LAUSR

3601 Background: High stromal content (CMS4) and immune infiltration (CMS1) are important features of the tumor micro-environment (TME) in early stage colorectal cancer, with poor and good RFS prognosis, respectively. However, it remains unclear how these TME features are influenced by intrinsic tumor epithelial expression profiles. We examined how the former can be predicted by the latter, and assessed their respective prognostic contributions. Methods: Eighteen public gene-expression datasets (n=3511) were used for discovery, and an adjuvant trial dataset (PETACC3 n=688) was used for validation. Whole-transcriptome hierarchical clustering was applied to identify intrinsic versus extrinsic co-expression modules, corroborated by single-cell profiling data. The epithelial gene subset were used to predict previously assigned CMS labels, producing new CRC classifier that we refer to as epithelial-CMS (eCMS). Results: Despite lacking stromal and immune response genes, eCMS predicted CMS with an accuracy of 82% [CI 78-85%] in the test dataset. Not only were the RFS prognostic values retained (eCMS1 vs non-eCMS1: HR=0.59 [CI 0.35–0.95, p=0.029], and eCMS4 vs non-e-CMS4: HR=1.82 [CI 1.38–2.41, p=0.0003]), but they were also significantly improved when compared to the original CMS classification (CMS1 vs non-CMS1: HR=0.76 [CI 0.49–1.16, p=0.2] and CMS4 vs non-CMS4: HR=1.43 [CI 1.07–1.91, p=0.017]. In a multivariable regression comparison, eCMS was superior to CMS (p=0.001 vs p=0.36). Each eCMS group was associated with specific combinations of up- and down-regulation of CRC-relevant pathways, such as CDX2, map-kinase, mucin secretion, DNA damage repair and cell proliferation. Conclusions: Stromal and immune infiltration patterns in the TME are strongly associated with tumor epithelial expression profiles, as captured by eCMS. This finding sheds new light on the interplay between intrinsic molecular features of CRCs and their TME, and illuminates new paths for potential combination therapies.