Related Experiment Video
Updated: Sep 20, 2025
13:21
Comprehensive DNA Methylation Analysis Using a Methyl-CpG-binding Domain Capture-based Method in Chronic Lymphocytic Leukemia Patients
Published on: June 16, 2017
10.2K
Comprehensive study of semi-supervised learning for DNA methylation-based supervised classification of central
Quynh T Tran1, Md Zahangir Alom1, Brent A Orr2
1Department of Pathology, St. Jude Children's Research Hospital, 262 Danny Thomas Place, MS 250, Memphis, TN, 38105-3678, USA.
BMC Bioinformatics
|June 8, 2022
Summary
Semi-supervised learning (SSL) with DNA methylation data significantly improves central nervous system (CNS) tumor classification accuracy. This approach leverages unlabeled data to enhance supervised models, especially for rare tumor types, boosting diagnostic precision.
Area of Science:
- Computational biology and bioinformatics
- Oncology and cancer research
- Machine learning in medicine
Background:
- Accurate pathological diagnosis is crucial for precision cancer medicine.
- Increasing tumor classes and limitations of traditional histopathology necessitate advanced methods.
- Supervised machine learning models using DNA methylation data require extensive labeled datasets for tumor classification.
Purpose of the Study:
- To explore the application of semi-supervised learning (SSL) on DNA methylation data for central nervous system (CNS) tumor classification.
- To improve the accuracy of supervised learning models by leveraging both labeled and unlabeled epigenetic data.
- To address the challenge of limited labeled data, particularly for rare tumor types.
Main Methods:
- Comprehensive evaluation of 11 SSL methods applied to DNA methylation data.
- Development of a novel combination approach using a self-training with editing support vector machine (SETRED-SVM) and L2-penalized multinomial logistic regression.
- Generation of high-confidence pseudo-labels from limited labeled instances for training machine learning models.
Main Results:
- The proposed SSL method significantly increased prediction accuracy across eight random forest and neural network models.
- Pseudo-labels generated by the SSL approach enhanced classification for 82 CNS tumor types and 9 normal controls.
- Demonstrated a substantial improvement in classification accuracy compared to traditional supervised methods alone.
Conclusions:
- The combination of SSL techniques and multinomial logistic regression effectively utilizes abundant unlabeled methylation data.
- This approach provides additional training examples, crucial for improving the prediction accuracy of supervised models, especially for rare tumor types.
- Offers a promising strategy to enhance the standardization and reduce costs in histopathological diagnosis.
