Skip to content

HCLS AI Factory — Demo Guide

Step-by-step walkthrough for demonstrating the HCLS AI Factory on DGX Spark.

License: Apache 2.0 | Author: Adam Jones | Date: February 2026

Demo Overview

Parameter Value
Demo Duration 15-20 minutes (live walkthrough)
Pipeline Mode demo (pre-configured VCP/FTD)
Hardware NVIDIA DGX Spark (GB10, 128 GB unified)
Target Gene VCP — Frontotemporal Dementia
End Result 100 ranked novel drug candidates

What the Audience Will See

  1. Raw DNA data (FASTQ) entering the platform
  2. GPU-accelerated variant calling (Parabricks) completing in minutes
  3. 11.7 million variants annotated and indexed in a vector database
  4. Interactive Claude RAG chat identifying VCP as a drug target
  5. BioNeMo generating 100 novel VCP inhibitors
  6. Ranked candidates with docking scores and drug-likeness profiles
  7. PDF report generated automatically

Pre-Demo Setup

Step 1: Verify Hardware

# Verify DGX Spark GPU
nvidia-smi

# Expected: GB10 GPU, 128 GB unified memory
# Verify ARM64 architecture
uname -m
# Expected: aarch64

Step 2: Set Environment Variables

# Copy and edit environment template
cp .env.example .env

# Required variables:
# ANTHROPIC_API_KEY=sk-ant-...     (for Claude RAG)
# NGC_API_KEY=nvapi-...            (for BioNeMo NIMs)

Step 3: Start All Services

# Start services in dependency order
./start-services.sh

# This starts:
# 1. Infrastructure (Milvus, monitoring)
# 2. Stage 1 services (Parabricks, genomics portal)
# 3. Stage 2 services (RAG engine, Streamlit chat)
# 4. Stage 3 services (BioNeMo NIMs, discovery UI)
# 5. Landing page (health monitor)

Step 4: Verify All Services Healthy

Open the landing page: http://localhost:8080

All 10 services should show green status:

Service Port Expected Status
Parabricks Portal 5000 GREEN
Milvus Vector DB 19530 GREEN
RAG API 5001 GREEN
Streamlit Chat 8501 GREEN
MolMIM NIM 8001 GREEN
DiffDock NIM 8002 GREEN
Discovery UI 8505 GREEN
Grafana 3000 GREEN
Prometheus 9099 GREEN
DCGM Exporter 9400 GREEN

Step 5: Pre-stage Demo Data

# Ensure reference genome and FASTQ data are available
ls /reference/GRCh38.fa          # 3.1 GB reference
ls /data/HG002_R1.fastq.gz      # ~100 GB read 1
ls /data/HG002_R2.fastq.gz      # ~100 GB read 2

Demo Script

Opening (1 minute)

Talking points:

  • "This is the HCLS AI Factory — a platform that transforms patient DNA into novel drug candidates in under 5 hours."
  • "Everything runs on this single DGX Spark — a $3,999 desktop workstation."
  • "Three stages: GPU-accelerated genomics, RAG-grounded target identification, and AI-driven drug discovery."

Show: Landing page at http://localhost:8080 — highlight the 10-service health grid.

Stage 1: Genomics (3-4 minutes)

Launch Pipeline

# Launch demo mode
python run_pipeline.py --mode demo

Show: Genomics portal at http://localhost:5000

Talking points:

  • "Patient DNA was sequenced by an Illumina instrument, producing ~200 GB of raw FASTQ data."
  • "NVIDIA Parabricks runs BWA-MEM2 alignment on the GPU — 20-45 minutes instead of 12-24 hours on CPU."
  • "Google DeepVariant then calls variants with >99% accuracy — 10-35 minutes on GPU."
  • "Total Stage 1: under 1 hour. On CPU, this would take 1-2 days."

Key numbers to mention: - ~200 GB FASTQ input (30× whole-genome sequencing) - ~11.7 million variants called - ~3.5 million pass quality filter (QUAL>30) - GRCh38 reference genome (3.1 GB)

Show: Grafana dashboard at http://localhost:3000 — GPU utilization spiking during alignment and variant calling.

Stage 2: RAG/Chat (5-6 minutes)

Show Annotation Pipeline

Talking points:

  • "The 3.5 million high-quality variants are annotated with three databases."
  • "ClinVar — 4.1 million clinical variants from NCBI. Tells us if a variant is pathogenic."
  • "AlphaMissense — 71 million AI-predicted pathogenicity scores from DeepMind."
  • "VEP — Ensembl Variant Effect Predictor for functional consequences."

Annotation funnel to walk through: - 11.7M variants → 3.5M (quality filter) → 35,616 (ClinVar match) → 6,831 (AlphaMissense match) → 2,412 (actionable)

Show Vector Database

Talking points:

  • "Every variant is embedded using BGE-small-en-v1.5 and indexed in Milvus."
  • "3.5 million variant embeddings, each 384 dimensions, searchable in milliseconds."
  • "This is what makes natural language queries possible over genomic data."

Show: Attu UI at http://localhost:8000 — show the genomic_evidence collection with 17 fields and 3.5M records.

Interactive Chat Demo

Show: Streamlit chat at http://localhost:8501

Type this query:

"What are the most promising drug targets in this patient's genome for neurodegenerative disease?"

Expected response from Claude:

Claude will identify VCP as the primary target with: - rs188935092 (chr9:35065263 G>A) — ClinVar Pathogenic - AlphaMissense score 0.87 (well above 0.564 pathogenic threshold) - HIGH impact missense variant in D2 ATPase domain - Known drug target — CB-5083 reached Phase I clinical trial - Druggability score 0.92

Follow-up query:

"Tell me more about the VCP protein structure and druggable binding sites."

Talking points:

  • "Claude is grounded in the actual variant evidence — not hallucinating."
  • "The knowledge base covers 201 genes across 13 therapeutic areas."
  • "171 of those genes (85%) are classified as druggable targets."

Stage 3: Drug Discovery (4-5 minutes)

Show Structure Retrieval

Talking points:

  • "VCP maps to UniProt P55072. The pipeline queries RCSB PDB for protein structures."
  • "Four structures found: 8OOI, 9DIL, 7K56, 5FTK."
  • "5FTK scores highest — it's an X-ray structure at 2.3 Å resolution with the CB-5083 inhibitor already bound."
  • "Having an inhibitor-bound structure means we know exactly where drugs bind."

Show Molecule Generation

Show: Discovery UI at http://localhost:8505

Talking points:

  • "BioNeMo MolMIM takes the CB-5083 seed compound and generates 100 novel analogs."
  • "It's a masked language model for molecules — masks parts of the SMILES string and regenerates novel chemistry."
  • "98 of 100 pass RDKit chemical validity checks."

Show Docking Results

Talking points:

  • "BioNeMo DiffDock predicts how each molecule binds to the VCP D2 ATPase domain."
  • "It uses diffusion modeling — no pre-defined grid needed."
  • "34 candidates score below -8.0 kcal/mol — excellent binding affinity."

Show Ranked Candidates

Talking points:

  • "Final ranking: 30% generation confidence, 40% docking affinity, 30% QED drug-likeness."
  • "Top candidate: composite score 0.89 — that's 39% better than the CB-5083 seed."
  • "The top candidate also has better QED (0.81 vs 0.62) and better docking (-11.4 vs -8.1)."
  • "87 of 98 pass Lipinski's Rule of Five for oral drug-likeness."

Key demo table:

Metric CB-5083 (Seed) Top Candidate Improvement
Dock Score -8.1 kcal/mol -11.4 kcal/mol +41%
QED 0.62 0.81 +31%
MW 487.2 Da 423.5 Da -13%
Composite 0.64 0.89 +39%

Show PDF Report

Talking points:

  • "A complete PDF report is generated automatically via ReportLab."
  • "It includes provenance: every tool version, every database version, every threshold."
  • "This is reproducible science — run the same pipeline, get the same results."

Closing (2 minutes)

Talking points:

  • "In under 5 hours, on a $3,999 desktop workstation, we went from raw DNA to 100 ranked drug candidates."
  • "This collapses a process that traditionally takes weeks to months."
  • "The same Nextflow pipelines scale from DGX Spark to DGX SuperPOD for enterprise deployments."
  • "Everything is open-source under Apache 2.0."

Scaling story:

Phase Hardware Scale
Phase 1 DGX Spark ($3,999) Proof build — what you just saw
Phase 2 DGX B200 Department — multiple concurrent patients
Phase 3 DGX SuperPOD Enterprise — thousands concurrent, federated

Cross-modal:

  • "The Imaging Intelligence Agent can trigger genomic analysis — a Lung-RADS 4B+ finding automatically initiates tumor profiling."
  • "Drug candidates feed back into integrated clinical reports."

Troubleshooting

Service Not Starting

# Check Docker container status
docker compose ps

# Check container logs
docker compose logs <service-name>

# Restart a specific service
docker compose restart <service-name>

BioNeMo NIM Not Ready

MolMIM and DiffDock NIMs require NGC API key and may take 2-5 minutes to initialize on first start.

# Check NIM health
curl http://localhost:8001/v1/health/ready
curl http://localhost:8002/v1/health/ready

Milvus Connection Issues

# Check Milvus status
curl http://localhost:19530/v1/vector/collections

# Check etcd and MinIO dependencies
docker compose logs milvus-etcd
docker compose logs milvus-minio

GPU Out of Memory

Monitor GPU memory during Parabricks execution. DeepVariant peaks at ~60 GB on the 128 GB unified memory system.

# Monitor GPU memory
watch -n 1 nvidia-smi

Quick Reference

Action Command / URL
Start services ./start-services.sh
Launch demo python run_pipeline.py --mode demo
Landing page http://localhost:8080
Genomics portal http://localhost:5000
Chat interface http://localhost:8501
Milvus UI http://localhost:8000
Discovery UI http://localhost:8505
Grafana http://localhost:3000
GPU monitoring nvidia-smi or http://localhost:3000

HCLS AI Factory — Apache 2.0 | Author: Adam Jones | February 2026