mechanism-aware multimodal framework that fuses three complementary views of each compound: (i) a molecular graph encoder learned from SMILES, (ii) a compact vector of coarse mechanistic proxies (e.g., CYP involvement and reactive-metabolite flags), and (iii) a 384-dimensional text embedding derived from short drug descriptors. A gated fusion network learns sample-specific modality weights and outputs five organ-level toxicity probabilities (hepatotoxicity, nephrotoxicity, cardiotoxicity, mitochondrial toxicity, and hERG/QT risk). On a materialized dataset of 250 drug instances, the full Graph+Mech+Text model achieves strong discrimination on a held-out test split (macro-AUROC ≈ 0.99 in a representative run; hERG/QT AUROC ≈ 0.93). Modality ablations reveal that a Graph-only model performs modestly (macro-AUROC ≈ 0.62), while adding mechanistic proxies improves performance (≈ 0.91), and incorporating drug-specific text pushes macro-AUROC to ≈ 0.98–0.99. To distinguish genuine multimodal synergy from shortcut learning, I perform a diagnostic stress test in which all textual inputs are neutralized. Under this condition, macro-AUROC collapses to 0.533 (95% bootstrap CI 0.490–0.566), approaching chance for several endpoints, and the learned gate concentrates mass almost entirely on the text modality. These findings make two contributions. First, they demonstrate a practical decision-support architecture for organ-specific toxicity prediction that produces interpretable probability profiles rather than hard labels. Second, they introduce a simple but powerful audit methodology—modality ablations plus text neutralization—that exposes when apparent multimodal gains are driven by text-based shortcuts rather than robust structural or mechanistic signals. I discuss implications for dataset design, external validation, and how richer, non-leaky clinical text sources such as FDA label sections could be incorporated in future work.