PDB-Monomeric-Structure-ESMFold2

Monomeric, protein-only PDB structure dataset for minimum ESMFold2-style training. Each row is one eligible single-chain biological assembly with a canonical amino-acid sequence input and all-atom protein labels in atom37.

Labels

• atom37_positions: residue x 37 x 3 coordinates, with zeros for missing atoms.
• atom37_mask: residue x 37 resolved-atom mask.
• aatype, residue_index, auth_seq_id, insertion_code, residue_name, ca_mask.
• Experiment metadata: resolution, r_work, r_free, method, release_date.

Scope

The dataset is PDB-only and protein-only. Ligands, waters, ions, nucleic acids, unsupported atoms, MSAs, ligand bonds, nucleotide fields, and AFDB distillation rows are not included.

ESMFold2-Min Loss Support

The schema supports deriving ESMFold2-style flat atom-list tensors in a collator: diffusion MSE, smooth LDDT, distogram, resolved/pLDDT, PAE, and PDE confidence losses can all be computed from resolved atom coordinates and masks. Default loss metadata recorded for this build:

• protein atom weight: 1x
• smooth LDDT cutoff: 15 A
• distogram bins: 64, range 2-22 A

References

• ESMFold2 preprint: https://www.biorxiv.org/content/10.64898/2026.06.03.729735v1
• Biohub ESM repository: https://github.com/Biohub/esm
• ESMFold2 model page: https://huggingface.co/biohub/ESMFold2

Split Policy

Rows are split by connected components over exact sequence duplicate groups, 30% MMseqs2 sequence clusters, Foldseek structure clusters, PDB ID, and biological assembly membership. The builder fails before upload if any leakage assertion fails.

Build Metadata

{
  "allow_incomplete_structure_clusters": true,
  "allow_incomplete_structure_edges": false,
  "contains_ligand_bonds": false,
  "contains_ligands": false,
  "contains_msas": false,
  "contains_nucleotides": false,
  "cpu_count": 30,
  "distogram_bins": 64,
  "distogram_max_angstrom": 22.0,
  "distogram_min_angstrom": 2.0,
  "foldseek_min_aligned_residues": 40,
  "foldseek_min_coverage": 0.8,
  "foldseek_min_tm": 0.5,
  "foldseek_path": "/usr/local/bin/foldseek",
  "foldseek_version": "Foldseek enables fast and sensitive comparisons of large structure sets. It reaches sensitivities similar to state-of-the-art structural aligners while being at least 20,000 times faster.\n\nPlease cite:\nvan Kempen, M., Kim, S.S., Tumescheit, C., Mirdita, M., Lee, J., Gilchrist, C.L.M., S\u00f6ding, J., and Steinegger, M. Fast and accurate protein structure search with Foldseek. Nature Biotechnology, doi:10.1038/s41587-023-01773-0 (2023)\n\nfoldseek Version: 941cd33ff0771cd2e3f144e3293e22a2b87e9fda\n\u00a9 Michel van Kempen, Stephanie Kim, Charlotte Tumescheit, Milot Mirdita, Jeongjae Lee, Cameron L. M. Gilchrist, Johannes S\u00f6ding, Martin Steinegger\n\nusage: foldseek <command> [<args>]\n\nEasy workflows for plain text input/output\n  easy-search       \tStructual search\n  easy-cluster      \tSlower, sensitive clustering\n  easy-rbh          \tFind reciprocal best hit\n  easy-multimercluster\tMultimer level cluster\n  easy-multimersearch\tMultimer level search\n\nMain workflows for database input/output\n  createdb          \tConvert PDB/mmCIF/tar[.gz]/DB files or directory/TSV to a structure DB\n  search            \tSensitive homology search\n  rbh               \tReciprocal best hit search\n  cluster           \tSlower, sensitive clustering\n  multimercluster   \tMultimer level cluster\n  multimersearch    \tMultimer level search\n\nInput database creation\n  databases         \tList and download databases\n  createindex       \tStore precomputed index on disk to reduce search overhead\n  createclusearchdb \tBuild a searchable cluster database allowing for faster searches\n\nUnite and intersect databases\n  createsubdb       \tCreate a subset of a DB from list of DB keys\n\nFormat conversion for downstream processing\n  convertalis       \tConvert alignment DB to BLAST-tab, SAM or custom format\n  compressca        \tCreate a new C-alpha DB with chosen compression encoding from a sequence DB\n  convert2pdb       \tConvert a foldseek structure db to a single multi model PDB file or a directory of PDB files\n  createmultimerreport\tConvert complexDB to tsv format\n\nPrefiltering        \n  expandmultimer    \tRe-prefilter to ensure complete alignment between multimers\n\nAlignment           \n  tmalign           \tCompute tm-score \n  structurealign    \tCompute structural alignment using 3Di alphabet, amino acids and neighborhood information\n  structurerescorediagonal\tCompute sequence identity for diagonal\n  aln2tmscore       \tCompute tmscore of an alignment database \n  scoremultimer     \tGet multimer level alignments from alignmentDB\n\nClustering          \n  clust             \tCluster result by Set-Cover/Connected-Component/Greedy-Incremental\n\nProfile databases   \n  result2profile    \tCompute profile DB from a result DB for both amino acid and 3di\n\n\nInvalid Command: --version\nDid you mean \"foldseek convertalis\"?",
  "local_shard_rows": 1000,
  "local_structure_layout": "structures/{split}-*.parquet",
  "max_ca_gap": 10.0,
  "min_sequence_length": 10,
  "mmcif_dir": "/data/mmcif",
  "mmseqs_path": "/usr/local/bin/mmseqs",
  "mmseqs_version": "MMseqs2 (Many against Many sequence searching) is an open-source software suite for very fast, \nparallelized protein sequence searches and clustering of huge protein sequence data sets.\n\nPlease cite: M. Steinegger and J. Soding. MMseqs2 enables sensitive protein sequence searching for the analysis of massive data sets. Nature Biotechnology, doi:10.1038/nbt.3988 (2017).\n\nMMseqs2 Version: 8cc5ce367b5638c4306c2d7cfc652dd099a4643f\n\u00a9 Martin Steinegger (martin.steinegger@snu.ac.kr)\n\nusage: mmseqs <command> [<args>]\n\nEasy workflows for plain text input/output\n  easy-search       \tSensitive homology search\n  easy-linsearch    \tFast, less sensitive homology search\n  easy-cluster      \tSlower, sensitive clustering\n  easy-linclust     \tFast linear time cluster, less sensitive clustering\n  easy-taxonomy     \tTaxonomic classification\n  easy-rbh          \tFind reciprocal best hit\n\nMain workflows for database input/output\n  search            \tSensitive homology search\n  linsearch         \tFast, less sensitive homology search\n  map               \tMap nearly identical sequences\n  rbh               \tReciprocal best hit search\n  linclust          \tFast, less sensitive clustering\n  cluster           \tSlower, sensitive clustering\n  clusterupdate     \tUpdate previous clustering with new sequences\n  taxonomy          \tTaxonomic classification\n\nInput database creation\n  databases         \tList and download databases\n  createdb          \tConvert FASTA/Q file(s) to a sequence DB\n  createindex       \tStore precomputed index on disk to reduce search overhead\n  createlinindex    \tCreate linsearch index\n  convertmsa        \tConvert Stockholm/PFAM MSA file to a MSA DB\n  tsv2db            \tConvert a TSV file to any DB\n  tar2db            \tConvert content of tar archives to any DB\n  db2tar            \tArchive contents of a DB to a tar archive\n  msa2profile       \tConvert a MSA DB to a profile DB\n\nHandle databases on storage and memory\n  compress          \tCompress DB entries\n  decompress        \tDecompress DB entries\n  rmdb              \tRemove a DB\n  mvdb              \tMove a DB\n  cpdb              \tCopy a DB\n  lndb              \tSymlink a DB\n  aliasdb           \tCreate relative symlink of DB to another name in the same folder\n  unpackdb          \tUnpack a DB into separate files\n  touchdb           \tPreload DB into memory (page cache)\n  gpuserver         \tStart a GPU server\n\nUnite and intersect databases\n  createsubdb       \tCreate a subset of a DB from list of DB keys\n  concatdbs         \tConcatenate two DBs, giving new IDs to entries from 2nd DB\n  splitdb           \tSplit DB into subsets\n  mergedbs          \tMerge entries from multiple DBs\n  subtractdbs       \tRemove all entries from first DB occurring in second DB by key\n\nFormat conversion for downstream processing\n  convertalis       \tConvert alignment DB to BLAST-tab, SAM or custom format\n  createtsv         \tConvert result DB to tab-separated flat file\n  convert2fasta     \tConvert sequence DB to FASTA format\n  result2flat       \tCreate flat file by adding FASTA headers to DB entries\n  createseqfiledb   \tCreate a DB of unaligned FASTA entries\n  taxonomyreport    \tCreate a taxonomy report in Kraken or Krona format\n\nSequence manipulation/transformation\n  extractorfs       \tSix-frame extraction of open reading frames\n  extractframes     \tExtract frames from a nucleotide sequence DB\n  orftocontig       \tWrite ORF locations in alignment format\n  reverseseq        \tReverse (without complement) sequences\n  translatenucs     \tTranslate nucleotides to proteins\n  translateaa       \tTranslate proteins to lexicographically lowest codons\n  splitsequence     \tSplit sequences by length\n  masksequence      \tSoft mask sequence DB using tantan\n  extractalignedregion\tExtract aligned sequence region from query\n\nResult manipulation \n  swapresults       \tTranspose prefilter/alignment DB\n  result2rbh        \tFilter a merged result DB to retain only reciprocal best hits\n  result2msa        \tCompute MSA DB from a result DB\n  result2dnamsa     \tCompute MSA DB with out insertions in the query for DNA sequences\n  result2stats      \tCompute statistics for each entry in a DB\n  filterresult      \tPairwise alignment result filter\n  offsetalignment   \tOffset alignment by ORF start position\n  proteinaln2nucl   \tTransform protein alignments to nucleotide alignments\n  result2repseq     \tGet representative sequences from result DB\n  sortresult        \tSort a result DB in the same order as the prefilter or align module\n  summarizealis     \tSummarize alignment result to one row (uniq. cov., cov., avg. seq. id.)\n  summarizeresult   \tExtract annotations from alignment DB\n\nTaxonomy assignment \n  createtaxdb       \tAdd taxonomic labels to sequence DB\n  createbintaxonomy \tCreate binary taxonomy from NCBI input\n  createdmptaxonomy \tCreate dmp files from binary taxonomy\n  createbintaxmapping\tCreate binary taxonomy mapping from tabular taxonomy mapping\n  addtaxonomy       \tAdd taxonomic labels to result DB\n  taxonomyreport    \tCreate a taxonomy report in Kraken or Krona format\n  filtertaxdb       \tFilter taxonomy result database\n  filtertaxseqdb    \tFilter taxonomy sequence database\n  aggregatetax      \tAggregate multiple taxon labels to a single label\n  aggregatetaxweights\tAggregate multiple taxon labels to a single label\n  lcaalign          \tEfficient gapped alignment for lca computation\n  lca               \tCompute the lowest common ancestor\n  majoritylca       \tCompute the lowest common ancestor using majority voting\n\nMulti-hit search    \n  multihitdb        \tCreate sequence DB for multi hit searches\n  multihitsearch    \tSearch with a grouped set of sequences against another grouped set\n  besthitperset     \tFor each set of sequences compute the best element and update p-value\n  combinepvalperset \tFor each set compute the combined p-value\n  mergeresultsbyset \tMerge results from multiple ORFs back to their respective contig\n\nPrefiltering        \n  prefilter         \tDouble consecutive diagonal k-mer search\n  ungappedprefilter \tOptimal diagonal score search\n  gappedprefilter   \tOptimal Smith-Waterman-based prefiltering (slow)\n  kmermatcher       \tFind bottom-m-hashed k-mer matches within sequence DB\n  kmersearch        \tFind bottom-m-hashed k-mer matches between target and query DB\n\nAlignment           \n  align             \tOptimal gapped local alignment\n  alignall          \tWithin-result all-vs-all gapped local alignment\n  transitivealign   \tTransfer alignments via transitivity\n  rescorediagonal   \tCompute sequence identity for diagonal\n  fwbw              \tForward Backward Alignment\n  alignbykmer       \tHeuristic gapped local k-mer based alignment\n\nClustering          \n  pickconsensusrep  \tSelect new representatives for each cluster based on consensus\n  clust             \tCluster result by Set-Cover/Connected-Component/Greedy-Incremental\n  clusthash         \tHash-based clustering of equal length sequences\n  mergeclusters     \tMerge multiple cascaded clustering steps\n\nProfile databases   \n  result2profile    \tCompute profile DB from a result DB\n  msa2result        \tConvert a MSA DB to a profile DB\n  msa2profile       \tConvert a MSA DB to a profile DB\n  sequence2profile  \tTurn sequence into profile by adding context specific pseudo counts\n  profile2pssm      \tConvert a profile DB to a tab-separated PSSM file\n  profile2neff      \tConvert a profile DB to a tab-separated list of Neff scores\n  profile2consensus \tExtract consensus sequence DB from a profile DB\n  profile2repseq    \tExtract representative sequence DB from a profile DB\n  convertprofiledb  \tConvert a HH-suite HHM DB to a profile DB\n\nProfile-profile databases\n  tsv2exprofiledb   \tCreate a expandable profile db from TSV files\n  convertca3m       \tConvert a cA3M DB to a result DB\n  expandaln         \tExpand an alignment result based on another\n  expand2profile    \tExpand an alignment result based on another and create a profile\n\nUtility modules to manipulate DBs\n  setextendeddbtype \tWrite an extended DB \n  view              \tPrint DB entries given in --id-list to stdout\n  apply             \tExecute given program on each DB entry\n  filterdb          \tDB filtering by given conditions\n  swapdb            \tTranspose DB with integer values in first column\n  prefixid          \tFor each entry in a DB prepend the entry key to the entry itself\n  suffixid          \tFor each entry in a DB append the entry key to the entry itself\n  renamedbkeys      \tCreate a new DB with original keys renamed\n\nSpecial-purpose utilities\n  diffseqdbs        \tCompute diff of two sequence DBs\n  summarizetabs     \tExtract annotations from HHblits BLAST-tab-formatted results\n  gff2db            \tExtract regions from a sequence database based on a GFF3 file\n  maskbygff         \tMask out sequence regions in a sequence DB by features selected from a GFF3 file\n  convertkb         \tConvert UniProtKB data to a DB\n  summarizeheaders  \tSummarize FASTA headers of result DB\n  nrtotaxmapping    \tCreate taxonomy mapping for NR database\n  extractdomains    \tExtract highest scoring alignment regions for each sequence from BLAST-tab file\n  countkmer         \tCount k-mers\n\nBash completion for modules and parameters can be installed by adding \"source MMSEQS_HOME/util/bash-completion.sh\" to your \"$HOME/.bash_profile\".\nInclude the location of the MMseqs2 binary in your \"$PATH\" environment variable.\n\nInvalid Command: --version\nDid you mean \"/usr/local/bin/mmseqs_avx2 reverseseq\"?",
  "official_sequence_threshold": 0.3,
  "platform": "Linux-6.8.0-1046-nvidia-x86_64-with-glibc2.41",
  "protein_atom_weight": 1.0,
  "repo_id": "Synthyra/PDB-Monomeric-Structure-ESMFold2",
  "residue_rows": 44845076,
  "schema": "atom37",
  "seed": 42,
  "sequence_thresholds": [
    0.3,
    0.5,
    0.7,
    0.9,
    0.95,
    1.0
  ],
  "sequence_verification_edges": null,
  "smooth_lddt_cutoff_angstrom": 15.0,
  "split_counts": {
    "test": 7384,
    "train": 132922,
    "valid": 7384
  },
  "structure_rows": 147690,
  "structure_shards": {
    "test": 8,
    "train": 133,
    "valid": 8
  },
  "test_frac": 0.05,
  "use_gpu": false,
  "valid_frac": 0.05
}

Leakage Assertions

{
  "chains": {
    "fields": {
      "chain_uid": {
        "pairwise": {
          "max_pairwise_overlap": 0,
          "unique_values": 147690
        },
        "partition": {
          "crossing_values": 0,
          "unique_values": 147690
        }
      },
      "seqs": {
        "pairwise": {
          "max_pairwise_overlap": 0,
          "unique_values": 51156
        },
        "partition": {
          "crossing_values": 0,
          "unique_values": 51156
        }
      },
      "sequence_cluster_30": {
        "pairwise": {
          "max_pairwise_overlap": 0,
          "unique_values": 16108
        },
        "partition": {
          "crossing_values": 0,
          "unique_values": 16108
        }
      },
      "sequence_hash": {
        "pairwise": {
          "max_pairwise_overlap": 0,
          "unique_values": 51156
        },
        "partition": {
          "crossing_values": 0,
          "unique_values": 51156
        }
      },
      "split_component": {
        "pairwise": {
          "max_pairwise_overlap": 0,
          "unique_values": 5971
        },
        "partition": {
          "crossing_values": 0,
          "unique_values": 5971
        }
      },
      "structure_cluster": {
        "pairwise": {
          "max_pairwise_overlap": 0,
          "unique_values": 8173
        },
        "partition": {
          "crossing_values": 0,
          "unique_values": 8173
        }
      }
    },
    "rows": 147690,
    "sequence_verification_edges": {
      "checked_edges": 0,
      "cross_split_edges": 0,
      "skipped_self_edges": 0,
      "skipped_unknown_edges": 0,
      "total_edges": 0
    },
    "structure_verification_edges": {
      "checked_edges": 0,
      "cross_split_edges": 0,
      "skipped_self_edges": 0,
      "skipped_unknown_edges": 0,
      "total_edges": 0
    }
  },
  "structures": {
    "atom37_slots": 37,
    "fields": {
      "pdb_id": {
        "pairwise": {
          "max_pairwise_overlap": 0,
          "unique_values": 98272
        },
        "partition": {
          "crossing_values": 0,
          "unique_values": 98272
        }
      },
      "seqs": {
        "pairwise": {
          "max_pairwise_overlap": 0,
          "unique_values": 51156
        },
        "partition": {
          "crossing_values": 0,
          "unique_values": 51156
        }
      },
      "sequence_cluster_30": {
        "pairwise": {
          "max_pairwise_overlap": 0,
          "unique_values": 16108
        },
        "partition": {
          "crossing_values": 0,
          "unique_values": 16108
        }
      },
      "sequence_hash": {
        "pairwise": {
          "max_pairwise_overlap": 0,
          "unique_values": 51156
        },
        "partition": {
          "crossing_values": 0,
          "unique_values": 51156
        }
      },
      "split_component": {
        "pairwise": {
          "max_pairwise_overlap": 0,
          "unique_values": 5971
        },
        "partition": {
          "crossing_values": 0,
          "unique_values": 5971
        }
      },
      "structure_cluster": {
        "pairwise": {
          "max_pairwise_overlap": 0,
          "unique_values": 8173
        },
        "partition": {
          "crossing_values": 0,
          "unique_values": 8173
        }
      }
    },
    "max_ca_gap": 10.0,
    "residues": 44845076,
    "rows": 147690,
    "shard_size": 1000,
    "shards": {
      "test": 8,
      "train": 133,
      "valid": 8
    }
  }
}