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

Area of Science:

• Microbiome research
• Computational biology
• Disease diagnostics

Background:

• Gut microbiome dysbiosis is linked to complex diseases, but computational diagnosis is challenging.
• Existing methods struggle with large, imbalanced microbiome datasets and capturing microbial interactions.
• Accurate differential diagnosis using microbiome data remains a significant unmet need.

Purpose of the Study:

• To develop a novel deep learning framework, AR-CDT Net, for accurate and robust classification of host disease states from gut microbiome data.
• To address limitations in predictive performance, robustness, and interaction capture of current computational approaches.
• To enable precise differential diagnosis by disentangling disease-specific microbial signatures.

Main Methods:

• Developed AR-CDT Net, a deep learning framework integrating Multi-Scale Deformable Convolution (MS-DConv) and Channel-wise Dynamic Tanh (CD-Tanh).
• Evaluated on a large cohort (>8000 samples, 8 phenotypes) for within-cohort performance.
• Validated cross-dataset generalization on independent cohorts, including a heterogeneous T2D cohort.

Main Results:

• AR-CDT Net outperformed nine representative models in within-cohort classification tasks.
• Achieved a significant AUC of 0.7921 in cross-dataset generalization on the T2D cohort, indicating transferable biological signals.
• Successfully disentangled disease-specific pathogenic signatures from shared dysbiotic backgrounds using dimensionality reduction and SHAP interpretation.

Conclusions:

• AR-CDT Net offers a robust and accurate deep learning approach for microbiome-based disease classification.
• The framework demonstrates effective generalization across datasets, capturing transferable microbial signals.
• AR-CDT Net provides interpretable insights, distinguishing disease-specific microbial patterns for improved differential diagnosis.