Benchmarks

How we compare. Honestly.

NOVA-3 is not training a new structure-prediction model — we wrap best-of-breed open foundations and add the workflow, library, validation, and quantum layers on top. Hard numbers benchmarking NOVA-3 vs each model below publish Q3 '26.

What's being measured · Q3 '26
  • Single-chain folding accuracy: RMSD on a held-out antibody structure set vs. PDB ground truth
  • Multi-specific complex prediction: interface RMSD and DockQ on known bispecific co-crystal structures
  • Ranking consistency check: rank-order (Spearman) agreement between the module-anchored affinity estimate and SPR-measured affinity on a held-out set, reported with n. Engineered-window arms (attenuated CD3 / conditional 4-1BB, where the design target is a range, not "tighter is better") are assessed against their target window, not monotonic correlation. This is a retrospective consistency check, not a claim of de-novo predictive calibration.

Q3 '26 is a target, not a committed date. Results publish when ready, not when the calendar says.

Capability matrix: NOVA-3 versus general-purpose folding models NOVA-3 wraps general folding models on raw fold accuracy and differentiates on composition, provenance, quantum corrections, and wet-lab data. CAPABILITY qualitative — not a benchmark NOVA-3 orchestration platform General folding models AF3 / Chai / ESMFold class of structure predictors Raw single-chain fold accuracy state-of-the-art structure prediction wraps best-of-breed state of the art ↑ SHARED FOUNDATION NOVA-3 DIFFERENTIATION ↓ Multi-specific composition curated module library — design, not just predict predict, not compose Cryptographic provenance Evidence Bundles, timestamp-anchored out of scope Quantum active-space corrections targeted refinement — no advantage claimed not offered Wet-lab DBTL data flywheel design–build–test–learn feedback into models no wet-lab loop yes wraps / partial not in scope qualitative comparison · no performance metrics implied
Figure — Where NOVA-3 differs. On raw single-chain fold accuracy, NOVA-3 does not out-fold AF3, Chai, or ESMFold — it wraps those best-of-breed predictors rather than beating them. Its differentiation sits elsewhere: multi-specific composition from a curated module library upstream, and cryptographic provenance, quantum active-space corrections, and a wet-lab DBTL data flywheel downstream. Strictly qualitative — no benchmark numbers are claimed.
Capability
NOVA-3
AlphaFold 3
Chai-1
Boltz-2
ESM-3
Single-chain folding accuracy (CASP15 monomer)
Q3 '26
SOTA
~AF3
~AF3
Good
Multimer complex prediction (AbDb / SAbDab)
Q3 '26
SOTA
SOTA
~AF3
—
Direct affinity prediction (Kd)
Q3 '26
No
Indirect
Yes (native)
No
Multi-specific composability (3+ arms)
Designed for this
Possible
Possible
Possible
No
Module Library composability (curated · multi-class)
Yes — unique
No
No
No
No
Quantum active-space correction
Yes — unique
No
No
No
No
Cryptographic provenance (OpenTimestamps)
Yes — unique
No
No
No
No
Wet-lab validation flywheel
Yes
No
Internal
No
No
Open-source weights
Wraps OS
No (commercial)
Yes · Apache-2.0
Yes
ESM C yes
Carbohydrate / glycan active-space
Yes — quantum-corrected
Limited
Limited
Limited
No

Honest framing: on raw folding accuracy, AlphaFold 3 and Chai-1 set the bar. NOVA-3 is not trying to beat them — we use them as building blocks. Our differentiation is upstream (multi-specific composability via the Module Library) and downstream (quantum-corrected refinement + cryptographic provenance + wet-lab flywheel). A pharma team using NOVA-3 gets AF3-grade folding plus the workflow that closes the loop to in-vitro PoC.

See the platform → Try live prediction