Interpretable three-dimensional deep learning identifies and reveals the spatial Microstructure of multi-enzyme...

Understanding the spatial mechanisms of multi-enzyme lignocellulose deconstruction is hindered by the lack of spatial quantification and nondestructive analytical methods. This study established an interpretable three-dimensional (3D) deep learning framework integrated with a microfluidic platform to characterize the microstructural evolution of lignocellulose under diverse enzyme treatments, including cellulase, lignin peroxidase, and laccase. Standardized microfluidic 3D spatial datasets were constructed using confocal microscopy imaging and used to train 3D convolutional neural networks integrated with attention mechanisms. The DenseNet121-CBAM model achieved an optimal balance between predictive performance and generalization. Interpretability analysis through explainable artificial intelligence effectively revealed distinct spatial degradation signatures. Specifically, treatment with cellulase alone resulted in surface-level degradation. By contrast, ligninolytic enzymes disrupted the lignin matrix, forming scale-like modifications that exposed the underlying cellulose skeleton. Crucially, during combined multi-enzyme treatment, this initial lignin disruption enhanced cellulose accessibility, enabling synergistic enzyme consortia to drive extensive sheet-like structural changes, severe particle fragmentation, and deep internal hollowing. Furthermore, cross-scale validation using scanning electron microscopy confirmed the physical relevance of the model-identified microstructural features. To conclude, this low-cost, nondestructive framework enables rapid, autonomous quantification of structural remodeling and established a robust foundation for monitoring enzymatic degradation dynamics in advanced biorefineries.