Tissue-of-Origin Classifier

Single-sample tissue-of-origin inference pipeline using an AutoGluon ensemble trained on AACR GENIE v18 data.

Quick Start

# Install
uv venv && uv sync

# 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

Prerequisites

• Python >= 3.11
• uv (for environment management)

Input Files

MAF (required)

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

SEG (optional)

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

SV (optional)

Tab-delimited structural variant file with columns: Sample_Id, Site1_Hugo_Symbol, Site2_Hugo_Symbol.

Model

• CatBoost_BAG_L2 via AutoGluon (22 tumor types, 4,320 features)
• Training performance: 85.0% balanced accuracy, 94.8% top-3 accuracy
• 7 feature modalities: mutations, structural variants, SBS-6 mutation spectrum, mutation frequency, clinical, TERT promoter, copy number alterations

Output

Self-contained HTML report with:

• Top-3 predictions with confidence scores
• Full probability distribution across 22 tumor types
• SHAP feature explanations (when background data available)
• Non-zero features by modality
• Input data summary