alphaPredict

AlphaFold2's per-residue pLDDT score has quietly become one of structural biology's most-used annotations: it marks which parts of a chain the network could place and, read with care, which parts may simply have no fixed structure to place. Getting it, though, normally means either running structure prediction or looking the protein up in a precomputed database — and the database only covers sequences someone has already folded. A designed protein, a disease variant, a synthetic construct, or an organism outside the deposited proteomes has no entry.

alphaPredict removes the structure prediction from that loop. A small bidirectional recurrent network was trained to regress AlphaFold2's pLDDT values directly from amino acid sequence, so any string of residues yields a per-residue confidence profile in milliseconds on a laptop CPU. The result is a predicted prediction: not a structure, and not a claim about one, but the confidence curve AlphaFold2 would likely report, available for sequences AlphaFold2 has never seen.

The network comes from Ryan Emenecker, Daniel Griffith and Alex Holehouse in the Department of Biochemistry and Molecular Biophysics and the Center for Science and Engineering of Living Systems at Washington University in St. Louis, and was released in July 2021, days after the first AlphaFold2 proteome predictions appeared. It shipped both inside the group's disorder predictor Metapredict and as its own MIT-licensed Python package.

#Key Features

• Sequence in, confidence out: alpha.predict(sequence) takes a bare amino acid string and returns one pLDDT-scale value per residue. No multiple sequence alignment, no template, no structure, and no lookup against a precomputed database.
• A frozen network, not a fitting procedure: the default checkpoint is loaded when the package is imported, and the public interface is prediction and plotting only — there is no training entry point and nothing accepts labelled data.
• Six networks in one install: the V1, V2 and V4–V7 checkpoints all ship inside the package, but only the default V7 is reachable from the public API — predict() takes a sequence and nothing else. Running an older network means using the backend directly and rebuilding the recurrent network by hand with that checkpoint's hidden size and layer count.
• An orthogonal read on disorder: predicted confidence anti-correlates with predicted disorder, which makes it a second, independently trained opinion to consult where a consensus disorder score is ambiguous.

#Technical Details

The architecture is the many-to-many bidirectional LSTM from PARROT, the Holehouse lab's sequence-annotation training toolkit: one-hot encoded residues over a 20-symbol alphabet feed two bidirectional LSTM layers with a hidden size of 100 per direction, and a single linear head maps the 200-dimensional concatenated hidden state at each position to one regression output. That is roughly 339,000 parameters, a 1.4 MB float32 checkpoint that travels inside the wheel. Training targets were AlphaFold2 per-residue pLDDT values taken from whole-proteome predictions: the default V7 network used about 363,000 sequences from 21 proteomes spanning bacteria, archaea, fungi, plants, metazoa and apicomplexan parasites, where the earlier networks described in the 2021 metapredict paper drew on nine proteomes. Reported accuracy is a mean per-residue error of about 8% for V7, against roughly 11.5% for V1, 9.5% for V2, and 8.5–11% for the intermediate networks.

#Applications

The practical use is triage. Before committing compute to full structure prediction, a predicted confidence profile shows which regions are likely to come back well resolved and which are likely to come back as low-confidence ribbon, which is useful for choosing construct boundaries, for scanning families, and above all for designed or non-natural sequences that no structure database covers. It also became an ingredient rather than only an endpoint: inverted, rescaled predicted pLDDT was combined with consensus disorder scores to build metapredict's hybrid predictor, whose outputs were then distilled into a single faster network.

#Impact

alphaPredict's reach has been largely through metapredict, where the same network remains available as the non-default pLDDT_version=1 option alongside a newer pLDDT network the authors describe as better on every metric. Its own limits are worth stating plainly. It emits one scalar per residue with no embeddings and no generative capability, so it annotates rather than represents; it inherits whatever biases AlphaFold2 carries, and low confidence is evidence about a structure prediction rather than a measurement of disorder. It has never been published as a standalone artifact, the pLDDT training set was not released, and feature development on the package ended with version 1.3 in October 2024.