Harnessing methylation signals inherent in long-read sequencing data for improved variant phasing

The authors present LongHap, a novel read-based phasing method that leverages native methylation signals from long-read sequencing data to significantly improve haplotype reconstruction accuracy and contiguity compared to existing tools.

The Gist

Imagine your genome is a massive, two-volume encyclopedia set. One volume is from your mother, and the other is from your father. For decades, scientists have been able to read the letters (the DNA sequence) in these books, but they've struggled to figure out which sentence belongs to which volume. This process is called phasing. Without it, it's like having a shredded book where you know the words, but you don't know which page they came from, making it hard to understand the full story.

Until now, scientists had to use expensive, complex tricks or rely on short, choppy snippets of text to guess the order. But a new tool called LongHap, developed by researchers at Princeton University, changes the game by using a hidden clue that was already there all along: methylation.

Here is how LongHap works, explained through simple analogies:

1. The Problem: The "Shredded" Books

Think of long-read sequencing (like PacBio and Oxford Nanopore) as a high-tech shredder that spits out very long strips of paper. These strips are great because they contain long sentences. However, sometimes the strips are so long and complex that the computer gets confused about where one sentence ends and another begins, or which "volume" (mom's or dad's) a specific word belongs to.

Existing tools try to solve this by looking only at the letters (A, C, T, G). If the letters are ambiguous, the chain of logic breaks, and the "phase block" (the continuous story) stops.

2. The Secret Weapon: The "Highlighter" (Methylation)

Here is the twist: Long-read machines don't just read the letters; they can also see if the letters have been "highlighted" with a chemical marker called methylation (specifically 5-methylcytosine).

Think of methylation like a highlighter pen that your cells use to mark certain pages.

• Sometimes, Mom's volume has a page highlighted in yellow, while Dad's volume has that same page left blank.
• Sometimes, the pattern is the opposite.

For a long time, scientists ignored these highlights because they were focused only on the text. LongHap is the first tool to say, "Wait a minute! Let's use the highlights to figure out which volume is which!"

3. How LongHap Works: The Detective Story

LongHap acts like a super-smart detective solving a mystery in three steps:

• Step 1: The First Guess (The Sequence)
  The tool first tries to connect the dots using just the letters. It builds a "skeleton" of the story. But sometimes, there are gaps where the letters don't give a clear answer.

• Step 2: The "Belief Propagation" (The Logic Puzzle)
  When the tool finds a tricky spot (like a complex genetic mutation), it doesn't just give up. It uses a mathematical trick called "belief propagation." Imagine you are solving a Sudoku puzzle. If you can't figure out one number, you look at the numbers surrounding it and use logic to deduce what must be there. LongHap does this for DNA, embedding difficult spots into the bigger picture to see how they fit.

• Step 3: The "Highlighter" Check (The Methylation)
  This is the magic step. If the tool still has a gap between two story segments, it looks at the "highlighter" marks.

  • Scenario: "I have a strip of paper. I can't tell if it belongs to Mom or Dad based on the letters. BUT, I see a highlight on this strip. I know from other parts of the book that Mom's pages are usually highlighted here, and Dad's are not. Therefore, this strip belongs to Mom!"
  • By using these methylation patterns, LongHap can bridge gaps that were previously impossible to cross, stitching together much longer, more accurate stories.

4. Why This Matters: The "Medical Map"

Why does this matter to you?

• Better Medicine: Many diseases are caused by specific combinations of genes on one side (either Mom's or Dad's). If you can't phase the genes correctly, you might miss a dangerous mutation. LongHap helps doctors see the full picture, especially in tricky, medically important genes (like the LIX1 gene mentioned in the paper) that were previously too hard to read.
• Efficiency: It's fast and doesn't require expensive new equipment. It just uses the data the machines are already producing.
• Accuracy: In tests, LongHap made fewer mistakes and created longer, more continuous "stories" of our DNA than any other tool currently available.

The Bottom Line

Think of LongHap as a new pair of glasses for geneticists. Before, they could read the text of our DNA, but the story was often fragmented. Now, by using the "highlighter marks" (methylation) that were hiding in plain sight, LongHap allows us to read the complete, unbroken story of our genetic inheritance, leading to better understanding of human history, evolution, and disease.

Technical Summary

1. Problem Statement

Accurate haplotype phasing (determining which alleles co-occur on the same chromosome) is critical for population genetics, clinical variant interpretation, and understanding complex traits. While long-read sequencing technologies (PacBio HiFi and Oxford Nanopore Technologies/ONT) have significantly improved the continuity of genome assembly, reconstructing chromosome-scale haplotypes remains challenging.

• Limitations of Current Tools: Existing read-based phasing tools (e.g., WhatsHap, HapCUT2, LongPhase) rely primarily on DNA sequence co-occurrence. They often struggle with:
  • Complex Variants: Accurately phasing insertions, deletions (INDELs), and structural variants (SVs) due to alignment ambiguities.
  • Phase Block Fragmentation: Gaps in coverage or low variant density break phase blocks, preventing chromosome-scale reconstruction.
  • Untapped Data: Long-read platforms natively detect 5-methylcytosine (5mC) epigenetic modifications, but current phasing algorithms do not integrate this information.
  • Existing Methylation Tools: The only existing tool, MethPhaser, operates post hoc (refining haplotypes after initial phasing) rather than integrating methylation into a unified probabilistic framework.

2. Methodology: LongHap

The authors developed LongHap, a read-based phasing method that seamlessly integrates DNA sequence and native methylation signals into a unified framework. The workflow consists of four main stages:

A. Input and Allelic Support Determination

• Inputs: Aligned reads (BAM), variant calls (VCF), and optional unphased methylation calls.
• Allele Inference: LongHap streams through BAM files to determine which allele (A or B) each read supports at heterozygous sites.
  • It handles "multiallelic" sites where both haplotypes differ from the reference.
  • Optimization: It first attempts to infer alleles via CIGAR strings and matching subsequences. If ambiguous, it performs local realignment against synthetic references (embedding alleles in surrounding sequence).
  • Dynamic Penalties: Gap open/extension penalties are dynamically adjusted based on sequencing technology (PacBio vs. ONT), variant type, and local sequence context (e.g., homopolymers) to account for specific error profiles.

B. Directed Acyclic Graph (DAG) Construction

• A DAG is constructed where vertices represent alleles at heterozygous sites, and edges represent transitions between adjacent sites.
• Edge Weights: Calculated based on the frequency of allele co-occurrence in reads (transition probabilities).
• Strand Bias Filtering: For ONT data, variants supported only by one DNA strand are excluded to reduce false positives.

C. Belief Propagation for Complex Variants

• Problem: Standard DAGs fail when variants (like SVs) have low read support or ambiguous alignments.
• Solution: LongHap employs Loopy Belief Propagation (LBP).
  • It constructs a subgraph embedding difficult-to-phase variants within a broader context of confidently phased flanking variants.
  • It introduces long-range edges connecting non-adjacent variants to propagate information across the difficult region.
  • This allows the algorithm to infer marginal probabilities for complex variants that would otherwise break the phase block.

D. Methylation-Based Gap Bridging

• Identification: LongHap dynamically identifies differentially methylated sites (DMS) where methylation patterns differ significantly between the two haplotypes (e.g., one haplotype is methylated, the other is not).
• Iterative Assignment:
  1. Reads are anchored based on sequence evidence.
  2. If a gap exists between phase blocks, LongHap checks if reads overlapping the gap show consistent methylation patterns with one of the flanking haplotypes.
  3. Reads are iteratively assigned to haplotypes based on these methylation signals, effectively "bridging" the gap.
  4. This process repeats until no more reads can be assigned or the gap is closed.

E. Final Phasing

• A Viterbi-like decoding scheme is used to determine the most likely haplotype path through the DAG, outputting a phased VCF with phase block coordinates and haplotagged reads.

3. Key Contributions

1. Unified Framework: First tool to integrate sequence and methylation data simultaneously during the phasing process, rather than as a post-processing step.
2. Advanced Handling of Complex Variants: Uses Loopy Belief Propagation to embed INDELs and SVs into the haplotype context, improving accuracy for structural variants.
3. Dynamic Methylation Utilization: Identifies phase-informative methylation sites on-the-fly to bridge gaps that sequence data alone cannot resolve.
4. Open Source: The tool is publicly available via GitHub.

4. Results

The authors benchmarked LongHap against WhatsHap, HapCUT2, LongPhase, and WhatsHap + MethPhaser using HG002 data (PacBio Revio HiFi, ONT Simplex, and UL-ONT Duplex).

• Switch Error Rates:
  • PacBio HiFi: LongHap achieved the lowest switch error rates (0.196% with methylation), outperforming LongPhase (0.192%) and WhatsHap (0.246%). It reduced errors by ~3.3% compared to WhatsHap + MethPhaser.
  • ONT Data: LongHap maintained lower error rates than WhatsHap and HapCUT2, though LongPhase had slightly lower errors (at the cost of phasing fewer sites).
• Phase Block Contiguity (N50):
  • Integrating methylation increased LongHap's phase block N50 by 31.9% (from 443 kb to 584 kb) on PacBio data.
  • On Ultra-Long ONT data, LongHap achieved chromosome-scale phasing with an average N50 of 80.7 Mb, significantly outperforming WhatsHap (5.82 Mb).
• Complex Variants: LongHap phased a higher fraction of INDELs and SVs (94–95%) compared to LongPhase (70–75%) while maintaining superior accuracy.
• Medically Relevant Genes (CMRGs):
  • LongHap fully phased the most challenging medically relevant genes (211 out of 273) when using methylation data.
  • Case Study (LIX1): The tool successfully phased the LIX1 gene, a region previously unphased by sequence data alone, by leveraging methylation signals to bridge a coverage gap.
• Computational Performance:
  • LongHap is faster than WhatsHap, HapCUT2, and MethPhaser.
  • It is slower than LongPhase but offers significantly better contiguity and variant coverage.
  • Memory usage is modest (1–9 GB depending on methylation integration).

5. Significance

• Maximizing Data Utility: LongHap demonstrates that modern long-read sequencing data contains rich epigenetic information that, when integrated, significantly enhances genetic analysis.
• Clinical Impact: By improving phasing in difficult, medically relevant regions (e.g., LIX1), LongHap facilitates more accurate interpretation of compound heterozygosity and disease mechanisms.
• Scalability: The method is computationally efficient enough for population-scale studies, offering a flexible framework that can potentially be extended to other epigenetic signals.
• Paradigm Shift: It moves beyond simply extending short-read phasing paradigms to long-reads, instead leveraging the unique multi-modal capabilities (sequence + methylation) of third-generation sequencing.