A computational model for quantifying instability of tandem repeats across the genome

This paper introduces a general-purpose computational model that leverages long-read sequencing data to accurately quantify genome-wide tandem repeat instability by characterizing read-to-consensus deviations, revealing that instability is primarily driven by repeat composition rather than length and enabling the detection of significant mosaicism in pathogenic expansions.

The Gist

The Big Picture: The "Fuzzy" Parts of Your DNA

Imagine your genome (your body's instruction manual) is a massive library. Most of the books in this library are written in perfect, clear sentences. But, there are some pages that look like a broken record stuck on repeat: "CAG-CAG-CAG-CAG..." or "GAA-GAA-GAA..."

These are called Tandem Repeats (TRs). They are like a chorus line of dancers doing the exact same move over and over.

The Problem:
Sometimes, these dancers get tired or confused. Instead of doing the exact same move, one might stumble, another might skip a step, or a third might do a slightly different dance move. In biology, this is called mosaicism or instability.

When these repeats get too long or too messy, they can cause serious diseases (like Huntington's disease or Fragile X syndrome). Scientists want to know: How "wobbly" or unstable are these repeat sections? If a section is extremely unstable, it's a red flag for disease.

The Challenge: Distinguishing "Real" Chaos from "Camera" Glitches

Until now, measuring this instability was like trying to film a chaotic dance floor with a shaky, low-quality camera.

• The Real Issue: The dancers are actually messing up (biological instability).
• The Fake Issue: The camera is jittery, or the lighting is bad (technical noise from the sequencing machine).

It was very hard to tell the difference. If you saw a messy dance, you didn't know if the dancer was sick or if your camera was just broken.

The Solution: A New "Instability Meter"

The authors of this paper built a new computational tool (a "meter") that works with long-read sequencing (a high-definition camera that can see the whole dance floor at once).

Here is how their new method works, step-by-step:

1. The "Consensus" Photo

First, the tool looks at all the photos (reads) of a specific repeat section and creates a "perfect average" photo. Let's call this the Consensus. It's like taking a group photo of the dancers and blurring them together to find the "ideal" pose.

2. Measuring the "Stumble"

Next, it compares every single individual photo to that perfect average.

• If a photo matches the average perfectly, the "stumble score" is zero.
• If a photo has a dancer who skipped a step or added an extra move, the tool measures exactly how different it is. This is the Divergence Rate.

3. Building a "Baseline" for Each Dance

The tool doesn't just look at one person; it looks at 256 different people (samples) and 617,000 different repeat sections.

• It realizes that some dances are naturally wobbly (like a complex jazz routine), while others are naturally steady (like a simple march).
• It builds a custom rulebook (a model) for each specific repeat location. It learns: "Okay, for the 'CAG' repeat at location X, it's normal to see a little bit of stumbling. But for the 'GAA' repeat at location Y, everyone should be perfect."

4. The "Outlier" Alarm

Now, when the tool looks at a new patient, it checks their repeat section against the custom rulebook.

• Normal: "This person's repeat is wobbling just as much as everyone else's. No alarm."
• Unstable: "Whoa! This person's repeat is wobbling way more than the rulebook says is normal. This is an outlier!"

What They Found

Using this new "Instability Meter" on a huge dataset of healthy and diseased cells, they discovered three big things:

1. Most repeats are actually quite stable. The "dancers" usually keep their rhythm.
2. Length isn't the main problem. You might think a longer line of dancers is more likely to trip, but the study found that purity matters more. If the line is perfectly identical (pure), it's more likely to get messy. If the line has some interruptions (like a different dancer stepping in), it actually stays more stable.
3. Sick people have "super-wobbly" repeats. When they looked at people with known genetic diseases, the disease-causing repeats were significantly more unstable than the rulebook predicted.

Why This Matters

Think of this tool as a smoke detector for your DNA.

Previously, we could only see if the house was on fire (the disease was already present). Now, this tool can detect the smoke (instability) before the fire gets out of control.

• For Doctors: It helps identify which patients are at high risk for diseases like Huntington's, even before symptoms appear.
• For Researchers: It gives them a standard way to measure how "unstable" a specific part of the genome is, helping them understand why some people get sick and others don't.

In short: They built a smart calculator that knows what "normal" looks like for every single repeat in your DNA, so it can instantly spot the ones that are going haywire and causing trouble.

Technical Summary

1. Problem Statement

Tandem repeats (TRs) are highly mutable genomic regions where somatic mosaicism (cell-to-cell variation in repeat length) is a critical factor in the pathogenesis and progression of repeat expansion disorders (e.g., Huntington's disease). While long-read sequencing (specifically PacBio HiFi) can resolve full-length TR alleles that short-read sequencing cannot, there is a lack of robust computational frameworks to:

• Quantify baseline instability across the genome for both simple and structurally complex loci.
• Distinguish between true biological mosaicism and technical noise inherent in sequencing data.
• Identify specific alleles with "unusually high" instability relative to their locus-specific baseline, which could serve as biomarkers for pathogenicity.

2. Methodology

The authors developed a general-purpose probabilistic model, implemented in the TRGT-instability tool, to quantify TR instability without explicitly separating biological mosaicism from technical noise. The workflow consists of the following steps:

• Data Input & Genotyping:

  • Utilized PacBio HiFi whole-genome sequencing (WGS) data from 256 Human Pangenome Reference Consortium (HPRC) cell line samples.
  • Used the TRGT (Tandem Repeat Genotyping Tool) to generate full-length consensus sequences for every TR allele and identify supporting reads.
  • Analyzed 617,007 TR loci derived from the adotto TR catalog v2.0 lite.

• Divergence Rate Calculation:

  • For each read supporting an allele, the edit distance between the read sequence and the allele's consensus sequence is calculated.
  • This is normalized by the maximum length of the repeat sequence in the read or consensus to produce a divergence rate (ranging from 0 to 1).

• Instability Profile Generation:

  • Divergence rates for reads supporting a specific allele are binned into 15 quantile-defined bins (based on empirical cumulative distribution function quantiles).
  • This creates an instability profile (a count vector) representing the distribution of read-to-consensus deviations for that allele.

• Locus-Level Modeling (Dirichlet-Multinomial):

  • A Dirichlet-Multinomial (DM) distribution is fitted to the instability profiles of all observed alleles at a specific locus across the cohort.
  • Multinomial component: Models variation in read depths.
  • Dirichlet component: Captures overdispersion (variability) in instability profiles across alleles.
  • Refinement: To account for potential mosaic outliers in the training set, the model fitting is iterative. Alleles with deviance scores in the top 5% are temporarily removed (up to 10% of alleles), and the model is refitted to the remaining "stable" alleles to establish a robust baseline.

• Statistical Testing:

  • To identify unusually unstable alleles, a query allele's profile is compared against the fitted locus-specific DM model.
  • A likelihood ratio test statistic is computed.
  • Significance is assessed via parametric bootstrap (100,000 draws), generating an empirical p-value to determine if the allele's instability significantly exceeds the locus-specific baseline.

3. Key Contributions

• General-Purpose Framework: Introduced a scalable method applicable to both simple repeats and complex, structurally heterogeneous loci (e.g., RFC1), without requiring prior knowledge of motif structure.
• Noise-Agnostic Modeling: Instead of attempting to disentangle technical noise from biological mosaicism (which is difficult), the model treats them as a combined signal to define a locus-specific baseline. This allows for the detection of outliers regardless of the noise source.
• Locus-Specific Baselines: Demonstrated that a single global instability metric is insufficient; instability varies significantly by locus, necessitating individual models for each TR.
• Open Source Tool: The methodology is implemented in the TRGT-instability tool (v1.0.0), available on GitHub.

4. Key Results

• Global Instability Characteristics:

  • Across 617,007 TRs in 256 samples, mean read divergence rates were low (mean: 0.0051, median: 0.0039), indicating high sequencing accuracy.
  • Drivers of Instability: Instability rates were strongly correlated with repeat purity (fraction of bases in perfect tracts) but only weakly correlated with overall repeat length. Perfect repeats are significantly more unstable than interrupted ones.
  • Model Fit: The Dirichlet-Multinomial model provided a good fit for the majority of loci (rejected in only 2% of cases when fit/evaluated on the same data; 7% in split-sample validation).

• Pathogenic Repeat Analysis:

  • Known Pathogenic Loci: Analysis of 71 known pathogenic repeats (e.g., PRNP, DMPK, FMR1) revealed locus-to-locus variability in baseline mosaicism. Some loci showed tight distributions (low instability), while others (like DMPK and FMR1) showed heavy-tailed distributions.
  • Expanded Alleles: In targeted PureTarget data from samples with known expansions (22 samples, 880 alleles), 17 out of 21 expanded alleles (covered by $\ge$10 reads) were identified as significantly unstable compared to their locus-specific baseline.
  • This confirms that pathogenic expansions often exhibit elevated instability relative to the population baseline, consistent with the observation that many pathogenic expansions are perfect or near-perfect repeats.

5. Significance

• Clinical Biomarker Potential: The ability to quantify instability genome-wide provides a new candidate biomarker for clinical trials in repeat expansion disorders. Unusually unstable alleles can be prioritized as candidate pathogenic variants in genetically undiagnosed individuals.
• Therapeutic Targeting: By establishing a baseline for instability, the model helps identify specific loci and alleles that may be targets for therapies aimed at reducing somatic mosaicism.
• Scalability: The method offers a practical, computationally efficient approach to study TR instability across the entire genome, moving beyond the limitations of short-read sequencing and manual inspection of specific loci.
• Foundation for Future Research: This work establishes a scalable framework for systematic studies of repeat instability, bridging the gap between raw long-read data and biological insight into repeat expansion disorders.