Predicted Structures · LiteFold
ESMAtlas
Predicted metagenomic protein structures from MGnify generated with ESMFold and packaged as WebDataset shards.
What ESMAtlas is
ESMAtlas is a resource of predicted atomic-level protein structures derived from environmental and metagenomic sequences in MGnify. Generated using Meta AI's ESMFold protein language model, the collection addresses the vast structural gap between known sequence space and the relatively small fraction of proteins with experimentally determined coordinates in public archives.
The dataset contains structural predictions for hundreds of millions of metagenomic proteins, packaged into WebDataset tar shards alongside metadata tables. It supports large-scale protein structure retrieval, structure-based search pipelines, structural diversity surveys, and machine learning representation training across microbial and uncultivated biology.
How it was made
Structures were predicted from MGnify protein sequences (release v2023_02) using the ESMFold model (esm.pretrained.esmfold_v1). Sequences longer than 1280 residues were excluded from folding. For each folded protein, the model produces coordinate records alongside two primary confidence metrics: the predicted local distance difference test (pLDDT) measuring per-residue local accuracy, and the predicted template modelling score (pTM) assessing overall topological confidence.
The coordinates are formatted as gzipped PDB files and packaged into WebDataset tarballs partitioned into discrete pTM and pLDDT bins. Accompanying Parquet and compressed SQLite metadata tables index all sequence records, detailing sequence length, fragment status, residue counts exceeding 0.7 pLDDT, CRC64 checksums, database source versions, and manifests for Foldseek databases and ESM-2 embeddings.
Key facts
- Prediction model
- ESMFold v1 (esm.pretrained.esmfold_v1)
- Source sequences
- MGnify release 2023_02
- Total metadata entries
- 773,846,840
- Maximum sequence length folded
- 1280 residues
- Storage format
- Gzipped PDB files in WebDataset tar shards
- Metadata formats
- Parquet and SQLite
- Publisher
- LiteFold
What you can use it for
- High-confidence structure retrieval
- Isolate high-confidence protein models by filtering the metadata or downloading specific shard bins where both pLDDT and pTM exceed 0.7.
- Metagenomic fold discovery
- Investigate structural novelty and fold diversity across uncultivated microbial sequences that lack close homologues in experimental structure databases.
- Structure-based sequence search
- Perform fast structural similarity comparisons across environmental proteins using the provided Foldseek database manifests.
- Structural representation learning
- Train and evaluate 3D geometric deep learning architectures and protein representations over massive collections of uncurated natural folds.
Explore the data
The schema, real rows, full-text search and a SQL filter — read in your browser, nothing downloaded.
Explore
Read the real rows without downloading anything
Reading rows…
LiveRead from LiteFold/ESMAtlas at the moment you asked. Nothing is cached or stored — every row above came from that request.
How the columns are distributed
Over 40,100 rows of default/train
__key__
string_text- min
- 22
- median
- 22
- mean
- 22
- max
- 22
__url__
string_label- hf://datasets/LiteFold/ESMAtlas@d8c02e20b8e914d13ce0f5f6bdff97ef45882945/v2023_02/full/tarballs/tm_.40_.50_plddt_.40_.50_00.tar.gz100%
Strengths and limits
Good at
- Expands structural coverage over hundreds of millions of metagenomic proteins that lack experimental structure determinations.
- Tar shards are organized into confidence bins, enabling targeted access to high-confidence models without processing the entire repository.
- Comprehensive metadata tables provide pre-computed quality metrics, fragment flags, and checksums for rapid dataset filtering.
Watch out for
- Coordinates are computational predictions rather than experimentally validated structures and should not be treated as ground truth.
- Proteins exceeding 1280 amino acids were omitted from structure generation.
- Many metagenomic sequences represent fragments, orphan proteins, or uncharacterised chains that require independent validation for active-site or interface analyses.
Questions
Can I look inside ESMAtlas without downloading it?
Yes. This page reads its schema, real rows and per-column statistics in the browser, with full-text search and a SQL filter. Nothing is downloaded to your machine.
What licence is ESMAtlas under?
Its authors declare cc-by-4.0. Read the terms on the original page before reusing the data.
What confidence criteria define high-quality predictions in this dataset?
Entries with an average pLDDT above 0.7 and a pTM above 0.7 are recommended as the standard high-confidence subset for downstream analyses.
Why do some records in the metadata lack folded PDB files?
While the metadata tracks 773,846,840 records, sequences with lengths exceeding 1280 residues were not folded by ESMFold.
How can specific quality subsets be downloaded without fetching all data?
The WebDataset tarballs are named according to specific pTM and pLDDT intervals, allowing selective loading of particular confidence tiers.
What auxiliary files are included alongside the PDB shards?
The repository includes Parquet and SQLite metadata tables, Foldseek database manifests, ESM-2 embedding manifests, and FASTA sequences for a 30% sequence-identity clustered subset.
How to cite it
Lin, Z., Akin, H., Rao, R., Hie, B., Zhu, Z., Lu, W., Smetanin, N., Verkuil, R., Kabeli, O., Shmueli, Y., dos Santos Costa, A., Fazel-Zarandi, M., Sercu, T., Candido, S., & Rives, A. (2023). Evolutionary-scale prediction of atomic-level protein structure with a language model. Science, 379(6637), 1123-1130. https://doi.org/10.1126/science.ade2574
Written from ESMAtlas's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.