DirectContacts2: A network of direct physical protein interactions derived from high throughput mass spectrometry experiments

Proteins carry out cellular functions by self-assembling into functional complexes, a process that depends on direct physical interactions between components. While tools like AlphaFold and RoseTTAFold have advanced structure prediction, they remain limited in scaling to the full human proteome. DirectContacts2 addresses this challenge by integrating diverse large-scale protrin interaction datasets, including AP/MS (BioPlex1–3, Boldt et al., Hein et al.), biochemical fractionation (Wan et al.), proximity labeling (Gupta et al., Youn et al.), and RNA pulldown (Treiber et al.), to predict whether ~26 million human protein pairs interact directly or indirectly.

Funding

NIH R00, NSF/BBSRC

Citation

Erin R. Claussen, Miles D Woodcock-Girard, Samantha N Fischer, Kevin Drew

References

Kevin Drew, Christian L. Müller , Richard Bonneau, Edward M. Marcotte (2017) Identifying direct contacts between protein complex subunits from their conditional dependence in proteomics datasets. PLOS Computational Biology 13(10): e1005625. https://doi.org/10.1371/journal.pcbi.1005625
Samantha N. Fischer, Erin R Claussen, Savvas Kourtis, Sara Sdelci, Sandra Orchard, Henning Hermjakob, Georg Kustatscher, Kevin Drew hu.MAP3.0: Atlas of human protein complexes by integration of > 25,000 proteomic experiments. Molecular Systems Biology 1–33 (2025) doi:10.1038/s44320-025-00121-5.
Huttlin et al. Dual proteome-scale networks reveal cell-specific remodeling of the human interactome Cell. 2021 May 27;184(11):3022-3040.e28. doi: 10.1016/j.cell.2021.04.011.
Huttlin et al. Architecture of the human interactome defines protein communities and disease networks. Nature. 2017 May 25;545(7655):505-509. DOI: 10.1038/nature22366.
Treiber et al. A Compendium of RNA-Binding Proteins that Regulate MicroRNA Biogenesis.. Mol Cell. 2017 Apr 20;66(2):270-284.e13. doi: 10.1016/j.molcel.2017.03.014.
Boldt et al. An organelle-specific protein landscape identifies novel diseases and molecular mechanisms. Nat Commun. 2016 May 13;7:11491. doi: 10.1038/ncomms11491.
Youn et al. High-Density Proximity Mapping Reveals the Subcellular Organization of mRNA-Associated Granules and Bodies. Mol Cell. 2018 Feb 1;69(3):517-532.e11. doi: 10.1016/j.molcel.2017.12.020.
Gupta et al. A Dynamic Protein Interaction Landscape of the Human Centrosome-Cilium Interface. Cell. 2015 Dec 3;163(6):1484-99. doi: 10.1016/j.cell.2015.10.065.
Wan, Borgeson et al. Panorama of ancient metazoan macromolecular complexes. Nature. 2015 Sep 17;525(7569):339-44. doi: 10.1038/nature14877. Epub 2015 Sep 7.
Hein et al. A human interactome in three quantitative dimensions organized by stoichiometries and abundances. Cell. 2015 Oct 22;163(3):712-23. doi: 10.1016/j.cell.2015.09.053. Epub 2015 Oct 22.
Huttlin et al. The BioPlex Network: A Systematic Exploration of the Human Interactome. Cell. 2015 Jul 16;162(2):425-40. doi: 10.1016/j.cell.2015.06.043.
Reimand et al. g:Profiler-a web server for functional interpretation of gene lists (2016 update). Nucleic Acids Res. 2016 Jul 8;44(W1):W83-9. doi: 10.1093/nar/gkw199.

Description of dataset files

|-- full

|-- humap3_full_feature_matrix_20220625.csv.gz

  This is the full feature matrix that contains all 26 million proteins pairs that DirectContacts2 makes 
  predictions on and the 324 features from >25,000 proteomic mass spectrometry experiments. The feature 
  matrix was originally developed in hu.MAP3.0.
  

|-- reference_interactions

|-- benchmark_test_INTRAandINTER_complex_NegativePairs_20250619.csv

  These are the test set negatives that contains both intra- and inter-complex negative pairs. The file
  contains a pair of proteins (defined by UniProt acession) per line. The set of structures (PDBs) these
  interactions come from all have at least 5 subunits.
  
|-- benchmark_test_IntraComplex_NegativePairs_pdbsize5_20240326.txt

  These are the test set negatives that contain only intra-complex negative pairs and are the negatives used
  in the test feature matrix provided when reading in the test feature matrix using the *datasets* package.
  The other feature matrix is provided under this same directory (*reference_interactions). All PDBs that
  these interactions come from have at least 5 protein subunits. The file is structured with one pair of 
  proteins (defined by UniProt acession) per line.
  
|-- benchmark_test_IntraComplex_PositivePairs_pdbsize5_20240326.txt

  These are the test set positives, which are only intra-complex pairs. The PDB structures from which the 
  interactions are derived all have at least 5 subunits. The file is structured with one pair of 
  proteins (defined by UniProt acession) per line.
  
|-- benchmark_train_INTERandINTRA_complex_NegativePairs_pdbsize3_20240326.txt

  These are the training set negatives, which contain both inter- and intra-complex pairs. The set of 
  structures (PDBs) these interactions come from all have at least 3 subunits. The file is structured with 
  one pair of proteins (defined by UniProt acession) per line.
  
|-- benchmark_train_IntraComplex_PositivePairs_pdbsize3_20240326.txt

  These are the test set positives, which are only intra-complex pairs. The PDB structures from which the 
  interactions are derived all have at least 3 subunits. The file is structured with one pair of 
  proteins (defined by UniProt acession) per line.
  

|-- test

|-- test_FeatureMatrix_pdbsize5_only_INTRA_complex_NegativePairs_20240326.csv.gz

  This is the feature matrix for the test set of interactions.
  

|-- train

|-- train_FeatureMatrix_pdbsize3_20240326.csv.gz

  This is the feature matrix for the training set of interactions.
  

Associated code

Additional code examples can be found on our GitHub, including: importing the DirectContacts2 model to make predictions, importing the training and testing data, or using the full feature matrix.

Usage

Accessing and using the model

DirectContacts2 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 DirectContacts2 model can be accessed through HuggingFace with huggingface_hub

 >>> from huggingface_hub import snapshot_download

>>> model_dir = snapshot_download(repo_id="sfisch/DirectContacts2_AutoGluon")

>>> predictor = TabularPredictor.load(f"{model_dir}/DirectContacts2_Autogluon_Model")

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('sfisch/DirectContacts2')

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

Accessing full feature matrix and all test/train interaction/complex 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="sfisch/DirectContacts2", filename='full/humap3_full_feature_matrix_20220625.csv.gz', 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_csv(full_file, compression="gzip")

Dataset card authors

Samantha Fischer (sfisch6@uic.edu)