Responsibility

A short, honest account of what RataFold tries to do, what it does not, and the choices behind the way we publish it.

What we're trying to do

AlphaFold and the structures it predicts are the foundation we build on; without that work, this project would not exist. AlphaFold gives the most-likely shape of a protein. We try to be useful in a small, complementary way — surfacing a curated set of alternative shapes that same backbone could plausibly adopt, with energies, structural distances, and an honest confidence check on each.

The hope is that these alternatives are useful starting points for downstream work — molecular dynamics validation, docking exploration, thinking about binding states or misfolded forms. We are explicit that they are candidates, not verified structures.

What we publish, per protein

The native AlphaFold prediction — full-atomic, with per-residue confidence preserved.
A ranked set of low-energy variants — full-atomic, with energy, RMSD to native, and an AlphaFold-confidence cross-check.
The metadata to interpret each one (stability label, disulfide preservation, category, AF reality-check verdict).

Released under CC-BY 4.0 — free to use, free to redistribute, free to integrate into any downstream pipeline. No login, no API key.

The approach, briefly

The search engine behind these variants is kept private while the outputs, schema, viewer, and audit framework are open. The contribution we hope to make is the combination of four ordinary ideas put together for this purpose:

A structured state space. Each protein is mapped into a discrete, finite, reproducible space whose size scales as 10^{(constant × residue-count)} — roughly 10⁸⁸³ for insulin and 10⁶¹⁸⁴ for the Amyloid-β precursor.
Sampling, not enumeration. Spaces of this size cannot be enumerated by any current computer. We sample them, biasing toward energetically reasonable neighborhoods and keeping a rolling pool of the most interesting candidates.
Atomic-grade reconstruction. Every accepted sample is rebuilt as a full-atom structure (backbone + sidechains) using the per-residue geometry measured in the AlphaFold model, then scored with a multi-term force field.
AlphaFold-aware honesty check. Each variant carries a verdict comparing it against the regions AlphaFold was most confident about (plDDT ≥ 70). Variants that preserve those regions earn an AF ✓; ones that scramble them are flagged so users can prioritize accordingly.

None of these ideas is new on its own. We are simply trying to put them together carefully and publish what falls out.

Why this is computationally hard

Trying to walk through a 10⁵⁰-conformation space by brute force is not feasible on current hardware. Some rough numbers, for context:

Hardware	Time to evaluate 10⁵⁰ conformations
Modern laptop core	~3.5 × 10⁴⁰ years
Single H100-class GPU	~1 × 10³⁷ years
One exascale supercomputer	~1 × 10³¹ years
All top-10 supercomputers combined	~3 × 10²⁹ years
Universe age ≈ 1.4 × 10¹⁰ years.

Anything that wants to reach beyond about 10³⁰ in conformational space has to do so by being clever about which points to evaluate — not by being faster per evaluation. Our work sits in that "be clever" tradition rather than competing on raw compute. The 30-second-per-protein figure is a consequence of the sampling discipline, not of any privileged hardware.

Where we fit among existing tools

Every tool below does something we cannot — we are listing them with respect.

Tool	What it does well
AlphaFold	State-of-the-art single-structure prediction. The native structures we display come from AlphaFold; we do not produce predictions ourselves.
Molecular Dynamics (GROMACS, AMBER, OpenMM)	Physics-based trajectories with real timesteps. The trusted way to validate any candidate structure, including ours.
Rosetta	Structure design and ab initio sampling. Decades of refinement; the reference point for high-quality protein modeling.
RataFold	A curated set of energy-ranked atomic-detail alternatives to each AlphaFold prediction, with an AF-confidence honesty check, in roughly 30 seconds per protein.

What we don't claim

We do not claim to find the global energy minimum — only the best of the candidates we visit.
We do not claim variants are experimentally real. They are computational candidates. Real verification requires NMR, cryo-EM, or X-ray crystallography.
We do not replace AlphaFold. We extend it. The native structure on every page is the AlphaFold prediction, unmodified.
We do not claim the search is exhaustive — no method can be at this scale.
We may be wrong. If a variant disagrees with experimental evidence, the experiment wins.

How we try to be a good citizen

Open data. Every record is downloadable — JSON and multi-model PDB. CC-BY 4.0. No login, no API key, no rate limit.
Audited releases. Each publish runs through a leak check so methodology fragments don't end up in published files. The audit script is open.
Reproducible provenance. Each record carries its generation timestamp, framework version, and source AlphaFold version, so results are pinnable.
Honest reporting. Acceptance rates, constraint-rejection counts, and AF-divergence verdicts are published alongside results — including the unflattering ones.

Why the search engine stays private

The search engine is kept private; the data, schema, viewer, and audit framework are open. This separation is pragmatic, not philosophical — it lets us improve the engine without breaking downstream consumers, and lets you build on the data without depending on us being around.

If you are doing academic work that needs methodology details to publish, please reach out. We will share enough to make your study reproducible against our outputs.

Thanks for reading. We are building this carefully and welcome corrections.