Gyujin Heo, Zhe Liu, Yujin Chung, Taesung Park
Developed DeepNTax, a deep neural network model that incorporates taxonomic information through regularization using taxonomic divergence degree between connected taxa and taxonomic rank level alongside abundance data. In the hepatocellular carcinoma dataset, DeepNTax showed competitive predictive performance with superior stability in Test AUC compared to baseline models, and in the colorectal cancer dataset, it achieved performance comparable to existing baselines with no statistically significant differences. Embedding taxonomic regularization in the model facilitated the identification of taxonomically grounded patterns, offering a structured mechanism for interpretability and a robust predictive framework for complex microbiome data.
High-throughput sequencing generates massive microbiome data, aiding the study of microbe-disease relationships. However, current analytical frameworks fail to precisely leverage taxonomic information, leading to suboptimal accuracy and interpretability due to community complexity and limited sample sizes. To explicitly integrate taxonomic information into a deep learning architecture, developing a structurally interpretable framework that achieves accurate prediction of clinically relevant features. We propose DeepNTax, a deep neural network model regularized using taxonomic information. Alongside abundance data, the model incorporates two key regularization components: the taxonomic divergence degree between connected taxa and the taxonomic rank level. In two real data applications, DeepNTax consistently provides better or competitive predictive performance compared to the comparison methods. In the hepatocellular carcinoma dataset, DeepNTax demonstrates competitive predictive performance with superior stability in Test AUC compared to baseline models. In the colorectal cancer dataset, the model achieves performance comparable to existing baselines, with no statistically significant differences in predictive power. By embedding taxonomic regularization for hierarchical representation learning, DeepNTax offers a robust predictive framework that maintains better or competitive performance while providing a structured mechanism for exploring microbial associations. This hierarchical approach facilitates the identification of taxonomically grounded patterns, offering a promising tool for interpretability in complex microbiome studies.