A protein-based machine learning approach to the identification of inflammatory subtypes in pancreatic ductal adenocarcinoma.

Background/objectives: The inherently immunosuppressive tumor microenvironment along with the heterogeneity of pancreatic ductal adenocarcinoma (PDAC) limits the effectiveness of available treatment options and contributes to the disease lethality. Using a machine learning algorithm, we hypothesized that PDAC may be categorized based on its microenvironment inflammatory milieu.

Methods: Fifty-nine tumor samples from patients naïve to treatment were homogenized and probed for 41 unique inflammatory proteins using a multiplex assay. Subtype clustering was determined using t-distributed stochastic neighbor embedding (t-SNE) machine learning analysis of cytokine/chemokine levels. Statistics were performed using Wilcoxon rank sum test and Kaplan-Meier survival analysis.

Results: t-SNE analysis of tumor cytokines/chemokines revealed two distinct clusters, immunomodulating and immunostimulating. In pancreatic head tumors, patients in the immunostimulating group (N = 26) were more likely to be diabetic (p = 0.027), but experienced less intraoperative blood loss (p = 0.0008). Though there were no significant differences in survival (p = 0.161), the immunostimulating group trended toward longer median survival by 9.205 months (11.28 vs. 20.48 months).

Conclusion: A machine learning algorithm identified two distinct subtypes within the PDAC inflammatory milieu, which may influence diabetes status as well as intraoperative blood loss. Opportunity exists to further explore how these inflammatory subtypes may influence treatment response, potentially elucidating targetable mechanisms of PDAC's immunosuppressive tumor microenvironment.