Platform

A five-layer stack for multi-specific antibody design.

Quantum-enhanced AI fused with a wet-lab data flywheel. Each layer feeds the next; each candidate carries a cryptographically anchored evidence trail. This page is the platform deep dive — the homepage version is the executive summary.

NOVA-3 architecture: five layers from Module Library to Application, with a wet-lab DBTL flywheel feeding back into the data layer. INPUT — TARGET, FORMAT & CONSTRAINTS 01 DATA Module Library 1,200+ VH/VL · 80+ targets 02 COMPUTE NOVA-Compute IBM Heron r3 · Quantinuum Helios 03 MODEL Generators & Heads Boltz-2 · ESM-3 · RFantibody 04 DBTL Wet-lab Loop SPR · BLI · ADCC · in-vivo 05 APPLICATION Application layer Dashboard · Mol* · evidence OUTPUT — TOP-24 RANKED CANDIDATES + OTS-ANCHORED EVIDENCE BUNDLE DBTL FLYWHEEL · +0.8% / CYCLE
Figure 1 — NOVA-3 stack. Each layer feeds the next, top to bottom. The wet-lab DBTL loop closes back into the Module Library: every assay result re-trains the corpus, compounding accuracy ~0.8% per program per cycle.
Request to delivery: input constraints, module retrieval, generate candidates, quantum active-space refinement (enrichment), rank to top-24, wet-lab DBTL validation, evidence bundle, and delivery — with a DBTL feedback loop from the wet-lab stage back into the Module Library. STEP 01 · INPUT Target request ANTIGEN · FORMAT · CONSTRAINTS CLDN18.2 · TRISPECIFIC · Kd<100pM STEP 02 · RETRIEVE Module Library PULL VALIDATED PARTS VH/VL · LINKERS · HINGES · EFFECTORS STEP 03 · GENERATE Candidates GENERATORS ASSEMBLE MULTI-SPECIFIC DESIGNS STEP 04 · ENRICHMENT Quantum refine CVaR-VQE ACTIVE-SPACE CORRECTS HARD CHEMISTRY ONLY STEP 05 · RANK Rank → top-24 RANKING HEAD SCORES Kd · pTM · DEVELOPABILITY STEP 06 · VALIDATE Wet-lab DBTL DESIGN · BUILD · TEST · LEARN SPR · BLI · ADCC STEP 07 · ATTEST Evidence Bundle PROVENANCE-ANCHORED SEALED RUN · ASSAY · PARAMS STEP 08 · DELIVER Delivery RANKED SET · PDB STRUCTURES BUNDLE · SUMMARY DBTL FLYWHEEL · RESULTS RE-TRAIN LIBRARY REQUEST → DESIGN · GENERATE · ENRICH FORWARD PASS RANK · VALIDATE · ATTEST → DELIVER RANKED SET PDB FOLDS EVIDENCE BUNDLE SUMMARY Quantum step corrects only hard chemistry — not a standalone designer. Schematic; “top-24” and “Kd<100pM” are illustrative examples. DESIGN / GENERATE QUANTUM ENRICH WET-LAB LOOP EVIDENCE / DELIVERY
Figure 2 — Request to delivery. One target request flows through module retrieval, generation, and a quantum active-space enrichment that corrects only the hard chemistry, then ranking to a top-24, wet-lab DBTL validation, a provenance-anchored evidence bundle, and final delivery — while assay results feed back into the Module Library so each program compounds.
01
Layer · Data

Module Library

1,200+ pre-validated VH/VL antibody pairs across 80+ therapeutic targets in oncology, neurology, and senolytics. Plus a curated library of linkers, hinges, and effectors. Each module is provenance-stamped, partner-IP-segregated, and schema-typed.
1,200+ VH/VL pairs 80+ targets Partner-IP carveouts
02
Layer · Compute

NOVA-Compute

AWS GPU for classical inference, IBM Heron r3 (ibm_boston · 156Q) and Quantinuum Helios (98Q, all-to-all) for CVaR-VQE / QAOA, Modal & Runpod for burst. Every job logged with backend, shots, calibration snapshot, and cost.
Quantum-hardware-explicit Simulator fallback Cost ledger
03
Layer · Model

Generators & Heads

Best-of-breed open foundations wrapped for multi-specific design: Boltz-2 (MIT license; commercial use permitted) and ESM-3 (EvolutionaryScale Cambrian license; commercial use requires a separate agreement with EvolutionaryScale), RFantibody (de-novo binders). AlphaFold 3 used for internal research benchmarking only under Isomorphic Labs' non-commercial research-use terms; production design runs use Boltz-2 / Chai-1. Three program-specific heads on top: ranking, structural confidence, in-silico Kd.
Three heads Mol* viewer Top-24 ranked
04
Layer · DBTL

Wet-lab Loop

Partner-lab adapters with IP-carveout enforcement. SPR/BLI, ADCC dose-response, in vivo efficacy. Every assay result flows back into the Module Library; flywheel learns +0.8% accuracy per program per cycle.
+0.8% / cycle SPR · BLI · ADCC In-vivo PoC
05
Layer · Application

Application layer

Next.js dashboard (run inspector, candidate ranking, structure viewer, cost ledger). Streamlit retained for engineering iteration. Public marketing site. Cream-mode BD data-room generator producing PDF + interactive room + zipped archive per partner.
Next.js 15 Mol* embed Signed S3 URLs
Module Library · live schema

A schema-typed corpus of curated modules.

1,200+ pre-validated VH/VL antibody pairs across 80+ therapeutic targets, plus linkers, hinges, and effectors. Validated subset is proprietary; the schema and exemplar entries are public.

Compose, not predict. ASSEMBLE A MULTI-SPECIFIC FROM VALIDATED, SCHEMA-TYPED PARTS MODULE LIBRARY curated · versioned · provenance per part BINDERS VH / VL pairs MOD-0492 MOD-0617 MOD-0701 LINKERS flexible peptide G4S (G4S)x3 EAAAK x2 HINGES pairing / stability IgG4-S228P KiH-T366W EFFECTORS costim / recruit 4-1BB costim CD3ε maps to ASSEMBLED · TRISPECIFIC build NV3-TRI-0231 · schema OK HINGE IgG4-S228P VH/VL 0492 arm A · target-1 VH/VL 0617 G4S arm B · target-2 EFFECTOR 4-1BB costim · agonist 3 specificities from curated, traceable parts 1 TRISPECIFIC = 7 VALIDATED MODULES · SCHEMA-CHECKED · PROVENANCE PER PART
Figure — Compose, not predict. NOVA-3 does not predict one molecule end-to-end; it assembles a multi-specific from schema-typed library parts — binders, linkers, hinges, effectors — so every chip carries its own provenance. Module IDs and the count of 7 are illustrative of the composition approach, not a real assembly record.
1,200+
Pre-validated VH/VL pairs
80+
Therapeutic targets
OTS
Provenance · every module
Module ID
Name
Target
Class
Origin
Status
MOD-0492
VHH-CLDN18.2-ECL1-α
CLDN18.2
Binder · VHH
Internal
Validated
MOD-0881
GD3-glycan-CB-quantum-12
GD3
Binder · novel
Quantum-corrected
Validated
MOD-1217
Linker-G4S-flex-stable-v6
—
Linker
Computational
Computational
MOD-1218
Hinge-IgG4-S228P
—
Hinge
Literature
Validated

Public exemplar subset shown. Full Library access requires Enterprise tier. Backed by GET /v1/library/search.

Modality gallery · formats

The molecular formats we design against.

NOVA-3 assembles candidates across the antibody and binder formats below. These are architectural cartoons of each format's domain layout — chain topology and valency, not folded model outputs. Live, pLDDT-coloured structures appear in the Predict console and the Playground.

IgG · ~150 kDa

Full-length antibody. Bivalent, monospecific; Fc confers half-life and effector function.

Bispecific IgG · asymmetric

Two distinct paratopes on one scaffold (knob-in-hole / charge-pair). Engages two targets at once.

Fab · ~50 kDa

Single antigen-binding fragment — paired VH-CH1 and VL-CL. No Fc, monovalent.

scFv · ~25 kDa

VH and VL on a single chain via a flexible linker. Compact building block for CARs and fusions.

VHH / nanobody · ~15 kDa

Single variable domain from heavy-chain-only antibodies. Smallest format; stable, reaches cryptic epitopes.

Tandem scFv · no Fc

Two scFv strung in series (BiTE-like). Two binding sites, small size; redirects effector cells.

Schematics, not outputs. Domain blocks are coloured by chain class — heavy copper, light verdigris, paratope sodium, second specificity magenta. Folded structures are coloured separately by pLDDT confidence.

NOVA-Compute

Quantum-augmented. Provenance-anchored. Pharma-aligned.

CVaR-VQE on IBM Heron r3 (ibm_boston · 156Q) for active-space corrections (production deployment Q3 '26); QAOA on Quantinuum Helios (98Q, all-to-all connectivity) for assignment problems; quantum Bayesian optimization on simulator. Every job's cost, calibration snapshot, and result is logged. Hardware fraction reported transparently. No claim of demonstrated quantum advantage in ground-state chemistry (cf. Lee et al., Nat Commun 2023); we engage with the open question.

NOVA-Compute allocation — illustrative split of where each compute class is used A horizontal proportional bar split into Classical GPU (the large majority), Quantum hardware (a small slice), and Simulator (a small slice), each annotated with its workload. NOVA-Compute · allocation by workload where each compute class runs · illustrative split (not measured share) scale 0–100% PROPORTIONAL ALLOCATION Classical GPU (AWS) generation · folding inference · ranking ~80% Quantum hardware ~12% Simulator ~8% quantum + simulator ≈ minority (~20%) WHAT EACH LANE DOES Classical GPU — AWS sequence generation, structure-folding inference, candidate ranking Quantum hardware CVaR-VQE active-space corrections · QAOA assignment Simulator quantum Bayesian optimization on simulator (small, exploratory) CLASSES classical quantum HW simulator Every job logs backend, shots, calibration & cost · hardware fraction reported per job · no quantum-advantage claim. illustrative split
Figure — Compute allocation. Classical GPU on AWS carries the bulk of NOVA-Compute — generation, folding inference and ranking. Quantum hardware (CVaR-VQE active-space corrections on IBM Heron r3, QAOA assignment on Quantinuum Helios) and the quantum-Bayesian-optimization simulator are small, targeted slices, not the workhorse. Proportions are an illustrative workload split, not a measured share, and imply no demonstrated quantum advantage.
Quantum spend · 30 days
$8,420
−12.4% vs prior 30d
Classical spend · 30 days
$3,180
+4.1% vs prior 30d
Hardware fraction
61%
Simulator fallback 39%
Job ID
Routine
Backend
Shots
Cost
Status
QJ-9871
CVaR-VQE · GD3 active space
Heron-3
8192
$420
✓ Done
QJ-9872
QAOA · linker assignment P1
H-Series
4096
$310
✓ Done
CJ-44021
ESM3 inference · top-24 P2
A100·24
—
$92
✓ Done
QJ-9873
qBO · CDR3 P3 — running
Sim-512
16384
$184
In flight
CJ-44022
AF3 fold · 24 candidates P3
H100·8
—
$74
✓ Done
NOVA-Dashboard

The internal platform UI.

Next.js 15 + tRPC + Mol*. Same FastAPI backend the Streamlit fallback uses. Behind Cloudflare Access; role-based ACL for engineering, BD-preview-guests, and partner labs.

Programs / NOVA-3·P3 / Run · RUN-P3-2026-0517-3F2A

Run inspector · top 24 candidates

EVB-P3-2026-0517-3F2A · 8,420 candidates evaluated · CVaR-VQE corrected
Top Kd (pM)
42.1
vs target <100 pM
Struct conf
0.94
pTM · top quartile
Quantum cost
$840
Heron-3 · 16k shots
Evidence
Anchored
Block 944,517 · OK
Rank
Candidate · CDR3 excerpt
Kd (pM)
Conf
Q-flag
Validated
01
CAND-P3-LEAD-08 · ARDYWG...QGN
42.1
0.94
VQE+
Yes
02
CAND-P3-LEAD-12 · ARGFTW...EVH
58.7
0.92
VQE+
Yes
03
CAND-P3-LEAD-19 · ARSPKV...NLY
71.4
0.89
VQE
Pending
04
CAND-P3-LEAD-21 · ARGYLN...QQS
88.2
0.87
—
Pending
05
CAND-P3-LEAD-27 · ARWEGY...HTF
94.1
0.85
VQE
No
Try live structure prediction → See benchmarks