stem.MAP: Identification of cell type specific protein complexes in mammalian embryonic stem cells.

Proteins self-assemble into macromolecular protein complexes, responsible for carrying out almost all cellular function. Therefore it is essential to characterize all protein-protein interactions in the cell. Major efforts utlilizing high-throughput proteomics [1-8], machine learning [9-11], and labor-intensive manual curation [12, 13] have increased the accuracy and coverage of protein complexes in various organisms, however, it is likely that only a subset of these complexes are present in a given cell or tissue type at a specific time. In fact, it has been shown that in two commonly use cell types, HCT116 and HEK293T [3-5], only share ~50% of their interactions, suggesting substantial cell-type-specific interaction remodeling. Resources of this magnitude require a considerable amount of experiments, and are therefore limited by their cost and laborious efforts to develop. Here we present a strategy to predict protein complexes in mouse embryonic stem cells using relatively low amounts of experiments (<250) compared to other methods, incorporating prior interaction information from cell-type agnostic complex maps, and pre-training engineering to identify 3,986 protein complexes covering 8,839 proteins. These predictions include stem cell specific chromatin remodeler complexes and novel stem cell interactions.

Funding

NIH R00, NIH R35, NSF/BBSRC, American Heart Association

Citation

Akshaya Rajaraman, Samantha N Fischer, Miles Woodcock-Girard, Riddhiman K. Garge, Kevin Drew

References

1. Krogan et al. Global Landscape of Protein Complexes in the Yeast Saccharomyces Cerevisiae. Nature. 2006. https://doi.org/10.1038/nature04670
2. Gavin et al. Proteome survey reveals modularity of the yeast cell machinery. Nature. 2006. https://doi.org/10.1038/nature04532
3. Huttlin et al. The BioPlex Network: A Systematic Exploration of the Human Interactome. Cell. 2015. https://doi.org/10.1016/j.cell.2015.06.043
4. Huttlin et al. Architecture of the Human Interactome Defines Protein Communities and Disease Networks. Nature. 2017. https://doi.org/10.1038/nature22366
5. Huttlin et al. Dual Proteome-scale Networks Reveal Cell-specific Remodeling of the Human Interactome. Cell. 2021. https://doi.org/10.1016/j.cell.2021.04.011
6. Hein et al. A human interactome in three quantitative dimensions organized by stoichiometries and abundances. Cell. 2015. https://doi.org/10.1016/j.cell.2015.09.053
7. Wan et al. Panorama of ancient metazoan macromolecular complexes. Nature. 2015. https://doi.org/10.1038/nature14877
8. McWhite et al. A pan-plant protein complex map reveals deep conservation and novel assemblies. Cell. 2020. https://doi.org/10.1016/j.cell.2020.02.049
9. Drew et al. Integration of over 9,000 mass spectrometry experiments builds a global map of human protein complexes. Mol Sys Biol. 2017. https://doi.org/10.15252/msb.20167490
10. Drew et al. Hu.MAP 2.0: Integration of over 15,000 proteomic experiments builds a global compendium of human multiprotein assemblies. Mol Sys Biol. 2021. https://doi.org/10.15252/msb.202010016
11. Fischer et al. hu.MAP3.0: Atlas of human protein complexes by integration of > 25,000 proteomic experiments. Mol Syst Biol. 2025. https://doi.org/10.1038/s44320-025-00121-5
12. Balu, et al. Complex portal 2025: predicted human complexes and enhanced visualisation tools for the comparison of orthologous and paralogous complexes. Nucleic Acids Research. 2025. https://doi.org/10.1093/nar/gkae1085
13. Steinkamp et al. CORUM in 2024: protein complexes as drug targets. Nucleic Acids Research. 2025. https://doi.org/10.1093/nar/gkae1033

Description of the dataset files

├── full/
  └── stem.MAP_all_pairs_full_feature_set_20240815.parquet
  File that contains all pairwise interactions (>96 Mil.) and any data across the 6 features.
  Proteins (in 'ID' column) are listed as UniProt accessions.
  Not all protein pairs have data in every feature. No present data is represented by NaN.

├── reference_interactions/

  ├── stem.MAP_benchmark_20231026_TEST_negative_PPIs.txt 
    File of test set negative interactions in the format of one protein pair per line using UniProt accessions
  
  ├── stem.MAP_benchmark_20231026_TEST_positive_PPIs.txt 
    File of test set positive interactions in the format of one protein pair per line using UniProt accessions
  
  ├── stem.MAP_benchmark_20231026_TRAIN_negative_PPIs.txt 
    File of train set negative interactions in the format of one protein pair per line using UniProt accessions
  
  └── stem.MAP_benchmark_20231026_TRAIN_positive_PPIs.txt 
    File of train set positive interactions in the format of one protein pair per line using UniProt accessions
  
├── test/
  └── stem.MAP_test_set_20240730.parquet
    File containing the test set feature matrix that was used for stem.MAP model evaluation. 
    Proteins are represented as UniProt accessions.
    Contains 2,519 positive interactions and 216,523 negative interactions.

└── train/
  └── stem.MAP_train_set_20240730.parquet
    File containing the train set feature matrix that was used to train the stem.MAP model.
    Proteins are represented as UniProt accessions.
    Contains 1,452 positive interactions and 8,827 negative interactions

Associated Code

Additional code examples can be found on our GitHub, including: importing the stem.MAP model to make predictions, importing the training and testing data, or using the full feature matrix.

Usage

Accessing the model

stem.MAP was constructed using AutoGluon an auto-ML tool. The module TabularPredictor is used to is used train, test, and make predictions with the model.

This can be downloaded using the following:

  $ pip install autogluon==0.8.2

Then it can be imported as:

>>> from autogluon.tabular import TabularPredictor

Note that to perform operations with our model the 0.8.2 version must be used. Alternatively, if disk space is a concern the user can just install autogluon.tabular

The stem.MAP model can be accessed through HuggingFace with huggingface_hub

 >>> from huggingface_hub import snapshot_download

>>> model_dir = snapshot_download(repo_id="DrewLab/stem.MAP_AutoGluon")

>>> predictor = TabularPredictor.load(f"{model_dir}/stem.MAP_AutoGluon_model_20240731")

Using the training and testing data

Both the train and test feature matrices can be loaded using the Huggingface datasets library.

This can be done from the command-line using:

  $ pip install datasets

When loading into Python use the following:

>>> from datasets import load_dataset
>>> dataset = load_dataset('DrewLab/stem.MAP')

Training and test feature matrices can then be accessed as separate objects:

>>> train = dataset["train"].to_pandas()
>>> test = dataset["test"].to_pandas()

Jupyter notebooks containing more in-depth examples of model training, testing, and generating predictions can be found on our GitHub

Using the full feature matrix and the interaction files

All other files, such as the full feature matrix, can be accessed via Huggingface_hub.

>>> from huggingface_hub import hf_hub_download
>>> full_file = hf_hub_download(repo_id="DrewLab/stem.MAP", filename='full/stem.MAP_all_pairs_full_feature_set_20240815.parquet', repo_type='dataset')

This just provides the file for download. Depending on your workflow, if you wish to use as a pandas dataframe for example:

>>> import pandas as pd
>>> full_featmat = pd.read_parquet(full_file)

Dataset card authors

Samantha Fischer (sfisch6@uic.edu)