Learn/Predict a structure

The alignment, and templates

A multiple sequence alignment lines up a protein with its relatives. Where two positions change together across species, they are usually in contact; hundreds of relatives give a model enough of those pairs to place every part of the fold. The depth of the alignment, how many diverse relatives it holds, is the single largest influence on a prediction of a natural protein.

Barnase and barstar by Boltz-1: with an alignment, and from the sequence alone

Metric	1BRS	Boltz-1, with an alignment	Boltz-1, single sequence
pLDDT		96.71	40.51

The same model, the same two sequences; the only difference is whether it was given the alignment of their relatives. The crystal structure is the reference, and the row of Cα distances says how far each prediction sits from it.

The transferrin receptor by AlphaFold 2: with an alignment, and from the sequence alone

Metric	AlphaFold 2, with an alignment	AlphaFold 2, single sequence
pLDDT	92.67	28.61
pTM	0.92	0.23
iPAE	5.362 Å	26.71 Å

A natural protein with thousands of relatives loses almost everything when they are taken away: the fold, the confidence, and the placement of all three domains.

When no alignment is needed

pLDDT ≥ 90 · very high confidence
70 to 89 · confident
50 to 69 · low confidence
below 50 · very low confidence

A de novo design has no relatives, so an alignment finds nothing. Single-sequence models fold designed proteins well, which is why ESMFold is the default for one chain and why design pipelines fold with it.

Templates

A template is a known structure of a related protein handed to the model as a hint. It helps when the alignment is shallow or when a known fold should anchor the prediction; when the alignment is deep the model mostly ignores it. AlphaFold 2, OpenFold 3 and OpenDDE have the switch. A sound habit on a hard target is to run with and without and keep the higher pLDDT.

Back: Confidence Next: Complexes Read the docs →