Duong, Hoang Phi Yen, Vo, Nghia Trong, Nguyen, Tuan Thanh ORCID: https://orcid.org/0000-0003-0055-8218 and Duong, Trung Q. (2026) Enhancing drug-induced liver injury prediction via multi-representation molecular images with Grad-CAM explainability and functional group attribution. IEEE Internet of Things. ISSN 2327-4662 (Online) (In Press)

Preview

PDF (Author's Accepted Manuscript) 53567 NGUYEN_Enhancing_Drug-Induced_Liver_Injury_Prediction_(AAM)_2026.pdf - Accepted Version Available under License Creative Commons Attribution. Download (12MB) | Preview

Abstract

Drug-induced liver injury (DILI) constitutes a critical challenge in pharmaceutical development and accounts for over 50% of acute liver failure cases. This study employs deep learning (DL) methodologies for DILI prediction utilizing convolutional neural networks (CNNs) and multi-representation molecular imaging. We employ ResNet (18, 34, 50) and EfficientNet (B0-B3) architectures on three molecular image representations, such as normal images (NM), heatmap 1 (HM1) derived from Crippen logP contributions, and heatmap 2 (HM2) extracted from bioconcentration factor analysis. Experimental outcomes on 475 compounds using 4-fold cross-validation demonstrate that heatmap representations substantially outperform conventional molecular images. ResNet-34 attains optimal performance with HM2, achieving an AUC of 0.8853 and a Recall of 0.8410. This significantly enhances performance compared to conventional images (AUC: 0.8604, Recall: 0.8295). ResNet-34 and EfficientNet-B0 with HM1 exhibits superior functional group identification capabilities with a Recall of 0.8499 and 0.8682. Grad-CAM visualization illustrates that HM1 effectively emphasizes specific functional groups while HM2 facilitates comprehensive molecular analysis. Functional group analysis indicates that nitro groups (93.75%), sulfones (92.31%), and urea derivatives (91.67%) are associated with the highest DILI risk, while sulfur-containing moieties broadly serve as toxicity indicators. In contrast, iodine-containing compounds and ketones exhibit notably lower toxicity rates. Our multi-representation methodology exhibits competitive performance while delivering enhanced interpretability through explainable AI techniques. This framework presents considerable potential for pharmaceutical toxicity assessment and diminishes dependence on animal testing protocols.

Item Type:	Article
Uncontrolled Keywords:	drug discovery, computational toxicology, cheminformatics, machine learning, Grad-CAM
Subjects:	Q Science > Q Science (General) Q Science > QA Mathematics > QA75 Electronic computers. Computer science R Medicine > R Medicine (General)
Faculty / School / Research Centre / Research Group:	Faculty of Engineering & Science Faculty of Engineering & Science > School of Computing & Mathematical Sciences (CMS)
Last Modified:	22 May 2026 13:32
URI:	https://gala.gre.ac.uk/id/eprint/53567

Actions (login required)

View Item

Downloads