BuriedBet · a mode for ages 15–18
TwoChains
The AlphaFold Database now holds predicted pairs: two chains modelled together. A pair model always puts the chains somewhere, so the question is whether to trust the join. You see each chain's confidence on its own, bet, then read what the pair model says about its own join and whether an experiment has seen the pair.
the Sketcher's drawing Prediction Computed by a model, with how sure it is. Not an experiment.
A predicted pair is a prediction about a prediction. Each chain's shape is predicted, and so is the way the two sit together. A confident join says the model is sure of the geometry. It does not say the two proteins meet in a cell.
Bet on the join
7 predicted pairs. For each: is the predicted join trustworthy? Then say how sure you are.
Is this predicted join trustworthy?
How sure?
the Photographer's picture Experiment Measured from real molecules.
HalfAFold
One chain on its own, shown by its single-chain confidence. Tap the 10-residue stretches where you think a partner would firm it up (or say none), then check against the pair model.
How every number is read
- Single-chain confidence: pLDDT from each protein's own AlphaFold DB model, in AlphaFold's bands: very high 90 and above, confident 70 to 90, low 50 to 70, very low below 50.
- ipTM (interface predicted TM-score, 0 to 1): the pair model's own score for how the chains sit together. The AlphaFold course at EMBL-EBI reads it this way: above 0.8 a confident join, below 0.6 likely a failed one, between them a grey zone that could be right or wrong. Your bet is matched against that verdict; grey-zone pairs are not scored.
- Between-chain error: the mean predicted aligned error (PAE, in Å) over every pair of residues on different chains. Low means the model is sure where one chain sits relative to the other.
- Interface: residues with a Cα atom within 8 Å of a Cα on the other chain, in the pair model.
- HalfAFold: a stretch firms up when its mean pLDDT is at least 10 points higher in the pair model than in the single-chain model. A firmer chain in a pair model does not prove the pair is real. Check ipTM too.
- Experiment: a PDB entry that holds both proteins together (for two copies of one protein: a biological assembly with two or more copies). If the entry holds only a piece of one protein (under half its residues), we say so: it checks the join site, not the whole pair. Found by a search of the RCSB PDB on 2026-10-08. No hit means our search found none, not that the pair never forms.
The pair models were made by NVIDIA with ColabFold v1.6.0 / AlphaFold-Multimer and OpenFold-TRT / AlphaFold-Multimer (2025-12-25 to 2026-02-20), and are served by the AlphaFold Database under CC BY 4.0. Snapshot 2026-10-08.
Sources
- Evans R, et al. (2021). Protein complex prediction with AlphaFold-Multimer. bioRxiv. doi:10.1101/2021.10.04.463034
- EMBL-EBI training, AlphaFold course: Confidence scores in AlphaFold-Multimer (the 0.8 / 0.6 ipTM reading).
- Jumper J, et al. (2021). Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589. doi:10.1038/s41586-021-03819-2 (pLDDT and PAE).
- AlphaFold Database complex API: alphafold.ebi.ac.uk/api/complex/<UniProt>; experimental assemblies: RCSB PDB.