FoldCompare · ModelDiff mode
ModelDiff: AlphaFold vs another model
FoldCompare sets a prediction beside an experiment. This mode sets it beside another prediction: the same sequence, folded by AlphaFold (AlphaFold Database) and by ESMFold.
- ESMFold predicts a structure from the single sequence, using a protein language model; AlphaFold (as in the AlphaFold Database) also uses an alignment of related sequences.
- We fit the two models on the residues both are sure of, then measure how far apart every residue sits.
- Two predictions agreeing is not an experiment: they can share a blind spot.
Haemoglobin subunit alpha UniProt P69905 · 142 residues
Both models are sure of 141 of 142 residues (pLDDT 70 or more in both). Fitted on those, they sit 0.31 Å apart (RMSD).
1 residues: one model is sure and the other is not. A model's doubt is about the model; on its own it does not mean the protein is disordered.
Across the whole chain, mean pLDDT is 98 for AlphaFold and 95 for ESMFold.
Open AlphaFold's model in the viewer · ESMFold model (PDB file)
AlphaFold DB model v6 for UniProt P69905 (CC BY 4.0) and ESMFold v1 via the ESM Metagenomic Atlas (source code MIT, facebookresearch/esm; ESM Atlas data CC BY 4.0), both from Fold Commons' snapshot of 2026-10-02. Two predictions, not an experiment.
Lysozyme C UniProt P61626 · 148 residues
Both models are sure of 131 of 148 residues (pLDDT 70 or more in both). Fitted on those, they sit 0.31 Å apart (RMSD).
7 residues: one model is sure and the other is not. A model's doubt is about the model; on its own it does not mean the protein is disordered.
Across the whole chain, mean pLDDT is 94 for AlphaFold and 92 for ESMFold.
10 residues: neither model is sure.
Over the whole chain, including the residues at least one model is unsure of, they sit 4.2 Å apart.
Open AlphaFold's model in the viewer · ESMFold model (PDB file)
AlphaFold DB model v6 for UniProt P61626 (CC BY 4.0) and ESMFold v1 via the ESM Metagenomic Atlas (source code MIT, facebookresearch/esm; ESM Atlas data CC BY 4.0), both from Fold Commons' snapshot of 2026-10-02. Two predictions, not an experiment.
Alpha-synuclein UniProt P37840 · 140 residues
The two models are sure of fewer than 20 of the same residues, so there is no shared core to fit on; fitted on the whole chain they sit 35.3 Å apart (RMSD).
88 residues: one model is sure and the other is not. A model's doubt is about the model; on its own it does not mean the protein is disordered.
Across the whole chain, mean pLDDT is 75 for AlphaFold and 33 for ESMFold.
52 residues: neither model is sure.
Open AlphaFold's model in the viewer · ESMFold model (PDB file)
AlphaFold DB model v6 for UniProt P37840 (CC BY 4.0) and ESMFold v1 via the ESM Metagenomic Atlas (source code MIT, facebookresearch/esm; ESM Atlas data CC BY 4.0), both from Fold Commons' snapshot of 2026-10-02. Two predictions, not an experiment.
GTPase KRas UniProt P01116 · 189 residues
Both models are sure of 175 of 189 residues (pLDDT 70 or more in both). Fitted on those, they sit 0.81 Å apart (RMSD).
3 residues: both models are sure of them, yet they sit 3 Å or more apart after the fit.
5 residues: one model is sure and the other is not. A model's doubt is about the model; on its own it does not mean the protein is disordered.
Across the whole chain, mean pLDDT is 92 for AlphaFold and 92 for ESMFold.
9 residues: neither model is sure.
Over the whole chain, including the residues at least one model is unsure of, they sit 4.7 Å apart.
Open AlphaFold's model in the viewer · ESMFold model (PDB file)
AlphaFold DB model v6 for UniProt P01116 (CC BY 4.0) and ESMFold v1 via the ESM Metagenomic Atlas (source code MIT, facebookresearch/esm; ESM Atlas data CC BY 4.0), both from Fold Commons' snapshot of 2026-10-02. Two predictions, not an experiment.
Green fluorescent protein UniProt P42212 · 238 residues
Both models are sure of 33 of 238 residues (pLDDT 70 or more in both). Fitted on those, they sit 0.42 Å apart (RMSD).
202 residues: one model is sure and the other is not. A model's doubt is about the model; on its own it does not mean the protein is disordered.
Across the whole chain, mean pLDDT is 97 for AlphaFold and 43 for ESMFold.
3 residues: neither model is sure.
Over the whole chain, including the residues at least one model is unsure of, they sit 16.0 Å apart.
Open AlphaFold's model in the viewer · ESMFold model (PDB file)
AlphaFold DB model v6 for UniProt P42212 (CC BY 4.0) and ESMFold v1 via the ESM Metagenomic Atlas (source code MIT, facebookresearch/esm; ESM Atlas data CC BY 4.0), both from Fold Commons' snapshot of 2026-10-02. Two predictions, not an experiment.
Cellular tumour antigen p53 UniProt P04637 · 393 residues
Both models are sure of 225 of 393 residues (pLDDT 70 or more in both). Fitted on those, they sit 9.8 Å apart (RMSD).
Piece by piece they agree more closely: residues 96 to 292, 0.88 Å; residues 330 to 355, 0.47 Å. So they place the pieces differently relative to each other. pLDDT is per residue; it does not say how pieces sit together (AlphaFold's PAE is about that).
186 residues: both models are sure of them, yet they sit 3 Å or more apart after the fit.
65 residues: one model is sure and the other is not. A model's doubt is about the model; on its own it does not mean the protein is disordered.
Across the whole chain, mean pLDDT is 75 for AlphaFold and 80 for ESMFold.
103 residues: neither model is sure.
Over the whole chain, including the residues at least one model is unsure of, they sit 21.0 Å apart.
Open AlphaFold's model in the viewer · ESMFold model (PDB file)
AlphaFold DB model v6 for UniProt P04637 (CC BY 4.0) and ESMFold v1 via the ESM Metagenomic Atlas (source code MIT, facebookresearch/esm; ESM Atlas data CC BY 4.0), both from Fold Commons' snapshot of 2026-10-02. Two predictions, not an experiment.
Reading the strips
Colour and pattern are each model's own confidence (pLDDT) per residue: solid = very high (90 and above), dots = confident (70 to 90), lines = low (50 to 70), crosses = very low (below 50). The grey bars are how far apart the two models put each residue after the fit; the line marks 3 Å.
Not here: the SARS-CoV-2 spike from FoldCompare's set (1,273 residues is longer than the 400 the ESMFold service folds). Next models to add: Boltz-2 and OpenFold3, which need computing offline first.