Deep feature fusion and hybrid CNN-SVM architecture for reliable and explainable drug toxicity prediction

Reducing animal testing in laboratories and promoting the green chemistry of designing new chemicals hinges on accurately predicting chemical toxicity. To assist in the sustainable design of drugs and materials, this paper proposes a machine learning approach for predicting chemical toxicity. The molecular, assay, and clinical levels of toxicity in three datasets—namely Tox21, ToxCast, and DILI—were integrated. The datasets toxicity level integration enables enhanced training of the machine learning models. Normalization of descriptors and fragmentation of the chemical compounds facilitated the use of advanced Graph Neural Networks (GNNs) and the ChemBERTa pre-trained transformer. Hyper deep learning molecular embeddings alongside ADMET and physicochemical descriptors empowered a feature extraction hybrid framework. A CNN-SVM deep learning architecture was developed, with the CNNs learning sophisticated feature configurations for the SVMs to fit non-linear boundaries. The framework surpassed standard classifiers with 91.2% accuracy, AUC-ROC = 0.934, and MCC = 0.781. Combined with standard classifiers, the SHAP analysis and activation heatmaps demonstrated interpretability, highlighting LogP, TPSA, hydrogen bond parameters molecular weight as primary LogP determinants.