Learn/Predict a structure

Compare, then go

A high score from one model is a claim. The same answer from two models is evidence, and an experimental structure, when one exists, is the measure of both. Overlay them: where predictors agree the fold is settled; where they spread, so should your confidence.

Six models' ectodomains, overlaid on the crystal structure

Metric	ESMFold	AlphaFold 2	Boltz-2	Chai-1	OpenFold 3	Protenix
pLDDT	73.56	92.67	90.81	85.19	88.89	91.12
pTM	0.8944	0.92	–	0.9214	0.9315	0.9547
Ranking score	–	–	–	0.1843	0.1863	0.1909

Three domains, and the distances say which model placed the apical domain where the crystal has it.

The five that take a complex, on barnase–barstar, overlaid on the crystal structure

Metric	AlphaFold 2	Boltz-2	Chai-1	OpenFold 3	Protenix
pLDDT	96.91	96.8	94.39	80.21	96.11
pTM	0.93	–	0.9411	0.7856	0.9772
ipTM	0.92	–	0.877	0.6594	0.9746
Ranking score	–	–	0.8898	0.6847	0.9751

An easy case: a small, well-studied complex that every predictor with an alignment gets to within an ångström. The spread is in the metrics, not the fold. The project library runs this comparison on every fold you make.

What to try first

One natural chain, well studied: ESMFold first, minutes. AlphaFold 2 or a diffusion model with the alignment on if pLDDT is low or the fold matters.
A designed sequence: Single sequence is enough. ESMFold, or ESMFold 2 fast.
Several protein chains: Boltz-2, Chai-1 or OpenFold 3, or AlphaFold 2 with the alignment on. Read ipTM and the PAE blocks; sample more if ipTM sits between 0.5 and 0.8.
A ligand, DNA or RNA: Boltz-2 or OpenFold 3. Affinity as a screen, not a number to report.
Always: Compare two predictors in the project library before trusting one, and overlay the experimental structure when there is one.

Open the Structure lab Compare folds in Work

Back: Sampling Sign in Read the docs →