🧬 PathoScan — Clinical Genomic Variant Intelligence

AI-powered system for pathogenicity prediction of genomic variants using XGBoost, SHAP explainability, and Gemini-based clinical insights.

---

🚀 Live Demo

👉 Frontend: https://pathoscan-zeta.vercel.app/
👉 Backend API: https://pathoscan-backend.onrender.com/docs

🚀 Overview

PathoScan is a full-stack AI system that:

• Predicts whether a genetic variant is Pathogenic or Benign
• Explains predictions using SHAP (Explainable AI)
• Generates clinical insights using Gemini AI
• Provides an intuitive web-based interface for researchers and clinicians

---

🧠 Core Features

• ⚡ High-performance ML model (XGBoost)
• 🔍 Explainable AI with SHAP
• 🤖 LLM-powered clinical interpretation (Gemini)
• 🌐 Interactive frontend dashboard
• 📊 Feature-level importance visualization

---

🧬 Data Preprocessing Pipeline

PathoScan leverages a multi-stage genomic data pipeline to transform raw datasets into structured features suitable for ML models.

---

🧬 1. Data Sources

PathoScan integrates multiple high-quality genomic databases:

• ClinVar – Clinical annotations of variants
• gnomAD – Population allele frequencies
• dbNSFP – Functional prediction scores (CADD, REVEL)
• GRCh38 – Reference human genome

📥 Dataset Source:
https://ftp.ncbi.nlm.nih.gov/

---

⚙️ 2. ClinVar Data Cleaning

Raw ClinVar data is curated into a reliable training dataset.

🔹 Filtering

• Retain only Single Nucleotide Variants (SNVs)
• Align variants to GRCh38

🔹 Label Standardization

• Pathogenic + Likely pathogenic → Pathogenic
• Benign + Likely benign → Benign

🔹 Data Cleaning

• Remove missing values
• Remove invalid alleles (na, -)

🔹 Feature Extraction

• chrom – Chromosome
• pos – Position
• ref – Reference allele
• alt – Alternate allele

🔹 Unique Identifier

variant_id = chrom_pos_ref_alt 🌍 3. gnomAD Processing

Provides population-level signals for rarity detection.

🔹 Extract: Allele Frequency (gnomad_af) Allele Count (AC) Total Alleles (AN) 🔹 Output: Stored in Parquet format Includes variant_id for merging 🔗 4. Data Integration Perform left join (ClinVar ← gnomAD) Missing values: gnomad_af = 0.0 🧠 5. dbNSFP Enrichment

Adds biological intelligence to predictions.

🔹 Features: CADD_phred – Deleteriousness score REVEL_score – Pathogenicity predictor 🔹 Processing: Chunk-based parsing (10GB+ files) Match using: chrom:pos:ref:alt 🧬 6. Reference Genome (GRCh38) Parse FASTA files Extract sequences: Chromosomes 1–22 X, Y chrM Store as Parquet for fast access ⚠️ Data Leakage (Critical ML Concept) 📌 What is Data Leakage?

Data leakage occurs when information from outside training data (like test data or future data) is used during training.

❌ Why It’s Dangerous Model appears highly accurate during testing But fails in real-world scenarios 🛡️ How PathoScan Prevents It Strict train/test separation Deduplication of variants Feature generation only from available biological signals No future or label-derived features used 🧠 Model Architecture Model: XGBoost Classifier Task: Binary classification (Pathogenic vs Benign) Input: Genomic + conservation + frequency features Output: Probability score 📊 Explainability (SHAP)

PathoScan uses SHAP to explain predictions:

Feature contribution per prediction Positive/negative impact visualization Ranked feature importance 🖥️ Application Screens

📊 Dashboard

📈 SHAP Explanation

🧾 Output

⚙️ Tech Stack 🔹 Backend FastAPI XGBoost SHAP Gemini API 🔹 Frontend HTML / CSS / JavaScript Interactive UI 🔹 Deployment Backend → Render Frontend → Vercel 🚀 Installation (Local) git clone https://github.com/Wrostdevil/Pathoscan cd project/backend

pip install -r requirements.txt uvicorn app:app --reload 🌐 API Endpoint POST /predict Example Input: { "chrom": "1", "pos": 123456, "ref": "A", "alt": "G", "gnomad_af": 0.001 } 🧠 Future Improvements 🔐 Authentication system 📊 Real-time SHAP visualization 🧬 Multi-variant analysis ☁️ Scalable cloud inference