Metagenomics has brought about a paradigm shift in our understanding of microbial populations. However, the incorporation of machine learning with metagenomic analysis is plagued by two closely associated challenges, that is, the compositional nature of sequencing data and the failure to interpret deep learning networks. Current XAI methods used with shallow classifiers depend on count-based features. Such an approach is in contrast with the fundamental tenets of standard machine learning techniques and results in unreliable feature attribution. While deep learning models have shown superior performance in taxonomy prediction and functional pathways, they are still considered black boxes. Key contributions made in our research include a number of novel bioinformatics pipeline developments that can resolve both of these issues and consist of: (i) An Aitchison geometry compositional transformation (CLR, Centered Log-Ratio) applied before training to mitigate the simplex constraint issue; (ii) Sequence classification model based on the transformer architecture and multiple heads; (iii) A method of attributions developed based on the modified SHAP algorithm. As can be seen from the results of our experiment using HMP2 IBD data set, the usage of SHAP attributions for explaining raw count data produces highly unreliable attribution maps (faithfulness AUROC = 0.63). At the same time, CLR SHAP provides a significant improvement compared to this approach (AUROC = 0.81).

Index Terms—metagenomics, compositional data analysis, explainable AI, SHAP, transformer, clinical bioinformatics, XAI.