Use ESMFold Online
Commercially Available ESMFold No-Code Web Server
ESMFold: Fast Protein 3D Structure Prediction from a Single Amino Acid Sequence
ESMFold is a protein structure prediction model that builds on the ESM-2 protein language model to predict atomic-level 3D structure directly from a single amino acid sequence, with no multiple sequence alignment (MSA), templates or database search required. It is described in a bioRxiv preprint that also presents the ESM Metagenomic Atlas, a large set of structures predicted with the model.
High-accuracy structure predictors typically rely on MSAs and homology searches, which add computing cost and limit use on orphan sequences and in high-throughput settings. The authors argue that language models trained on sequences alone can learn information about structure as they scale, and ESMFold is built to exploit that signal for fast prediction.
Feature | Details |
|---|---|
Input | A single protein sequence; no MSA or templates |
Language model | ESM-2, transformer trained with masked language modeling on UniRef sequences (8M to 15B parameters) |
Structure module | Folding blocks plus an equivariant transformer, with three recycling steps |
Outputs | Atomic coordinates and confidence scores |
Speed | 14.2 seconds for a 384-residue protein on a V100 GPU |
How ESMFold Predicts Structure Without an MSA
Language model representations: ESM-2 reads the sequence and produces internal representations that encode evolutionary and structural patterns learned during training.
Folding head: Folding blocks alternate between sequence and pairwise representations.
Structure module: An equivariant transformer converts these into atomic coordinates, refined over three recycling steps.
Confidence estimates: The model outputs confidence scores such as pLDDT and pTM, which flag reliable predictions.
Training data
The folding head was trained on about 325K PDB structures (roughly 25K clusters) plus about 12M AlphaFold2-predicted structures, using AlphaFold-style losses.
Accuracy and Speed Compared with AlphaFold2 and RoseTTAFold
Model | CAMEO TM-score | CASP14 TM-score |
|---|---|---|
ESMFold | 0.83 | 0.68 |
AlphaFold2 (full search pipeline) | 0.88 | 0.85 |
RoseTTAFold | 0.82 | Not reported in this summary |
On a V100 GPU, a 384-residue protein took 14.2 seconds, about 6x faster than a single AlphaFold2 model, with speedups up to about 60x on shorter sequences. The authors folded 617M MGnify90 sequences in about two weeks on roughly 2,000 GPUs, and about 225M predictions were high confidence (mean pLDDT and pTM above 0.7). Of 1M sampled high-confidence structures, 76.8% had under 90% sequence identity to UniRef90, and 12.6% had no PDB match with TM-score above 0.5.
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Discovery with ESMFold on Tamarind Bio
High-throughput structure prediction: Fast, MSA-free inference suits screening large sets of sequences.
Metagenomic exploration: The authors used the model to predict structures for hundreds of millions of metagenomic sequences, including many with no close PDB match.
Orphan and novel sequences: Because no alignment is needed, sequences with few homologs can still be folded.
Quick structural triage: Confidence scores help prioritize which predictions merit follow-up analysis.
How to Use ESMFold on Tamarind Bio
Open the tool: Log in to tamarind.bio and open ESMFold.
Enter the sequence: Paste the amino acid sequence; no MSA or templates are needed.
Check parameters: Review any options exposed in the interface, such as recycling settings, if exposed.
Run the prediction: Submit the job and wait for the result.
Download outputs: Download the predicted structure file (atomic coordinates) and confidence information.
Interpret confidence: Use pLDDT and pTM to judge reliability; the authors treat values above 0.7 as high confidence.
Use downstream: Feed structures into docking, design or function analysis workflows.
Things to Keep in Mind
ESMFold accuracy tracks language model perplexity, and low-accuracy regions look similar to AlphaFold outputs.
On benchmarks it scored below AlphaFold2 run with its full search pipeline.
ESMFold was not trained on protein complexes; on 2,978 multimer complexes it matched AlphaFold-Multimer's DockQ category for 53.2% of chain pairs.
Source: bioRxiv preprint, DOI 10.1101/2022.07.20.500902, https://www.biorxiv.org/content/10.1101/2022.07.20.500902v2 ; ESM Metagenomic Atlas: https://esmatlas.com