Mutant Fragment Size Distribution (mFSD)

Command: krewlyzer mfsd

Plain English: mFSD compares fragment sizes at known mutation sites. Fragments carrying the mutation (ALT) tend to be shorter than healthy fragments (REF).

Use case: MRD monitoring - track tumor DNA by comparing mutant vs. wild-type fragment sizes.

Purpose

Compares the size distribution of mutant vs. wild-type fragments at variant sites, with support for all small variant types and 4-way fragment classification.

Processing Flowchart

flowchart LR
    BAM["sample.bam"] --> RUST["Rust Backend"]
    VCF["variants.vcf/maf"] --> RUST
    REF["Reference FASTA"] --> RUST

    RUST --> MFSD["mFSD.tsv"]

    subgraph "Per Variant"
        RUST --> CLASS{"Classify Reads"}
        CLASS --> REF_F["REF fragments"]
        CLASS --> ALT_F["ALT fragments"]
        CLASS --> NONREF["NonREF"]
        CLASS --> N_F["N fragments"]
    end

    subgraph "Statistics"
        REF_F --> KS["KS Test"]
        ALT_F --> KS
        KS --> PVAL["p-value, Delta"]
    end

Biological Context

Mutant ctDNA fragments are typically shorter (~145bp) than wild-type cfDNA (~166bp) due to: - Altered nucleosome positioning in tumor cells - Different chromatin accessibility - Enhanced apoptosis patterns

This module quantifies this difference using high-depth targeted sequencing data.

Variant Types Supported

Usage

# Basic usage with VCF
krewlyzer mfsd -i sample.bam -V variants.vcf -o output_dir/

# With MAF file and GC correction
krewlyzer mfsd -i sample.bam -V variants.maf -o output/ \
    -r hg19.fa --correction-factors factors.csv

# With per-variant distributions
krewlyzer mfsd -i sample.bam -V variants.vcf -o output/ \
    --output-distributions

Dual BAM Support

For optimal results with duplex sequencing panels (e.g., MSK-ACCESS), use separate BAMs:

Via run-all CLI

# Provide dedicated duplex BAM for mFSD
krewlyzer run-all -i sample.all_unique.bam --mfsd-bam sample.duplex.bam \
    -r hg19.fa -o out/ --assay xs2 --variants sample.maf

Via Nextflow Samplesheet

sample,bam,mfsd_bam,meth_bam,vcf,bed,maf,single_sample_maf,assay,pon,targets
ACCESS_001,/path/to/sample.all_unique.bam,/path/to/sample.duplex.bam,,,/path/to/variants.maf,true,XS2,,

CLI Options

Duplex Sequencing Support

mFSD supports duplex consensus sequencing for ultra-sensitive variant detection. When --duplex is enabled, fragments from high-confidence duplex families receive higher weight in statistical calculations.

Supported Duplex Formats

XS2: fgbio/Picard

Uses SAM auxiliary tags set by fgbio CallDuplexConsensusReads:

Example SAM record:

read1   0   chr1   1000   60   100M   *   0   150   ACGT...   IIII...   cD:i:5   aD:i:8   bD:i:7

XS1: Marianas

Encodes family size in the read name per Marianas documentation:

Marianas:UMI1+UMI2:contig:start:posCount:negCount:contig2:start2:pos2:neg2

Family size = posCount + negCount (e.g., 4 + 3 = 7)

Duplex Weighting Formula

When --duplex is enabled, each fragment receives a weight based on its duplex family size:

Interpretation in Duplex Mode

Example:

Variant at chr1:156845927
  ALT_Count    = 393
  ALT_Weighted = 691  (ratio = 1.76)
  → Indicates high-confidence duplex families

Log-Likelihood Ratio (LLR) Scoring

For low fragment count scenarios (common in duplex/panel sequencing), traditional KS tests are unreliable. mFSD provides LLR scoring as a probabilistic alternative.

LLR Formula

Using Gaussian models: - Healthy: μ = 167bp, σ = 35bp (nucleosomal periodicity) - Tumor: μ = 145bp, σ = 25bp (sub-nucleosomal fragments)

LLR Output Columns

LLR Interpretation Guide

Clinical Example: MRD Detection

Variant at TP53:chr17:7577539
  ALT_Count = 3           # Too few for KS test
  KS_Valid  = FALSE       # KS test unreliable
  ALT_LLR   = 4.2         # Positive = tumor-like fragments
  REF_LLR   = -89.5       # Negative = healthy-like REF population

  → Interpretation: ALT fragments show tumor signature despite low count

Cross-Species and Assay Support

The LLR model uses Gaussian distributions for healthy and tumor fragment length peaks. These parameters can be customized for different species or library preparations.

Built-in Presets

Biological Rationale

Fragment length peaks vary across species due to: - Nucleosome spacing differences - Canine nucleosomes are more tightly packed - Chromatin structure - Different histone modifications - Library preparation - ssDNA preps capture shorter fragments

flowchart LR
    subgraph "Human cfDNA"
        H_HEALTHY["Healthy: ~167bp"] 
        H_TUMOR["Tumor: ~145bp"]
    end

    subgraph "Canine cfDNA"
        C_HEALTHY["Healthy: ~153bp"]
        C_TUMOR["Tumor: ~135bp"]
    end

    H_HEALTHY --> DELTA1["Δ = -22bp"]
    H_TUMOR --> DELTA1
    C_HEALTHY --> DELTA2["Δ = -18bp"]
    C_TUMOR --> DELTA2

Python API Access

from krewlyzer._core import LLRModelParams

# Use built-in presets
human_params = LLRModelParams.human()
canine_params = LLRModelParams.canine()
ssdna_params = LLRModelParams.ssdna()

# Custom parameters
custom = LLRModelParams(
    healthy_mu=160.0,
    healthy_sigma=32.0,
    tumor_mu=140.0,
    tumor_sigma=20.0
)

Fragment Classification

Fragments are classified into 4 categories:

flowchart TB
    READ["Fragment at variant site"] --> CHECK{"Base at variant?"}
    CHECK -->|"Matches REF"| REF["REF category"]
    CHECK -->|"Matches ALT"| ALT["ALT category"]
    CHECK -->|"Other base"| NONREF["NonREF category"]
    CHECK -->|"N base"| N["N category"]

Formulas

KS Test (Kolmogorov-Smirnov)

Where: - \(F_1(x)\) = CDF of ALT fragment sizes - \(F_2(x)\) = CDF of REF fragment sizes

Size Delta

Expected: - Healthy: \(\approx 0\) - Cancer: \(< 0\) (ALT shorter)

VAF Proxy

Output Format

Main Output: {sample}.mFSD.tsv

Optional: {sample}.mFSD.distributions.tsv

With --output-distributions:

Chrom  Pos    Ref  Alt  Category  Size  Count
chr1   12345  A    T    REF       145   3
chr1   12345  A    T    REF       166   12
chr1   12345  A    T    ALT       142   2

Clinical Interpretation

MRD Settings

• Low fragment counts (1-2) produce NA
• ALT_Confidence: HIGH (≥5), LOW (1-4), NONE (0)
• KS_Valid: TRUE if REF and ALT ≥2 fragments each

See Also

• PON Models – GC correction factors
• Citation – Scientific references