Derivation of Distinct Prognostic Tumor Cell States in PDAC via Single-Cell RNA Sequencing

Abstract

Pancreatic ductal adenocarcinoma (PDAC) is an aggressive cancer type with a poor prognosis. There is a critical need for robust biomarkers to better risk-stratify patients and personalize treatment. Here, we utilize single-cell RNA sequencing (scRNA-seq) and CIBERSORTx deconvolution to genomically classify PDAC based on distinct tumor cell states in order to more accurately prognosticate patients.scRNA-seq was conducted on 24 PDAC patient samples prior to any systemic therapy at Washington University (19 core needle pancreatic biopsies from the time of diagnosis and 5 surgical PDAC resections) and from the public domain (n = 34). The resulting scRNA-seq data from 58 patients was integrated and clustered in order to identify distinct tumor cell states. Following gene expression profiling analysis to characterize each distinct tumor cell state, CIBERSORTx was used to deconvolve 125 bulk RNA-seq early-stage PDAC tumor samples from The Cancer Genome Atlas (TCGA), and each tumor was classified based on the dominant malignant cell state present. Following this, survival analysis was performed to compare between different groupings of TCGA patients based on their dominant malignant cell states.scRNA-seq data from 58 public and in-house samples were clustered and cell types were annotated based on known expression markers. We then clustered the malignant cell cluster (Epcam+) into 5 distinct tumor cell states. Fractions of these 5 malignant cell states were then quantified in TCGA via deconvolution using CIBERSORTx. TCGA samples were categorized into 5 groups based on the dominating malignant (mal) cell state fraction value - ranging from mal 0 to mal 4. In this fashion, we identified 39 patients as mal 0, 9 as mal 1, 6 as mal 2, 2 as mal 3, and 69 as mal 4. Gene expression profiling analysis revealed that the mal 0 cell state is representative of the "squamous-like" or "basal" subtype reported in prior literature, while mal 1-4 more likely represent the previously described "classical" subtype. We then aimed to perform survival analysis. To achieve meaningful statistical power, we grouped mal 1-3 into one category (called mixed) containing 17 patients. Kaplan-Meier analysis revealed the following median survival times (in months): mal 0 (9 months), mal 4 (16 months), and mixed (15 months). Log-rank statistical testing revealed that patients classified as mal 0 experienced significantly worse overall survival than those classified as other mal subtypes (P < 0.05).We analyzed scRNA-seq data from 58 independent PDAC patients, which we used to identify 5 distinct tumor cell states. These tumor cell states appear to correspond with the previously reported "squamous-like" or "basal" vs. "classical" subtypes, correlate significantly with overall survival, and could potentially inform more personalized treatment strategies including radiotherapy in the future.