Structural provenance infrastructure
Proteins move, and a drug binds to one of the states — so a single predicted shape is usually the wrong one. Every conformation we return is physically valid, carries a proof anyone can check, and lands in an index you can search by shape.
No alignment needed Independently verifiable Searchable by shape
Problem to Solve
AlphaFold2, ESMFold and RoseTTAFold2 keep getting cheaper, faster and more accurate. The prediction itself is becoming a commodity. Four things none of them do — and each gap is a product line.
A structure arrives and you trust it. No proof it satisfies physical constraints, none that it came from the model claimed, none that it sits in the native basin. An unverified assertion.
What we do
Every result carries a proof of physical validity and model identity that a third party can check without our infrastructure, and without learning the coordinates or the model. That is also what makes a prediction sellable rather than merely publishable.
Proteins are dynamic ensembles. Drugs bind to conformational states, not to a snapshot. A single predicted structure is frequently the wrong one for the binding event you care about.
What we do
N conformations per sequence, each proven. A different output rather than a faster version of the same one, and the one ensemble docking actually needs.
Minutes per prediction. Engineering needs upwards of 10,000 per design cycle, which puts one iteration in days. Cycle time is the bottleneck, not accuracy.
What we do
Milliseconds instead of minutes closes the loop: generate candidates, predict every structure, verify, filter, iterate. At minutes per prediction that loop does not run at all. The speed difference is a change in kind.
200 million public structures, none searchable by shape. You cannot ask for everything within 2 Å RMSD of a given pocket. That question has no product behind it.
What we do
Shape becomes a searchable key, and every prediction inserts itself into the index on the serve path. Existing methods treat each prediction as an independent event, so nothing accrues; here the index compounds with use.
The encoding is not protein-specific: the same scheme applies to RNA folds, small-molecule conformations and protein–ligand complexes, so the index that holds proteins can hold binding poses. Cross-domain structural search becomes a matter of volume rather than research.
The Solution
Every request follows the same path, and every result leaves with three things a prediction normally lacks: more than one state, a proof of physical validity, and a place in a searchable index. What follows is what the pipeline guarantees; each step carries a collapsed technical note naming the method, for readers evaluating the science.
A sequence is sufficient. No multiple-sequence alignment, no homologs, no evolutionary history — which is why orphan and de novo proteins are in scope rather than out of it.
ESM-2 protein language model embeddings map sequence to per-residue representations. No multiple-sequence alignment, so no dependence on homolog availability.
One sequence yields N distinct conformational states rather than a single answer. The output is a landscape, and the count is a parameter the caller sets.
Embeddings project onto a learned conformational manifold (NIRVANA). Sampling the manifold yields N sets of backbone phi/psi torsions rather than a single point estimate.
Each state resolves to complete atomic coordinates — backbone and side chains — with exact geometry rather than an approximation refined afterwards.
NeRF (Natural Extension Reference Frame) converts torsions to exact backbone coordinates; side-chain chi angles are drawn from a Dunbrack rotamer library.
Every structure reduces to a fixed-length code with one property that matters: similar shapes produce similar codes. This is what makes shape a searchable key, and it works the same way on proteins, RNA folds and small-molecule conformations.
Morton (Z-order) encoding of the 3D coordinates yields a fixed-length, locality-preserving structural barcode. Near-in-space maps to near-in-code, which is what makes Hamming distance a usable shape metric.
Proves: physical constraints hold — bond lengths, bond angles, chirality, no clashes — the result came from the model claimed, and the state sits in the native basin. Reveals: nothing about the coordinates or the model itself.
A STARK proof over the barcode establishes bond lengths, bond angles, chirality and absence of clashes, plus a model hash for identity and native-basin membership. Zero-knowledge, so coordinates and model weights are never revealed.
Every result joins the searchable index automatically, on the serve path, with its proof attached. The index is not a separate product to be populated later — it compounds as a side effect of ordinary use.
Entries land in a Matryoshka lattice — a Morton-ordered B-tree supporting k-Hamming-distance neighbour queries in sub-millisecond time. Individual proofs aggregate recursively into one master proof.
Provenance at scale
Building a structural database is assumed to cost months of accelerator time. It does, if you predict the structures. We don't. The structures already exist — what they lack is provenance and an index.
Public databases already hold more than 800 million predicted structures, sitting there as unproven, unsearchable plaintext. We download them, encode each one, prove physical validity in batches, insert every entry into the index, and aggregate all of it into a single master proof.
Nothing is predicted. Everything is provenanced.
| Step | Time | Compute |
|---|---|---|
| Acquire public databases (800M structures) | 5.5 hours | Bandwidth only |
| Batch proof generation | ~2 hours | 100 cores |
| Structural encoding | ~13 minutes | 20 cores |
| Index insertion | ~13 minutes | 20 cores |
| Master proof aggregation | ~3 minutes | 1 core |
| Total | ~8 hours | ~$10 compute + $200 storage |
AlphaFold DB — 200M structures, months of accelerator time, unproven, unsearchable, static.
ESMFold Atlas — 600M structures, months of GPU, unproven, unsearchable, static.
VeraSolve — 800M structures, eight hours, proven, searchable by shape, and compounding with use.
Every prediction method that exists feeds this index rather than competing with it. That is the position: the provenance layer, not another predictor.
Products
Nobody wants a proof or an encoding for its own sake. They want faster discovery, cheaper protein engineering and structural search at scale. Four capabilities, with the group that would use each and what they use today instead.
Ask for every known structure with a binding pocket within 2 Å of a given one and get an answer in under a millisecond across the full index, each hit carrying its proof.
Who uses it: drug discovery teams doing lead identification. Today that means DALI or Foldseek — minutes per query, no provenance, a smaller database.
Structural Search API
N conformations per protein instead of one — the dynamic landscape rather than a single state. This is why single-structure predictions underperform in docking: the predicted state often is not the binding-relevant one.
Who uses it: computational chemistry groups in pharma, who currently run molecular dynamics for days per protein to get the same thing.
Ensemble Prediction API
Submit 10,000 candidate sequences, get structures back in seconds, filter on structural criteria, iterate. The same loop takes days on current tooling, which is why closed-loop design is rare rather than standard.
Who uses it: protein engineering and synthetic biology firms, enzyme design groups, anyone running directed evolution or de novo design.
Batch Prediction Pipeline
Hundreds of millions of structures, searchable by shape, every entry's physical validity checked. Clients query by structure rather than by sequence, and the index improves as it is used.
Who uses it: bioinformatics groups, academic labs, biotech. Anyone asking whether a fold has been seen before, or what else looks like this.
Index Query Subscription
Target markets
Every one of these is blocked by the same three constraints: prediction is too slow, returns a single structure, and cannot be searched by shape. Drug discovery is where we start, because the groups working there already know ensemble docking works and cannot afford to generate the ensembles.
Lead market — drug discovery
Ensemble docking hit rates from published literature (Amaro et al., 2018; Scown et al., 2020).
Today. One predicted structure, docked against. Hit rates run 1–5% because the binding-relevant conformation frequently is not the predicted one. Generating an ensemble by molecular dynamics takes roughly 20 days per protein, so it is not done at scale.
With VeraSolve. 20 conformational states per protein in milliseconds, docked against all 20. Same docking compute, 20× the conformations.
Who uses it: computational chemistry teams at large pharma. They already know ensemble docking works — generating the ensembles is what stops them.
| Market | Size | Value at stake | Constraint removed | Who uses it |
|---|---|---|---|---|
| Orphan proteins | ~8,000 targets | $80–160B est | Prediction without homologs, where alignment-based methods fall below the docking threshold | Pharma hard-to-drug programmes |
| Drug safety | $15B global | $50–100M per avoided Phase II failure | Whole-index pocket search before trials begin rather than after | Safety, regulatory affairs, CROs |
| Enzyme design | $25B global | 10× design cycles target | 10,000 candidate structures in seconds, so cycles run in a day not a fortnight | Industrial enzyme and synthetic biology firms |
| Genomics | $80B precision medicine | $500M–2B est | Structural variant classification at population scale, which no tool does today | Sequencing providers, biobanks, clinical genetics |
| Biosurveillance | $25B global | Hours → milliseconds | Structural threat assessment before the sequencer finishes | Public health and biodefense agencies |
Comparison
The useful comparison is not accuracy against accuracy. It is whether a result can be verified, whether structures can be searched by shape, whether you get an ensemble, and whether a prediction can be sold without exposing the model.
| Capability | AlphaFold2 | ESMFold | AlphaFold DB | VeraSolve |
|---|---|---|---|---|
| Alignment-free prediction | No | Yes | N/A | Yes |
| Conformational ensemble | Single structure | Single structure | No | N states |
| Verifiable predictions | No | No | No | Yes M1 |
| Structural similarity search | No | No | Keyword only | Yes M2 |
| Model provenance proof | No | No | No | Yes M1 |
| Sellable without exposing the model | No | No | N/A | Yes M3 |
| Real-time design loops | Minutes | Seconds | N/A | Milliseconds target |
| Compounding index | Independent runs | Independent runs | Static | Yes M0 |
| Encoding beyond proteins | Proteins only | Proteins only | Proteins only | Any 3D structure |
Roadmap
Each milestone has a binary gate stated as a measurable outcome. Work does not proceed until the gate passes. Figures below are targets, not results.
M0 · 2–3 weeks
Turn predicted states into full 3D coordinates, encode each structure, and insert every served result into the index automatically.
Gate: reconstruction within 1.0 Å backbone RMSD of PDB natives. Encoding and index insertion present on every serve path.
M1 · 1–2 weeks
Resolve side-chain geometry, then raise the proof from threshold checks to full geometric constraints over the encoded structure.
Gate: full-atom structures with valid stereochemistry. Proof covers bond lengths, bond angles, chirality and absence of clashes.
M2 · 3–4 weeks
Extend the model across fold classes and expose the query path: return every indexed structure within a given distance of a query shape.
Gate: GDT-TS above 40 on held-out proteins target. Similarity query returns structurally near entries with proofs attached.
M3 · 4–6 weeks
Handle sequences the model has not seen, and settle a standard proof format a third party can verify without our infrastructure.
Gate: five CASP free-modeling targets above GDT-TS 60 target. Proof format verifies end to end outside our stack.
M4 · 2–3 weeks
Run the CASP14 free-modeling set against AlphaFold2 and ESMFold, with emphasis on orphan targets where alignment-based methods degrade. Publish the proofs alongside the numbers.
Gate: published GDT-TS and wall-clock comparison, independently verifiable proofs, orphan-protein results, working search demonstration.
Structure prediction is the first application; provenance is the platform. Two kinds of collaboration would move it fastest — independent benchmarking against the CASP14 gate, and pressure-testing the proof format so results verify outside our stack. Every claim above is written so it can be falsified, which is the point.