Improved deconvolution of circulating tumor DNA from ultra-low-pass whole-genome methylation sequencing using CelFiE-ISH

This study demonstrates that the CelFiE-ISH framework, when applied to ultra-low-pass whole-genome methylation sequencing from Oxford Nanopore data, can detect circulating tumor DNA in epithelial cancers at fractions as low as 1.7–3.1% by utilizing a refined set of epithelial-specific methylation markers.

The Gist

Imagine your bloodstream is a vast, busy ocean. Normally, this ocean is filled with tiny boats representing healthy cells from your body (like white blood cells, red blood cells, and liver cells). However, if you have cancer, a few "rogue" boats from the tumor also drift into the ocean. These are called circulating tumor DNA (ctDNA).

The goal of this research is to find those rogue boats in a sea of millions of healthy ones, even when the cancer is small and the rogue boats are very few.

Here is a simple breakdown of what the scientists did and why it matters, using some everyday analogies:

1. The Challenge: Finding a Needle in a Haystack

The researchers used a special microscope called Oxford Nanopore sequencing to look at the DNA in the blood. This microscope is unique because it can read a "chemical sticker" on the DNA called methylation. Think of methylation like a name tag or a uniform that tells you exactly what kind of cell the DNA came from (e.g., "I am from the colon," "I am from the lung," or "I am from the blood").

The problem? In early cancer, the "rogue" name tags are extremely rare. If you have 100,000 healthy name tags, you might only have 1 or 2 cancer ones. Previous methods often got confused, thinking healthy cells were cancer cells, or they simply missed the cancer entirely.

2. The First Fix: "Clipping" the Noise

The team used a computer program called CelFiE-ISH to sort these name tags. However, they noticed the computer was sometimes "hallucinating." When it saw a blurry or confusing name tag, it would guess, "Maybe this is a cancer cell?" and count it, even if it wasn't. This made healthy people look like they had cancer.

The Solution: They added a rule called "Clipping."

• The Analogy: Imagine a bouncer at a club. If a person's ID is slightly blurry and the bouncer isn't 100% sure they are a VIP, the bouncer used to let them in anyway. Now, the rule is: "If you aren't 95% sure, cut them off (clip them) and don't let them in."
• The Result: This stopped the computer from falsely accusing healthy people of having cancer. It cleaned up the data significantly.

3. The Big Breakthrough: Grouping the "Siblings"

The next problem was that the computer was trying to be too specific. It was trying to distinguish between a "colon cell from the left side" and a "colon cell from the right side." But with so few cancer DNA fragments, the computer didn't have enough clues to tell them apart, so it got confused.

The Solution: They decided to stop looking at individual cell types and instead look at families.

• The Analogy: Instead of trying to identify exactly which specific sibling (e.g., "Uncle Bob from Ohio") is in the room, they just asked, "Is there anyone from the Smith Family in the room?"
• They grouped all "epithelial" cells (the type of cells that line our organs like the colon, lung, and breast) into one big "Pan-Epithelial" family.
• Why it worked: Even if the computer couldn't tell which organ the cancer came from, it could easily tell, "Hey, this DNA definitely belongs to the Epithelial Family, and healthy blood doesn't have many of those!" This made it much easier to spot the cancer.

4. The Results: Seeing the Invisible

By using this "Family Grouping" method and the "Clipping" rule, the researchers achieved two major things:

1. Lower Detection Limit: They could now detect cancer when it made up as little as 1.7% to 3% of the total DNA in the blood. Before, the limit was usually around 3%. This is a big deal because it means they can catch cancer earlier.
2. Accuracy: They tested this on patients with colon, breast, lung, and pancreatic cancer. The method worked well, especially for advanced cancers, and was much better at ignoring healthy people who didn't have cancer.

Why This Matters

Think of this like upgrading a metal detector at an airport.

• Old Metal Detector: It beeped at everything (coins, belt buckles, and guns), causing lots of false alarms and missing small, hidden guns.
• New Metal Detector (This Study): It has a smarter filter. It ignores the belt buckles (healthy cells) and focuses only on the shape of a gun (cancer cells). It can even spot a tiny, hidden gun that the old detector would have missed.

The Bottom Line:
This study shows that by being smarter about how we group the DNA clues (looking for families rather than individuals) and being stricter about what we count as a "match," we can find cancer in the blood much earlier and more accurately. This brings us one step closer to a simple blood test that can catch cancer before it becomes a big problem.

Technical Summary

1. Problem Statement

• Challenge in Liquid Biopsy: Detecting circulating tumor DNA (ctDNA) in blood plasma is critical for cancer management, but the tumor fraction is often minute (typically <5% at diagnosis).
• Limitations of Current Methods:
  • Copy Number Alteration (CNA): Standard tools like ichorCNA have a detection floor of approximately 3% tumor fraction. They also fail to detect cancers with few or no somatic CNAs.
  • Methylation Deconvolution: While DNA methylation is a highly informative biomarker, existing deconvolution algorithms (e.g., CelFiE, CelFiE-ISH, UXM) struggle at Ultra-Low-Pass Whole-Genome Sequencing (ULP-WGS) depths (~0.25x coverage).
  • Specific Issues:
    1. Marker Sparsity: At low coverage, many informative markers lack even a single tumor-derived read.
    2. Over-inflation: Algorithms tend to over-estimate the fraction of rare or absent cell types (e.g., epithelial cells in healthy controls) due to low-confidence read assignments.
    3. Granularity vs. Sensitivity: Using highly specific markers for individual cell types (e.g., distinguishing lung alveolar from bronchial cells) reduces sensitivity in low-coverage data compared to using broader lineage markers.

2. Methodology

The study utilized Oxford Nanopore Technologies (ONT) sequencing, which natively profiles DNA methylation without chemical conversion, preserving DNA integrity.

• Cohort:
  • 20 advanced colorectal cancer (CRC) patients.
  • 22 non-cancer healthy controls.
  • Additional validation cohorts: Breast (n=66), Lung (n=6), and Pancreatic (n=10) cancers.
• Sequencing: ULP-WGS at ~0.22x–0.3x coverage (median ~3.2M fragments).
• Computational Framework:
  • Deconvolution Algorithms: Tested UXM, CelFiE, CelFEER, and CelFiE-ISH (a fragment-level, haplotype-aware model).
  • Reference Atlas: Used a reference of 31 healthy cell types derived from whole-genome bisulfite sequencing (WGBS).
• Key Technical Innovations:
  1. Probabilistic Clipping: Modified CelFiE-ISH to set any cell-type assignment probability <0.05 to zero. This filters out "noise" from low-confidence reads that cause false positives in healthy controls.
  2. Pan-Epithelial Marker Strategy: Instead of treating 31 distinct cell types separately, the authors "collapsed" cell types into 7–8 broad lineages (e.g., grouping all epithelial subtypes into a single Pan-Epithelial group).
  3. Marker Expansion: Increased the number of markers per lineage from the standard 250 to 1,000 and finally to a "Max" set of 2,825 markers per lineage to maximize feature representation in low-coverage data.
  4. In Silico Dilution: To simulate early-stage cancer (low tumor fractions), CRC samples were computationally diluted with healthy plasma reads to create tumor fractions ranging from 100% down to ~1.7%.

3. Key Results

A. Impact of Probabilistic Clipping

• Noise Reduction: The default CelFiE-ISH model over-estimated epithelial fractions in healthy controls. Applying the clipping strategy (threshold <0.05) reduced epithelial estimates in healthy controls to near-zero levels while maintaining accuracy in cancer samples.
• Accuracy: Clipped CelFiE-ISH achieved a lower Root Mean Squared Error (RMSE = 0.099) compared to the default version (RMSE = 0.126) when compared against ichorCNA tumor fractions.

B. Impact of Pan-Epithelial Markers

• Superior Sensitivity: Switching from 31 specific cell types to collapsed pan-epithelial lineages significantly improved performance.
  • RMSE Improvement: Median RMSE dropped from 0.126 (31-cell model) to 0.082 (pan-epithelial model).
  • Correlation: $R^2$ improved from 0.684 to 0.810.
• Marker Quantity: Increasing markers from 250 to 1,000+ further reduced error rates, particularly in low-tumor-fraction scenarios. The "Max markers" set (2,825) yielded the best results.

C. Limit of Detection (LoD)

• Performance at Low Fractions: Using the clipped CelFiE-ISH with pan-epithelial markers and max markers:
  • 1.7% – 3.1% Tumor Fraction: Achieved good classification (AUROC 0.85–0.95).
  • >3.1% Tumor Fraction: Achieved near-perfect separation (AUROC 0.97–0.99).
• Comparison: This performance matches or exceeds the ~3% LoD of established CNA-based tools like ichorCNA, but with the added benefit of detecting cancers lacking CNAs.

D. Validation Across Cancer Types

• The optimized pipeline successfully detected ctDNA in Breast, Lung, and Pancreatic cancers.
• Stage IV cancers were readily detectable across all types.
• Early-stage (Stage I/II) breast cancers were harder to detect, consistent with lower ctDNA shedding, but the method showed stage-dependent stratification.

4. Key Contributions

1. Algorithmic Refinement: Introduced a "clipping" mechanism to CelFiE-ISH to eliminate false-positive cell type assignments in low-coverage data, significantly improving specificity in healthy controls.
2. Strategic Marker Selection: Demonstrated that lineage-level (pan-epithelial) markers outperform highly specific cell-type markers in ULP-WGS contexts, where read sparsity prevents fine-grained discrimination.
3. Scalability: Showed that increasing marker density (up to ~2,800 markers per lineage) is critical for lowering the detection limit in ultra-low-pass sequencing.
4. Platform Validation: Validated that ONT-based methylation sequencing can achieve CNA-level sensitivity (~3% LoD) while providing orthogonal methylation data for molecular subtyping.

5. Significance and Future Outlook

• Clinical Utility: This approach provides a robust method for monitoring tumor burden in patients with low ctDNA fractions, potentially bridging the gap between advanced and early-stage detection.
• Complementary to CNA: Since methylation and CNA features are derived from the same sequencing run, they offer two independent signals. This is crucial for cancers with low CNA burden (e.g., some breast or prostate cancers) where CNA-only methods fail.
• Molecular Monitoring: Unlike CNA, methylation profiling can detect molecular subtype switches (e.g., ER+ to ER- in breast cancer, or NSCLC to SCLC transformation), which are drivers of therapeutic resistance.
• Future Directions: The authors suggest that future iterations could integrate copy-number data directly into the deconvolution model or utilize 5-hydroxymethylation (5hmC) to further push the detection limits of ULP-WGS.

Conclusion: The study establishes that by combining probabilistic clipping, lineage-specific marker aggregation, and high-density marker sets, CelFiE-ISH can effectively deconvolve ctDNA from ultra-low-pass ONT sequencing, achieving a detection limit of 1.7–3.1% and offering a powerful tool for cancer management and monitoring.