Abstract
Introduction
The ability to anticipate drug-induced toxicity prior to extensive animal experimentation remains a major unmet need in early drug discovery. Conventional preclinical toxicology workflows are resource intensive and often rely on iterative dose finding studies, motivating the development of predictive frameworks that can integrate diverse chemical, biological, and experimental information. This work introduces a multimodal, AI-assisted approach for preclinical toxicity prediction that combines chemical structure, physicochemical properties, biological drug sensitivity profiles, dosing characteristics, experimental metadata, and structural similarity.
Methods
Using 64 oncology drugs from the Responses to Oncology Agents and Dosing Models to Aid Preclinical Studies (ROADMAPS) dataset, we integrated four toxicity classes; defined by weight loss and mortality outcomes along with NCI 60 GI50 values, molecular descriptors, physicochemical properties, and dosing variables. Missing GI50 measurements were imputed using a Generative Adversarial Imputation Network (GAIN). Chemical structures were encoded using 2048 bit Morgan fingerprints, and pairwise Tanimoto similarity was computed to capture structural relationships. Five machine learning algorithms - Gaussian Naive Bayes, K Nearest Neighbor, Random Forest, XGBoost, and Multi Layer Perceptron; were evaluated across four experiments designed to (i) establish baseline performance, (ii) quantify the contribution of structural similarity, and (iii) assess synthetic data augmentation.
Results
Random Forest delivered the strongest baseline performance, achieving roughly 85% accuracy under a 95/5 train–test split and maintaining balanced sensitivity, specificity, AUC, and MCC. When Tanimoto derived structural similarity features were incorporated, the XGBoost model showed an additional ~2% improvement at an 85/15 split, demonstrating that structural context provides a meaningful discriminatory signal beyond traditional descriptor only representations. In contrast, synthetic data augmentation- including both SMOTE based oversampling and GAN generated samples did not yield measurable performance gains, reinforcing that model improvement in this setting depends more on high quality real experimental data than on synthetic expansion of limited datasets. Together, these findings highlight the value of structural similarity information and underscore the need for additional real toxicity observations to further enhance predictive robustness.
Conclusion
This study establishes a multi modal AI framework capable of integrating complementary chemical, biological, dosing, and structural information to enhance the prediction of preclinical drug toxicity. By enabling earlier identification of toxicity risk, the approach can inform dose selection, guide experimental prioritization, and reduce unnecessary animal use in alignment with the 3Rs - Replacement, Reduction, and Refinement. Future work should focus on validating this framework using larger, chemically diverse, and externally independent datasets to assess generalizability, robustness, and translational utility for real world preclinical drug development.