Background: The increased multi-omics information on carefully phenotyped patients in studies of complex diseases requires novel methods for data integration. Unlike continuous intensity measurements from most omics data sets, phenome data contain clinical variables that are binary, ordinal and categorical.
Results: In this paper we introduce an integrative phenotyping framework (iPF) for disease subtype discovery. A feature topology plot was developed for effective dimension reduction and visualization of multi-omics data. The approach is free of model assumption and robust to data noises or missingness. We developed a workflow to integrate homogeneous patient clustering from different omics data in an agglomerative manner and then visualized heterogeneous clustering of pairwise omics sources. We applied the framework to two batches of lung samples obtained from patients diagnosed with chronic obstructive lung disease (COPD) or interstitial lung disease (ILD) with well-characterized clinical (phenomic) data, mRNA and microRNA expression profiles. Application of iPF to the first training batch identified clusters of patients consisting of homogenous disease phenotypes as well as clusters with intermediate disease characteristics. Analysis of the second batch revealed a similar data structure, confirming the presence of intermediate clusters. Genes in the intermediate clusters were enriched with inflammatory and immune functional annotations, suggesting that they represent mechanistically distinct disease subphenotypes that may response to immunomodulatory therapies. The iPF software package and all source codes are publicly available.
Conclusions: Identification of subclusters with distinct clinical and biomolecular characteristics suggests that integration of phenomic and other omics information could lead to identification of novel mechanism-based disease sub-phenotypes.