Overview of Machine Learning in Microbiome Research

This chapter provides an overview of machine learning in microbiome research. Except for the Summary, there are six sections. Section 2.1 introduces two approaches to microbiome prediction using machine learning: learning the microbiome directly or learning multi-omics integration. Section 2.2 simply lists three categories of the most often used machine learning methods in microbiome research – (1) supervised machine learning, (2) unsupervised machine learning, and (3) deep learning – and briefly introduces related Python and R packages. Section 2.3 introduces some benefits of machine learning of microbiome data, including (1) improving model performance, (2) mitigating the issue of high dimensionality, (3) mitigating the issues of high dimensionality and dependency, (4) mitigating the issue of heterogeneity, (5) mitigating the issue of nonlinearity, and (6) facilitating interpretability by ranking feature importance. Section 2.4 describes the limitations of machine learning microbiome data from three perspectives: (1) machine learning models correlation rather than causation, (2) training data size matters in machine learning, and (3) machine learning does not ensure the effectiveness and accuracy of its results. Section 2.5 describes the challenges and issues of machine learning microbiome data, including (1) challenges due to unique characteristics of microbiome data, (2) issues of overfitting, (3) model generalization, (4) assessing algorithms, (5) minimizing predictive errors, (6) performance metrics, and (7) data preprocessing and feature engineering. Section 2.6 provides some remarks on machine learning in microbiome research from five perspectives: (1) model complexity versus interpretability, (2) universal superiority versus best suitability, (3) effects of parameters and hyperparameters, (4) misuse of performance metrics, and (5) insights from assessing data processing methods and machine learning models.