A single-sample gene-expression classifier for the intrinsic consensus molecular subtypes in colorectal cancer.

e15587 Background: An analysis of colorectal cancer epithelial cells showed two intrinsic subtypes called iCMS2 and iCMS3, in addition to the bulk consensus subtypes (CMS1-4). The intrinsic subtypes are prognostically important. We present a robust method for calculating the iCMS subtypes in individual samples, as opposed to batches, applicable in a clinical setting. Methods: We selected the subset of iCMS genes (N = 201) with the strongest epithelial expression in colorectal cancer, aiming to reduce the signal from other cell types. We then calculated gene expression centroids by taking the per-gene mean expression of 8 random samples from three non-batch corrected datasets (1779 samples total) that were previously used to derive the iCMS subtypes. Thirty-two centroids were calculated per dataset and per iCMS, for a total of 192 centroids. The single sample classifier (SSC) calculates the most likely iCMS class based on the class of the nearest centroid by non-parametric correlation distance. A confident call is made when the correlation is at least 0.1 and distance from the nearest centroid of the opposite iCMS class is greater than 0.05. Results: The SSC achieved 98.3% agreement with the reference calls in training data. The SSC was applied to unseen data for validation, by classifying non-batch corrected samples one at a time. The same data were clustered with the reference published iCMS public dataset and the reference iCMS was estimated at the batch level by proximity to reference samples in the hierarchical tree structure. Concordance between reference and SSC calls was 91.4% in non-batch corrected data and only modestly improved to 93.9% when batch correction was applied, showing the robustness of this approach. Concordance remained acceptable even when non-confident SSC calls were included (86.9%). The proportion of calls that were judged confident with the SSC was 77.9%, slightly lower than the proportion of confident reference calls (87.6%). In addition, the SSC was applied to data from the VELOUR trial, for which reference iCMS calls were also available. The SSC iCMS was prognostic (HR 0.601 for OS in iCMS2 vs iCMS3, P < 0.00001) and performed at least as well as the reference iCMS (HR 0.669, P = 0.0004). Conclusions: The SSC reproduces the iCMS classification and appears robust to batch effect. The SSC calls are prognostic in clinical trial data. The SSC R package is available for download (https://github.com/CRCrepository/iCMS.SSC). Table: see text