Motivation Human diseases are characterized by multiple features such as their pathophysiological, molecular, and genetic changes. The rapid expansion of such multi-modal disease-omics space provides an opportunity to re-classify diverse human diseases and to uncover their latent molecular similarities, which could be exploited to repurpose a therapeutic-target for one disease to another. Results Herein, we probe this underexplored space by soft-clustering 6,955 human diseases by multi-modal generative topic modeling. Focusing on chronic kidney disease and myocardial infarction, two most life-threatening diseases, unveiled are their previously underrecognized molecular similarities to neoplasia and mental/neurological-disorders, and 69 repurposable therapeutic-targets for these diseases. Using an edit-distance based pathway-classifier, we also find molecular pathways by which these targets could elicit their clinical effects. Importantly, for the 17 targets, the evidence for their therapeutic usefulness is retrospectively found in the pre-clinical and clinical space, illustrating the effectiveness of the method, and suggesting its broader applications across diverse human diseases. Availability The code reported in this paper is available at: https://github.com/skozawa170301ktx/MultiModalDiseaseModeling Supplementary information Supplementary data are available at Bioinformatics Advances online.

Human diseases are multifactorial - hence it is important to characterize diseases on the basis of multiple disease-omics. However, the capability of the existing methods is largely limited to classifying diseases based on a single type or a few closely related omics data. Herein, we report a topic model framework that allows for characterizing diseases according to their multiple omics data. We also show that this method can be utilized to predict potential biomarkers and/or therapeutic targets. In this study, we illustrate a computational concept of this augmented topic model and demonstrate its prediction performance by a leave one-disease features out cross-validation scheme. Furthermore, we exploit this method together with human disease tissue/organ-transcriptome data and identify putative biomarkers and/or therapeutic targets across 79 diseases. In conclusion, this method and the prediction framework shown reported herein provide important tools for understanding complex human diseases and also facilitate diagnostic and/or therapeutic development.