ESMFold FAQ

These answers cover how ESMFold works, which sequences it supports, and how to read its confidence outputs — reach for this page when you're deciding whether ESMFold fits your target or interpreting a pLDDT or PAE result.

Frequently asked questions

What is ESMFold and how does it work?

ESMFold is a protein structure prediction method that uses a protein language model trained on millions of protein sequences. Unlike traditional structure prediction methods, ESMFold predicts 3D structures directly from single amino acid sequences without requiring multiple sequence alignments (MSA) or evolutionary information.

How does ESMFold compare to other folding methods?

ESMFold offers several key differences from MSA-based methods:

• Speed: ESMFold is significantly faster than MSA-based methods because it skips the time-consuming MSA generation step
• No MSA required: ESMFold works from single sequences, making it ideal for novel proteins, designed sequences, or orphan proteins without detectable homologs
• Accuracy trade-off: MSA-based methods are more accurate, and the gap is not small on hard targets. ESMFold's authors report an average TM-score of 0.83 on CAMEO (194 structures) and 0.68 on CASP14 (51 structures), against AlphaFold2 at 0.88 and 0.85 on the same sets. The gap narrows to near-parity on sequences the language model finds unsurprising and widens on difficult or orphan ones — how ESMFold results compare to the bench has the detail
• Confidence correlation: ESMFold's pLDDT scores correlate well with prediction accuracy and can reliably identify disordered regions

Choose ESMFold when speed is critical, when MSA generation is impractical, or when working with novel/designed sequences.

What sequences are supported?

Single-chain proteins with standard amino acids, 10-2,500 residues. Confidence is best in the 50-500 residue range; longer sequences risk reduced confidence in some regions.

Are outputs deterministic?

Yes, given the same sequence and settings. Set seed if you need explicit reproducibility.

Why is pLDDT low?

Likely disordered or low-complexity regions, very long linkers, or sequences outside training distribution. Consider trimming or splitting domains.

How do I read PAE?

Low PAE between residues or domains implies stable relative placement. High off-diagonal PAE means uncertain domain orientation.

Can I submit multiple sequences?

Prefer one per run, especially for long proteins. For short peptides, batching is possible but watch aggregate length.

Where is pLDDT stored?

In the B-factor column of the PDB when returnBfactors is enabled (default true).

Should I relax the structure?

Optional: run a quick minimization if you need cleaner geometries for docking or MD; focus on high pLDDT cores.

Read next