e15647 Background: High inter-patient heterogeneity and spatial diversity within tumors represent significant challenges to improving outcomes in colorectal cancer (CRC). Despite elucidation of consensus molecular subtypes (CMS), little progress has been made in translating this knowledge into clinical gain. Intratumoral CMS heterogeneity remains a relatively unexplored area and understanding of the interplay between subtypes holds promise for improving prognostication. We leverage spatial transcriptomics (ST) to enhance understanding of CMS heterogeneity and spatial distribution, with the goal of advancing personalized therapy through high resolution molecular sub-typing. Methods: ST using the 10x Genomics Visium platform was performed on paraffin-embedded CRC specimens for which comprehensive clinicopathological and long-term clinical outcome data were available. ST data were deconvoluted with respect to CMS single-cell RNA sequencing (scRNA-seq) data, yielding maps of CMS landscapes at a resolution of 55 microns. A pre-trained deep learning model was used to identify distinct CRC tissue classes (tumor epithelium, stroma, etc.) within H&E images, enabling spatial correlation of histomorphology with CMS. Cellular composition within these CMS maps was quantified through deconvolution of six major cell types using reference scRNA-seq data. Pseudo-bulk samples were generated from the ST data to evaluate CMS behaviors and compute pathway activities. Results: 40 CRC patients were assayed using ST and H&E profiling (13, 8, 14, and 5 from stages 1-4, respectively). Median follow-up was 112 months (Range 1–192). Intratumoral CMS heterogeneity was evident in all patients, with diverse CMS compositions. Using Simpson diversity index to quantify this heterogeneity, we observed significant correlation of heterogeneity with disease specific survival (DSS) (p = .03). CMS landscapes displayed spatial patterns that were distinct among tissue categories (average F1-score: testing = .92, validation = .89. Tissue categories based on deep learning from H&E images). Tumors with higher levels of CMS1 or 4 were associated with worse outcome. Stratification by the composite CMS biomarker (CMS1+CMS4)/(CMS2+CMS3) yielded significant separation of the Kaplan-Meier curve for DSS (p = .018). Conclusions: This study demonstrates that CRC tumors contain significant intratumoral CMS heterogeneity that is not detectable using bulk classification methods. Spatial patterns in CMS transcriptional topographies correlate with H&E-derived morphology, suggesting the potential for estimating CMS composition from histopathological features. Importantly, we show that spatially composite CMS-derived biomarkers possess strong prognostic potential. The capacity to decode CMS topographic patterns and associate these with clinical outcome offers great potential to advance personalized therapy for CRC.
Read full abstract