Skip to content

Stage 2: Evidence RAG

02

Finding the needle in 11.7 million variants

Interactive queries in seconds

What This Stage Does

Stage 1 produced 11.7 million variants. Most are harmless — they're just what makes you, you. The challenge is finding the ones that actually cause disease and can be targeted with drugs.

Stage 2 uses Retrieval-Augmented Generation (RAG) to:

  1. Annotate — Cross-reference every variant against clinical databases
  2. Embed — Convert variants into searchable vector representations
  3. Reason — Use AI to interpret evidence and identify targets

The Funnel

This is the heart of what the platform does — narrowing millions of variants to actionable targets:

flowchart TD
    A["11,700,000 variants discovered"]
    B["3,561,170 pass quality filters"]
    C["35,616 match ClinVar records"]
    D["6,831 have pathogenicity predictions"]
    E["2,400 high-impact, disease-causing"]
    F["847 in druggable genes"]

    A -->|"Quality filter (QUAL>30)"| B
    B -->|"ClinVar annotation"| C
    C -->|"AlphaMissense scoring"| D
    D -->|"Impact classification"| E
    E -->|"Druggability filter"| F

    style A fill:#B3E5FC,stroke:#0288D1
    style B fill:#81D4FA,stroke:#0288D1
    style C fill:#4FC3F7,stroke:#0277BD
    style D fill:#29B6F6,stroke:#0277BD,color:#fff
    style E fill:#039BE5,stroke:#01579B,color:#fff
    style F fill:#00B4D8,stroke:#0077B6,color:#fff

From 11.7 million down to 847. That's the power of AI-driven annotation.

Knowledge Base

The RAG system is grounded in curated biomedical knowledge:

Source Coverage
ClinVar 4.1 million clinically studied variants
AlphaMissense DeepMind's pathogenicity predictions
Gene Panel 201 genes across 13 therapeutic areas
Druggability 171 genes (85%) with known drug targets

Therapeutic Areas Covered

Neurology · Oncology · Cardiovascular · Metabolic · Immunology · Rare Disease · Ophthalmology · Dermatology · Respiratory · Hematology · Musculoskeletal · Endocrine · Infectious Disease

How RAG Works

Retrieval-Augmented Generation combines search with AI reasoning:

flowchart LR
    Q["Natural Language\nQuery"]
    E["BGE Embedding\n384 dimensions"]
    M["Milvus Vector Search\n3.56M variants"]
    R["Retrieved Evidence\nClinVar + AlphaMissense"]
    C["Claude AI\nGrounded Reasoning"]
    A["Grounded Answer\nwith citations"]

    Q --> E --> M --> R --> C --> A

    style Q fill:#E1BEE7,stroke:#7B1FA2
    style E fill:#B3E5FC,stroke:#0288D1
    style M fill:#00B4D8,stroke:#0077B6,color:#fff
    style R fill:#B2DFDB,stroke:#00796B
    style C fill:#FFE082,stroke:#FFA000
    style A fill:#C8E6C9,stroke:#388E3C
  1. You ask a question in natural language
  2. Vector search finds the most relevant variants and annotations
  3. Evidence is retrieved from the knowledge base
  4. Claude reasons over the evidence to provide grounded answers

The AI can only cite variants and evidence that actually exist in this patient's data. It's not hallucinating — it's reasoning over real genomic evidence.

Example Query

Question: "What pathogenic variants does this patient have in genes associated with neurodegeneration?"

Response: The system identifies a VCP variant (chr9:35,065,263 G>A) with:

  • ClinVar classification: Pathogenic
  • AlphaMissense score: 0.87 (threshold: 0.564)
  • Associated disease: Frontotemporal Dementia
  • Druggability: Yes — VCP is a known therapeutic target

This variant becomes the input for Stage 3: Drug Discovery.

Technology Stack

- **Milvus 2.4** — Vector database for 3.56M variant embeddings - **Anthropic Claude** — AI reasoning over retrieved evidence - **ClinVar** — NIH clinical variant database - **AlphaMissense** — DeepMind pathogenicity predictions - **Streamlit** — Interactive chat interface - **Flask** — REST API for programmatic access

By the Numbers

Metric Value
Variants indexed 3.56 million
Embedding dimensions 384
Query latency < 2 seconds
Knowledge sources 4 (ClinVar, AlphaMissense, Gene Panel, PDB)
Genes covered 201
Druggable targets 171 (85%)

Why This Matters

Traditional variant interpretation requires:

  • Manual literature review — Hours per variant
  • Expert curation — Scarce clinical geneticists
  • Fragmented tools — Separate databases, no unified view

The RAG pipeline delivers:

  • Instant answers — Seconds, not hours
  • Grounded reasoning — Every claim traceable to evidence
  • Unified view — All annotations in one place
  • Natural language — No bioinformatics expertise required