Integrated hierarchical fusion of GRU and meta-features for enhanced identification and interpretation of calcium ion inhibitors

The flow of calcium ions across cell membranes is essential for a wide range of physiological processes, including hormone secretion, epithelial function, the maintenance of electrochemical gradients, and the generation of electrical impulses. Calcium ion inhibitors have emerged as promising therapeutic agents for the treatment of various conditions such as cancer, muscle weakness, renal dysfunction, endocrine disorders, impaired cell function, and cardiac arrhythmias. Therefore, accurate identification and characterization of these inhibitors is crucial in regulating ion flow through calcium channels. In this study, we develop a neural network-based model for the identification of calcium ion inhibitors. To construct a comprehensive predictive framework, we employed seven peptide-encoding strategies and five distinct machine learning classifiers, resulting in 35 baseline models. From each, we extracted 35-dimensional probabilistic outputs and 35-dimensional predicted class labels, which were concatenated into a unified 70-dimensional meta-feature vector. In addition, we used the AAindex matrix to encode the physicochemical and biostructural properties of the peptide sequences. A gated recurrent unit (GRU) model was applied to capture long-range dependencies within this matrix, producing another 70-dimensional feature vector. The two 70-dimensional vectors were then concatenated into a final 140-dimensional representation, integrating physicochemical, biostructural, biochemical, and sequence-derived features. This 140D feature vector was used to train a two-layer multilayer perceptron (MLP) classifier. On the independent test set, our proposed model (PredCaII) achieved an accuracy of 0.941, MCC of 0.883, and AUC of 0.990, outperforming the state-of-the-art Multi-Branch-CNN method (accuracy: 0.751, MCC: 0.431, AUC: 0.766) and all individual baseline models. Furthermore, we assessed the interpretability and robustness of the model using Shapley additive explanations (SHAP) and Integrated Gradients. These interpretability analyses provided deep insight into the model’s decision-making behavior, highlighting the most influential compositional and physicochemical characteristics. The experimental results demonstrated the proposed method offers a powerful and interpretable computational framework for the discovery and screening of potential calcium ion inhibitors.