Madheswaran Monnishkaran, Jaganathan Keerthana, Shanmugam Lakshmanan
Department of Mathematics, School of Advanced Sciences, Vellore Institute of Technology, Chennai, Tamil Nadu 600 127, India.
Department of Computer and Information Science, Northumbria University Newcastle, Newcastle upon Tyne NE1 8ST, U.K.
J Chem Inf Model. 2025 Jul 28;65(14):7540-7553. doi: 10.1021/acs.jcim.5c01128. Epub 2025 Jul 18.
Acute dermal toxicity testing is essential for assessing the safety of chemicals used in pharmaceuticals, pesticides, cosmetics, and industrial chemicals. Conventional toxicity testing methods rely significantly on animal tests, which are resource-intensive and time-consuming and raise ethical issues. To address these issues and support the 3Rs principle (replacement, reduction, and refinement) in animal testing, this study investigates whether a multimodal deep learning framework based on the fusion of heterogeneous molecular representations can yield a reliable and accurate model for the prediction of acute dermal toxicity. This study proposes TriModalToxNet, a novel architecture that extracts features from three distinct molecular representations: 2D molecular images through a 2D convolutional neural network, SMILES embeddings via a 1D convolutional neural network, and molecular fingerprints via a fully connected neural network. These extracted features are then concatenated and passed into a deep neural network for classification. For comparative purposes, this study also evaluates BiModalToxNet, a baseline model using only 2D molecular images and fingerprints. The models are trained and tested on a curated data set consisting of 3845 compounds derived from experimental rat and rabbit acute dermal toxicity studies. The proposed model is evaluated using multiple standard performance metrics such as area under the receiver operating characteristic curve, sensitivity, Matthews correlation coefficient, and accuracy derived from stratified 10-fold cross-validation and external validation. TriModalToxNet achieved an area under the receiver operating characteristic curve of 95% and a sensitivity of 91.2% in cross-validation. External validation was also conducted to further demonstrate the robustness and generalizability of the model. These results show that multimodal methods can attain better predictive performance than traditional single-modality methods. This TriModalToxNet framework highlights the potential for integration into regulatory frameworks, pharmaceutical screening pipelines, and advancing the field toward more ethical and efficient chemical safety assessment.
急性皮肤毒性测试对于评估用于药品、农药、化妆品和工业化学品的化学物质的安全性至关重要。传统的毒性测试方法在很大程度上依赖动物试验，这既耗费资源又耗时，还引发了伦理问题。为了解决这些问题并支持动物试验中的3R原则（替代、减少和优化），本研究调查了基于异构分子表示融合的多模态深度学习框架是否能够生成用于预测急性皮肤毒性的可靠且准确的模型。本研究提出了TriModalToxNet，这是一种新颖的架构，它从三种不同的分子表示中提取特征：通过二维卷积神经网络提取二维分子图像特征，通过一维卷积神经网络提取SMILES嵌入特征，以及通过全连接神经网络提取分子指纹特征。然后将这些提取的特征连接起来并传入深度神经网络进行分类。为了进行比较，本研究还评估了BiModalToxNet，这是一个仅使用二维分子图像和指纹特征的基线模型。这些模型在一个经过整理的数据集上进行训练和测试，该数据集由来自实验大鼠和兔子急性皮肤毒性研究的3845种化合物组成。使用多个标准性能指标对所提出的模型进行评估，如受试者操作特征曲线下面积、灵敏度、马修斯相关系数以及分层10折交叉验证和外部验证得出的准确率。在交叉验证中，TriModalToxNet的受试者操作特征曲线下面积达到95%，灵敏度达到91.2%。还进行了外部验证以进一步证明该模型的稳健性和通用性。这些结果表明，多模态方法比传统的单模态方法具有更好的预测性能。这个TriModalToxNet框架凸显了整合到监管框架、药物筛选流程中的潜力，并推动该领域朝着更符合伦理和高效的化学安全评估方向发展。
