Bwanya Brian, Lodhi Saad, de Kok Theo M, Ladeira Luiz, Verheijen Marcha Ct, Jennen Danyel Gj, Caiment Florian
Department of Translational Genomics, GROW Research Institute for Oncology and Developmental Biology, Maastricht University, Maastricht 6229 ER, the Netherlands.
Biomechanics Research Unit, GIGA Institute, University of Liège, Avenue de l'Hôpital, 11, B34 +5, Liège 4000, Belgium.
Toxicology. 2025 Nov;517:154237. doi: 10.1016/j.tox.2025.154237. Epub 2025 Jul 18.
The transition toward new approach methodologies for toxicity testing has accelerated the development of computational models that utilize transcriptomic data to predict chemical-induced adverse effects. Here, we applied supervised machine learning to gene expression data derived from primary human hepatocytes and rat liver models (in vitro and in vivo) to predict drug-induced hepatic steatosis. We evaluated five machine learning classifiers using microarray data from the Open TG-GATEs database. Among these, support vector machine (SVM) consistently achieved the highest performance, with area under the receiver operating characteristic curve (ROC-AUC) of 0.820 in primary human hepatocytes, 0.975 in the rat in vitro model, and 0.966 in the rat in vivo model. To gain mechanistic insights, we functionally profiled the top-ranked predictive genes. Enrichment analyses revealed strong associations with lipid metabolism, mitochondrial function, insulin signalling, oxidative stress, all biological processes central to steatosis pathogenesis. Key predictive genes such as CYP1A1, PLIN2, and GCK mapped to lipid metabolism networks and liver disease annotations, while others highlighted novel transcriptomics signals. Integration with differentially expressed genes and known steatosis markers highlighted both overlapping and distinct molecular features, suggesting that machine learning models capture biologically relevant signals. These findings demonstrate the potential of machine learning models guided by transcriptomic data to identify early molecular signatures of drug-induced hepatic steatosis. The support vector machine model's strong predictive accuracy across species highlights its promise as a scalable and interpretable tool for chemical risk assessment. As data limitations in human toxicology persist, expanding high-quality transcriptomic resources will be critical to further advance non-animal approaches in regulatory toxicology.
向毒性测试新方法学的转变加速了利用转录组数据预测化学诱导不良反应的计算模型的发展。在此，我们将监督式机器学习应用于源自原代人肝细胞和大鼠肝脏模型（体外和体内）的基因表达数据，以预测药物诱导的肝脂肪变性。我们使用来自Open TG-GATEs数据库的微阵列数据评估了五种机器学习分类器。其中，支持向量机（SVM）始终表现出最高的性能，在原代人肝细胞中的受试者操作特征曲线下面积（ROC-AUC）为0.820，在大鼠体外模型中为0.975，在大鼠体内模型中为0.966。为了获得机制性见解，我们对排名靠前的预测基因进行了功能分析。富集分析揭示了与脂质代谢、线粒体功能、胰岛素信号传导、氧化应激的强烈关联，这些都是脂肪变性发病机制的核心生物学过程。关键预测基因如CYP1A1、PLIN2和GCK映射到脂质代谢网络和肝脏疾病注释，而其他基因则突出了新的转录组学信号。与差异表达基因和已知脂肪变性标志物的整合突出了重叠和不同的分子特征，表明机器学习模型捕获了生物学相关信号。这些发现证明了由转录组数据指导的机器学习模型识别药物诱导肝脂肪变性早期分子特征的潜力。支持向量机模型在不同物种间的强大预测准确性突出了其作为化学风险评估的可扩展和可解释工具的前景。由于人类毒理学中的数据限制仍然存在，扩大高质量转录组资源对于进一步推进监管毒理学中的非动物方法至关重要。
