Documentation
Hypothetical protein structures

Hypothetical protein structures

Note: All structures available here have been integrated into the AlphaFold Database.

For more detailed visual analysis and downloads, please see https://alphafold.ebi.ac.uk

You cannot search the AFDB using the ATB/Bakta locus tags - see below for more information on how to map the ATB/Bakta tags to AFDB accessions.

Preprocessing

Note: All .pdb files in this dataset use the locus_tag from Bakta, which contains the SRA accession before the ‘.’ delimited e.g. SAMEA7561354.JGMNBG_20630 comes from SRA accession SAMEA7561354.

To begin: all 275,945,762 hypothetical proteins from the Bakta annotations were taken.

Then, only proteins belonging to genomes passing QC (i.e. PASS) from Aggregated/work_in_progress_2025-05/species_calls.tsv.gz here were filtered to prioritise highly confident genome assemblies and protein calls.

This left 224,184,215 proteins remaining for downstream consideration.

These were combined and deduplicated at 100% seqid over 100% sequence length.

Then, this set was filtered to keep only proteins below 3000AA, due to VRAM limitations of available GPUs for folding.

We have not released the additional longer proteins (work in progress, contact George Bouras if interested).

Overall, this left 17,711,165 unique proteins.

Taxonomy

We used the sylph species predictions from Aggregated/work_in_progress_2025-05/species_calls.tsv.gz. These were generated using the GTDB GTDB 214.0 release (28 April 2023 database).

We also mapped these to NCBI taxids using the bac120_metadata_r214.tsv and ar53_metadata_r214.tsv release from GTDB 214.

The NCBI taxonomy calls are what you will see in AlphaFold Database.

Predictions

Multiple sequence alignments were generated in chunks of 250,000 proteins using ColabFold v1.5.5 using MMseqs2 version 15.6f452 with the colabfold_search command. The default parameters and databases (uniref30_2302 and colabfold_envdb_202108 ) were used.

Each AlphaFold2 implemented via ColabFold v1.5.5 was then run with each MSA as input using colabfold_batch with ptm model 1 only (--num-models 1) for maximum throughput, using otherwise the default parameters ( 3 recycles, no templates or relaxation ).

All predictions were conducted on AMD MI250x GPUs on Setonix at the Pawsey Supercomputing Research Centre in Perth, Australia using the specific container tag rocm6_cpuTF available at https://quay.io/repository/sarahbeecroft9/colabfold?tab=tags.

Files

Foldseek Database

We provide a Foldseek database (created using foldseek createdb) via OSF (https://osf.io/x693m/overview) in the Foldseek_Databases component.

Note that due to the size restrictions on individual file, it was split into 1GB parts for OSF

To remake it: cat foldseek_db.tar.gz_parta* > foldseek_db.tar.gz && tar -xzf foldseek_db.tar.gz

Foldcomp Database

We provide a Foldcomp database (created using foldseek createdb) via OSF (https://osf.io/x693m/overview) in the Foldcomp_Databases component.

Note that due to the size restrictions on individual file, it was split into 1GB parts for OSF

To remake it: cat foldcomp_db.tar.gz_parta* > foldcomp_db.tar.gz && tar -xzf foldcomp_db.tar.gz

Predictions

If you want to full, non-lossy compressed predictions, we provide .pdb format structures via OSF.

Hosting other AF2 output files (MSAs, pLDDT and PAE .json confidence metrics) take many many TBs of storage and are beyond our capacity to host on OSF. These can be accessed via https://alphafold.ebi.ac.uk

All .pdb files are batched into 142 batches .tar.gz files at https://osf.io/x693m/overview.

Each .tar.gz file contains the subdirectories with the SRA accessions of each protein in the batch. The .pdb files are then nested inside the subdirectories.

These are themselves in 9 Sets. Each Set contains 16 batches other than Set 9 which contains 14.

FASTA

The FASTA component contains all 275,945,762 unfiltered hypothetical proteins in FASTA format.

Note: you will also need to put it back together cat all_hypotheticals_under_3000AA_unfiltered.fastq.gz_part* > all_hypotheticals_under_3000AA_unfiltered.fastq.gz_part.gz

Metadata

To map between the ATB (i.e. Bakta with SRA accession) locus tags, the AlphaFold Database tags and the batch tarball (above), please see ATB_AFDB_map_with_batches.tsv.gz

This provides a 3 column tsv file that maps the ATB/Bakta locus tag (column 1) to the AlphaFold database accession 16 digit code (column 2) OSF batch (column 3).

e.g.

head -n1 ATB_hypothetical_protein_structures_batch.tsv
SAMEA7561354.JGMNBG_20630   0000000000710001    batch001

We also provide the representative protein (i.e. - what actually ended up being folded) , sylph GTDB and mapped NCBI species calls for all 224,184,215 member proteins before deduplication.

This provides a 5 column tsv file ATB_every_protein_with_species.tsv.gz with the representative (column 1), member (column 2), sylph GTDB species name of the member (column 3), mapped NCBI taxonomy species name (column 4) and NCBI taxid (column 5)

e.g.

head -n1 ATB_every_protein_with_species.tsv
SAMEA7561354.JGMNBG_20625   SAMEA7561354.JGMNBG_20625   Mycobacterium tuberculosis  Mycobacterium tuberculosis H37R 83332

Note: you will also need to put it back together cat ATB_every_protein_with_species.tsv.gz_parta* > ATB_every_protein_with_species.gz

For any questions regarding the protein structure predictions, please contact George Bouras george.bouras@adelaide.edu.au

Biosynthetic gene clustersWhatsGNU protein allele frequencies