SGPPI: structure-aware prediction of protein-protein interactions in rigorous conditions with graph convolutional network SGPPI, a structure-based deep learning framework for predicting PPIs using graph convolutional networks (GCN). In particular, SGPPI focuses on protein patches on protein-protein binding interfaces and extracts structural, geometric and evolutionary features from the residue contact map to predict PPIs. We demonstrate that our model outperforms traditional machine learning methods and state-of-the-art deep learning-based methods using a non-representation-bias benchmark data set. Moreover, our model trained on human data can be reliably transferred to predict yeast PPIs, indicating that SGPPI can capture converging structural features of protein interactions across various species.
In SGPPI, we used GCN to cpature the hidden features of protein structures. The graph used here is the residue contact map with the threshold of 10Å. Features of the node including the pssm profiles, second structure and Jet2 features. To use SGPPI, users should prepare the adjacency matrix of the graph and the feature list of the residues.
We provided the positive and negative samples for three baseline datasets: Profkernelppi human dataset, HuRI dataset and filtered Pan’s dataset. You can find all the datasets used in SGPPI from three datasets folders and the form of the dataset is as follows:
| Protein A | Protein B |
|---|---|
| O15015 | P53582 |
| Q9NZC7 | Q8IY17 |
| Q9UK11 | O43795 |
Each data set contains two columns, which are the input two proteins. In the SGPPI model, we set the label of the positive sample to 1 and the label of the negative sample to 0.
torch (==1.5.0)
scipy (==1.5.2)
scikit-learn (==0.24.2)
dgl (0.7.2)
numpy (==1.19.1)
SGPPI regard the protein as collection of protein interface patches and integrated the global and local structural features of each residue in these patches. Besides, a comprehensive set of protein sequence and structural features are considered: a) evolutionary information of the residue through position-specific scoring matrices (PSSMs); b) location in the underlying protein secondary structure; c) global and local geometrical descriptors.
To use SGPPI, you should first calculate all the needed features of proteins. We have published the calculated features of both human and yeast proteins, you can find them at figshare. The features mainly include the following files: .atomAxs, .axs, .clusters, .cv , .cvlocal and .pssm.
| Features | Description |
|---|---|
.atomAxs |
accessibility at atomic level |
.axs |
accessibility at residue level |
.clusters |
potential protein interaction interface |
.cv |
global circular variances |
.cvlocal |
local circular variances |
.pssm |
position-specific scoring matrices |
Use feature_extract.py to generate the features of the corresponding protein. Please note: dssp.txt and all the feature files should be in the same file directory as "feature_extract.py".
python feature_extract.py –i protein_name –o protein_featuresSGPPI consider a contact if the geometrical distance of any two residues’ Cα atoms is less than a certain threshold (default 10 Å), allowing us to represent a protein structure by an undirected graph of the included surface/patch residues. Use adjmatrix_extract.py to generate the adjacent matrix of the corresponding protein structure. Please note: Human_RSA0.2.pkl or Yeast_RSA0.2.pkl should be in the same file directory as "adjmatrix_extract.py". Use the parameter -s to select the species
python adjmatrix_extract.py –i pdb_file –o adjacent_matrix -s humanBefore starting, users should prepare two files: sample_adj.pkl and sample_fea.pkl corresponding to the dictionary of sample adjacency matrix and sample features. Use SaveToDict.py generate these two files. Before running the script, you need to prepare a list of all the samples’ names, and then modify the sampleList and rootdir in the script. If your sample contains two proteins: P14859 and Q5SXM2, please modify the list in the script to sampleList = ['P14859','Q5SXM2'] and then run the script:
python SaveToDict.pyAfter running, two .pkl files (sample_adj.pkl and sample_fea.pkl) will be obtained, which will be used for the training and prediction of the model in the training step.
After all the above documents are ready, the training of the model can be started.
Use train_model.py to train the SGPPI model.
python train_model.py -e 20 -l 0.0005 -o model0In the source code, we provide the seed for users to repeat our results. The default dataset is the first cross validation data of HuRI dataset. If you want to modify the data set, please change to the target dataset path at the corresponding position of the source code.
At the same time, we also uploaded all the models used in the paper, which you can find in three models folders or at figshare. To use these models directly, please use pytorch load as follows:
torch.load("model.pt")We would like to thank the DGL team for the source code of GCN part.