Frozen 72-CpG DNA Methylation Classifier for Colorectal Tumor Detection
Overview
This repository contains a frozen 72-CpG DNA methylation classifier developed for colorectal tumor detection using high-dimensional Illumina HumanMethylation450 data.
The project emphasizes statistical reliability, leakage-safe validation, reproducibility, and external validation in high-dimensional biomedical machine-learning settings.
Author
Enock Kumi Ackaah
Research Areas
- Biomedical machine learning
- Colorectal cancer
- DNA methylation
- High-dimensional data
- Predictive modeling
- Statistical reliability
- Reproducible machine learning
Development Data
The model was developed using GSE101764.
After quality control and preprocessing, the discovery dataset contained:
- 256 methylation profiles
- 153 patient groups
- 409,990 CpG sites
- 147 mucosal samples
- 109 tumor samples
Raw Illumina HumanMethylation450 IDAT files were processed using quality-control procedures, probe filtering, functional normalization, and beta/M-value extraction.
Model Development
A patient-grouped, leakage-safe validation framework was used.
Feature selection and preprocessing were performed using training data only.
Across five outer validation folds, 72 CpG sites were consistently selected in all five folds.
These CpGs were used to construct the final frozen classifier with:
- Fixed CpG feature order
- Fixed preprocessing/scaling parameters
- Fixed logistic-regression coefficients
- Fixed regularization settings
- Classification threshold of 0.50
External Validation
Independent validation was performed using GSE131013.
The external dataset contained 239 valid profiles:
- 95 colorectal tumor samples
- 96 adjacent-normal samples
- 48 healthy-mucosa samples
External validation performance:
- ROC-AUC: 0.9842
- Accuracy: 0.9707
- Balanced accuracy: 0.9703
- Sensitivity: 0.9684
- Specificity: 0.9722
- F1 score: 0.9634
- Brier score: 0.0531
For matched tumor–adjacent-normal pairs, 89 of 91 pairs showed the correct probability ordering.
Intended Use
This model is intended for:
- Biomedical machine-learning research
- Methodological evaluation
- Reproducibility studies
- Reliability assessment of high-dimensional prediction models
- Educational and research demonstrations
Important Limitation
This model is not intended for clinical diagnosis or direct patient-care decision making.
Performance may vary across populations, laboratories, methylation platforms, preprocessing pipelines, and independent cohorts.
Further prospective and multi-cohort validation would be required before any clinical application.
Reproducibility
This repository is intended to provide the frozen model and supporting files required to reproduce predictions using the predefined 72-CpG signature.
Raw GEO data are not redistributed in this repository.
Software Citation
Ackaah, E. K. (2026).
Frozen 72-CpG DNA Methylation Classifier for Colorectal Tumor Detection
Version 1.0.0.
Zenodo DOI: 10.5281/zenodo.21819138
License
MIT License