str_mut_signatures

STR Mutation Signature Analysis.

Python package for analysis of Short Tandem Repeat (STR) mutation signatures from VCF files. It extracts somatic STR mutation events from paired tumor–normal VCFs, builds count matrices, filters them, performs NMF-based signature decomposition, and projects new samples onto learned STR mutation signatures.

Contents

• Installation
• Quick start
• Input format
• Annotating standard VCFs
• Matrix construction
• Filtering mutation matrices
• NMF signatures and projection
• Visualization
• Command line interface
• Python API
• Output
• Contributing
• License

Installation

From PyPI

The package is available through PyPI. Install with:

pip install str_mut_signatures

From source

git clone https://github.com/acg-team/str_mut_signatures
cd str_mut_signatures
pip install -e .

Development installation

pip install -r requirements_dev.txt

Quick start

Python Library Usage

Basic pipeline

from str_mut_signatures import (
    parse_vcf_files,
    build_mutation_matrix,
    filter_mutation_matrix,
    run_nmf,
    save_nmf_result,
    load_nmf_result,
    project_onto_signatures,
    plot_exposures,
    plot_pca_samples,
    plot_signatures,
)

# 1) Parse annotated paired tumor–normal VCF files into a long table
mutations = parse_vcf_files("vcf_directory/")

# 2) Build a mutation count matrix
# ru:
#   None     -> ignore motif
#   "length" -> use only motif length (LEN1, LEN2, ...)
#   "ru"     -> use full repeat unit sequence (e.g. AT, AAT)
#   "AT"     -> AT-rich vs non-AT-rich classification
# ref_length:
#   include reference repeat length as a feature component
# change:
#   include tumor–normal repeat-length change
matrix = build_mutation_matrix(
    mutations,
    ru="length",
    ref_length=True,
    change=True,
)

# 3) Filter the matrix (e.g. manual thresholds)
matrix_filt, summary = filter_mutation_matrix(
    matrix,
    feature_method="manual",
    min_feature_total=10,
    min_samples_with_feature=3,
    min_sample_total=0,
)

# 4) Run NMF
# Optional clustering is performed here and stored in nmf_res.groups
nmf_res = run_nmf(
    matrix_filt,
    n_signatures=5,
    random_state=11,
    max_clusters=4 # <= 1 disables clustering
)

# Access signatures and exposures
signatures = nmf_res.signatures   # features x K
exposures  = nmf_res.exposures    # samples x K
groups     = nmf_res.groups       # samples x 1 ("group")

# 4a) Plot mutation signatures
fig = plot_signatures(nmf_res, top_n=20)

# 4b) Visualize sample exposures to signatures (ordered by group, then total exposure)
figs = plot_exposures(
    nmf_res,
    stacked=True,
    max_samples_per_fig=40,  # paginate if many samples
)

# figs["absolute"]   -> list of figures with absolute exposures
# figs["proportion"] -> list of figures with exposures normalized per sample

# 4c) Visualize samples in PCA space based on exposures (colored by stored groups)
coords, var_ratio, ax = plot_pca_samples(nmf_res)

# coords   -> DataFrame with PC1, PC2, ...
# var_ratio-> variance explained by PC1 and PC2
# ax       -> matplotlib Axes with the scatter plot

# 5) Save NMF result for reuse
save_nmf_result(nmf_res, "nmf_results")

# 6) Load NMF result and project new samples
nmf_loaded = load_nmf_result("nmf_results")
new_exposures = project_onto_signatures(
    new_matrix=new_counts_df,
    signatures=nmf_loaded.signatures,
    method="nnls",
)

Command Line

1. Extract somatic STR mutation counts from paired tumor–normal VCFs:

str_mut_signatures extract \
    --vcf-dir data/vcfs/ \
    --out-matrix counts_raw.tsv \
    --ru length \
    --ref-length \
    --change

This produces a count matrix (TSV) with:

• rows = samples
• columns = STR mutation features such as:

LEN{motif_length}_{ref_length}_{change}

For example: LEN1_10_+1 means:

• motif length = 1 bp
• reference repeat length = 10 copies
• tumor has +1 copy relative to normal.

2. Filter the matrix to remove extremely rare features/samples:

str_mut_signatures filter \
    --matrix counts_raw.tsv \
    --out-matrix counts_filtered.tsv \
    --feature-method elbow

3. Run NMF to learn STR mutation signatures:

str_mut_signatures nmf \
    --matrix counts_filtered.tsv \
    --outdir nmf_results \
    --n-signatures 5

This writes:

• nmf_results/signatures.tsv – STR mutation signatures (features x K)
• nmf_results/exposures.tsv – sample exposures (samples x K)
• nmf_results/metadata.json – parameters and metadata.

4. Project new samples onto existing signatures:

str_mut_signatures project \
    --matrix new_counts.tsv \
    --nmf-dir nmf_results \
    --out-exposures new_exposures.tsv

Input format

To be processed by str_mut_signatures, VCF files must:

1. Contain paired samples (normal and tumor) per record.
2. Be annotated with STR-specific fields that describe the repeat motif and allele-level repeat copy numbers.

Supported sources

str_mut_signatures currently supports STR VCFs produced by:

• GangSTR
• conSTRain
• VCFs annotated with the strvcf_annotator tool

as long as they satisfy the requirements below.

Required structure

• Each VCF record must contain at least two samples:

  • Sample 1 (first column after FORMAT): normal
  • Sample 2 (second column after FORMAT): tumor

  By default, str_mut_signatures assumes this order and computes somatic changes as tumor vs normal. Only loci with differences between tumor and normal are used (somatic STR mutations).

• Required annotations:

  INFO fields

  • RU: Repeat unit / motif (e.g. A, AT, AAT).
  • REF: Reference repeat count (copy number of the motif in the reference genome).

  FORMAT fields (copy number)

  One of the following per sample:

  • REPCN: Per-allele repeat copy number (used by GangSTR and strvcf_annotator).
  • REPLEN: Per-allele repeat length in repeat units (used by conSTRain).

  str_mut_signatures automatically detects which field is present and uses it as the source of allele copy numbers.

Example schemas

GangSTR / strvcf_annotator-style

##INFO=<ID=RU,Number=1,Type=String,Description="Repeat unit">
##INFO=<ID=REF,Number=1,Type=Integer,Description="Reference repeat count">
##FORMAT=<ID=REPCN,Number=R,Type=Integer,Description="Per-allele repeat copy number">
#CHROM POS  ID REF ALT QUAL FILTER INFO        FORMAT        NORMAL     TUMOR
chr1   100 .  A   AT  .    .     RU=A;REF=10  GT:REPCN      0/0:10,10  0/1:10,11

conSTRain-style

##INFO=<ID=RU,Number=1,Type=String,Description="Repeat unit">
##INFO=<ID=REF,Number=1,Type=Integer,Description="Reference repeat count">
##FORMAT=<ID=REPLEN,Number=R,Type=Integer,Description="Per-allele repeat length in repeat units">
#CHROM POS  ID REF ALT QUAL FILTER INFO        FORMAT        NORMAL       TUMOR
chr1   100 .  A   AT  .    .     RU=A;REF=10  GT:REPLEN     0/0:10,10    0/1:10,11

From such records, str_mut_signatures:

• Extracts allele-level repeat copy numbers from REPCN or REPLEN.
• Compares NORMAL vs TUMOR allele counts at each locus.
• Identifies loci where tumor repeat copy number differs from normal.
• Encodes the net repeat-length change as tumor–normal (e.g. +1).
• Uses only these somatic STR events for downstream count matrices and signature extraction.

Annotating standard VCFs

If your VCFs lack RU, REF, or copy-number fields, you can annotate them using the tool strvcf_annotator:

• Takes standard VCF + STR reference.
• Produces STR-annotated VCFs compatible with str_mut_signatures.

For details see: strvcf_annotator.

Matrix construction

build_mutation_matrix provides a flexible way to define the feature space (columns) using simple flags.

Core components

Given:

• RU: repeat unit sequence (e.g. A, AT, AAT)
• len(RU): motif length
• REF: reference repeat count
• change: tumor–normal repeat-length change at that locus

you can select:

• ru:
  • None:
    • Do not use motif information.
  • "length":
    • Use only motif length.
    • Features start with LEN{motif_length}.
    • Example: LEN1_10_+1 for motif length 1, REF=10, change=+1.
  • "ru":
    • Use full repeat unit sequence.
    • Example: A_10_+1, AT_20_-2.
  • "AT":
    • Collapse motifs into AT-rich vs non-AT-rich:
    • AT_rich: motif consists only of A and T.
    • non_AT_rich: motif contains any C or G.
• ref_length (bool):
  • If True, include the reference repeat length as part of the feature key.
• change (bool):
  • If True, include tumor–normal change as part of the feature key.
  • Only somatic events (non-zero change) are counted.
  • If False, no change term is added and loci are not filtered by change.

Examples

1. Motif length + ref length + somatic change:

m = build_mutation_matrix(
    mutations,
    ru="length",
    ref_length=True,
    change=True,
)
# Columns: LEN{motif_length}_{ref_length}_{change}
# e.g. LEN1_10_+1

2. Full motif + change only:

m = build_mutation_matrix(
    mutations,
    ru="ru",
    ref_length=False,
    change=True,
)
# Columns: {RU}_{change}
# e.g. AT_+2

3. AT-rich vs non-AT-rich, with ref length and change:

m = build_mutation_matrix(
    mutations,
    ru="AT",
    ref_length=True,
    change=True,
)
# Columns: AT_rich_10_+1, non_AT_rich_20_-2, ...

4. Motif length only (no change, e.g. for presence/absence-style summaries):

m = build_mutation_matrix(
    mutations,
    ru="length",
    ref_length=False,
    change=False,
)
# Columns: LEN1, LEN2, ...

Filtering mutation matrices

Large STR feature spaces can be sparse. filter_mutation_matrix provides several strategies to reduce noise before NMF.

Supported methods

from str_mut_signatures import filter_mutation_matrix

filtered, summary = filter_mutation_matrix(
    matrix,
    feature_method="manual",
    min_feature_total=10,
    min_samples_with_feature=3,
    min_sample_total=0,
    feature_percentile=0.9,  # used for percentile method
)

Methods:

• feature_method="manual"
  • Keep features with:
    • total count across samples >= min_feature_total
    • present (non-zero) in at least min_samples_with_feature samples.
  • Drop samples with total counts < min_sample_total.
• feature_method="elbow"
  • Compute feature totals.
  • Use an "elbow" heuristic to choose a count threshold.
  • Keep features above that threshold.
  • Apply min_samples_with_feature and min_sample_total as in manual mode.
• feature_method="percentile"
  • Compute feature totals.
  • Keep features above a chosen percentile of totals (e.g. feature_percentile=0.9 keeps the top 10% by total count).
  • Apply min_samples_with_feature and min_sample_total as in manual mode.

The function returns:

• filtered – filtered count matrix.
• summary – small dataclass with filtering statistics (e.g. numbers of features/samples before/after, thresholds used).

NMF signatures and projection

NMF decomposition

NMF is used to decompose the filtered matrix into:

• Signatures: STR mutation patterns (features x K)
• Exposures: how much each sample uses each signature (samples x K)

from str_mut_signatures import run_nmf

nmf_res = run_nmf(
    matrix,
    n_signatures=5,
    init="nndsvd",
    max_iter=200,
    random_state=0,
    alpha_W=0.0,
    alpha_H=0.0,
    l1_ratio=0.0,
)

signatures = nmf_res.signatures  # DataFrame: features x K
exposures  = nmf_res.exposures   # DataFrame: samples x K
params     = nmf_res.model_params

Saving and loading NMF results

You can save and reload NMF results in a stable format (TSV + JSON):

from str_mut_signatures import save_nmf_result, load_nmf_result

save_nmf_result(nmf_res, "nmf_results")

nmf_loaded = load_nmf_result("nmf_results")
# nmf_loaded.signatures, nmf_loaded.exposures, nmf_loaded.model_params

Projecting new samples

Given a previously learned set of signatures, you can compute exposures for new samples (e.g. a new cohort or single sample):

from str_mut_signatures import project_onto_signatures

new_exposures = project_onto_signatures(
    new_matrix=new_counts_df,
    signatures=nmf_loaded.signatures,
    method="nnls",  # non-negative least squares
)

Rows in new_exposures are new samples, columns are signatures.

Visualization

Overview

After fitting NMF, str_mut_signatures provides convenience functions to:

• Visualize signature exposures per sample (bar plots).
• Visualize samples in PCA space (scatter plots of PC1 vs PC2).

These plots are useful for:

• Checking whether signatures separate known biological groups.
• Detecting outlier samples or batch effects.
• Inspecting clusters of samples with similar STR mutation patterns.
• Communicating results in figures for manuscripts or presentations.

Exposure bar plots

Use plot_exposures to visualize how much each sample uses each signature.

Key features:

• Stacked or grouped bars.
• Optional pagination when there are many samples.
• Samples are ordered by: 1. group (ascending) 2. total exposure within each group (descending)

from str_mut_signatures import plot_exposures

# nmf_res is the result of run_nmf(...)
figs = plot_exposures(
    nmf_res,
    stacked=True,
    max_samples_per_fig=40,  # paginate if many samples
)

# figs["absolute"]   -> list of figures with absolute exposures
# figs["proportion"] -> list of figures with exposures normalized per sample

Typical use:

• Absolute exposures: how much each signature contributes in raw counts.
• Proportions: composition of each sample (bars sum to 1), easier for comparing relative contributions across samples.

Signature plots

Use plot_signatures to visualize STR mutation signatures.

Each signature is shown as a bar plot of its top contributing features.

from str_mut_signatures import plot_signatures

fig = plot_signatures(
    nmf_res,
    top_n=25,
)

This is useful for:

• Interpreting biological meaning of signatures
• Comparing mutation patterns across signatures
• Selecting signatures for downstream analysis

PCA of samples

Use plot_pca_samples to see how samples cluster in PCA space based on their exposures (or any other sample x feature matrix).

Key features:

• Takes NMFResult directly and computes PCA internally.
• Samples are colored by NMFResult.groups["group"]

from str_mut_signatures import plot_pca_samples
# PCA on exposures, color by "group"
coords, var_ratio, ax = plot_pca_samples(nmf_res)

# coords      -> DataFrame with PC1, PC2, ...
# var_ratio   -> fraction of variance explained by each PC
# ax          -> matplotlib Axes with the scatter plot

Why use PCA plots?

• To see whether samples separate according to known phenotypes
• To detect subgroups that may correspond to novel STR-driven biology.
• To identify outliers or potential QC issues (samples far away from the main cloud).

Using custom groups

You can replace or augment groups manually:

nmf_res.groups["group"] = clinical_df.loc[nmf_res.exposures.index, "MSI_status"]

All plotting functions will automatically use the updated groups.

Command line interface

Global options

• -v / --verbose: Enable verbose logging.
• --version: Show package version.

Extract

str_mut_signatures extract \
    --vcf-dir PATH \
    --out-matrix OUTPUT.tsv \
    [--ru {none,length,ru,AT}] \
    [--ref-length] \
    [--change]

Key options:

• --vcf-dir: Directory with STR-annotated, paired tumor–normal VCF files.
• --ru:
  • none: ignore motif.
  • length: use motif length (LEN1, LEN2, ...).
  • ru: use full motif sequence.
  • AT: use AT-rich vs non-AT-rich labeling.
• --ref-length: Include reference repeat length in feature labels.
• --change: Encode tumor–normal repeat-length change and restrict to somatic events.
• --out-matrix: Output TSV with samples as rows and STR mutation features as columns.

Filter

str_mut_signatures filter \
    --matrix INPUT.tsv \
    --out-matrix FILTERED.tsv \
    [--feature-method {manual,elbow,percentile}] \
    [--min-feature-total INT] \
    [--min-samples-with-feature INT] \
    [--min-sample-total INT] \
    [--feature-percentile FLOAT]

Examples:

# Simple manual thresholds
str_mut_signatures filter \
    --matrix counts_raw.tsv \
    --out-matrix counts_filtered.tsv \
    --feature-method manual \
    --min-feature-total 10 \
    --min-samples-with-feature 3 \
    --min-sample-total 0

# Percentile-based filtering
str_mut_signatures filter \
    --matrix counts_raw.tsv \
    --out-matrix counts_filtered.tsv \
    --feature-method percentile \
    --feature-percentile 0.9

NMF

str_mut_signatures nmf \
    --matrix counts_filtered.tsv \
    --outdir nmf_results \
    --n-signatures 5 \
    [--max-iter 200] \
    [--random-state 0] \
    [--init nndsvd] \
    [--alpha-W 0.0] \
    [--alpha-H 0.0] \
    [--l1-ratio 0.0]

Outputs:

• nmf_results/signatures.tsv – signatures (features x K)
• nmf_results/exposures.tsv – exposures (samples x K)
• nmf_results/metadata.json – parameters and metadata

Project

str_mut_signatures project \
    --matrix NEW_COUNTS.tsv \
    --nmf-dir nmf_results \
    --out-exposures NEW_EXPOSURES.tsv

• --matrix: New count matrix (samples x features).
• --nmf-dir: Directory with an existing NMF result (signatures.tsv, metadata.json).
• --out-exposures: Output TSV with new sample exposures.

Python API

Main functions

from str_mut_signatures import (
    parse_vcf_files,
    build_mutation_matrix,
    filter_mutation_matrix,
    run_nmf,
    save_nmf_result,
    load_nmf_result,
    project_onto_signatures,
)

• parse_vcf_files(vcf_dir) → DataFrame of per-locus STR mutation data.
• build_mutation_matrix(mutations, ...) → samples x features count matrix.
• filter_mutation_matrix(matrix, ...) → filtered matrix + summary.
• run_nmf(matrix, n_signatures, ...) → NMFResult(signatures, exposures, model_params).
• save_nmf_result(result, outdir) / load_nmf_result(outdir) for persistence.
• project_onto_signatures(new_matrix, signatures, method="nnls") → new exposures.

Visualization helpers

from str_mut_signatures import (
    plot_exposures,
    plot_pca_samples,
    plot_signatures,
)

• plot_signatures(result, ...) Plot STR mutation signatures (feature loadings).
• plot_exposures(result, ...) Plot per-sample exposures, ordered and grouped using result.groups.
• plot_pca_samples(result, ...) PCA of samples, colored by result.groups.

Output

Typical outputs include:

• Count matrices (TSV): samples x STR mutation features.
• Filtered matrices (TSV): reduced feature space for robust NMF.
• NMF signatures and exposures (TSV).
• Metadata (JSON) describing NMF runs and parameters.
• Figures (from visualization helpers) for exploratory analysis and reporting.

These can be used to:

• Characterize somatic STR mutation processes.
• Compare STR signatures across cohorts.
• Associate STR signatures with clinical or genomic features.
• Apply learned STR signatures to new datasets.
• Visualize and communicate STR mutation signatures in manuscripts and presentations.

Contributing

Contributions are welcome!

For major changes, please open an issue first to discuss what you’d like to change.

Please ensure:

1. All tests pass (including integration tests).
2. Code follows existing style and module structure.
3. New features include unit tests and, where appropriate, integration tests.
4. Documentation and examples are updated.

License

MIT License