SATYA Protein Verification

DeepMind gave the world prediction.
SATYA gives the world verification.

SATYA verifies predicted protein structures for structural coherence using cellular sheaf cohomology. Deterministic. No GPU. No training data. Run any structure below and get a signed receipt.

SATYA Protein is the structural biology application of the Invariant verification stack.
Generated is not verified. Same engine. Different evidence. Same receipt.

Upload a structure Try demo cases Download SATYA

The Ladder

Each protein proves a different architectural property.

Same engine. Same sheaf. Same receipts. From a 10-residue peptide to the largest human protein. Each entry in the ladder adds a structural challenge the previous entries did not have.

#	Protein	PDB	Residues	Time	RMSD	Verdict
1	Chignolin	1UAO	10	0.0 s	0.09 A	SAFE
2	Insulin	4INS	51	0.2 s	0.03 A	SAFE
3	BPTI	5PTI	58	0.2 s	0.04 A	SAFE
4	Aβ42 fibril	5OQV	126	0.9 s	0.02 A	SAFE
5	Spike RBD	6M0J	194	1.7 s	0.03 A	SAFE
6	TIM barrel	1YPI	247	2.4 s	0.04 A	SAFE
7	Titin (300 Ig)	1WAA x300	26,700	22.0 s	per-domain	SAFE

What Each Structure Proved

Infrastructure, not hype.

1ChignolinLoop closure and beta-hairpin geometry
2InsulinMulti-chain architecture with cross-chain disulfide bonds
3BPTIDisulfide-rich topology with long-range cysteine networks
4Aβ42 fibrilDisease-relevant interchain beta-sheet aggregate topology
5Spike RBDPathogenic topology with flexible receptor-binding loops
6TIM barrelO(n) scaling on the most common fold in biology
7TitinDomain-decomposed verification of the largest human protein

Upload & Verify

Upload a structure. Get an obstruction report.

Not a confidence score. A reproducible obstruction map. Every run produces a signed, replayable receipt. Every obstruction signature can become an atlas entry.

Drop candidate structure here Accepts .pdb, .cif, .mmcif

Reference

Use PDB IDFetch reference from RCSB

Upload reference file

Self-consistency onlyNo reference. Checks internal structural coherence.

Runs on local hardware. Structures are not stored after verification. Nothing is transmitted to any cloud service.

Live Verification

Click a structure. The engine runs. See the result.

This is not a mockup. Each click calls the SATYA verification engine, runs sheaf-cohomology verification, and returns a signed receipt. Sixteen cases range from native structures to catastrophic failures. Try the Corrupted Decoy first.

Checking engine... Click any case below to run a live verification.

Native vs Native PDB 1UAO Model 1 against itself. Perfect self-consistency. Click to verify

NMR Conformer Model 2 from the NMR ensemble. Real variation, still valid. Click to verify

Corrupted Decoy Residues 4-7 displaced by 6 A. Controlled corruption. Click to verify

Real Partial Fold September 2025 MD simulation output. Strands never closed. Click to verify

Extended Chain Fully unfolded starting chain. Global structural failure. Click to verify

Frozen Trajectory 500 frames, zero atomic motion. Preflight rejects it. Click to verify

Insulin: Native 4INS chains A+B. 51 residues, 3 disulfide bonds. Click to verify

Insulin: Broken SS Chain B displaced. Interchain disulfides broken. Click to verify

BPTI: Native 5PTI. 58 residues, 3 long-range disulfide bonds. Click to verify

BPTI: Broken SS Residues 28-40 displaced. CYS30-CYS51 broken. Click to verify

Abeta42: Fibril 5OQV. 3 chains, 126 residues, fibril topology. Click to verify

Abeta42: Broken Middle chain displaced. Interchain contacts broken. Click to verify

Spike RBD: Native 6M0J chain E. 194 residues, 4 disulfides. Click to verify

Spike RBD: Broken Loop RBM loop displaced. C480-C488 SS broken. Click to verify

TIM Barrel: Native 1YPI. 247 residues. Most common fold in biology. Click to verify

TIM Barrel: Unsprung Barrel opened. Beta-strands 4-5 displaced. Click to verify

The Protocol

Break it. Detect it. Repair it. Certify it.

For each protein in the ladder, we ran the same four-step protocol. The structures are real PDB entries. The decoys are synthetic corruptions designed to test specific failure modes. The refinement is deterministic. Every step produces a signed receipt.

1. Parse

Extract residue-level geometry from the PDB file. Multi-chain, disulfide bonds, backbone dihedrals, SG atom positions. No external dependencies beyond NumPy.

2. Break

Generate a structurally corrupted decoy. Displace residues, break disulfide bonds, unspring barrels, shear interchain contacts. Each decoy targets the specific structural property that protein tests.

3. Verify

Build a cellular sheaf from the constraint graph. Compute the constant-sheaf connectivity diagnostic (Laplacian kernel), the spectral gap (lambda_2), and Dirichlet energy. Localize obstructions to specific residue regions. Emit verdict: SAFE, REVIEW_REQUIRED, or UNSAFE.

4. Refine

Minimize structural inconsistency over the sheaf constraint graph using analytical gradients. Spatial cell list for O(n) steric detection. Checkpoint verification with SATYA receipts. Track whether refinement improves structural coherence and reduces localized obstruction.

The Math

Cellular sheaf cohomology on protein constraint graphs.

SIGMA maps a protein structure into a cellular sheaf over its residue-level constraint graph. Residues carry local geometric state. Constraints such as backbone bonds, native contacts, disulfide bonds, and steric exclusions define compatibility conditions between neighboring regions.

The sheaf Laplacian's spectrum reveals global structural coherence. The kernel dimension encodes topological invariants. The spectral gap (lambda_2) measures how well-connected the constraint network is. The Dirichlet energy of the deviation section measures total structural inconsistency. Local constraint violations that no local adjustment can reconcile raise the coboundary and concentrate that energy on the affected residue regions.

For a native structure verified against itself: deviation = 0, coboundary = 0, Dirichlet energy = 0, coherence = 1.0. For a corrupted decoy: nonzero deviation, inflated coboundary, and Dirichlet energy that localizes the damage to specific residue regions. The result is deterministic, auditable, and reproducible.

SATYA adds an independent deterministic verification layer to protein-structure pipelines. Small local checks can run in microseconds; full protein receipts complete in seconds on CPU.

Run It On Your Own Structures

Add a verification gate before the expensive step.

You just ran it. The demo above is the real engine, not a mockup. Download SATYA and the same verification runs on any PDB or mmCIF structure on your own machine. Nothing leaves your network.

Before synthesis Rank candidates by structural coherence. Flag failures before wet-lab spend.

Before partner review Attach a signed verification receipt to every structure in a partner package.

Before publication Independent structural verification for supplementary materials and depositions.

Before downstream ML Filter incoherent structures from training sets, databases, and prediction pipelines.

Download SATYA Try another structure

Request a private structure audit

DeepMind gave the world prediction.SATYA gives the world verification.