Skip to content

Stage 1: GPU Genomics

01

From raw sequencing data to millions of variants

120 – 240 minutes on DGX Spark

Prerequisite: Stage 0 — Data Acquisition

Before running Stage 1, all required data must be downloaded via setup-data.sh. This includes HG002 FASTQ files (~200 GB), the GRCh38 reference genome, and BWA-MEM2 index. Stage 0 is a one-time step.

What This Stage Does

When a patient's DNA is sequenced, the machine produces raw data — billions of short DNA fragments stored in FASTQ files, typically around 200 GB per patient.

Stage 1 transforms this raw data into actionable genetic information:

  1. Alignment — Each DNA fragment is mapped back to its position on the human reference genome (like assembling a 3-billion-piece puzzle)

  2. Variant Calling — The pipeline identifies where this patient's DNA differs from the reference — these differences are called variants

  3. Quality Filtering — AI-powered models (DeepVariant) distinguish real variants from sequencing errors with >99% accuracy

By the Numbers

Metric Value
Input size ~200 GB FASTQ
Reads aligned 800M – 1.2B
Variants discovered 11.7 million
High-quality variants 3.56 million
Accuracy >99% (DeepVariant)
Runtime 120 – 240 minutes

The Speed Advantage

Step Traditional (CPU) HCLS AI Factory (GPU)
Alignment (BWA-MEM2) 12 – 24 hours 1 – 2 hours
Variant Calling 8 – 12 hours 1 – 2 hours
Total 1 – 2 days 2 – 4 hours

GPU acceleration via NVIDIA Parabricks delivers 10–50x speedup over traditional CPU pipelines.

Technology Stack

- **NVIDIA Parabricks 4.6** — GPU-accelerated GATK best practices - **BWA-MEM2** — Burrows-Wheeler Aligner for read mapping - **Google DeepVariant** — AI-powered variant calling - **Nextflow** — Workflow orchestration - **Docker** — Containerized, reproducible execution

How It Works

flowchart LR
    FASTQ["FASTQ\n200 GB raw reads"]
    ALIGN["BWA-MEM2\nGPU Alignment"]
    DV["DeepVariant\nAI Variant Calling"]
    VCF["VCF\n11.7M variants"]
    GPU["⚡ NVIDIA GPU"]

    FASTQ --> ALIGN --> DV --> VCF
    GPU -.->|accelerates| ALIGN
    GPU -.->|accelerates| DV

    style FASTQ fill:#B3E5FC,stroke:#0288D1
    style ALIGN fill:#76B900,stroke:#5a8f00,color:#fff
    style DV fill:#76B900,stroke:#5a8f00,color:#fff
    style VCF fill:#FFE082,stroke:#FFA000
    style GPU fill:#1B2333,stroke:#76B900,color:#76B900

Detailed HCLS AI Factory Pipeline Logical Diagram

Click to view the full pipeline logical diagram

Output

The stage produces a VCF file (Variant Call Format) containing:

  • Chromosome position of each variant
  • Reference allele (what the reference genome has)
  • Alternate allele (what the patient has)
  • Quality scores (confidence in the call)
  • Genotype (heterozygous or homozygous)

This VCF file becomes the input for Stage 2: Evidence RAG, where we determine which variants actually matter.

Why GPU Acceleration Matters

Traditional genomics pipelines run on CPU clusters and take 1–2 days per patient. For a hospital processing hundreds of patients, this creates bottlenecks.

GPU acceleration changes the equation:

  • Same accuracy — FDA-cleared Parabricks matches CPU results
  • 10x faster — Hours instead of days
  • Lower cost — One DGX Spark vs. a CPU cluster
  • Real-time capability — Results while the patient is still in clinic