Poster #46: Akil Talakola

DNA Methylation-Based Classification of Consensus Molecular Subtypes in Colorectal Cancer

Akil Talakola, BS Candidate¹; Ivy D. Tran, MD/PhD Candidate¹; Jayashri Ghosh, PhD¹ ¹Lewis Katz School of Medicine at Temple University, Temple University, Philadelphia, PA, USA

Colorectal cancer is the second most common fatal cancer in the United States and can be divided into four consensus molecular subtypes (CMS1 through CMS4) with distinct biology, prognosis, and treatment response. Standard subtype identification requires RNA sequencing, which is unstable and degrades during clinical sample handling, while DNA methylation is more stable and could serve as a practical alternative. The best published methylation classifier (Van den Berg et al., 2021) concluded that DNA methylation cannot reliably classify CMS4, the mesenchymal subtype associated with the poorest prognosis, and their success was limited to binary discrimination of CMS2 and CMS3. This project aimed to build an elastic net methylation classifier trained on The Cancer Genome Atlas (TCGA) for all four CMS subtypes. Illumina 450K methylation data from 274 TCGA colorectal tumors were used. CMS labels were assigned by CMSclassifier's single sample predictor (SSP) using matched RNA sequencing data. 49 tumors that SSP could not confidently classify were excluded, leaving 225 labeled samples. After deduplication, filtering for CpGs with SD > 0.15, and removing probes with missing values, 61,375 CpGs remained. A stratified 80/20 train-test split produced 181 training and 44 test tumors. A multinomial elastic net classifier was trained with ten-fold cross-validation. The model achieved 88.6% overall test accuracy across all four subtypes. Notably, the model reached 57.1% accuracy on CMS4, exceeding the 47% ceiling reported by Van den Berg et al. Annotation of the 388 model-selected CpGs identified relevant features: FOXF2, a transcription factor driving epithelial-mesenchymal transition in CMS4; NDRG4, a Cologuard biomarker in CMS1; and enrichment for cell adhesion and tight junction pathways in CMS3, consistent with its enterocyte-like biology. Future work will expand training with additional cohorts, apply cross-cohort validation with batch correction, and test biology-guided feature selection methods to further improve CMS4 accuracy.