-
-
Helixstreet: the consent-native pharmacogenomics layer (overview)
-
Ed25519-signed phenotype attestation explained live by Nemotron 3 Super on Nebius Token Factory
-
One-tap consent revocation: the next query halts at the Consent Gate with a signed REFUSED_CONSENT_REVOKED receipt
-
Per-step latency, tokens and USD cost across Nemotron 3 Nano, Super and Ultra
-
Verified NVIDIA Clara Parabricks job log on a Nebius L40S Serverless GPU job (6.05 s)
-
Persisted 5-run-per-tier Token Factory benchmark and the Nebius resources actually used
-
Verbatim CPIC guideline passage via Tavily Extract plus Tavily Search citations
-
RCSB 3D structure viewer and live OpenFDA label evidence
-
Architecture: Nebius Token Factory, Nebius Serverless GPU Jobs, Object Storage and the consent-native enclave
-
Command Center overview with the 6-step Living Twin strip
⚡ 30-Second "Aha Moment" Summary: Up to 99% of people carry at least one actionable pharmacogenomic variant (Van Driest 2014; Ji 2016), yet genomes are almost never checked at the pharmacy counter because a leaked genome can never be rotated like a password. Helixstreet (https://nebius.helixstreet.io) is a consent-native pharmacogenomics layer: a 10 kb chr20 slice of a patient's genome (GIAB NA12878) is called in 6.05 seconds by NVIDIA Clara Parabricks (
pbrun deepvariant) on Nebius L40S GPU (aijob-e00p9xmknz4yphfvmm), sealed inside a salted SHA-256 Merkle vault, and queried by a 3-tier NVIDIA Nemotron 3 (Nano 30B·Super 120B·Ultra 550B) fleet on Nebius Token Factory (+ Tavily Search & Extract for live FDA/CPIC/NCT evidence). Clinicians receive a 1 KB Ed25519-signed phenotype attestation with ranked Level-1A drug alternatives while0 bytesof raw genome are released to requesters. (Variant summaries are sent to Nebius Token Factory for inference; raw BAM/VCF never leave the bucket; enable Token Factory Zero Data Retention for production.)
Inspiration
Between 91% and 99% of us carry at least one actionable pharmacogenomic (PGx) variant, depending on the gene panel (Van Driest et al., Clin Pharmacol Ther 2014, PMID 24253661; Ji et al., J Mol Diagn 2016, PMID 26947514). Adverse drug reactions cause about 6.5% of hospital admissions (Pirmohamed et al., BMJ 2004, PMID 15231615) and rank among the leading causes of death in hospitalised patients (Lazarou et al., JAMA 1998, PMID 9555760). Yet pre-emptive PGx testing is still rare at the point of prescribing or dispensing — for two structural reasons:
- Today's PGx report is a dead PDF locked in one lab. A patient sequenced at a cardiology clinic in 2026 has no way for their genome to automatically warn an outpatient psychiatrist prescribing an SSRI in 2028, or an oncologist ordering fluoropyrimidine chemotherapy in 2030, or to re-evaluate their risk when CPIC or European DPWG guidelines change.
- A leaked genome can never be rotated like a password. Centralizing raw BAM or VCF files inside hospital EHRs, pharmacy chains, or cloud AI SaaS providers creates an irreversible biological liability under EU GDPR Art. 9(2)(a), Swiss nFADP Art. 6, and the European Health Data Space (EHDS 2027).
We asked a simple question: What if a hospital, pharmacy, or clinical AI never needed to see your genome to get a mathematically provable, guideline-grounded answer to whether a drug is safe for you?
Helixstreet is a consent-native pharmacogenomics layer. It turns a person's genome into a lifelong, provable, revocable Living Twin that answers drug-safety questions inside a sealed software vault and releases only an Ed25519-signed phenotype attestation — with 0 bytes of raw genome released to requesters.
What it does
Live Demo Command Center (research-grade, synthetic data): https://nebius.helixstreet.io
Same container on Nebius AI Cloud (Serverless AI endpoint aiendpoint-e01mb5sjet7c8rys8h, cpu-d3, eu-west1; running during the judging window Dec 1–15): https://port8080-e6d5g79dfjp897z.tunnel.applications.eu-west1.nebius.cloud
Helixstreet unites a 6-Step Consent-Native Living Twin with an NVIDIA Stack (Clara Parabricks 4.3 + 3-Tier Nemotron 3) running on Nebius AI Cloud and Nebius Token Factory:
1. The 6-Step Consent-Native Living Twin Lifecycle (0 Bytes Raw Genome Released to Requesters)
- Step 1 — Patient Genome Vault & Standing Consents (
EHDS 2027&W3C Verifiable Credential): A patient's called variants are sealed in a salted SHA-256 Merkle commitment tree (minimal-disclosure Merkle proofs, not zero-knowledge). The patient grants granular, purpose-bound, time-limited standing consents (e.g.,cns-pharm-basel-001for Demo Pharmacy Basel (Synthetic Community Pharmacy), purposedrug_safety_dispensing_check) and can export an Ed25519-signed W3C Verifiable Credential (.vc.json) Sovereign Genome Wallet Pass. - Step 2 — Pre-Agent Consent Gate (
POST /ask): When a pharmacist or hospital EHR asks "Is clopidogrel 75 mg/day safe to dispense post-stent?", the Consent Gate checks(patient_id, requester, purpose, expires_at, status == ACTIVE)before any agent is invoked or any vault genotype is read. - Step 3 — Enclave Phenotype Computation, Knowledge Graph & Tavily Extract: Inside the enclave, Helixstreet computes
CYP2C19 *2/*2 → Poor Metabolizer, traverses our Pharmacogenomic Knowledge Graph (13 drug-gene pairs · 39 phenotype rules · hand-curated from CPIC + DPWG), fetches a Verbatim passage (Tavily Extract) from the published CPIC guideline on PubMed Central (https://api.tavily.com/extractwith 24h SQLite caching), enriches with live Tavily Clinical Guidelines & ClinicalTrials.gov searches, and generates a grounded clinical explanation (no-invention prompt; every number, drug and PMID comes from the signed attestation payload) via NVIDIA Nemotron 3 Super (nvidia/nemotron-3-super-120b-a12b) returning:- Primary recommendation: Avoid Clopidogrel (lack of active thiol metabolite conversion → stent thrombosis risk).
- Ranked Therapeutic Alternatives:
#1 Prasugrel 10 mg/day(RxNorm 613391, unaffected by CYP2C19),#2 Ticagrelor 90 mg BID(RxNorm 1116632). - Guideline Summary & Verbatim passage (Tavily Extract):
CPIC 2022/2024 + DPWG 2024 · PMID: 35034351 · PMC9287492 · Evidence Level 1A.
- Step 4 — Ed25519 RFC-8032 Signed Attestation & Tamper-Evident Receipt Ledger: The enclave signs the 1 KB attestation with its Ed25519 key and appends a hash-chained receipt (
raw_genome_bytes_exposed: 0) verifiable in one click (POST /verify) or offline viapython scripts/verify_attestation.py. - Step 5 — One-Tap Consent Revocation & Provable Refusal (
POST /consents/revoke): When the patient revokes consent, the very next query halts at the Consent Gate (vault_accessed: False,agent_invoked: False) and writes an Ed25519-signedREFUSED_CONSENT_REVOKEDreceipt to the ledger. - Step 6 — Always-On Living Twin (
POST /fhir/MedicationRequest,POST /guidelines/update,GET /guidelines/watch,POST /rwe/outcome): Months or years later, when an outpatient clinic orders Escitalopram 20 mg/day via an HL7 FHIR R4MedicationRequest, the Living Twin wakes from persistent phenotype memory unprompted, intercepts theCYP2C19 *2/*2QTc-prolongation overdose risk (50% dose cap <=10 mg/day), and logs a consented Closed-Loop Real-World Evidence (RWE) outcome (adr_prevented = True). Its second stream is a live guideline watch powered by Tavily: a domain-restricted Tavily Search for the governing guideline source (cpicpgx.org,clinpgx.org,pharmgkb.org,knmp.nl), ranked to prefer the canonical CPIC guideline page, a Tavily Extract of the guideline text (falling back to the CPIC publication on PubMed Central when a publisher page blocks automated fetching), and a SHA-256 fingerprint over the top source and passage compared with the previous check — so the twin can tell the pharmacist whether the governing guideline actually changed. Every alert states itsexecution_mode(live_tavily_watch,caller_supplied_update, orsimulated_guideline_updatewhen no live source exists).
2. Simulated Scenarios & Roadmap Extensions (execution_mode Explicitly Labelled)
To keep every metric and API response strictly auditable, Helixstreet explicitly labels simulated scenarios (execution_mode: "simulated_scenario", data_provenance: "illustrative constants, not measured") and roadmap mocks:
- Simulated scenario: Agent-to-Agent (A2A) Minimal-Disclosure Regimen Negotiation (
POST /a2a/negotiate): A Hospital EHR / ICU Prescribing Agent (did:web:demo-hospital.example:cardiology-icu-agent) proposes a 4-drug cardiac discharge bundle (clopidogrel 75mg,simvastatin 40mg,escitalopram 20mg,warfarin 5mg) to the patient's Living Twin (did:helix:ch:pgx-demo-001). The enclave deterministically evaluates each drug against the patient's sealed vault genotypes and the13 drug-gene pairs · 39 phenotype rulesKnowledge Graph, returning an Ed25519-signed counter-regimen with0 bytesraw genome released to requesters. - Simulated scenario: Hypothetical 10-year timeline (
2026 → 2036,GET /twin/timeline/{patient_id}): Illustrates how a single sequencing run and persistent Living Twin can protect a patient across cardiology, psychiatry, and oncology prescribing events over a decade. - Simulated scenario: Federated Enclave Cohort & Population Equity Aggregator (
POST /federated/cohort-query): Illustrates how 4 regional biobank/clinical vault nodes (EUR,EAS/SAS,AFR/AMR) can aggregate Ed25519-signed phenotype counts with 0 raw genomes pooled, highlighting higherCYP2C19loss-of-function carrier prevalence in East/South Asian cohorts. - Roadmap (not implemented in this build): Local SFT JSONL export of consented RWE outcomes (
GET /nebius/datalab-export/{patient_id},status: "local_jsonl_export_only") and placeholder target specifications fornvidia/RFD4-Proteina(POST /bionemo/rfd4-design,execution_mode: "mock_not_executed").
How we built it
Helixstreet is architected on Nebius AI Cloud (eu-north1) and Nebius Token Factory:
| Layer | NVIDIA & Partner Technology | Nebius Infrastructure | Verified Role in Helixstreet |
|---|---|---|---|
| Secondary Genomics | NVIDIA Clara Parabricks 4.3 (pbrun deepvariant) (nvcr.io/nvidia/clara/clara-parabricks:4.3.0-1) |
Nebius Serverless GPU Jobs (gpu-l40s-a, 48 GB L40S, 1gpu-16vcpu-64gb) + Nebius S3 (s3://helixstreet-nebius-data) + SecretStash (mbsec-e00zxbcp465mhbrhwf) |
Verified job aijob-e00p9xmknz4yphfvmm called 78 variant records (72 PASS) from the GIAB NA12878 chr20:10,000,000-10,010,000 BAM in 6.049 seconds (replayed as parabricks_na12878_replay_fixture in standalone demo mode). |
| Consent & Clinical Fleet | 3-Tier NVIDIA Nemotron 3 (Nano 30B · Super 120B · Ultra 550B) |
Nebius Token Factory (https://api.tokenfactory.nebius.com/v1/) via custom Google ADK BaseLlm adapter (nebius_llm.py) |
5 specialized agents route tasks across Nano ($0.06/$0.24), Super ($0.30/$0.90), and Ultra ($1.00/$3.00) with per-step latency, token, and cost telemetry (37.2% measured savings on the two seeded runs, measured 2026-10-25: $0.17111 vs $0.27243 All-Ultra; the dashboard recomputes the live aggregate over every run since the last restart). |
| Clinical Grounding & 3D Structures | Tavily Search & Extract API + CPIC / OpenFDA / Europe PMC + RCSB PDB (3Dmol.js) |
24h SQLite Extract Cache + guideline_watch fingerprints + Process-Safe WAL Memory Bank |
Extracts Verbatim passage (Tavily Extract) from PubMed Central CPIC guidelines, matches live ClinicalTrials.gov (NCT) biomarker studies, runs the Tavily-powered live guideline watch (domain-restricted Search for the governing guideline → Extract of the guideline text → fingerprint diff) behind the Living Twin's Stream B, and renders experimental RCSB PDB crystal structures (4GQS, 1H7X, 1OG5, 7FFE, 5VEU, 8HND). |
Right-Sized 3-Tier NVIDIA Nemotron 3 Fleet Routing
| Fleet Agent | NVIDIA Model on Nebius Token Factory | Catalog Rate (In / Out per 1M) | Responsibility |
|---|---|---|---|
Tier 1 · helixstreet_intake |
nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B |
$0.06 / $0.24 | Pre-agent Consent Gate, deterministic HIPAA Safe Harbor PII redaction (regex_safe_harbor_v1, 0 PII tokens enter any LLM), and raw BAM/VCF cryptographic fingerprinting (SHA-256, CRC32C, MD5). |
Tier 1 · helixstreet_variant_caller |
nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B |
$0.06 / $0.24 | Dispatches pbrun deepvariant on Nebius L40S GPU (--volume s3://helixstreet-nebius-data:/mnt/nebius), writes VARIANT_CALLING_RUNNING, and goes dormant (burning zero idle LLM tokens). |
Tier 3 · helixstreet_analyst |
nvidia/Nemotron-3-Ultra-550b-a55b |
$1.00 / $3.00 | Wakes as soon as the VCF is ready (chained after the ingest turn, or via /storage-event when a live GPU job lands), parses VCF calls, queries Ensembl 1000 Genomes Phase 3 (ALL, AFR, AMR, EAS, EUR, SAS), reconciles 0-based vs 1-based genomic coordinates, and seals the salted SHA-256 Merkle commitment root. |
Tier 3 · helixstreet_pharmacist |
nvidia/Nemotron-3-Ultra-550b-a55b |
$1.00 / $3.00 | Evaluates star-allele zygosity across CPIC Level-A genes, queries CPIC REST API + OpenFDA + Europe PMC + Tavily Search (cpicpgx.org, fda.gov, clinicaltrials.gov), and synthesizes research-grade medication insights. |
Tier 2 · helixstreet_steward |
nvidia/nemotron-3-super-120b-a12b |
$0.30 / $0.90 | Synthesizes multi-agent outputs into a strict Digital Twin JSON schema, verifies the Merkle vault root, and seals the SHA-256 hash-chained audit ledger. |
Challenges we ran into
- Deterministic Clinical Facts vs. LLM Expressiveness: A clinical system can never let an LLM "guess" a star-allele diplotype or invent a drug dose. We solved this by separating deterministic enclave computation (VCF genotype parsing, CPIC/DPWG knowledge-graph traversal, Tavily Extract passage retrieval, and Ed25519 signing) from Nemotron 3 clinical synthesis, locking Nemotron under a strict no-invention system prompt where every number, drug alternative, and PMID must come from the verified knowledge-graph payload.
- Reconciling 1-Based VCF Coordinates with 0-Based GA4GH / Ensembl Population APIs: Off-by-one coordinate mismatches silently corrupt allele-frequency lookups across super-populations. Our
helixstreet_analyst(Nemotron 3 Ultra) explicitly reconciles 1-based VCFPOSwith Ensembl's 1000 Genomes Phase 3 REST endpoints acrossAFR,AMR,EAS,EUR, andSAS. - Zero-Idle-Token Asynchronous GPU Orchestration: Running variant calling on an aligned BAM takes seconds to minutes on GPU; keeping an LLM agent session open during GPU execution wastes tokens and risks timeouts. We built a two-stage hand-off (
ingest_pipeline→ VCF ready →analysis_pipeline: chained deterministically by the service in replay mode, or woken by the Nebius S3/storage-eventwebhook when a live GPU job's output lands) backed by a persistent Memory Bank so zero LLM tokens are burned while Parabricks runs on the L40S GPU.
Accomplishments that we're proud of
- Real End-to-End Parabricks Execution on Nebius L40S (
aijob-e00p9xmknz4yphfvmm): Executednvcr.io/nvidia/clara/clara-parabricks:4.3.0-1(pbrun deepvariant) on Nebiusgpu-l40s-aineu-north1mounted tos3://helixstreet-nebius-data, calling 78 variant records (72 PASS calls) on GIAB NA12878 in 6.049 seconds. - Mathematically Verifiable Zero-Genome-Released Architecture: Every
/askresponse carries an RFC-8032 Ed25519 signature verifiable against a trusted public key, a SHA-256 payload digest, a live NVIDIA Nemotron 3 Super clinical summary with token usage, a Verbatim passage (Tavily Extract), and a hash-chained Consent Receipt provingraw_genome_bytes_exposed = 0, backed by 92 automated unit, integration and cryptographic tests. - Runtime Telemetry-Derived Economics & 5-Run/Tier Benchmarks: Every agent turn logs exact prompt/completion tokens, latency, and USD cost against official Nebius Token Factory rates (
$0.06/$0.24,$0.30/$0.90,$1.00/$3.00), demonstrating a measured 37.2% token cost reduction across our two seeded runs (40.4%ondemo-na12878; the live dashboard recomputes the figure over every run since the last restart) without a single hardcoded savings constant.
What we learned
- Right-sizing across the Nemotron 3 family is a game-changer for multi-agent economics: Using
Nemotron 3 Nano 30Bfor deterministic intake and GPU job dispatch while reservingNemotron 3 Ultra 550Bfor deep population-genetics and pharmacogenomic reasoning delivered a 37.2% measured cost reduction ($0.17111vs$0.27243across 2 runs) and709.3 msmedian Tier-1 latency (5 runs/tier, measured 2026-10-25). - Consent-native architecture unlocks clinical adoption: Framing genomic privacy not as a legal obstacle but as a cryptographic primitive (
Ed25519attestations + minimal-disclosure Merkle commitments + W3C Verifiable Credentials) turns a one-time lab test into a lifelong clinical utility ready for the European Health Data Space (EHDS 2027).
What's next for Helixstreet
- Hardware Confidential Computing (NVIDIA H100 / Blackwell Confidential GPUs on Nebius): Binding the Helixstreet enclave's Ed25519 signing key directly to hardware TEE remote attestation quotes (
nv-attestation-sdk). - Full CPIC + DPWG + PharmVar Coverage (120+ Drug-Gene Pairs): Expanding from our curated
13 drug-gene pairs · 39 phenotype rulesslice to the complete CPIC & European DPWG catalog, including complex structural star-alleles (CYP2D6copy-number / hybrid tandem rearrangements). - Swiss EPD & German ePA Pharmacy Dispensing Pilots: Integrating our HL7 FHIR R4
MedicationRequestwatcher (POST /fhir/MedicationRequest) and W3C Verifiable Credential passport (GET /wallet/vc/{patient_id}) with community pharmacy dispensing software.