DNA methylation-based epigenetic signatures predict somatic genomic alterations in gliomas

Jie Yang, Qianghu Wang, Ze-Yan Zhang, Lihong Long, Ravesanker Ezhilarasan, Jerome M. Karp, Aristotelis Tsirigos, Matija Snuderl, Benedikt Wiestler, Wolfgang Wick, Yinsen Miao, Jason T. Huse, Erik P. SulmanNature Communications 2022 · DOI

TL;DR

Researchers have developed the Unified Diagnostic Pipeline (UniD), a machine learning framework that uses DNA methylation data to accurately predict key genetic mutations and gene expression subtypes in gliomas. By replacing a battery of expensive, slow tests with a single epigenetic assay, UniD offers a faster and more cost-effective path to precision neuro-oncology.

What Problem Does This Paper Address?

Classifying gliomas (the most common primary brain tumors) is no longer just about looking under a microscope. Modern treatment depends on identifying specific "somatic alterations"—genetic mutations like IDH, TERTp, and ATRX, or chromosomal deletions like 1p/19q.

Currently, clinicians must run a "patchwork" of different tests to find these: Next-Generation Sequencing (NGS) for mutations, FISH for chromosome status, and IHC for protein expression. This process is:

  1. Slow and Costly: Multiple assays take weeks and thousands of dollars.
  2. Tissue-Hungry: Each test consumes precious tumor samples.
  3. Technically Fragile: Clinical samples are often preserved in paraffin (FFPE), which degrades RNA, making gene expression subtyping—a key predictor of treatment response—nearly impossible in a standard clinical setting.

Methods

The authors developed UniD, a supervised machine learning pipeline. Instead of looking at the genome (the blueprint), they look at the epigenome (the "switches" that turn genes on or off via DNA methylation).

  • The Platform: They used the Illumina Infinium Methylation array, a standard tool that measures methylation at hundreds of thousands of sites across the genome.
  • The Models: They trained Random Forest and Elastic Net models to recognize the specific epigenetic "fingerprints" left behind by genetic mutations.
  • Innovation: Unlike previous "unsupervised" methods that group tumors by general similarity, UniD is supervised. It is specifically trained to answer the exact questions clinicians ask: "Is IDH mutated?" or "Is 1p/19q deleted?"

Datasets

The study utilized a robust set of data for training and rigorous independent validation:

  • Training/Discovery: Data from The Cancer Genome Atlas (TCGA), including 644 glioma samples (both Glioblastoma and Low-Grade Glioma).
  • External Validation: The NOA-04 cohort, an independent dataset from a multicenter phase III randomized trial. This is a high bar for any diagnostic model, as it tests how the algorithm performs on "real-world" data from different institutions.

Key Insights and Ideas

  • The Epigenetic-Genetic Link: The most profound "aha moment" is the confirmation that specific genetic mutations leave distinct, non-overlapping marks on the methylome. A mutation in ATRX creates a different epigenetic signature than a mutation in TERTp, even though both affect telomere maintenance.
  • Functional Superiority: In several cases, UniD predicted an ATRX "mutant" signature even when DNA sequencing called it "wildtype." Further analysis showed these tumors had low ATRX expression, suggesting the epigenetic signature might be a more accurate reflection of the tumor's actual biology than sequencing alone.
  • RNA Proxy: Because DNA is much more stable than RNA in clinical samples, UniD’s ability to predict gene expression subtypes using DNA methylation is a major breakthrough for clinical feasibility.

Key Results

UniD demonstrated remarkable accuracy across the board in the test sets:

  • IDH Mutation: 100% accuracy (AUC = 1.0).
  • TERTp Mutation: 98.31% accuracy (AUC = 1.0).
  • 1p/19q Co-deletion: 99.21% accuracy (AUC = 0.997).
  • ATRX Mutation: 90.48% accuracy (AUC = 0.995).
  • Gene Expression Subtyping: 72.2% accuracy. While lower than the binary mutations, this is significant because no other clinical-grade assay currently exists for this classification.

The pipeline proved to be faster and required significantly less tissue than the traditional multi-assay approach.

Limitations and Future Work

While UniD is a major step forward, the authors note a few areas for growth:

  • Rare Subsets: Some specific "discordant" cases (where methylation and sequencing disagreed) had small sample sizes, making it hard to draw definitive conclusions about why they differed.
  • Expansion: The current model is optimized for gliomas. The authors suggest this approach could be applied to other malignancies where genetic testing is currently fragmented.
  • Clinical Integration: While the validation on the NOA-04 cohort is strong, moving this into routine frontline clinical practice will require prospective studies to prove it speeds up the "time-to-treatment" in a hospital setting.

Summary updated

Abstract

Molecular classification has improved diagnosis and treatment for patients with malignant gliomas. However, classification has relied on individual assays that are both costly and slow, leading to frequent delays in treatment. Here, we propose the use of DNA methylation, as an emerging clinical diagnostic platform, to classify gliomas based on major genomic alterations and provide insight into subtype characteristics. We show that using machine learning models, DNA methylation signatures can accurately predict somatic alterations and show improvement over existing classifiers. The established Unified Diagnostic Pipeline (UniD) we develop is rapid and cost-effective for genomic alterations and gene expression subtypes diagnostic at early clinical phase and improves over individual assays currently in clinical use. The significant relationship between genetic alteration and epigenetic signature indicates broad applicability of our approach to other malignancies.
BibTeX
@inproceedings{yang2022dna,
  author = {Jie Yang and Qianghu Wang and Ze-Yan Zhang and Lihong Long and Ravesanker Ezhilarasan and Jerome M. Karp and Aristotelis Tsirigos and Matija Snuderl and Benedikt Wiestler and Wolfgang Wick and Yinsen Miao and Jason T. Huse and Erik P. Sulman},
  title = {{DNA methylation-based epigenetic signatures predict somatic genomic alterations in gliomas}},
  booktitle = {Nature Communications},
  year = {2022},
  doi = {10.1038/s41467-022-31827-x},
}
Figure