Integration of multi-omics datasets enables molecular classification of COPD

Discussion

The primary objective of the current study was to investigate whether integration of large-scale omics data from multiple molecular levels and anatomical locations increases the power and accuracy of molecular classification in complex disease, here exemplified by COPD. Our proof-of-concept investigations applying SNF network integration [14] to nine omics data blocks from 52 female subjects from the Karolinska COSMIC cohort clearly show that integrating multi-omics data greatly improves the accuracy of unsupervised classification. The mean accuracy of prediction of the three groups of COPD, smokers with normal lung function and never-smokers increased in a near-linear fashion for the 303 evaluated network combinations, from a mean accuracy of 28% for the single-omics platforms to 90% for septuple omics integration. However, a large degree of variation was observed depending on the specific omics platforms included. Larger n-tuple omics did not automatically equate with better prediction for the specific omics combinations, and 100% accurate classification was achieved with sextuple omics integration (figure 3, figure E7). An accuracy >85% was reached by 25 different quadruple to septuple omics combinations, indicating that there is a large degree of plasticity in the combination of omics data required to optimise the accuracy of prediction (figure 5). Notably, the seven networks with >95% accuracy of prediction all contained omics datasets from several different molecular levels (miRNA, proteomes, metabolomes and eicosanoids) and anatomical locations (BAL cells, BAL fluid and serum), implying that combining data from different anatomical compartments and molecular levels is advantageous.

Smoking alone induces significant alterations of up to 50% of all biomolecules in the lung, as demonstrated in the BAL cell and BEC proteomes of the same cohort [28, 29]. As such, the true challenge in this cohort is to distinguish the subtle molecular effects associated with COPD pathology from the confounding effects of acute smoking in the group of current-smoker COPD patients. None of the single-omics data blocks had the power to separate out the molecular alterations due to mild-to-moderate COPD from the confounding effects of current smoking in an unsupervised fashion (figure 3, figure E7). In our previous investigations of the single-omics data blocks [17, 21, 30] as in most COPD investigations, stratification by current smoking status in combination with supervised analysis has therefore been mandatory to identify COPD-related alterations. In contrast, quintuple to septuple omics integration provided the power to classify COPD diagnosis from both never-smoker and current-smoking controls with 100% accuracy (figure 2a). It should be emphasised that this was achieved in an unsupervised, data-driven manner, with subgroups as small as n=6. The ability to sub-phenotype the COPD group based on chronic bronchitis diagnosis with 100% accuracy following integration of two or more omics platforms using subgroups as small as n=4 subjects (figure E9) further demonstrates the clinical utility of these methods. The integration methods employed here may thus enable systems medicine-based approaches to be performed in small, focused cohorts, which is desirable given the prohibitive costs of performing nonuple omics analyses on larger cohorts. Integration of the full arsenal of omics characterisations from multiple compartments could provide the power to detect rare molecular sub-phenotypes of disease from statistically relatively small cohorts, which is the general case for translational multi-omics studies.

The homogeneous COPD population selected for these proof-of-principle investigations, consisting of female patients with mild-to-moderate COPD who were free from co-morbidities or treatment, reflects a somewhat artificial scenario that poses limitations on the ability to extrapolate the findings to a more general disease population. COPD is a heterogeneous disease that can consist of 10–15 distinct molecular sub-phenotypes [35–37], few of which have been defined to date [16]. It is clear that the current clinical characterisation scheme does not provide the necessary resolution to classify or even identify these disease subgroups [16]. While the end goal of the unsupervised workflow presented here is to perform molecular sub-classification of all or at least several of the existing COPD sub-phenotypes, any study evaluating the performance of this method absolutely must focus on a COPD subgroup that can be expected to have molecular similarities to ours. The existence of a female-dominated molecular COPD phenotype in the Karolinska COSMIC cohort has been well established in our findings from supervised analyses of the single-omics data blocks [17, 21, 23, 30]. The female part of the Karolinska COSMIC cohort thus provides a set of molecularly distinct ground-truth study groups to evaluate the unsupervised classification.

Missing data is a common issue in studies of human subjects, particularly in multi-omics, multi-compartment studies such as the Karolinska COSMIC study. It is common for omics data blocks to be missing from a given subject for a number of reasons, e.g. omitted sampling of selected biospecimens because of safety considerations for the patient, or for individual omics experiments to be excluded owing to quality control criteria. The result is a data matrix with gaps, where every subject has some missing omics data blocks (figure E1). The original SNF data integration approach does not accommodate missing data blocks [14]. As such, developing approaches for dealing with missing data in the network construction was a secondary aim of the study. Out of three evaluated approaches, the most liberal of the three in terms of allowing inclusion of subjects with missing data (unequal sample size strategy; figure E5) performed the best. Accordingly, this strategy appears to provide a robust way to allow inclusion of all subjects and data collected in spite of the missing data block issue, which is inevitable in this type of translational study with invasive sampling at the site of inflammation.

Although the utilised cross-validation design assures that the accuracy is estimated independent of the training set, the relative homogeneity and limited sample size of this cohort poses some constraints. A larger cohort may facilitate the true goal of multi-omics data integration: to identify previously unknown, molecularly distinct sub-phenotypes of health and disease. The results of these proof-of-principle investigations may provide some guidance in the design of future systems medicine studies, in which five to seven omics data blocks collected from complementary molecular levels and anatomical locations, with cohort sizes of six to ten subjects times the expected number of distinct molecular subgroups, may represent a good starting point for molecular sub-phenotyping of complex diseases. For example, the current diagnostic criteria for COPD represent a range of sub-phenotypes, driven by sex, smoking history, premature birth, environmental exposures etc. If we were to postulate that these aetiologies give rise to 15 molecularly distinct phenotypes, each represented by a different subset of biomarkers and pharmaceutical targets, then based on the subgroup sizes indicated from this cohort (table 1), an investigation using one omics dataset of choice would require 30 study subjects per COPD subgroup, i.e. 450 patients plus relevant control groups. The increased molecular resolution afforded by septuple omics integration reduces the required number of study subjects to n=6 per subgroup to be identified. Thus, for the postulated study design aiming to identify 15 molecularly distinct subgroups of COPD, it would be sufficient to include 90 COPD patients. This dramatic reduction in the number of patients needed to achieve the study aims will greatly increase the clinical feasibility of molecular phenotyping studies. A significant bottleneck in clinical studies is often the ability to recruit a sufficient patient population in a timely fashion. The use of septuple omics integration can vastly reduce the time necessary for the cohort collection, facilitating the identification of unknown molecular sub-phenotypes.

In conclusion, these examples from the integration of nine omics data blocks from the Karolinska COSMIC cohort demonstrate an extraordinary increase in statistical power and accuracy of group classification, achieved by integrating data from multiple molecular levels and anatomical locations. For the first time, we have quantified the improvements in statistical power afforded by multi-omics integration, with the classification powers increasing on average from 28% to 90% with septuple omics integration. From the perspective of computational systems medicine, the mechanism of disease is not caused by independent subsets of genes/proteins/metabolites identifiable by traditional univariate statistics, but rather by their interactions, which makes it vital to develop a system-level understanding of disease. As demonstrated here, bridging and integrating data from multiple molecular levels and anatomical compartments relevant for disease pathology from the same individual could at last provide the statistical power of unsupervised classification. The unsupervised aspect is mandatory to facilitate identification of unknown molecular sub-phenotypes of complex diseases such as COPD and asthma. The unsupervised identification of molecularly distinct subgroups of disease represents a first, crucial step in elucidating treatable traits and biomarker subsets. The molecularly distinct subgroups identified by SNF can then be interrogated for handprints of diagnostic or prognostic biomarkers by supervised multivariate modelling approaches, i.e. orthogonal projections to latent structures, that provide a filter of variables of interest. In addition, we are currently developing multi-omics integration approaches at the pathway level. These leverage on the increased power achieved by integrating across several molecular levels in the downstream steps of identifying mechanistic features and treatable traits associated with identified patient subgroups. Combining the identified subsets of biomarkers, or “handprints of disease” [38], from multiple molecular levels may bring forth a long-awaited paradigm shift in precision- and personalised medicine.