Related Experiment Video
Updated: Sep 19, 2025
07:50
Genome-Wide Analysis of DNA Methylation in Gastrointestinal Cancer
Published on: September 18, 2020
5.7K
crossNN is an explainable framework for cross-platform DNA methylation-based classification of tumors
Dongsheng Yuan1,2, Robin Jugas3, Petra Pokorna3
1Department of Experimental Neurology, Charité - Universitätsmedizin Berlin, corporate member of Freie Universität Berlin and Humboldt-Universität zu Berlin, Berlin, Germany.
Nature Cancer
|June 6, 2025
Summary
A new machine learning framework, crossNN, accurately classifies over 170 tumor types using DNA methylation data from various platforms. This approach enhances diagnostic accuracy and computational efficiency for cancer classification.
Area of Science:
- Oncology
- Genomics
- Bioinformatics
- Machine Learning
Background:
- DNA methylation-based tumor classification is crucial for diagnostics.
- Current methods using fixed feature spaces lack cross-platform compatibility, especially with diverse DNA sequencing technologies.
- This limits the adaptability and scalability of existing tumor classifiers.
Purpose of the Study:
- To introduce crossNN, a novel neural network framework for tumor classification.
- To enable accurate classification using sparse methylomes across different platforms and sequencing depths.
- To develop a robust and scalable pan-cancer classifier.
Main Methods:
- Developed crossNN, a neural network-based machine learning framework.
- Trained crossNN on sparse methylome data from various platforms, including nanopore and targeted bisulfite sequencing.
- Compared crossNN performance against other deep and conventional machine learning models.
Main Results:
- crossNN accurately classifies tumors using sparse methylomes, outperforming existing models in accuracy and computational efficiency.
- A pan-cancer classifier trained with crossNN can discriminate over 170 tumor types across all organ sites.
- Validation on over 5,000 tumors showed high precision: 99.1% for brain tumors and 97.8% for the pan-cancer model.
Conclusions:
- crossNN offers a robust, scalable, and explainable solution for DNA methylation-based tumor classification.
- The framework overcomes platform limitations, enabling accurate classification across diverse epigenome data.
- This advance has significant implications for improving cancer diagnostics and research.
