Protein structure · open showcase

Fold-Protein

Fold-Protein investigates whether a protein backbone can be constructed from the project’s structural rules without learned weights. This page makes one committed result inspectable: a 76-residue ubiquitin reconstruction, its coordinates, residue contacts and the measurements produced when it is compared with the experimental structure.

Checking the source machine…
0.989121repository TM-score for ubiquitin
0.260858 ÅCα distance-matrix RMSD
76residues in the replayed witness
0trained weights or gradient steps

The constructed witness

Read the backbone in three dimensions

The rotating trace is the committed 76-residue Cα witness. It lets you inspect the geometry that the release replay scores against experimental ubiquitin.

Loading the interactive replay…

Residue contact explorer

Move through the folded witness

Scrub across all 76 Cα coordinates. The map lights every non-neighbour residue within 8 Å of the selected point, exposing the packed geometry behind the score.

Proof by construction

From experimental target to byte-exact witness

This release constructs the ubiquitin backbone on the theorem-forced 24×24 dihedral lattice at near-experimental resolution. The native coordinates forward-force the successful state path, and the same theory now drives the advance into blind sequence-to-structure selection.

TargetTake the experimental 1UBQ Cα coordinates
LatticeMap construction onto 24×24 dihedral states
WitnessCommit all 76 states and source hashes
ReplayRebuild the PDB and score it independently

What the score means

Super Parity, with its evidence visible

Construction

The native fold is contained

The 24-lattice construction reaches the experimental backbone with a TM-score of 0.9891211351 and 0.2608575408 Å distance-matrix RMSD.

  • Committed state path and witness PDB
  • Exact source hashes and release manifest
  • Independent byte-exact replay
Expansion

Blind selection from sequence

The next stage selects a complete structure directly from its sequence using the same declared rules, without fitting values to the known answer, and preserves each blind run for comparison.

  • Zero trained networks or gradient fitting
  • Named, replayable blind-run protocols
  • Every advance committed as a new evidence surface

Evidence package

Every number points to a file

manifest.jsonsequence, lattice states, hashes, and metrics
replay.pyindependent release reconstruction and score check
witness.pdbthe constructed three-dimensional backbone
1ubq.pdbthe experimental development target
NeRF buildercommitted backbone construction code
TM scorerrepository scoring implementation

Reproduce it

Replay the witness

python3 verify/replay_ubiquitin_24_lattice.py

Read the work

The method and the result

The current V3 biology publication places this construction inside the wider model. The repository documentation preserves the reconstruction method, comparison scripts and exact witness files used on this page.

Source documentationRead the repository guide on GitHub