The state of FoldEngine on 25 October 2026. This page follows the work in progress and changes with it: every result below has already been measured and recorded, failed tests included. Performance figures are not here: they are on the dashboard and in the registry, with their limits next to them.
FoldEngine is the app. Underneath are four engines, kept apart so that each one can be checked on its own.
Molecular dynamics on the Apple silicon GPU
Simulates proteins in explicit water, written in Metal. The same input given twice gives the same bits. Production runs use a frozen bundle, binary and kernels together, each with its sha256. A second program on the CPU recomputes its forces to check it; it does not yet redo a whole trajectory.
Folding, on your Mac
Predicts the structure of a protein from its sequence, on the Mac itself: the sequence does not leave the device. A predicted structure remains a prediction, to be checked before you rely on it.
Research you can verify
Takes a question through competing hypotheses to a prediction written in advance, then to the run and its review. Every package is deposited with the fingerprints of its files and passes a check before it starts. The claims ledger closes every day with a root that, since 25 October, you can query from here.
Protein–ligand poses
The engine that looks for how a small molecule sits on a protein. It is being built inside the research work and has no surface in the app yet. We will not say what it can do until it has a test behind it.
Mac, iPhone and iPad
The Mac computes; iPhone and iPad follow the runs live through your private iCloud. No account, no ads, no tracking. It is not on the App Store yet. New features are tried with non-expert people before they ship.
Every test has its prediction and its threshold, written in advance. Here is the result, including when the prediction fails. A test marked “not evaluated” has not passed: something needed to run it is missing, and we say what.
The benchmark run behind the promoted figure was repeated: it comes back byte for byte. That means it is reproducible, not that it is faster or more correct.
The SHAKE test the frozen engine runs at start-up reads outside its buffer twice: the prediction “no out-of-bounds access” fails, the prediction “the run is valid” passes. Whether it touches the bits of the runs is still to be worked out. Another SHAKE defect, which only distorted the residual printed at start-up, is fixed in the live source; the frozen bundles stay as they are.
Seventeen checks on the engine's forces: 11 pass, none fail, 6 were not evaluated.
The sum of the drifts of each energy term matches the drift of the total, in all eight replicas and in four tests out of four. Limit: the total is the same sum of the same terms, so the test checks how the terms are recorded, not the physics.
The bonds of the test system were 0.18 kcal/mol away from the published files. Four competing hypotheses, each with the measurement that could knock it down: three fail, the fourth holds. The cause is the rounding of the coordinates to three decimals in the Protein Data Bank file.
A check that places a structure among a thousand experimental structures from the Protein Data Bank passes on chirality and atomic clashes. The tests on backbone angles and rotamers are not evaluated: the tool does not compute them yet.
On the alanine dipeptide the sampling and analysis tests pass; one is not evaluated, because a reference value to read at the source is missing. An erratum is written: a surface from an earlier version belonged to the D enantiomer. Until that test exists, the map remains an internal tool.
One criterion of the map that groups proteins by similarity fails on two pairs. It will be tried again after removing the purification tags from the sequences, with a new prediction.
A simulation state taken on an M4 Pro and resumed on an M2 reaches the same final file. This holds for this test; across GPU families identity is not assumed, it is compared.
Two completed studies were closed by re-reading the fingerprints of every signed file: in one, 180 fingerprints and no difference. From now on a new package does not start without first declaring how much computing it needs.
The app used too much energy while idle. The two causes suspected first were innocent; the real one was a figure animated every second in the overview. With it removed, it was measured again, before and after, on the same build.
A programme that screens a hundred features and fifty innovations. Of the thirteen features chosen first, all have taken their first step: six pass or hold, six fail as written, one is under review. The twelve innovations written up as specifications have taken their first step, and more than one has failed. The rest are waiting.
Every result has its deposited prediction, its commit and, where it changes a published figure, its erratum. The entries of the public registry carry only fingerprints and status; the texts of the claims stay inside the project.
Possible uses, not promises. For each: what exists today and what is missing.
FoldEngine is not a medical device, and its results can be wrong. Before trusting one, read what it does not do.
FoldEngine