Development of a consensus molecular classifier for pancreatic ductal adenocarcinoma

Abstract

Background: Pancreatic ductal adenocarcinoma (PDAC) presents a significant challenge, with a 5-year survival rate of approximately 10%. Tumor heterogeneity contributes to the limited effectiveness of treatments. Several tumor and stroma molecular classifiers have attempted to clarify this heterogeneity with moderate agreement. Recognizing the complexity introduced by this extensive array of taxonomies, this study aims to develop a consensus molecular classifier by including both tumor and stroma features.

Methods: We analyzed mRNA expression data from 514 PDAC samples, applying batch correction, filtering out low-expression genes, and using variance-stabilizing transformation. Tumor and stroma profiles were classified with previously published systems, while stroma compartments were estimated through virtual microdissection. For each classifier, multiple machine learning models were trained and optimized, with the top performers used to assign subtypes. A consensus classifier was created by building subtype similarity networks and applying a Markov clustering algorithm, and robustness was evaluated through resampling. Associations between consensus classes and overall survival were examined using multivariate Cox models.

Results: The results indicated that Elastic-Net emerged as the superior model. We identified two classes for tumor components (Consensus Classical and Consensus Non-classical) and stroma components (Consensus Normal-Immune and Consensus Activated-ECM). The consensus Random Forest achieved a balanced accuracy of 96.33 and 98.92%, respectively. Across cohorts, the PDAConsensus algorithm identified tumor subtypes with distinct prognostic value, with Consensus Non-classical tumors showing poorer survival than Consensus Classical tumors. Associations for stroma consensus classes were weaker and less consistent across datasets.

Conclusions: We developed a robust consensus classifier for PDAC that integrates tumor and stroma features. The classifier is accessible through the R package PDACMOC (PDACMolecularOmniClassifier, https://github.com/pavillos/PDACMOC , https://doi.org/10.5281/zenodo.17161373 ) and a Shiny app ( https://pdacmoc.cnio.es/ ). This classifier offers a biologically grounded framework that integrates existing systems, allows for single-sample classification, and improves prognostic stratification. By enabling subtype-specific therapies and better patient stratification in clinical trials, it can help guide precision medicine and enhance outcomes in PDAC management.

Keywords: Consensus molecular classifier; Machine Learning; PDAC; Personalized medicine; Transcriptomics.

Fig. 1

Circus plots display the distribution of subtypes across various classification systems. ADEX (aberrantly differentiated endocrine-exocrine), QM (quasi-mesenchymal)

Fig. 2

Consensus classes are derived from applying the Markov Clustering Algorithm over the network of Cohen’s kappa indices between the subtypes of each classification

Fig. 3

Parallel sets illustrate how samples classified as Consensus Classical or Consensus Non-classical and as Consensus Normal-immune or Consensus Activated-ECM are categorized by other classifiers

Fig. 4

Correlation between NonClassicalScore and ActivatedECMScore

Fig. 5

Cox Proportional Hazards models of tumor and stroma classifiers in the validation datasets. ADEX (aberrantly differentiated endocrine-exocrine), ECM (extracellular matrix), ICGC (International Cancer Genome Consortium), QM (quasi-mesenchymal