RosettaCommons/MIP

Microbiome Immunity Project: Protein Universe

~200,000 predicted structures for diverse protein sequences from 1,003 representative genomes across the microbial tree of life and annotate them functionally on a per-residue basis.

Quickstart Usage

Install HuggingFace Datasets package

Each subset can be loaded into python using the Huggingface datasets library. First, from the command line install the datasets library

$ pip install datasets

Optionally set the cache directory, e.g.

$ HF_HOME=${HOME}/.cache/huggingface/
$ export HF_HOME

then, from within python load the datasets library

>>> import datasets

Load model datasets

To load one of the MIP model datasets, use datasets.load_dataset(...):

>>> dataset_tag = "rosetta_high_quality"
>>> dataset_models = datasets.load_dataset(
  path = "RosettaCommons/MIP",
  name = f"{dataset_tag}_models",
  data_dir = f"{dataset_tag}_models")['train']
Resolving data files: 100%|█████████████████████████████████████████| 54/54 [00:00<00:00, 441.70it/s]
Downloading data: 100%|███████████████████████████████████████████| 54/54 [01:34<00:00,  1.74s/files]
Generating train split: 100%|███████████████████████| 211069/211069 [01:41<00:00, 2085.54 examples/s]
Loading dataset shards: 100%|███████████████████████████████████████| 48/48 [00:00<00:00, 211.74it/s]

and the dataset is loaded as a datasets.arrow_dataset.Dataset

>>> dataset_models
Dataset({
    features: ['id', 'pdb', 'Filter_Stage2_aBefore', 'Filter_Stage2_bQuarter', 'Filter_Stage2_cHalf', 'Filter_Stage2_dEnd', 'clashes_bb', 'clashes_total', 'score', 'silent_score', 'time'],
    num_rows: 211069
})

which is a column oriented format that can be accessed directly, converted in to a pandas.DataFrame, or parquet format, e.g.

>>> dataset_models.data.column('pdb')
>>> dataset_models.to_pandas()
>>> dataset_models.to_parquet("dataset.parquet")

Load Function Predictions

Function predictions are generated using DeepFRI across

>>> dataset_function_prediction = datasets.load_dataset(
  path = "RosettaCommons/MIP",
  name = f"{dataset_tag}_function_predictions",
  data_dir = f"{dataset_tag}_function_predictions")['train']
Downloading readme: 100%|████████████████████████████████████████| 15.4k/15.4k [00:00<00:00, 264kB/s]
Resolving data files: 100%|██████████████████████████████████████| 219/219 [00:00<00:00, 1375.51it/s]
Downloading data: 100%|█████████████████████████████████████████| 219/219 [13:04<00:00,  3.58s/files]
Generating train split: 100%|████████████| 1332900735/1332900735 [13:11<00:00, 1684288.89 examples/s]
Loading dataset shards: 100%|██████████████████████████████████████| 219/219 [01:22<00:00,  2.66it/s]

this loads the >1.3B function predictions for all 211069 targets across 6315 GO and EC ontology terms. The predictions are stored in long format, but can be easily converted to a wide format using pandas:

>>> import pandas
>>> dataset_function_prediction_wide = pandas.pivot(
  dataset_function_prediction.data.select(['id', 'term_id', 'Y_hat']).to_pandas(),
  columns = "term_id",
  index = "id",
  values = "Y_hat")
>>> dataset_function_prediction_wide.shape
(211069, 6315)

Dataset Details

Dataset Description

Large-scale structure prediction on representative protein domains from the Genomic Encyclopedia of Bacteria and Archaea (GEBA1003) reference genome database across the microbial tree of life. From a non-redundant GEBA1003 gene catalog protein sequences without matches to any structural databases and which produced multiple-sequence alignments of N_eff > 16 and all putative novel domains between 40 and 200 residues were extracted. For each sequence 20,000 Rosetta de novo models and up to 5 DMPfold models were generated. The initial output dataset (MIP_raw) of about 240,000 models were curated to high-quality models comprising about 75% of the original dataset (MIP_curated): Models were filtered out if (1) Rosetta models had >60% coil content or DMPFold models with >80% coil content, (2) the averaging the pairwise TM-scores of the 10 lowest-scoring models was less than 0.4, and (3) if the Rosetta and DMPfold models had TM-score less than 0.5. Functional annotations of the entire dataset were created using structure-based Graph Convolutional Network embeddings from DeepFRI. The highest quality structure for each sequence for both Rosetta and DMPFold, is included in this dataset; the entire dataset of more than 5 billion Rosetta models and 1 million DMPFold models is available upon request.

• Acknowledgements: We kindly acknowledge the support of the IBM World Community Grid team (Caitlin Larkin, Juan A Hindo, Al Seippel, Erika Tuttle, Jonathan D Armstrong, Kevin Reed, Ray Johnson, and Viktors Berstis), and the community of 790,000 volunteers who donated 140,661 computational years since Aug 2017 of their computer time over the course of the project. This research was also supported in part by PLGrid Infrastructure (to PS). The authors thank Hera Vlamakis and Damian Plichta from the Broad Institute for helpful discussions. The work was supported by the Flatiron Institute as part of the Simons Foundation to J.K.L., P.D.R., V.G., D.B., C.C., A.P., N.C., I.F., and R.B. This research was also supported by grants NAWA PPN/PPO/2018/1/00014 to P.S. and T.K., PLGrid to P.S., and NIH - DK043351 to T.V. and R.J.X.

• License: cc-by-4.0

Dataset Sources

• Repository: https://github.com/microbiome-immunity-project/protein_universe
• Paper: Koehler Leman, J., Szczerbiak, P., Renfrew, P. D., Gligorijevic, V., Berenberg, D., Vatanen, T., … Kosciolek, T. (2023). Sequence-structure-function relationships in the microbial protein universe. Nature Communications, 14(1), 2351. doi:10.1038/s41467-023-37896-w
• Zenodo Repository: https://doi.org/10.5281/zenodo.6611431

Uses

Exploration of sequence-structure-function relationship in naturally ocurring proteins. The MIP database is complementary to and distinct from the other large-scale predicted protein structure databases such as the EBI AlphaFold database because it consists of proteins from Archaea and Bacteria, whose protein sequences are generally shorter than Eukaryotic.

Out-of-Scope Use

While this dataset has been curated for quality, in some cases the predicted structures may not represent physically realistic conformations. Thus caution much be used when using it as training data for protein structure prediction and design.

Source Data

Sequences were obtained from the Genomic Encyclopedia of Bacteria and Archaea (GEBA1003) reference genome database across the microbial tree of life:

1,003 reference genomes of bacterial and archaeal isolates expand coverage of the tree of life We present 1,003 reference genomes that were sequenced as part of the Genomic Encyclopedia of Bacteria and Archaea (GEBA) initiative, selected to maximize sequence coverage of phylogenetic space. These genomes double the number of existing type strains and expand their overall phylogenetic diversity by 25%. Comparative analyses with previously available finished and draft genomes reveal a 10.5% increase in novel protein families as a function of phylogenetic diversity. The GEBA genomes recruit 25 million previously unassigned metagenomic proteins from 4,650 samples, improving their phylogenetic and functional interpretation. We identify numerous biosynthetic clusters and experimentally validate a divergent phenazine cluster with potential new chemical structure and antimicrobial activity. This Resource is the largest single release of reference genomes to date. Bacterial and archaeal isolate sequence space is still far from saturated, and future endeavors in this direction will continue to be a valuable resource for scientific discovery.

Citation

@article{KoehlerLeman2023,
  title = {Sequence-structure-function relationships in the microbial protein universe},
  volume = {14},
  ISSN = {2041-1723},
  url = {http://dx.doi.org/10.1038/s41467-023-37896-w},
  DOI = {10.1038/s41467-023-37896-w},
  number = {1},
  journal = {Nature Communications},
  publisher = {Springer Science and Business Media LLC},
  author = {Koehler Leman,  Julia and Szczerbiak,  Pawel and Renfrew,  P. Douglas and Gligorijevic,  Vladimir and Berenberg,  Daniel and Vatanen,  Tommi and Taylor,  Bryn C. and Chandler,  Chris and Janssen,  Stefan and Pataki,  Andras and Carriero,  Nick and Fisk,  Ian and Xavier,  Ramnik J. and Knight,  Rob and Bonneau,  Richard and Kosciolek,  Tomasz},
  year = {2023},
  month = apr
}

Dataset Card Authors

Matthew O'Meara (maom@umich.edu)