Nature Medicine: doi: /nm.3967

Transcription

1 Supplementary Figure 1. Network clustering. (a) Clustering performance as a function of inflation factor. The grey curve shows the median weighted Silhouette widths for varying inflation factors (f [1.6, 10]). The red circle indicates the optimal inflation factor of 3.8, which yields the highest average weighted Silhouette width. Representative consensus matrices for 3 (f = 1.6), 4 (f = 3.8) and 5 (f = 10) clusters are shown in the heatmaps in (b), (c) and (d), respectively.

2 Supplementary Figure 2. Distribution of unlabeled samples as percent of data set size from the consensus subtype labeling. Median unlabeled percentage is 20%. The outlier data set (GSE20916) with the corresponding large percentage of unlabeled samples is a study comprised of multiple tissue types, including micro and macro dissected normal mucosa, adenoma, colonic crypt, carcinoma, and adenocarcinoma. Adenocarcinoma samples were extracted for the analysis in our study.

3 Supplementary Figure 3. Classifier performance metrics. (a) Performance metrics for all validation samples (sensitivity, specificity accuracy). (b) Performance metrics on all validation samples, reported per data set. (c) Performance metrics for validation samples (sensitivity, specificity accuracy) in data sets profiled on the Affymetrix and RNAseq expression platforms. (d) Performance metrics on validation samples, reported per data sets in data sets profiled on the Affymetrix and RNAseq expression platforms. (e) Performance metrics for validation samples (sensitivity, specificity accuracy) in data sets profiled on the Agilent expression platform. (f) Performance metrics on validation samples, reported per data sets in data sets profiled on the Agilent expression platform. (g) Performance metrics measured on 3 holdout data sets (GSE39582, GSE17536, TCGA). (h) Summary of performance metrics. Classifier was trained using all fresh-frozen samples (Affymetrix/RNAseq) and validated on the PETACC3 (FFPE) samples.

4 Supplementary Figure 4. Classifier specificity analysis. Relationship between posterior probability threshold on percentage of unclassified samples (solid grey line) and specificity (CMS colored dash lines), where specificity is measured from the validation data split.

5 a b c d Supplementary Figure 5 Unlabeled sample analysis. (a) Principal components analysis of labeled (CMS classified) and unclassified samples. Labeled CMS samples are represented with centroid (colored dot) and 1 standard deviation (colored dashed line). Unclassified samples are represented as black dots. (b) Projected data in three-dimensional latent space using factor analysis model. Black dots represent all four CMS classified samples and red dots represent unclassified samples. (c) Clustering of Random Forest posterior probabilities of unclassified samples. Posterior probabilities less than 0.2 are thresholded to 0; (d) Distribution of mixed subtypes in unclassified samples.

6 Supplementary Figure 6. TCGA methylation analysis. (a) Hierarchical clustering plot for the TCGA 450K DNA methylation data set. Labels for batch and tissue source site (TSS) are plotted below. (b) Heatmap of TCGA DNA methylation 450K data set. Shown is heatmap representation of DNA methylation β-values of 1,486 most variable probes (among probes present in the 27K beadchip) (standard deviation >0.18, 10% of the probes) across 301 tumor samples and 38 normal samples (NT) with dark blue indicating low DNA methylation and yellow indicating high DNA methylation. The RPMM-based cluster assignments are indicated above the heatmap: light green cluster 1/CIMP-H (n = 51); green cluster 2/CIMP-L (n = 65); light blue cluster 3 (n = 115) and blue, cluster 4 (n = 70). (c) Heatmap of differentially methylated genes between the four CMS groups. The heatmap was produced with gene-level methylation data standardized per gene across all samples. Differentially methylated genes were ordered based on hierarchical clustering with Ward s linkage.

7 c WNT TP53 RTK/MAPK PI3K TGF beta d 1.00 Integrative analysis - mutation, copy number (n=168) e purity 0.50 purity Classified Unclassified CMS1 CMS2 CMS3 CMS4 Supplementary Figure 7. TCGA integrative analysis. (a) Key genomic/epigenomic markers in TCGA data set: SCNA high (>Q1 for non-zero GISTIC score events); hypermutation (>180 events in exome sequencing); MSI and CIMP cluster. (b) Distribution of targeted mutations in TCGA data set. (c) Activation of canonical signaling pathways in CRC from integrative mutation, copy number events (high-level focal amplifications, homozygous deletions), and significant up- or down-regulation of gene expression (TCGA data set). RTK/MAPK: receptor tyrosine kinase/ mitogen activated protein kinase. (d) ABSOLUTE purity estimates from the TCGA samples, comparing the CMS classified and unclassified sample. (e) Distribution of ABSOLUTE scores across the CMS groups in the TCGA cohort.

8 Supplementary Figure 8. TCGA microrna analysis. (a) Analysis of batch-effect in TCGA microrna data using hierarchical clustering. Labels for batch and tissue source site (TSS) are plotted below. (b) Comparison of differentially expressed CMS micrornas in TCGA data set 1 (n = 197) and TCGA data set 2 (n = 200): top left (CMS1), top right (CMS2), bottom left (CMS3), bottom right (CMS4). (c) Heatmap of top 110 differentially expressed mirnas across all CMS groups in TCGA.

9 Supplementary Figure 9. Tumor vs. normal comparison. (a) Principal component analysis including unmatched normal samples from the GSE39582 cohort (n = 19, black dots). Dashed blue line is optimal hyperplane separating tumor from normal samples. (b) Corresponding distribution (by CMS + normal) of samples from separating hyperplane. (c) Principal component analysis including unmatched normal samples from the PETACC-3 cohort (n = 64, black dots). Dashed blue line is optimal hyperplane separating tumor from normal samples. (d) Corresponding distribution (by CMS + normal) of samples from separating hyperplane.

10 Supplementary Figure 10. Distribution of CMS groups across different tumor sites. Right colon from cecum to transverse, left colon from splenic flexure to sigmoid, and rectum (n = 2,651).

11 Supplementary Figure 11. Analysis of clinical outcomes. (a) Kaplan-Meier analysis on the PETACC-3 data set: overall-survival (left), Relapse- free survival (center), and survival after recurrence (right). (b) Timedependent AUC analysis, comparing the clinical + molecular model (MSI/BRAF mutations/kras mutations) with the clinical + molecular + CMS model. See Supplementary Table 13 for details.

12 Supplementary Figure 12. Distribution of subtype labels as reported in original subtyping publication and in consortium samples. Overall, the same distribution of subtype labels is seen in individual subtyping efforts (green) and aggregated consortium data (orange).

13 Supplementary Figure 13. Data aggregation workflow. All data sets are combined into a single, normalized gene expression matrix. This matrix is used in the development of a consensus subtype classifier.