Tissue-of-Origin Classifier

Overview

Predict the tissue of origin for a tumor sample using somatic mutations, copy-number alterations, structural variants, and clinical features. The classifier integrates 7 feature modalities into 4,320 engineered features and predicts across 22 tumor types.

Output is a self-contained HTML report with top-3 predictions, full probability distributions, SHAP explanations, and a summary of the input data.

Performance

Evaluation

Feature modalities

• Somatic mutations (gene-level)
• Copy-number alterations (arm & gene-level)
• Structural variants
• SBS-6 mutation spectrum
• Mutation frequency
• TERT promoter status
• Clinical features (age, sex)

Installation

Requires Python ≥ 3.11 and uv.

git clone https://github.com/viktorlj/tissue-classifier.git
cd tissue-classifier
uv venv && uv sync

Usage

# Basic prediction
tissue-classifier predict \
    --maf sample.maf \
    --seg sample.seg \
    --age 65 --sex Male

# With all options
tissue-classifier predict \
    --maf sample.maf \
    --seg sample.seg \
    --age 65 --sex Male \
    --genome hg19 \
    --output ./results \
    --sample-id PATIENT_001

# Validate input files
tissue-classifier validate --maf sample.maf --seg sample.seg

# Show model info
tissue-classifier info

Input files

MAF (required)

Tab-delimited mutation annotation file with columns: Hugo_Symbol, Chromosome, Start_Position, End_Position, Reference_Allele, Tumor_Seq_Allele2, Variant_Classification, Variant_Type, Tumor_Sample_Barcode.

SEG (optional)

Copy-number segmentation file with columns: ID, chrom, loc.start, loc.end, seg.mean.

SV (optional)

Structural variant file with columns: Sample_Id, Site1_Hugo_Symbol, Site2_Hugo_Symbol.

Related repositories