LongcallD: joint calling and phasing of small, structural and mosaic variants from long reads

LongcallD is a unified framework that leverages long-read sequencing and local multiple-sequence alignment to simultaneously call and phase small variants, structural variants, and low-fraction mosaic variants, thereby overcoming the limitations of existing disconnected tools and improving accuracy in complex genetic analysis.

The Gist

Imagine your genome is a massive, ancient library containing the instructions for building a human being. For years, scientists tried to read this library using short-read sequencing. This is like trying to understand a complex novel by cutting it into tiny, 10-word snippets. You can read the words, but you can't tell which sentence they belong to, or if a whole paragraph is missing. You also can't easily tell if two similar-looking sentences are actually different chapters.

Then came long-read sequencing (like PacBio and Oxford Nanopore). This is like having a scanner that can read entire pages, or even whole chapters, in one go. Suddenly, you can see how sentences connect, spot missing paragraphs, and understand the full story.

However, there was a problem: the computer programs trying to analyze these long pages were still thinking like they were reading short snippets. They treated "small typos" (single letter changes), "missing paragraphs" (structural variants), and "which page belongs to which story" (phasing) as three completely separate puzzles. This led to confusion, especially in the library's most chaotic sections, like the Tandem Repeats (areas where the text repeats itself over and over, like "ATATATAT...").

Enter LongcallD.

The New Librarian: LongcallD

Think of LongcallD as a super-smart, unified librarian who looks at the entire long page at once and solves all three puzzles simultaneously. Here is how it works, using some everyday analogies:

1. Sorting the "Clean" from the "Messy"

Imagine the library has two types of rooms:

• Clean Rooms: The text is clear, and the words are easy to read.
• Messy Rooms: The text is scribbled, repeated, or torn (these are the "noisy regions" like homopolymers and repeats).

Old tools tried to read the Messy Rooms the same way they read the Clean Rooms, which led to mistakes. LongcallD is smart enough to say, "Okay, this area is messy. I need a different strategy."

2. The "Haplotype" Detective (Phasing)

In a diploid human, you have two copies of every book (one from mom, one from dad). Sometimes, a typo appears on Mom's copy but not Dad's.

• Old Tools: They would see a typo and guess which copy it belongs to, often getting it wrong in the messy sections.
• LongcallD: It acts like a detective who looks at the whole long page. If it sees a known typo near the start of the page, it knows, "Ah, this whole page belongs to Mom's copy." It then uses that knowledge to correctly interpret the messy, scribbled parts of that same page. It groups all the "Mom" pages together and all the "Dad" pages together before trying to read the difficult parts.

3. Reconstructing the Story (Consensus)

In the Messy Rooms, the scanner might get confused by the repeating text. LongcallD takes all the "Mom" pages and all the "Dad" pages and lines them up perfectly (like a choir singing in harmony). By comparing them, it can figure out the true story, filtering out the scanner's static noise. This allows it to find big missing chunks (Structural Variants) that other tools miss because they were too busy looking at tiny words.

4. Finding the "Whispers" (Mosaic Variants)

Sometimes, a mutation happens after you are born, affecting only a small group of cells (like a whisper in a crowded room). These are called mosaic variants.

• Old Tools: They often ignore these whispers because they look like background noise or scanner errors.
• LongcallD: Because it has already organized the pages into "Mom" and "Dad" groups, it can listen very carefully. If it hears a whisper that fits perfectly with the "Mom" group but not the "Dad" group, it knows, "This isn't noise; this is a real, rare mutation!" It can even detect a mutation supported by just one single read if the context is right.

Why This Matters

The paper shows that LongcallD is a game-changer for two main reasons:

1. It solves the "Repeat" Problem: It is much better at reading the chaotic, repeating parts of our DNA (where many diseases hide) than previous tools.
2. It's a "One-Stop Shop": Instead of running three different programs to find small errors, big errors, and phase information, you run LongcallD once. It does it all together, making the results more accurate and consistent.

In a nutshell: If previous tools were like three different people trying to assemble a giant jigsaw puzzle while wearing blindfolds, LongcallD is a single expert who takes off the blindfolds, sorts the pieces by color (haplotypes), and assembles the whole picture at once, revealing the hidden details that were previously invisible.

Technical Summary

1. Problem Statement

Long-read sequencing technologies (PacBio HiFi and Oxford Nanopore) offer the unique ability to span complex genomic regions, bridging small variants (SNPs, indels) and large structural variants (SVs) within single reads. This provides inherent phasing information crucial for resolving haplotypes. However, current computational tools suffer from significant limitations:

• Disconnected Workflows: Small variant calling, SV detection, and phasing are typically treated as independent problems using distinct algorithms.
• Alignment Artifacts: Small variant callers (e.g., Clair3, DeepVariant) often fail in low-complexity regions (LCRs) or near SVs because they lack awareness of SVs, leading to spurious calls or missed variants due to inconsistent alignments.
• SV Calling Limitations: Existing SV callers often ignore read phasing or fail to resolve SVs in repetitive regions (e.g., tandem repeats) without realignment. Furthermore, they struggle to distinguish true low-variant allele fraction (VAF) mosaic variants from sequencing artifacts.
• Lack of Integration: No existing tool tightly integrates small variant calling, SV calling, and phasing into a unified framework that leverages the full length of long reads.

2. Methodology: The LongcallD Framework

LongcallD is a unified framework that utilizes local multiple-sequence alignment (MSA) to simultaneously call and phase small and structural variants. Its workflow consists of four main stages:

A. Region Classification (Clean vs. Noisy)

LongcallD explicitly partitions the genome into two types of regions based on alignment quality and complexity:

• Clean Regions: Areas with accurate, concordant alignments. High-confidence germline SNPs and small indels are called here via direct allele counting.
• Noisy Regions: Areas prone to errors or alignment inconsistencies, including:
  • Homopolymers and short tandem repeats (STRs).
  • Variable number tandem repeats (VNTRs).
  • Large insertion/deletion SV boundaries (characterized by read clipping).

B. Haplotype Tagging and Partitioning

• Clean Region Phasing: Heterozygous variants identified in clean regions are used to locally partition long reads into haplotypes (H1 and H2) using an iterative haplotype-extension strategy.
• Noisy Region Handling: Reads overlapping noisy regions are extracted and grouped by their assigned haplotype.

C. Haplotype-Aware Consensus Generation

For each noisy region, LongcallD performs:

1. Haplotype-Aware MSA: If a valid phase set exists (reads covering both haplotypes), it uses abPOA (partial order alignment) to generate haplotype-specific consensus sequences.
2. Haplotype-Agnostic MSA: If no phase set is available, it performs MSA on all reads, followed by K-medoids clustering to distinguish heterozygous from homozygous regions.
3. Variant Calling: The consensus sequences are realigned to the reference using WFA (Wavefront Alignment) to generate variant calls, which are then integrated with clean-region calls.

D. Iterative Phasing and Mosaic Detection

• Iterative Refinement: The framework alternates between assigning haplotypes to variants and reads until convergence, ensuring consistency across the genome.
• Mosaic Variant Calling: LongcallD detects low-VAF mosaic variants (SNVs and large indels) by leveraging the established haplotype structure.
  • Filtering: It applies stringent, context-aware filters to distinguish true mutations from artifacts.
  • MEI Detection: Specifically for Mobile Element Insertions (MEIs), it requires canonical features like Target Site Duplications (TSDs) and poly(A) tails, allowing detection supported by a single read.

3. Key Contributions

• Unified Framework: First tool to jointly call and phase small variants, SVs, and mosaic variants from long reads in a single pipeline.
• Noise-Aware Strategy: The explicit separation of "clean" and "noisy" regions allows for tailored algorithms (counting vs. MSA), significantly improving accuracy in difficult genomic contexts.
• Superior Mosaic Detection: By integrating germline phasing, LongcallD can filter out artifacts that appear on both haplotypes, enabling the detection of mosaic variants supported by as few as one or two reads.
• Mobile Element Insertion (MEI) Sensitivity: Capable of detecting single-read supported MEIs by modeling specific sequence signatures (TSDs, poly-A tails).

4. Results and Benchmarks

The authors benchmarked LongcallD against state-of-the-art tools (Clair3, DeepVariant, Sniffles2, SVDSS, PAV, dipcall, etc.) using the HG002 T2T Q100 v1.1 reference (HiFi and ONT data) and tumor-normal mixture datasets.

• Small Variant Calling:
  • Achieved competitive accuracy to DeepVariant and Clair3.
  • Demonstrated the lowest False Negative Rates (FNR) for SNPs and indels in homopolymer and non-homopolymer regions among alignment-based callers.
• Structural Variant (SV) Calling:
  • Tandem Repeats: LongcallD achieved the best FDR (3.2%) and FNR (9.8%) in tandem repeat regions, outperforming all other alignment-based SV callers.
  • Haplotype Reconstruction: Using aardvark (a haplotype-centric benchmark), LongcallD achieved perfect genotype and basepair F1 scores in >80% of tandem repeat regions, significantly outperforming SV-only callers which often produce inconsistent haplotypes.
• Mosaic Variant Detection:
  • SNVs: In a 1:49 tumor-normal mixture (COLO829), LongcallD achieved the highest precision (52.0%) and recall (33.1%), identifying 4,343 true positives missed by other tools.
  • Large Indels: In a 1:10 mixture (H2009), LongcallD showed superior recall (78.2%) compared to Sniffles2 (46.8%) for mosaic indels, while maintaining higher precision.
  • MEIs: Achieved 95.5% precision and 92.7% recall for mosaic MEIs by leveraging sequence feature modeling.
• Performance: LongcallD is computationally efficient, processing a 40x HiFi dataset in ~19 minutes with <20GB RAM, significantly faster than assembly-based pipelines.

5. Significance

• Clinical and Evolutionary Applications: LongcallD provides a robust foundation for resolving complex genetic architectures, particularly in repetitive regions where accurate haplotype reconstruction is critical for understanding disease mechanisms and evolutionary history.
• Mosaic Variant Discovery: The ability to detect low-VAF mosaic variants with high precision is a major advancement for cancer genomics and the study of somatic mosaicism in development.
• Paradigm Shift: The paper demonstrates that treating small variants, SVs, and phasing as a unified problem yields superior results compared to the traditional "pipeline" approach of running separate tools. This is especially vital for long-read data, where the length of the read inherently connects these variant types.

In summary, LongcallD represents a significant leap forward in long-read variant calling by unifying the detection of diverse variant types and leveraging phasing to resolve complex genomic regions and detect rare mosaic events with unprecedented accuracy.