DrewLab/stem.MAP_AutoGluon

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

Associated Code

Code examples for using the stem.MAP model can be found on our GitHub Our feature matrices used for testing, training, and making predictions can be found at our HuggingFace dataset All predicted protein complexes and pairwise predictions can be found on our website

Using the stem.MAP 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

Or simply:

  $ pip install autogluon.tabular==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

To use the model and make predictions, we show two full code examples using the full feature matrix and the test feature matrix in jupyter notebooks.

All feature matrices can be pulled using the 'datasets' module from HuggingFace and examples of that are seen on our GitHub and on our stem.MAP HuggingFace dataset

Model card authors

Samantha Fischer (sfisch6@uic.edu)