FinaleMe: Predicting DNA methylation by the fragmentation patterns of plasma cell-free DNA

Yaping Liu, Sarah C. Reed, Christopher Lo, Atish D. Choudhury, Heather A. Parsons, Daniel G. Stover, Gavin Ha, Gregory Gydush, Justin Rhoades, Denisse Rotem, Samuel S. Freeman, David Katz, Ravi Bandaru, Haizi Zheng, Hailu Fu, Viktor A. Adalsteinsson, Manolis KellisbioRxiv 2024 · DOI

TL;DR

Researchers have developed FinaleMe, a Hidden Markov Model (HMM) that predicts DNA methylation levels directly from standard whole-genome sequencing (WGS) of cell-free DNA. By analyzing non-random fragmentation patterns rather than using destructive chemical treatments, FinaleMe allows scientists to extract "hidden" epigenetic information and determine tissues-of-origin from millions of existing clinical datasets.

What Problem Does This Paper Address?

Cell-free DNA (cfDNA)—small fragments of DNA shed into the bloodstream by dying cells—serves as a powerful "liquid biopsy" for detecting cancer and monitoring fetal health. While the genetic sequence is useful, the epigenetic layer (specifically DNA methylation) is often more informative because it acts as a molecular "return address," revealing which tissue the DNA originated from.

The current gold standard for measuring methylation is Whole-Genome Bisulfite Sequencing (WGBS). However, bisulfite treatment is a "harsh" process that chemically degrades the majority of the DNA sample, requiring high input amounts and driving up costs. Meanwhile, millions of cfDNA samples are already sequenced using standard WGS (for non-invasive prenatal testing or cancer screening), but this standard method is traditionally considered "blind" to methylation. The authors sought to bridge this gap: can we infer methylation states simply by looking at how the DNA is broken?

Methods

The authors developed FinaleMe (FragmentatIoN AnaLysis of cEll-free DNA Methylation), a non-homogeneous Hidden Markov Model. The core premise is that cfDNA fragmentation is not random; it is dictated by nucleosome positioning, which is itself influenced by DNA methylation.

Key technical features include:

  • Non-homogeneous HMM: Unlike standard HMMs that assume constant transition probabilities, FinaleMe incorporates the physical distance between CpG sites to adjust the probability of switching between methylated and unmethylated states.
  • Three-Feature Input: The model looks at three specific signals for every CpG site:
  • Fragment length: Methylated DNA often results in different fragment sizes.
  • Normalized coverage: Local "depth" of sequencing.
  • Distance to center: The position of the CpG site relative to the center of the DNA fragment.
  • Bayesian Prior: The model utilizes a methylation prior derived from healthy genomic DNA to refine its predictions during the decoding (Viterbi) step.

Datasets

The study validated the model using a robust collection of paired data:

  • Matched High-Coverage Data: 80 pairs of deep WGS (~16-39X) and WGBS (~10-15X) from healthy donors and prostate cancer patients.
  • Ultra-Low-Pass (ULP) Data: Matched WGS and WGBS at very shallow depths (~0.1X) from 77 individuals, including healthy donors and patients with breast or prostate cancer.
  • Public Benchmarks: Existing WGBS datasets from healthy individuals were used for initial hypothesis testing and correlation analysis.

Key Insights and Ideas

  • Fragmentation as an Epigenetic Proxy: The most striking "aha moment" is the confirmation that cfDNA carries a "wavy" methylation pattern at mono-nucleosomal lengths (around 167bp). This pattern is absent in randomly sheared genomic DNA, proving that the biological process of cell death "imprints" epigenetic information into the physical structure of the fragments.
  • The Power of CpG-Rich Regions: The model is exceptionally accurate in CpG islands and promoters. Because these regions are critical for gene regulation and tissue identity, the model’s strength aligns perfectly with the most clinically relevant parts of the genome.
  • Democratizing Epigenetics: By enabling methylation inference from standard WGS, FinaleMe effectively turns "trash" data into "treasure." Researchers can now re-analyze massive existing repositories of WGS data to look for epigenetic biomarkers of disease without needing new samples or expensive bisulfite runs.

Key Results

  • High Predictive Accuracy: In CpG-rich regions, FinaleMe achieved an auROC of 0.91 for predicting the methylation status of single CpG sites.
  • Strong Correlation: When looking at 1kb genomic windows, the Pearson correlation between predicted methylation (from WGS) and ground truth (from WGBS) was 0.89.
  • Tissue-of-Origin Deconvolution: The model successfully estimated the fraction of tumor-derived DNA in plasma. These estimates were highly consistent with results obtained from specialized copy-number variation tools (like ichorCNA) and actual WGBS.
  • Robustness at Low Coverage: Remarkably, the model maintained its ability to distinguish cancer samples from healthy ones even at ultra-low-pass coverage (0.1X), which is the standard for cost-effective clinical screening.

Limitations and Future Work

While FinaleMe is a significant step forward, the authors note a few hurdles:

  • CpG-Poor Regions: The model’s performance drops significantly in regions of the genome with low CpG density. These areas are often "partially methylated domains" in cancer, which remain difficult to predict using this method.
  • Prior Bias: The reliance on a healthy genomic DNA prior could potentially mask some highly specific cancer-related methylation changes or lead to overfitting in certain contexts.
  • Future Directions: The team aims to improve predictions in CpG-poor regions and explore whether this fragmentation-based approach can be combined with other "fragmentomic" features (like end-motifs) to further increase sensitivity for early cancer detection.

Summary updated

Abstract

Analysis of DNA methylation in cell-free DNA (cfDNA) reveals clinically relevant biomarkers but requires specialized protocols and sufficient input material that limits its applicability. Millions of cfDNA samples have been profiled by genomic sequencing. To maximize the gene regulation information from the existing dataset, we developed FinaleMe, a non-homogeneous Hidden Markov Model (HMM), to predict DNA methylation of cfDNA and, therefore, tissues-of-origin directly from plasma whole-genome sequencing (WGS). We validated the performance with 80 pairs of deep and shallow-coverage WGS and whole-genome bisulfite sequencing (WGBS) data.
BibTeX
@misc{liu2024finaleme,
  author = {Yaping Liu and Sarah C. Reed and Christopher Lo and Atish D. Choudhury and Heather A. Parsons and Daniel G. Stover and Gavin Ha and Gregory Gydush and Justin Rhoades and Denisse Rotem and Samuel S. Freeman and David Katz and Ravi Bandaru and Haizi Zheng and Hailu Fu and Viktor A. Adalsteinsson and Manolis Kellis},
  title = {{FinaleMe: Predicting DNA methylation by the fragmentation patterns of plasma cell-free DNA}},
  year = {2024},
  doi = {10.1101/2024.01.02.573710},
  howpublished = {bioRxiv},
}
Figure