🧬 GenomicsGPT

AI-Powered Genetic Variant Interpretation Platform

GenomicsGPT is an end-to-end variant interpretation pipeline that combines clinical database lookups, machine learning pathogenicity prediction, and LLM-powered clinical narrative generation.

Key Results

Architecture

User Input (HGVS, VCF, rsID, etc.)
        │
        ▼
┌─────────────────┐
│  Variant Parser  │  ← 7 input formats supported
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│ Data Aggregator  │  ← ClinVar API + Ensembl VEP
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│   ML Engine      │  ← XGBoost + LightGBM ensemble
│  (AUC = 0.9949) │     40 features, SHAP explainability
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│  LLM Narrative   │  ← Llama 3 (local) or Claude API
│    Engine        │     Structured clinical reports
└─────────────────┘

Model Performance

SHAP Explainability

Top predictive features:

1. cons_synonymous — Silent mutations strongly predict benign
2. is_lof — Loss-of-function variants strongly predict pathogenic
3. gene_path_ratio — Gene constraint score
4. cons_intronic — Deep intronic variants lean benign
5. num_submitters — Expert review signal

LLM Clinical Report Generation

The LLM engine takes ML predictions + database evidence and generates structured clinical interpretation reports following ACMG/AMP guidelines.

Example output (BRCA1 c.5266dupC → Llama 3, 24 seconds):

Classification: Pathogenic (confidence: 0.998). ClinVar consensus supports this prediction with 3 submissions reviewed by expert panel.

ACMG Criteria: PVS1 (frameshift → premature stop codon), PM2 (extremely rare in gnomAD, AF=0.000004, below BA1 threshold).

Clinical Implications: Increased risk for hereditary breast and ovarian cancer syndrome. Recommend genetic counseling and cascade testing for at-risk family members.

Report sections: Variant Summary, Classification, Evidence Summary, Molecular Mechanism, Population Data, Clinical Implications, ACMG Criteria, Limitations.

Two backends supported:

• Ollama (free, local) — runs Llama 3 on your GPU
• Claude API (paid) — higher quality output via Anthropic

# Free — local Llama 3
ollama serve  # in another terminal
python demo_report.py

# Or with Claude API
export ANTHROPIC_API_KEY="sk-ant-..."
python demo_report.py "BRAF V600E"

Feature Ablation Study

Molecular consequence features independently achieve 0.97 AUC, confirming the model learns biological patterns (LoF → pathogenic, synonymous → benign) rather than memorizing gene-specific statistics.

ML Pipeline

The classifier is trained on 1.69 million labeled ClinVar variants (GRCh38) with 40 engineered features across 9 categories: variant type, molecular consequence, loss-of-function flags, allele length, position, chromosome, review quality, gene constraint, and HGVS complexity.

Project Structure

genomicsgpt/
├── demo_report.py                     # LLM demo — generates clinical reports
├── notebooks/
│   └── 03_model_training.ipynb        # Full ML pipeline with visualizations
├── src/genomicsgpt/
│   ├── variant_parser/                # 7-format variant input parser
│   ├── data_aggregator/               # ClinVar + Ensembl API clients
│   ├── ml_engine/                     # Pathogenicity classifier
│   ├── llm_engine/                    # Clinical narrative generation
│   │   └── report_generator.py        # Ollama + Claude backends
│   ├── rag_engine/                    # Literature search (planned)
│   └── models.py                      # Core data models
├── data/models/
│   ├── xgb_model.pkl                  # Trained XGBoost model
│   ├── lgb_model.pkl                  # Trained LightGBM model
│   ├── metrics.json                   # Evaluation metrics
│   └── plots/                         # 7 evaluation visualizations
├── tests/                             # Unit and integration tests
└── train_pipeline.py                  # Self-contained training script

Quick Start

# Parse a variant
python -m genomicsgpt parse "BRAF V600E"

# Generate a clinical report (requires Ollama running)
ollama serve  # in another terminal
python demo_report.py "BRCA1 c.5266dupC"

# Reproduce the ML training
pip install xgboost lightgbm shap seaborn scikit-learn pandas
python train_pipeline.py

# Run the interactive notebook
jupyter notebook notebooks/03_model_training.ipynb

Tech Stack

• ML: XGBoost, LightGBM, scikit-learn, SHAP
• LLM: Llama 3 (Ollama), Claude API (Anthropic)
• Data: ClinVar (NCBI), Ensembl VEP
• Visualization: matplotlib, seaborn
• APIs: ClinVar E-utilities, Ensembl REST, Ollama REST
• Testing: pytest (38 tests)

Author

Samir Kerkar — github.com/skerk001

Data Scientist | B.S. Mathematics UC Irvine