Figures
- FIGURE 1
Nine omics data blocks collected from multiple molecular levels (mRNA, microRNA, proteomes and metabolomes) and multiple anatomical locations (airway epithelium, lung resident immune cells, airway exudates, exosomes and serum) from subjects from the Karolinska COSMIC cohort were used to explore how integration of multiple omics data blocks can improve the statistical power of group classification. The detailed methods used for sample collection as well as the analytical platforms used for data collection are described in the supplementary methods. The overlap of omics datasets for the 52 included subjects is shown in figure E1. DIGE: 2-D difference gel electrophoresis proteomics; iTRAQ: isobaric tags for relative and absolute quantitation proteomics; TMT: tandem mass tag proteomics; BALF: bronchoalveolar lavage fluid; BAL: bronchoalveolar lavage; BEC: bronchial epithelial cell; exosome: exosomes from BALF.
- FIGURE 2
a) Accuracy of group prediction as a function of the number of omics datasets included in the SNF-mediated omics integration using nine omics datasets from the Karolinska COSMIC cohort (see figure 1). Values are displayed as mean accuracy±se (solid line) as well as maximum accuracy (dashed line) for all possible omics combinations for each respective number of omics platforms, ranging from single to septuple omics integration. The presented data are based on the unequal sample size strategy (for other sampling strategies, see figure E5). b) Individual power curves corresponding to mean accuracy levels for each omics n-tuple. The graphs are showing group size (n) versus accuracy of group prediction for each respective omics n-tuple, ranging from single to septuple omics integration. Solid black horizontal line indicates 95% accuracy level, dashed vertical lines indicate n required at 95% accuracy level for single versus septuple omics integration. c) Heatmap displaying the mean accuracy levels achieved for subgroup sizes of n=1–5 for each n-tuple omics integration. Accuracy was calculated as the similarity network fusion-based accuracy compared to classification by chronic obstructive pulmonary disease diagnosis (according to the Global Initiative for Chronic Obstructive Lung Disease criteria) and current smoking status (defined by exhaled carbon monoxide monitoring).
- FIGURE 3
The best performing subject similarity network, consisting of a sextuple omics integration similarity network fusion similarity network (centre), provided 100% correct classification of the three subject groups of Healthy, Smokers and COPD. Similarity networks for each of the included single-omics data (periphery) are shown for reference. Nodes represent subjects. The networks are displayed with subjects clustered according to network similarity. The accuracy of 100% is based on 10 000-times leave-one-out cross-validation permutation test using training data iteratively selecting six samples from each group. The same network displayed as a fixed-position network, with clustering according to the sextuple fused network preserved for all seven networks to facilitate visual comparison, is available in figure E7. Blue: healthy never-smoker; yellow: smoker with normal spirometry; red: smoker with COPD. BAL: bronchoalveolar lavage; DIGE: 2-D difference gel electrophoresis proteomics; iTRAQ: isobaric tags for relative and absolute quantitation proteomics; BALF: bronchoalveolar lavage fluid.
- FIGURE 4
Example of accuracy increasing with omics n-tuple in similarity network fusion integration. The accuracy of prediction is indicated by the node size, ranging from the smallest (5% accuracy; bronchoalveolar lavage fluid (BAL) isobaric tags for relative and absolute quantitation proteomics (iTRAQ) single omics) to the largest representing 100% accuracy (sextuple omics). The n-tuple of omics datasets fused is shown from single (bottom) to sextuple omics (top). The n-tuple is also indicated by colour coding in grey to blue. BALF: bronchoalveolar lavage fluid; DIGE: 2-D difference gel electrophoresis proteomics.
- FIGURE 5
Thirty different similarity network fusion multi-omics network combinations reached an accuracy of group prediction >85%, calculated as the normalised mutual information compared to chronic obstructive pulmonary disease (COPD) diagnosis and smoking status (three groups: Healthy, Smoker, COPD). Seven of the network combinations reached an accuracy >95% (grey line). a) Accuracy of prediction achieved for the respective combination; b) omics datasets included in the specific fused network corresponding to the respective bar graph shown above. Please note that each bar represents the accuracy of a single, specific network, hence the lack of error bars. BAL: bronchoalveolar lavage cells; DIGE: 2-D difference gel electrophoresis proteomics; BALF: bronchoalveolar lavage fluid; iTRAQ: isobaric tags for relative and absolute quantitation proteomics; exo: exosomes isolated from BALF; BEC: bronchial epithelial cells; TMT: tandem mass tag proteomics.
Tables
- TABLE 1
Subgroup size required to reach 80% and 95% accuracy of group classification
n-tuple# n at 80%¶ n at 95%¶ Single omics 1 18 30 Dual omics 2 11 18 Triple omics 3 8 13 Quadruple omics 4 7 11 Quintuple omics 5 6 9 Sextuple omics 6 5 8 Septuple omics 7 4 6 #: the number of omics datasets included in the respective integration; e.g. quadruple analysis integrated four different omics datasets; ¶: accuracy of group classification for all included subjects, across all three groups.
Supplementary Material
Please note: supplementary material is not edited by the Editorial Office, and is uploaded as it has been supplied by the author.
Supplementary material ERJ-01930-2017_Supplement
