Metapredict

Ask several disorder predictors whether a stretch of a protein is folded and they will not agree. That disagreement is why metapredict began, in 2021, as a consensus method: run a panel of established predictors and combine their scores. Consensus buys robustness at the cost of runtime, and the best predictors take minutes to hours per protein — fine for one protein of interest, hopeless for a proteome.

Metapredict V3 removes the consensus machinery from inference entirely. It is a single network trained to reproduce a combined disorder score: the original consensus score merged residue by residue with inverted pLDDT confidence values drawn from AlphaFold DB. pLDDT is informative because regions AlphaFold 2 models with low confidence are frequently disordered, but reading it off directly would demand a structure prediction for every new sequence. Training a network to recover the merged signal from amino acids alone collapses that dependency: the shipped checkpoint maps sequence to per-residue disorder in one pass, with neither an ensemble of predictors nor a structure as input.

The work comes from Lotthammer, Hernández-García, Griffith, Weijers, Holehouse and Emenecker at Washington University in St. Louis and Wageningen University & Research, in a bioRxiv preprint posted in November 2024 that has not completed peer review. It is the same group behind STARLING, which builds conformational ensembles for the regions metapredict identifies.

#Key Features

• Distilled consensus in one forward pass: One network reproduces a predictor panel combined with AlphaFold2 confidence, so neither an external predictor nor a structure is needed at inference.
• Proteome-scale throughput: The human reference proteome of 20,594 sequences is annotated in 60–90 seconds on CPU and under 5 seconds on GPU, roughly 250 times faster than the original implementation.
• Community-benchmark accuracy: Judged by Matthews correlation coefficient on the Disprot-PDB set from CAID1 and the Disorder-PDB set from CAID2, V3 ranks among the most accurate entries in both rounds.
• Holds up on designed sequences: Though trained only on natural proteins, it correctly calls folded and disordered regions for 61 of 63 synthetic sequences from four protein-design studies.
• Five ways to run it: A Python package, command-line tool, Docker container, Colab notebook and hosted web server drive the same checkpoint, with the V1 and V2 networks kept available for older analyses.

#Technical Details

The V3 network is a bidirectional recurrent neural network with long short-term memory units (BRNN-LSTM), trained through parrot-lightning, an adaptation of the PARROT framework. One-hot encoded amino acids pass through two bidirectional LSTM layers of hidden size 52 into a single linear regression output. Hyperparameters came from Bayesian optimization with Optuna; training used stochastic gradient descent with Nesterov momentum of 0.9968, a learning rate of 0.01427, and a batch size of 256. Training data came from Swiss-Prot, deduplicated and then restricted to sequences with a corresponding AlphaFold2 structure, giving 455,666 sequences split 70:15:15. Targets were built by scaling pLDDT between 0.35 and 0.95, inverting it, taking the larger of that and the consensus score where either exceeds 0.5 and the smaller where both fall below, then applying a Savitzky-Golay filter and a 25-residue moving average. Residues above 0.5 are called disordered. The backend was also rewritten in Cython with batched, GPU-enabled prediction, which supplies most of the speedup.

#Applications

The headline demonstration is a survey no earlier version could have attempted: disorder across 23,129 UniProt reference proteomes and 88,082,402 sequences. Prokaryotes carry the least disorder, though Actinobacteria and the PVC superphylum stand out above other bacteria; bird proteomes are strikingly uniform in their low disorder content regardless of taxonomic rank; and viral proteome disorder tracks host, with phages of E. coli (host proteome disorder fraction 0.078) far less disordered than viruses of humans (0.378). A separate pass over the 1,608 human transcription factors found them 57% disordered on average, against 24% for the whole proteome. At the other end of the scale, the web server serves the everyday case of checking one protein before designing a construct.

#Impact

Metapredict's influence rests less on any single accuracy figure — the paper notes that the top ten disorder predictors are roughly equivalent in practice — than on making that accuracy cheap enough to use everywhere. Disorder annotation shifts from a step run on selected proteins to one run across a whole dataset by default, and evolutionary-scale questions about disorder in the Tree of Life become tractable on a laptop. The package is MIT-licensed with the weights inside it, so the model is installed rather than downloaded, and it is a common dependency in IDR pipelines. Two caveats: the training targets are predictions rather than experimental annotations, so systematic errors in the consensus scores or AlphaFold2 confidence propagate into the network, and the output is a single per-residue scalar, with no embeddings or generative capability for downstream reuse.