Knowledge guided multi-level network inference

Constantly decreasing costs of high-throughput profiling on many molecular levels generate vast amounts of so-called multi-omics data. Studying one biomedical question on two or more omic levels provides deeper insights into underlying molecular processes or disease pathophysiology. For the majority of multi-omics data projects, the data analysis is performed level-wise, followed by a combined interpretation of results. Few exceptions exist, for example the pairwise integration for quantitative trait analysis. However, the full potential of integrated data analysis is not leveraged yet, presumably due to the complexity of the data and the lacking toolsets. Here we propose a versatile approach, to perform a multi-level integrated analysis: The Knowledge guIded Multi-Omics Network inference approach, KiMONo. KiMONo performs network inference using statistical modeling on top of a powerful knowledge-guided strategy exploiting prior information from biological sources. Within the resulting network, nodes represent features of all input types and edges refer to associations between them, e.g. underlying a disease. Our method infers the network by combining sparse grouped-LASSO regression with a genomic position-confined Biogrid protein-protein interaction prior. In a comprehensive evaluation, we demonstrate that our method is robust to noise and still performs on low-sample size data. Applied to the five-level data set of the publicly available Pan-cancer collection, KiMONO integrated mutation, epigenetics, transcriptomics, proteomics and clinical information, detecting cancer specific omic features. Moreover, we analysed a four-level data set from a major depressive disorder cohort, including genetic, epigenetic, transcriptional and clinical data. Here we demonstrated KiMONos analytical power to identify expression quantitative trait methylation sites and loci and show its advantage to state-of-the-art methods. Our results show the general applicability to the full spectrum multi-omics data and demonstrating that KiMONo is a powerful approach towards leveraging the full potential of data sets. The method is freely available as an R package (https://github.com/cellmapslab/kimono).