ESMFold outputs and confidence
Read an ESMFold result — structure.pdb with pLDDT in the B-factor column plus pae.json — and what each pLDDT band supports, from above 90 down to below 50.
ESMFold returns a predicted structure (structure.pdb) alongside per-residue pLDDT confidence scores and a PAE matrix. Read the pLDDT first: it tells you whether to trust the model for your next step — high-confidence cores (>90) are the only regions worth examining in detail, while low-confidence tails should be trimmed or masked before downstream docking or simulation. Note that even a high-confidence predicted receptor costs downstream accuracy: see how ESMFold results compare to the bench.
Output artifacts
structure.pdb: predicted 3D coordinates with pLDDT stored in the B-factor column (if enabled).pae.json/pae.png: per-residue predicted aligned error; helps assess domain placement.metadata.json: includes sequence, model version, and runtime parameters.
Understanding pLDDT confidence scores
pLDDT (predicted Local Distance Difference Test) is a per-residue confidence metric that ESMFold provides for each amino acid in the predicted structure. The score ranges from 0 to 100 and indicates how confident the model is in the local structure around each residue. This metric is derived from the ESM-2 language model's internal representations and correlates well with the actual accuracy of the predicted coordinates.
pLDDT score interpretation
- 90–100: Very high confidence. On the CAMEO test set, ESMFold predictions above this threshold had a median all-atom RMSD95 of 1.42 Å and backbone RMSD95 of 0.94 Å against the experimental structure (author-reported). These regions are the ones worth using for detailed structural analysis and mechanistic reasoning — but docking into a predicted receptor still costs accuracy relative to a crystal structure.
- 70–90: Confident prediction. Backbone geometry is generally reliable, but side-chain rotamers may vary. Suitable for most structural biology applications with appropriate caution.
- 50–70: Low confidence. Treat these regions as provisional. Often corresponds to flexible loops, surface-exposed regions, or areas with limited structural constraints. Use for general topology only.
- <50: Very low confidence. These regions are likely intrinsically disordered or unstructured. Do not rely on the predicted geometry. Consider these regions as placeholders rather than accurate predictions.
Using pLDDT for quality assessment
- Mean pLDDT: A useful summary metric; values above 70 generally indicate a reliable overall fold
- pLDDT distribution: Examine the per-residue plot to identify confident cores vs. flexible regions
- Domain identification: Sharp transitions in pLDDT can indicate domain boundaries
- Comparison with known structures: High pLDDT regions should align well with experimental structures when available
Interpreting PAE
- Low PAE between domains suggests stable relative placement.
- High off-diagonal PAE indicates uncertain domain orientation; domain splitting may be needed.
Quality checks
- Inspect pLDDT coloring in a viewer; focus on high-confidence cores.
- Trim or mask low-confidence tails before downstream docking or simulation.
- Validate secondary structure plausibility with external tools if pLDDT is marginal.
Canvas Visualization
When an ESMFold job completes, a result card appears on the canvas.
- Card size: 450×560px
- Border color: Green (#4CAF50)
- Visualization: Interactive 3D Mol* viewer with rotate, zoom, and pan
What You See
- pLDDT score (0–100) with confidence category
- pTM score (structure-level metric)
- Residue count
- Color modes: confidence (pLDDT), chain, secondary structure, hydrophobicity
- Representation modes: cartoon, ball-and-stick, surface
- Sequence bar with residue highlighting
Interpreting Canvas Metrics
| Metric | Good Value | Concern |
|---|---|---|
| pLDDT > 90 | Very high confidence — median all-atom RMSD95 1.42 Å on CAMEO | A predicted receptor still costs pose accuracy downstream |
| pLDDT 70–90 | Confident — backbone reliable | Side-chain rotamers may vary |
| pLDDT 50–70 | Low confidence — general topology only | Loops and flexible regions |
| pLDDT < 50 | Very low — likely disordered | Do not use for detailed analysis |
Read next
Was this page helpful?
ESMFold quickstart
Fold a sequence with ESMFold in 10-60 seconds — paste it or fetch it by UniProt ID, keep 3 recycles, then read the pLDDT coloring and download PDB and PAE.
How ESMFold results compare to the bench
What an ESMFold structure has been shown to agree with experimentally, how it compares to MSA-based predictors, and the pose accuracy it costs downstream.