Multimodal AI Generates Virtual Rats: Predicting Hepatic Burden of Oral Drugs

Background & Aims

Drug-induced liver injury (DILI) remains a major cause of drug attrition, boxed warnings, and post-marketing withdrawal. However, its low incidence and high inter-individual variability limit the sensitivity of preclinical animal studies. We aimed to develop a multimodal virtual toxicology framework capable for simulating rat hepatic responses and improving detection of rare hepatotoxicity signals.

Methods

We developed Virtual Rat Experiment Model (VREM), a multimodal conditional generative adversarial network integrating chemical structure descriptors, bioactivity and network toxicology-informed pathway perturbation features. A modality-aware mixture-of-experts architecture adaptively fused heterogeneous biological information under different dose and exposure-duration conditions. The model was trained using 10,205 rat experimental records and externally validated using 1,453 records. Performance was evaluated using Pearson correlation, cosine similarity, root mean square error (RMSE), and Experimental Consistency Score (ECS).

Results

VREM faithfully reproduced biomarker distributions and consistently outperformed QSAR models. Key liver-injury biomarkers showed strong agreement between predicted and observed values, including ALT (adjusted R² = 0.9803), AST (adjusted R² = 0.9740), and ALP (adjusted R² = 0.9612). ECS values exceeded 0.90 for ALT, AST, and ALP across 10 independent generation runs. External validation demonstrated robust generalization despite substantial dataset distribution shifts, with significantly higher cosine similarity (P = 3.36×10^-31) and lower RMSE (P = 4.71×10^-58) than background distributions. Large-scale virtual experiments on 30 boxed-warning oral drugs identified potential DILI signals in 22 compounds, compared with only one compound detected in preclinical toxicology studies.

Conclusions

VREM provides a scalable multimodal digital twin framework for virtual hepatotoxicity experimentation. The model enables high-fidelity simulation of hepatic responses and substantially improves detection of low-frequency hepatotoxicity signals difficult to capture in conventional preclinical studies.