Koutroumpa Nikoletta-Maria, Varsou Dimitra-Danai, Kolokathis Panagiotis D, Antoniou Maria, Papavasileiou Konstantinos D, Papadopoulou Eleni, Papadiamantis Anastasios G, Tsoumanis Andreas, Melagraki Georgia, Velimirovic Milica, Afantitis Antreas
Entelos Institute, Nicosia 2102, Cyprus.
School of Chemical Engineering, National Technical University of Athens, Athens 15780, Greece.
Comput Struct Biotechnol J. 2025 Aug 6;29:222-235. doi: 10.1016/j.csbj.2025.08.001. eCollection 2025.
Assessing chemical toxicity is essential for understanding potential risks to human health. However, ethical, financial, and scientific challenges have driven the demand for non-animal testing methods. This study introduces a computational framework that leverages diverse molecular representations, including MACCS keys, Morgan fingerprints, and Mordred descriptors, to predict skin sensitization, irritation/corrosion, and acute dermal toxicity. Different molecular representations for skin toxicity-related endpoints were first evaluated using three machine learning algorithms (Random Forest, Support Vector Machine, and k-Nearest Neighbors), then combined into a unified input space for training a fully connected neural network (FCNN). Comparative analyses indicate that this multi-view FCNN model offers superior or comparable predictive performance relative to single-representation models, achieving area under the curve (AUC) values of up to 0.91 for irritation/corrosion, 0.88 for sensitization, and 0.82 for acute dermal toxicity on test sets. Additional validation on known toxicants further confirms the framework's robustness, correctly identifying 0.86 of skin sensitizers, 0.89 of irritants, and 0.86 of dermally toxic compounds. Shapley Additive exPlanation (SHAP) analyses highlight the most influential molecular features, providing mechanistic insights and enhancing model transparency. To promote broader adoption and reduce reliance on animal testing, the developed models are freely available through the SbD4Skin (Safe by Design for Skin) web platform (https://eoscloud.entelos.eu/ssbd4chem/sbd4skin/), offering a user-friendly tool for chemical risk assessment and regulatory decision-making. The dataset and model developed in this study have been FAIRified and made available in machine-actionable and modelling-ready formats, supporting transparency, reuse, and regulatory acceptance.
评估化学毒性对于了解对人类健康的潜在风险至关重要。然而，伦理、财务和科学方面的挑战推动了对非动物测试方法的需求。本研究引入了一个计算框架，该框架利用多种分子表示法，包括MACCS键、摩根指纹和莫德雷德描述符，来预测皮肤致敏性、刺激/腐蚀性和急性皮肤毒性。首先使用三种机器学习算法（随机森林、支持向量机和k近邻）评估与皮肤毒性相关终点的不同分子表示法，然后将其组合成一个统一的输入空间，用于训练全连接神经网络（FCNN）。比较分析表明，这种多视图FCNN模型相对于单表示模型具有优越或相当的预测性能，在测试集上，刺激/腐蚀性的曲线下面积（AUC）值高达0.91，致敏性为0.88，急性皮肤毒性为0.82。对已知毒物的进一步验证进一步证实了该框架的稳健性，正确识别出0.86的皮肤致敏剂、0.89的刺激物和0.86的皮肤毒性化合物。夏普利加性解释（SHAP）分析突出了最具影响力的分子特征，提供了机理见解并提高了模型透明度。为了促进更广泛的采用并减少对动物测试的依赖，可以通过SbD4Skin（皮肤设计安全）网络平台（https://eoscloud.entelos.eu/ssbd4chem/sbd4skin/）免费获得所开发的模型，该平台为化学风险评估和监管决策提供了一个用户友好的工具。本研究中开发的数据集和模型已符合FAIR原则，并以机器可操作和可用于建模的格式提供，支持透明度、重用和监管认可。
