Virtual Picrosirius Red Generation Enables Spatial Analysis of Collagen Biology in Colorectal Cancer
Jennifer Choi
Lay Summary:
We found that collagen organization is linked to nearby gene activity in patterns that differ from tumor to tumor, revealing biological relationships that broader analyses can miss.
Abstract:
Collagen architecture is extensively remodeled in colorectal cancer, but Picrosirius Red staining is not routinely available in clinical pathology. We used a framework to generate virtual Picrosirius Red from routine H&E and integrate quantitative collagen features with spatial transcriptomics. Generated collagen features were quantified with Cabana and integrated with 18 transcriptional programs across 233,613 Visium spots from 41 CRC specimens. Spatial relationships were examined using MISTy, MEFISTO, and tumor distance analysis. The diffusion model preserved collagen architecture better than Pix2pix- median Spearman correlation with real PSR increased from 0.559 to 0.667, with improvement in 22 of 34 variable features. Cohort wide collagen–molecular associations were highly heterogeneous and none remained significant after multiple testing correction. Spatial models, however, revealed recurring structure- MISTy showed strong contributions from neighboring collagen context, while MEFISTO identified 28 shared collagen–molecular factors across 21 specimens, most frequently involving ECM and fibroblast programs. Tumor distance analysis further suggested spatial variation across the tumor–stromal interface. Overall, virtual collagen staining from H&E enables quantitative analysis of collagen architecture and reveals spatially organized, tumor specific relationships between extracellular matrix structure and molecular activity in CRC.
Q&A:
Bios: Jennifer Choi
Program Track: Advanced Research
GitHub Username:
choichoi0 -Jennifer Choi
What was your favorite seminar? Why?
ST Conditioned Synthetic Tissue Patch Generation because it showed how spatial transcriptomic information can condition image generation to produce synthetic tissue patches that preserve biologically meaningful features. I was especially interested in the connection between molecular measurements and tissue morphology, since it demonstrated how generative models can do more than reproduce realistic pathology images. -Jennifer Choi
If you were to summarize your summer internship experience in one sentence, what would it be?
I learned how computational pathology can extract quantitative tissue architecture from routine histology images and connect those visual patterns to spatial gene expression -Jennifer Choi