FoldCompare
Also: ModelDiff, AlphaFold set beside another prediction (ESMFold).
How close is an AlphaFold prediction to the experimental structure? Pick a protein below (or enter your own pair), and this tool fetches the experimental structure from the RCSB PDB and AlphaFold's prediction from the AlphaFold Database, superposes them, and reports the Cα RMSD, the per-residue divergence, and — honestly — whether the disagreement is meaningful or just where the model was unsure. Everything runs in your browser; nothing is installed and nothing about you is collected.
Also: BlindTest, a weekly issue: mark where a prediction will be wrong from its confidence, then check against a new experiment.
Also: TrustCurve, how often AlphaFold's confidence holds up against experiment, band by band.
Also: StickSquare, what an RMSD number means, built from real fits.
This is the tool — the core superposition + honest read-out. A native version, when it ships, adds a 3D divergence-painted overlay, a numerical stat table, saved comparisons, and offline use; it never gates the web.
Per-residue divergence
Distance (Å) between the aligned experimental and predicted Cα, residue by residue. Tall, dark ink bars mark where the two structures disagree most (the bar's height and depth of ink both say how far, never AlphaFold's colours); faded bars mark residues AlphaFold predicted with low confidence (pLDDT < 70) — where a disagreement is ambiguous, because the model was unsure there too.
Split view
Why the per-residue differences matter
For p53, the default pair, the experimental structure 1TUP covers the DNA-binding domain bound to DNA, so the comparison is made there, where AlphaFold is confident and the two usually agree closely, often to within one or two ångströms. Large differences that fall where pLDDT is low, or where a crystal holds DNA or a partner, are usually honest disagreement about a flexible or bound region, not a failed prediction.
Methods & limits
- Alignment: rigid-body Kabsch superposition over matched Cα atoms (the same v1.0 algorithm as the native app), computed in your browser from the fetched coordinates.
- Residue matching is by number, within the experimental chain that best overlaps the prediction. There is no sequence realignment yet — if the two number their residues differently, coverage drops and the pair may not align. Coverage is reported honestly.
- Cα only. Side-chain, backbone-torsion, and all-atom differences are not measured; RMSD is over Cα, the standard backbone summary.
- Partial structures compare partially. A crystal of one domain, or one chain of an assembly, aligns only over the residues both structures share — the coverage percentage tells you how much.
- Predictions are models, not experiments. pLDDT colouring shows where AlphaFold was confident (very high and confident: dark and light blue) and where it was not (low and very low: yellow and orange); the verdict weights divergence by that confidence.
About this tool
Experimental structures come from the RCSB Protein Data Bank; predictions from the AlphaFold Database (EMBL-EBI / Google DeepMind). Both are fetched directly by your browser and rendered with Mol* via PDBe Mol*, self-hosted here — no third-party CDN, no tracking, no server. AlphaFold DB data is released under CC BY 4.0; please cite the AlphaFold and RCSB PDB papers when you use a structure. This tool is part of the free, non-profit Fold Commons project.