Abstract C015: Clinically usable classification of cancer-associated fibroblast subtypes in pancreatic cancer

Abstract

Abstract Background: Cancer-associated fibroblasts (CAFs) in pancreatic adenocarcinoma (PDAC) are known to play a significant role in regulating tumor progression, invasion, and metastasis. Multiple studies using experimental strategies, as well as single-cell RNA-sequencing (scRNAseq) technology, have shown the existence of subpopulations of PDAC CAF cells, with divergent physical and biological characteristics. However, how to translate CAF cell subpopulations to the clinic remains unclear. Methods: We developed a computational method for scRNAseq, SCISSORS, to identify rare cell clusters and determine exemplar genes. Using SCISSORS, we have defined human exemplar genes of the previously described myCAF and iCAF cell subpopulations using scRNAseq data. These exemplar genes were then used to develop a single-sample classifier (SSC) for PDAC CAFs that we call CAPER in bulk RNAseq data. CAPER was trained using penalized logistic regression in 4 datasets (n= 372) and validated in 8 datasets (n= 593). Results: SCISSORS was able to identify rare cells such as previously described apCAFs (0.59% of cells) and basal-like cell clusters (0.092% of cells) that were missed in the previous analysis. We hypothesized that the sensitivity of SCISSORS would allow us to define exemplar genes that may be applied to bulkRNAseq. Indeed, the top ranked 25 exemplar genes defined by SCISSORS were associated with overall survival (OS) (10 datasets, n=948, p=0.017, log-rank test). To develop a clinically usable classifier, we used the SCISSORS exemplar genes to develop the SSC CAPER which consists of 9 gene pairs. CAPER CAF calls in the validation datasets showed high reproducibility compared with cluster-based labels (AUC: 0.956). In a meta-analysis, patients with CAPER myCAF subtypes showed shorter OS (median OS: 18.0m) compared to CAPER iCAF subtypes (median OS: 28.7m) (p<0.001, log-rank test). Multivariable analysis found that CAPER CAF subtypes were independently prognostic when included with PurIST tumor subtype calls (CAPER p<0.001, PurIST p<0.001, Cox regression). Interestingly, patients subtyped as basal-like for the tumor and myCAF for the CAF showed the worst outcome with a median OS of 11.2 months; while patients subtyped as classical for the tumor and iCAF for the CAF showed the best outcome with a median OS of 30.2 months (p<0.001, log-rank test). Conclusions: CAF genes have largely been evaluated in scRNAseq. We have defined CAF exemplar genes that can be applied to bulk RNAseq, and have developed a clinically usable classifier similar to the now CLIA-approved PurIST that we call CAPER that is robust and accurate for the classification of CAF subtypes in patients. The ability to evaluate CAF subtypes in bulk gene expression provides a powerful tool that may be applied to the plethora of data from patients and clinical trials. CAF subtypes are independently prognostic compared to tumor-intrinsic subtypes, supporting important biological implications, and may facilitate the design of future clinical trials. Citation Format: Xianlu L. Peng, Elena Kharitonova, Joseph Kearney, Ian McCabe, Madison Jenner, Ashley B. Morrison, Alina Iuga, Naim U. Rashid, Jen Jen Yeh. Clinically usable classification of cancer-associated fibroblast subtypes in pancreatic cancer [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Pancreatic Cancer; 2023 Sep 27-30; Boston, Massachusetts. Philadelphia (PA): AACR; Cancer Res 2024;84(2 Suppl):Abstract nr C015.