MOTIVATION: Protein-protein interactions (PPIs) are central to cellular functions, and predicting mutation-induced changes in binding affinity (ΔΔG) remains challenging. Although existing computational methods integrate sequence- and structure-derived features and thus implicitly capture certain sequence-structure relationships, they typically fuse these modalities through simple concatenation, without explicitly modeling their multidimensional and multiscale interdependencies. RESULTS: Here, we introduce IGMI, an interpretable graph-based model that explicitly encodes multi-level feature interactions across 1D sequences, 2D contact maps, 3D structures, and residue- and atom-level representations. By recalibrating cross-dimensional and cross-scale dependencies, IGMI enables more accurate estimation of both local and long-range mutation effects. Across multiple benchmark datasets, IGMI consistently outperforms state-of-the-art methods in accuracy, robustness, and interpretability. Macro- and micro-level analyses further reveal biologically plausible patterns, distinguishing direct interface perturbations from indirect structural reorganizations. Complementary analyses under different data splitting strategies indicate that the model learns generalizable affinity-related interaction patterns, rather than relying on split-specific information. IGMI provides a reliable and interpretable framework for modeling mutation-induced affinity changes, supporting applications in protein engineering and therapeutic design. AVAILABILITY AND IMPLEMENTATION: IGMI is implemented in PyTorch and released under an open-source license. The full codebase, training scripts, and evaluation utilities are available at https://github.com/ShiweiWu-545/IGMI.git. An archival snapshot containing all source code, pre-trained weights, processed datasets, and reproducibility scripts is available on Zenodo (https://doi.org/10.5281/zenodo.17563574). CONTACT: [email protected]; [email protected]; [email protected]. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

With the rapid development of deep neural networks, significant progress has been made in the analysis of physiological signals such as EEG, greatly improving classification performance. However, the inherent “black-box” nature of these models limits our ability to understand how different brain regions coordinate to drive decisions-especially given that modern EEG architectures (e.g., EEGNet) use gating mechanisms that activate distinct computational pathways for different subjects. To address this, we introduce the MODET-based Interaction Backtracking Method (MIBM), which converts a standard EEGNet into a shallow, wide, and intrinsically interpretable Multi-Order Descartes Expansion Neural Network (MODENN). This “deep-to-broad” transformation makes high-order feature interactions explicit and traceable. Experimentally, the MIBMtransformed EEGNet preserves-or even surpasses-the original model's accuracy across subjects. By exposing each subject's gated pathways as a clear computation graph, MIBM substantially enhances transparency and trustworthiness, marking a paradigm shift from traditional single-feature attribution toward a holistic, network-level interpretability aligned with systems neuroscience.