12 Longitudinal studies are crucial for discovering casual relationships between the 13 microbiome and human disease. We present Microbiome Interpretable Temporal 14 Rule Engine (MITRE), the first machine learning method specifically designed for 15 predicting host status from microbiome time-series data. Our method maintains 16 interpretability by learning predictive rules over automatically inferred time-periods 17 and phylogenetically related microbes. We validate MITRE's performance on 18 semi-synthetic data, and five real datasets measuring microbiome composition 19 over time in infant and adult cohorts. Our results demonstrate that MITRE performs 20 on par or outperforms "black box" machine learning approaches, providing a 21 powerful new tool enabling discovery of biologically interpretable relationships 22 between microbiome and human host. 23 24 Keywords: microbiome, time-series, longitudinal, machine learning 25 26 65 66 In previous work, we presented the MDSINE [14] algorithm, which infers dynamical 67 systems models from microbiome time-series data in order to predict population 68 dynamics of the microbiome over time. Our present work, MITRE, addresses a 69 different question: can we predict the status of the host given microbiome time-70 series data. From the machine learning perspective, MDSINE is an unsupervised 71 model, whereas MITRE is a supervised model. The key distinction is that MDSINE 72 models the microbiome, whereas MITRE does not, and instead models host 73 outcomes. Like other supervised models, MITRE focuses on finding only the 74 essential features (in this case, microbial clades and relevant time-windows) to 75 explain the outcome, rather than attempting to explain the microbiome data itself. 76This architecture is ideal for highly heterogeneous datasets with many 77 "distractors," which are the reality for longitudinal studies of the human 78 microbiome. Supervised machine learning classifiers are employed in many 79 biomedical predictive modeling applications, including forecasting (predicting a 80 future outcome, such as onset of disease, based on past data) and diagnosis or 81 subtyping (predicting which category a subject belongs to based on all available 82 data). 83 84 MITRE's unique contributions are its modeling of the special properties of 85 microbiome time-series data (phylogenetic and temporal relationships), and its 86 process variance parameter, which is empirically estimated from the real data as