Research Article
Machine Learning Toxicity Prediction of Drug-Induced Liver Injury (DILI) with Grad-CAM Explainability
@INPROCEEDINGS{10.1007/978-3-032-44094-5_12, author={Hoang Phi Yen Duong and Nghia Trong Vo and Vu Phong Pham and Tuan Thanh Nguyen and Trung Q. Duong}, title={Machine Learning Toxicity Prediction of Drug-Induced Liver Injury (DILI) with Grad-CAM Explainability}, proceedings={Industrial Networks and Intelligent Systems. 12th EAI International Conference, INISCOM 2026, Da Nang, Vietnam, February 26--27, 2026, Proceedings}, proceedings_a={INISCOM}, year={2026}, month={11}, keywords={DILI Machine learning Toxicity prediction Drug discovery and development Grad-CAM}, doi={10.1007/978-3-032-44094-5_12} }- Hoang Phi Yen Duong
Nghia Trong Vo
Vu Phong Pham
Tuan Thanh Nguyen
Trung Q. Duong
Year: 2026
Machine Learning Toxicity Prediction of Drug-Induced Liver Injury (DILI) with Grad-CAM Explainability
INISCOM
Springer
DOI: 10.1007/978-3-032-44094-5_12
Abstract
This study focuses on predicting drug-induced liver injury (DILI) toxicity using two-dimensional image representations of chemical molecules. Four models are employed, i.e., ResNet-18, EfficientNet-B3, simple convolutional neural network (CNN), and support vector machine (SVM), by using the k-fold cross-validation method to ensure the reliability of the results. We shows that EfficientNet-B3 achieves the highest performance in the area under the receiver operating characteristic (ROC) curve (AUC) of 0.844 ± 0.054, demonstrating a stronger and more stable classification ability compared to the remaining models. In addition, gradient-weighted class activation mapping (Grad-CAM) is used to visually analyze how the model focuses on important features in chemical molecules when making predictions. The Grad-CAM results indicate that EfficientNet-B3 is capable of focusing on regions of chemical significance related to toxicity and can generalize productively to multiple functional groups at once, such as amine (-NH), halogen (Cl, Br, F), and aromatic rings, or focusing on the entire formula with small molecules. Conversely, in some cases of incorrect predictions, Grad-CAM indicates the model focusing on unimportant regions, followed by reducing the accuracy of the results. Our findings show that EfficientNet-B3 has high potential in predicting DILI toxicity based on molecular imaging and also demonstrates the value of Grad-CAM in understanding how the model works.