AR-CDT NET: a deep deformable convolutional network for gut microbiome-based disease classification

Advances in metagenomic sequencing have increasingly implicated gut microbiome dysbiosis in numerous complex diseases, yet its application for precise differential diagnosis remains a major challenge. Existing computational approaches often show limited predictive performance and insufficient robustness when applied to large-scale, imbalanced microbiome datasets, and they typically lack mechanisms to effectively capture microbial community-level or functional guild interactions. To address these limitations, we developed AR-CDT Net, a novel deep learning framework that integrates a Multi-Scale Deformable Convolution (MS-DConv) module with a Channel-wise Dynamic Tanh (CD-Tanh) activation function to achieve more accurate and robust classification of host disease states. Evaluated on a large-scale cohort comprising over 8000 samples spanning eight disease phenotypes, AR-CDT Net demonstrated highly competitive within-cohort performance, outperforming nine representative models across the majority of classification tasks. Importantly, in a stringent cross-dataset generalization test, the model was trained on the highly imbalanced primary multi-disease cohort and validated on relatively balanced independent external cohorts. It achieved a statistically significant AUC of 0.7921 on the highly heterogeneous external T2D cohort, confirming that AR-CDT captures transferable biological signals rather than dataset-specific artifacts. Furthermore, by combining dimensionality reduction with SHAP-based interpretation of our One-vs-Rest (OvR) classifiers, AR-CDT disentangles disease-specific pathogenic signatures from the shared dysbiotic background among clinically distinct yet microbially similar diseases.