Proteins · a check
FitForPurpose
An AlphaFold model can be good enough for one question and wrong for another. Paste up to 10 UniProt accessions, say what you need the model for, and get plain cautions. Each one names the number behind it. These are transparent rules, not a new predictor: all of them are listed below.
the Sketcher's drawing Prediction Computed by a model, with how sure it is. Not an experiment.
The rules
- Always: the model's sequence is compared with today's UniProt sequence; if it changed, residue numbers may not mean the same thing. The whole chain's mean pLDDT and its share below 50 are shown.
- A mutation: needs a residue number. Caution if that residue's pLDDT is below 70 (low or very low). A confident shape still says nothing about what a change does. See TolerancePaint.
- A pocket: caution if the residues you give have a mean pLDDT below 70, or more than 25% of them are below 50. See PocketScout.
- Domain boundaries: give the first residue of the second part. We average the predicted aligned error (PAE) between the two sides, both ways. Caution at 15 Å or more (trust each part, not their arrangement); below 5 Å the two sides are placed with confidence. These two cuts are ours, chosen to be plain, not taken from a paper. Chains longer than 1400 residues are not checked here. See DomainCuts and TwoIslands.
- A complex: the AlphaFold DB model is one chain on its own, so partners, metal ions and bound molecules are missing. We also ask the AlphaFold DB's complex list whether a predicted pair exists and show its best ipTM (above 0.8 a confident join, below 0.6 likely a failed one, per the EMBL-EBI AlphaFold course). See WhatsMissing.
pLDDT bands: very high 90 and above, confident 70 to 90, low 50 to 70, very low below 50 (Jumper et al. 2021). Your browser fetches the numbers straight from the AlphaFold Database and UniProt. Nothing you paste is stored or sent anywhere else.
Sources
- Jumper J, et al. (2021). Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589. doi:10.1038/s41586-021-03819-2
- EMBL-EBI training, AlphaFold course: Confidence scores in AlphaFold-Multimer.
- UniProt: accession number format.