Valiant Quantum — Drug Candidate Selection E2E

⬡ R&D Submission Batch

Step 1 — Candidate Batch

mRNA Sequence Candidates

12 mRNA sequence variants submitted by an R&D team for clinical screening prioritization. Each candidate is characterized by predicted immunogenicity, structural stability, and synthesis cost index. The task: select the optimal subset for Phase I advancement.

Candidates Submitted

Phase I Slots

To Rank Out

▦

Candidate Universe

mRNA variants — oncology antigen targeting (12 candidates, 9 qubits post-vault)

{}

Single Candidate Payload

JSON structure submitted to Garm

{
  "candidate_id": "BNT-V007",
  "sequence_hash": "sha256:e3b0c44298fc",  // proprietary — hashed before leaving Vault
  "immunogenicity_score": 0.87,              // predicted T-cell response strength
  "structural_stability": 0.74,              // RNAfold MFE normalized stability score (0=unstable, 1=highly stable)
  "synthesis_cost_index": 0.52,              // relative manufacturing cost (0=cheap, 1=expensive)
  "target_antigen": "KRAS-G12D",
  "ip_classification": "PROPRIETARY"
}

⬥ THE QUANTUM CASE

Selecting the optimal 3 from 12 candidates involves evaluating 220 possible combinations (C(12,3)). Classical brute-force is trivial here — but at 50+ candidates the space grows exponentially, reaching 10¹⁴ combinations for selecting 5 from 50.

The quantum advantage emerges when constraints are added: cross-antigen coverage requirements, manufacturing capacity coupling, regulatory co-development restrictions. These turn candidate selection into a constrained quadratic binary optimization problem (QUBO) that maps directly to quantum hardware.

This demo selects 3 from 9 qualifying candidates (post-vault filter) using QAOA on IQM’s 20-qubit Garnet processor — the same algorithmic structure that scales to 500+ candidate portfolios on D-Wave annealing hardware.

⬡ Vault — AES-256-GCM + IP Anonymizer

Step 2 — Vault Screening

IP Protection & Compliance

Before any candidate data reaches quantum hardware, Garm’s Vault applies AES-256-GCM field-level encryption and strips proprietary identifiers. Sequence hashes replace raw sequences. GDPR-compliant anonymization runs before data leaves the enterprise perimeter.

🔒

Vault Operations

Applied to each candidate payload

sequence_hashHASH (SHA-256)

candidate_idREMAP → UUID

target_antigenREMAP → token

immunogenicity_scorePRESERVE

structural_stabilityPRESERVE

synthesis_cost_indexPRESERVE

✓

Compliance Checks

GxP + data sovereignty enforcement

GDPR Article 25✓ Privacy by design

GxP Data Integrity✓ Audit trail active

IP classification✓ PROPRIETARY stripped

Processing regionEU-only enforced

EncryptionAES-256-GCM

RNAfold MFE filternormalized score ≥ 0.40

VAULT PROCESSING LOG

⬡ garm-engine / qubo.py

Step 3 — QUBO Construction

Quadratic Optimization Matrix

Candidate selection reformulated as a QUBO problem: maximize immunogenicity, penalize synthesis cost, enforce budget constraint (exactly k candidates selected). Mathematically identical to Markowitz portfolio optimization.

Q ij = -α\cdotimm i δ ij + β\cdotcost i δ ij + ρ\cdotcov ij + λ(1 - 2k) [diagonal terms] Q ij = 2λ [off-diagonal: budget coupling]

QUBO Parameters

Computed from vault-cleared candidates

Matrix size9 × 9

Immunogenicity weight (α)0.60

Cost penalty (β)0.25

Stability coupling (ρ)0.15

Phase I slots (k)3

Penalty strength (λ)auto

▨

Q Matrix Semantics

What each element encodes

Diagonal Q[i][i]: Immunogenicity incentive (−α·imm_i) + cost penalty (β·cost_i) + stability risk (ρ·σ_ii) + budget penalty (λ(1−2k))

Off-diagonal Q[i][j]: Budget coupling (2λ) + co-expression risk coupling (ρ·σ_ij)

Why this works: The same Markowitz math that selects a low-risk stock portfolio selects a high-immunogenicity, low-cost drug candidate set. The physics is identical — only the domain changes.

Full Q Matrix (9×9)

Computed from anonymized candidate data

// Computing...

⬡ Arvak Compiler — QASM3 → IQM Native

Step 4 — Circuit Generation

OpenQASM 3.0 Circuit

The QUBO matrix compiles into a QAOA ansatz circuit. Arvak transpiles it to IQM Garnet’s native gate set (CZ, PRX, U3) and topology (20 superconducting qubits in a star graph).

Qubits

Circuit Depth

Total Gates

Garnet

Target Processor

●

Gate Distribution

QAOA ansatz decomposition

IQM

IQM Garnet Compilation

Arvak compiles for Garnet’s star topology

Native gates: CZ, PRX, U3

Topology: 20 qubits, star graph (central resonator)

Backend: Scaleway QaaS — Frankfurt, Germany (EU data sovereignty)

Transpilation: RZZ → CZ + PRX decomposition via Arvak

Speed: 1000× faster than Qiskit transpilation (measured on ibm_torino benchmark)

OpenQASM 3.0 Circuit

Sent to Nathan for analysis, then to Arvak for IQM Garnet compilation

// Generating...

⬡ Nathan — arvak.io (live)

Step 5 — Research Analysis

Nathan Circuit Analysis

The QAOA circuit is sent to Nathan on arvak.io for classification, scholar paper matching, and optimization suggestions across 6,900+ indexed arXiv papers. Live API call.

Calling Nathan API at arvak.io...

⬡ Aleta Pricing Engine & HAL Contract

Step 6 — Pricing & Contract

Transparent Cost & SLA

Before any quantum job executes, Garm computes a transparent price using the Aleta Index and generates a HAL Contract–compliant SLA. The enterprise sees the full cost breakdown and signs off on terms — including IP protection and EU data sovereignty — before hardware time is consumed.

€

Aleta Pricing Breakdown

aleta-pricing service — live formula computation

Problem typeOPTIMIZATION (QAOA)

Qubits (n)9

Circuit depth (d)—

Shots (s)1,024

Backend multiplier1.15 (IQM Garnet)

Problem multiplier0.85 (combinatorial opt.)

Runtime factor (τ)—

Base cost (C_base)—

Platform fee (22%)—

Vault surcharge€12.00 (IP protection)

Total —

◆

Aleta Index Reference

Independent benchmark — quantum computing cost transparency

The Aleta Index is Garm’s independent benchmark for quantum computing costs — the quantum equivalent of Bloomberg Terminal pricing for derivatives. It provides a market reference point so enterprises can compare across providers.

Index value€60.00 / Aleta unit

This job—

Reference benchmarksVQE H&sub2;O · QAOA MaxCut-50 · Grover 2²&sup0; · QML Iris · Heisenberg Sim

Provider price sourcepricing_current DB table (live)

📋

HAL Contract SLA

hal-contract.org — open quantum interoperability standard

SLA classEnterprise

Result delivery300s timeout, guaranteed retry

Data retentionResult only — raw shots purged after delivery

Processing regionEU-only (Scaleway Frankfurt)

ReproducibilityCircuit hash committed to audit log

IP clauseActive

Sequence dataNever leaves Vault unencrypted

GDPR basisArt. 28 (processor contract)

GxP complianceAudit trail immutable, HMAC-signed

HAL version1.0 — hal-contract.org

⬥ THE BUSINESS MODEL

Garm charges a platform fee (22% ≤€100 / 20% €100–500 / 18% >€500) plus a fixed vault surcharge for IP protection (€12/job). The Vault surcharge is the premium enterprise differentiator: classical HPC doesn’t offer integrated IP protection with GxP audit trails.

As quantum hardware improves — more qubits, higher fidelity, lower cost/shot — Garm’s brokerage value increases. Every hardware provider’s progress makes the routing problem more valuable to solve. Garm is hardware-agnostic and benefits from all of them.

⬡ IQM Garnet — Scaleway QaaS (EU)

Step 7 — Hardware Execution

IQM Garnet — 20-Qubit QPU

Garm routes the compiled circuit to IQM Garnet via Scaleway’s quantum cloud. The QPU is physically located in Finland (IQM) and operated via Scaleway’s EU data centers. EU data sovereignty guaranteed end-to-end.

Physical Qubits

~99.5%

2Q Gate Fidelity

Data Sovereignty

⟐

Submission Pipeline

Garm → Arvak → Scaleway QaaS → IQM

QASM3 → OpenQASM — Arvak normalizes circuit representation

Topology mapping — Fetch IQM Garnet star graph from Scaleway API

Arvak compile — Transpile to IQM native gates (CZ, PRX, U3) at 1000× Qiskit speed

POST /v1/circuits — Submit to Scaleway Quantum API, Frankfurt region

Poll for results — GET result every 5s until COMPLETED (timeout 300s)

API

Scaleway QaaS Request

Payload sent to api.scaleway.com/qaas/v1alpha1

{
  "name": "bnt-candidate-selection",
  "platform_id": "p-garnet-20q",
  "circuit": {
    "serialization_type": "OPENQASM_V3",
    "circuit_serialized": "<compiled QASM3>"
  },
  "options": {
    "shots_count": 1024,
    "backend": "garnet"
  }
}

⬥ ARVAK / GARM STACK OVERVIEW

The four layers of the Valiant Quantum stack collaborate on every job:

ARVAK — Compiler Layer

Quantum circuit compiler. Transpiles QASM3 to native hardware gates 1000× faster than Qiskit. Supports 7 backends: IQM, IBM, D-Wave, Quantinuum, IonQ, Rigetti, Simulator.

GARM — Brokerage Layer

The enterprise API. Classifies problems (18 problem types, AI-powered), routes to optimal hardware, handles pricing, SLA, GDPR vault, audit trail. The “Stripe for quantum.”

NATHAN — Research Layer

AI research assistant. 6,900+ indexed papers. Analyzes circuits, recommends hardware-specific optimizations, matches to published algorithms. Lives at arvak.io/nathan.

HAL CONTRACT — Standard Layer

Open quantum interoperability standard (hal-contract.org). Defines how quantum computers talk to orchestration systems. European standardization play. Every Garm SLA is HAL-compliant.

⬡ Measurement Interpretation — Phase I Candidates

Step 8 — Results

Quantum Candidate Selection

The most frequent measurement bitstring maps back to selected candidates. Each bit position corresponds to a vault-cleared candidate. Results shown are from a 1,024-shot IQM Garnet simulation matching hardware noise characteristics.

Candidates Selected

Phase I advancement batch

Selection Probability

of 1,024 measurement shots

▨

Bitstring Distribution

Top measurement outcomes (1,024 shots)

✓

Selected Candidates — Phase I Batch

From best measurement bitstring — anonymized IDs resolved post-vault

Rank	Candidate ID	Immunogenicity	Stability	Cost Index	Selection Prob.

⬥ WHAT GARM DELIVERS

The R&D team submitted 12 candidate IDs and 3 numeric scores per candidate. They received 3 prioritized candidates for Phase I — without exposing a single proprietary sequence, without managing quantum hardware subscriptions, and with a full GxP-compliant audit trail. The quantum computation ran on European infrastructure, never leaving EU jurisdiction.

⬡ End-to-End Summary

Step 9 — Pipeline Summary

Full Pipeline Report

Complete end-to-end trace: R&D batch submission through Vault IP protection, QUBO construction, QAOA circuit generation, Nathan research analysis, and IQM Garnet execution on European quantum hardware.

⬥ PIPELINE SUMMARY

⬥ LIVE vs. SIMULATED

LIVE Steps 1–5: Vault processing, QUBO construction, QAOA circuit generation, and Nathan analysis are computed live in this demo. The Nathan API call goes to the real service on arvak.io.

HARDWARE-READY Step 6: The generated QASM3 circuit is valid and ready for submission to IQM Garnet via Scaleway QaaS. Requires Scaleway API credentials and Garnet QPU availability.

SIMULATED Step 7: Measurement results are from a noise-model simulation matching IQM Garnet characteristics (CZ gate fidelity ~99.5%, T1 ~50μs). Real hardware results will vary.

⬥ THE SCALING ARGUMENT

This demo selects 3 from 9 candidates on IQM Garnet (20 qubits). Today’s demo is a proof of structure, not a claim of quantum advantage at this scale.

The same Garm pipeline — without code changes — routes to:

20 qubits

IQM Garnet (today)

9–18 candidates

133 qubits

IBM Heron (ibm_torino)

100+ candidates

5,000+ vars

D-Wave Advantage

500+ candidates (annealing)

Garm selects the optimal backend automatically based on problem size, qubit estimate (from ONNX classifier), and backend availability. The R&D team never needs to know which hardware ran their job.