Authors

Bas Voermans^1,2, Marcus de Goffau^2,1, Max Nieuwdorp¹ and Evgeni Levin², ¹University of Amsterdam, Netherlands, ²HORAIZON Technology BV, Netherlands

Abstract

In recent years, machine learning, especially deep learning, has garnered substantial attention in the biomedical field. For instance, deep learning has become a preferred method for medical image analysis tasks. However, in other areas like fecal metagenomics analysis, the application of deep learning remains underdeveloped. This can be attributed to the tabular nature of metagenomics data, feature sparsity, and the complexity of deep learning techniques, which often lead to perceived inexplicability. In this paper, we introduce Microbe2Pixel, an innovative technique that applies deep neural net- works to fecal metagenomics data by transforming tabular data into images. This transformation is achieved by inferring location from the taxonomic information inherently present in the data. A significant advantage of our method is the use of transfer learning, which reduces the number of samples required for training compared to traditional deep learning. Our method aims to develop a local model-agnostic feature importance algorithm that provides interpretable explanations. We evaluate these explanations against other local image explainer methods using quantitative (statis- tical performance) and qualitative (biological relevance) assessments. Microbe2Pixel outperforms all other tested methods from both perspectives. The feature importance values align better with current microbiology knowledge and are more robust concerning the number of samples used to train the model. This is particularly significant for the application of deep learn- ing in smaller interventional clinical trials (e.g., fecal microbial transplant studies), where large sample sizes are unattainable and model interpretability is crucial.

Keywords

metagenomics, interpretable deep learning, local explanations.