A PyTorch implementation of NodeCoder Pipeline, a Graph Convolutional Network (GCN) - based framework for protein residue characterization. This work was presented at NeurIPS MLSB 2021: Residue characterization on AlphaFold2 protein structures using graph neural networks. link to paper

Link to GitHub repository.

Abstract:

Three-dimensional structure prediction tools offer a rapid means to approximate the topology of a protein structure for any
protein sequence. Recent progress in deep learning-based structure prediction has led to highly accurate predictions that have
recently been used to systematically predict 20 whole proteomes by DeepMind’s AlphaFold and the EMBL-EBI. While highly convenient,
structure prediction tools lack much of the functional context presented by experimental studies, such as binding sites or 
post-translational modifications. Here, we introduce a machine learning framework to rapidly model any residue-based
classification using AlphaFold2 structure-augmented protein representations. Specifically, graphs describing the 3D structure of
each protein in the AlphaFold2 human proteome are generated and used as input representations to a Graph Convolutional Network
(GCN), which annotates specific regions of interest based on the structural attributes of the amino acid residues, including their
local neighbors. We demonstrate the approach using six varied amino acid classification tasks.

Table of Contents

🧬 What does NodeCoder Pipeline do?
⚙️ Installing NodeCoder
🔌 NodeCoder Usage
🗄️ Graph data files
📂 Output files
🤝 Collaborators
🔐 License
📄 Citing this work

🧬 What does the NodeCoder Pipeline do?

---

The NodeCoder is a generalized framework that annotates 3D protein structures with predicted tasks such as binding sites. The NodeCoder model is based on Graph Convolutional Network. NodeCoder generates proteins' graphs from ALphaFold2 augmented proteins' structures where the nodes are the amino acid residues and edges are inter-residue contacts within a preset distance. The NodeCoder model is then trained with generated graph data for users task of interest like: Task = ['y_Ligand']. When running inference, NodeCoder takes the Protein ID like EGFR_HUMAN and for the proteins that are already in the database, input graph data files are created from the AlphaFold2 protein structure in addition to calculating some structure-based and sequence-based residue features. The input graph data will then be given to the trained model for prediction of multiple tasks of interest such as binding sites or post-translational modifications.

⚙️ Installing NodeCoder

---

Required dependencies

The codebase is implemented in Python 3.8.5 and package versions used for development are:

numpy              1.19.2
pandas             1.2.4
scipy              1.6.3
torch              0.4.1
torchvision        0.9.1
torchaudio         0.8.1
torch_scatter      2.0.6
torch_sparse       0.6.9
torch_cluster      1.5.9
torch_spline_conv  1.2.0
torch-geometric    1.7.0  
scikit-learn       0.24.2
matplotlib         3.3.3
biopython          1.77
freesasa           2.0.5.post2
loguru             0.6.0

Installation steps

Here is the step-by-step NodeCoder installation process:
Method 1 - install test.pypi package

1. Before installing NodeCoder, we highly recommend to create a virutal Python 3.8.5 environment using venv command, or Anaconda. Assuming you have anaconda3 installed on your computer, on your Terminal run the following command line:

$ conda create -n <your_python_env> python=3.8.5

2. Make sure your virtual environment is active. For conda environment you can use this command line:

$ conda activate <your_python_env>

3. Now you can install NodeCoder with this command line:

$ pip install --extra-index-url https://test.pypi.org/simple/ NodeCoder

Method 2 - install from GitHub repository
Follow above-mentioned steps 1-2, and continue with the following steps: 3. Clone the repository:

$ git clone https://github.com/NasimAbdollahi/NodeCoder.git

4. Make sure you are in the root directory of the NodeCoder package ~/NodeCoder/ (where setup.py is). Now install NodeCoder package with following command line, which will install all dependencies in the python environment you created in step 1:

$ pip install .

🔌 NodeCoder Usage

---

NodeCoder package can be employed for train and inference. Here we describe how to use it:

🗂️ Preprocessing raw data

link to paper NodeCoder uses AlphaFold2 modeled protein structures as input. AlphaFold protein structure database provides open access to protein structure predictions of human proteome and other key proteins of interest. Prediction labels can be obtained from BioLip database and Uniprot database.

Once you downloaded the protein databases, first step is to run the NodeCoder's featurizer module to process these raw data sets and extract node features and labels. When using NodeCoder's featurizer module, preprocess_raw_data, you will need to specify the directories you have the datasets saved:

alphafold_data_path
uniprot_data_path
biolip_data_path
biolip_data_skip_path

The featurizer module will create two files for every protein in the selected proteome: *.features.csv and *.tasks.csv . These files are saved in NodeCoder/data/input_data/featurized_data/TAX_ID/ directory in separate folders of features and tasks . For example if user choose human proteome, 9606, then the following tree structure will be generated:

data/input_data/featurized_data/
└── 9606
    ├── features
    └── tasks

The command line to run the featurizer module is:

$ python NodeCoder/preprocess_raw_data.py

To use NodeCoder as python package, import preprocessing module as:

>>> from NodeCoder import preprocess_raw_data 
>>> preprocess_raw_data.main(alphafold_data_path='.', uniprot_data_path='.', biolip_data_path='.', biolip_data_skip_path='.')

The default species/proteome is HUMAN, but user can change it with the following parameters:

>>> preprocess_raw_data.main(alphafold_data_path='.', uniprot_data_path='.', biolip_data_path='.', biolip_data_skip_path='.',
                              TAX_ID='9606', PROTEOME_ID='UP000005640')

🗃️ Generate graph data

The next step after running the featurizer is to generate graph data from the features and tasks files. NodeCoder has a graph-generator module that generate protein graph data by taking a threshold for distance between amino acid residues. The threshold distance is required to be defined by user in Angstrom unit to create the graph contact network, threshold_dist = 5. Graph data files are saved in this directory ./data/input_data/graph_data_*A/ with the following tree structure (the example here is for 8A cut-off distance and 5 folds for cross-validation):

data/input_data/graph_data_8A/
└── 9606
    └── 5FoldCV

The command line to run the graph generator module is:

$ python NodeCoder/generate_graph_data.py

To use NodeCoder as python package, import generate_graph_data module as:

>>> from NodeCoder import generate_graph_data 
>>> generate_graph_data.main()

Where, user can specify the following parameters

>>> generate_graph_data.main(TAX_ID='9606', PROTEOME_ID='UP000005640', threshold_dist=5, cross_validation_fold_number=5)

Note that for cross-validation setting, separate graphs are created for each fold.

🧠 Train NodeCoder

To train NodeCoder's graph-based model, user can use train.py module. Script parser.py has the model parameters used for training the model. User would need to use the following parameters in train.py script to specify the task/tasks of interest and the cutoff distance for defining the protein contact network:

Task = ['y_Ligand']
threshold_dist = 5

Command line to train NodeCoder:

$ python NodeCoder/train.py

To use NodeCoder as python package, import train module as:

>>> from NodeCoder import train 
>>> train.main()

Where, user can specify the following parameters

>>> train.main(threshold_dist=5, multi_task_learning=False, Task=['y_Ligand'], centrality_feature=True,
         cross_validation_fold_number=5, epochs=1000)

Here is a list of available training tasks (residue labels/annotations) :

'y_CHAIN', 'y_TRANSMEM', 'y_MOD_RES', 'y_ACT_SITE', 'y_NP_BIND', 
'y_LIPID', 'y_CARBOHYD', 'y_DISULFID', 'y_VARIANT', 'y_Artifact', 
'y_Peptide', 'y_Nucleic', 'y_Inorganic', 'y_Cofactor', 'y_Ligand'

🤖 Inference with NodeCoder

To use trained NodeCoder for protein functions prediction, user needs to run predict.py script. User would need to use the following parameters in predict.py script to specify the protein of interest, functions of interest and the cutoff distance for defining the protein contact network:

Task = ['y_Ligand']
threshold_dist = 5

Command line to run inference with NodeCoder:

$ python NodeCoder/predict.py

To use NodeCoder as python package, import train module as:

>>> from NodeCoder import predict 
>>> predict.main()

Where, user can specify the following parameters

>>> predict.main(protein_ID='KI3L1_HUMAN', threshold_dist=5, trained_model_fold_number=1, multi_task_learning=False,
         Task=['y_Ligand'], centrality_feature=True, cross_validation_fold_number=5, epochs=1000)

The user shall make sure the model with the desired parameters should have been trained already, otherwise the user would need to first train the model then use this pipeline for prediction.

🗄️ Graph data files

---

When graph data is generated from featurized data, files are saved in this directory ./data/input_data/graph_data_*A/ . Specific sub-directories are created depends on user choice of cutoff distance for protein contact network, proteom, and number of cross-validation folds. This helps user to keep track of different test cases. For every fold, following data files are created for train and validation sets:

data/input_data/graph_data_5A/9606/
└── *FoldCV
    ├── train_1_ProteinFileNames.csv
    ├── train_1_ProteinFiles.csv
    ├── train_1_edge_features.csv
    ├── train_1_edges.csv
    ├── train_1_features.csv
    ├── train_1_nodes_ProteinID.csv
    ├── train_1_target.csv
    ├── ....
    ├── validation_1_ProteinFileNames.csv
    ├── validation_1_ProteinFiles.csv
    ├── validation_1_edge_features.csv
    ├── validation_1_edges.csv
    ├── validation_1_features.csv
    ├── validation_1_nodes_ProteinID.csv
    ├── validation_1_target.csv
    ├── ...

Generated graph data files includes:

*_features.csv

Nodes' feature vectors are concatenated to create a long list of features for all nodes in the graph. The first row is a header:

node_id	feature_id	value
node number in the graph	one int. value of 0-45	...

*_edges.csv

It contains the list node-pairs for all edges in the graph:

id1	id2
index of node 1	index of node 2

*_edge_features.csv

It contains the edge features for all edges in the graph. Currently, three different features are calculated for the edges; however, the current model only uses the squared reciprocal of edge length as weight.

id1	id2	edge_length	edge_cosine_angle	edge_sequence_distance
index of node 1	index of node 2	Euclidean distance between the node pair	Cosine distance between the node pair	Sequence distance between the node pair

*_target.csv

It the target labels for all nodes in the graph. The number of columns depends on the number of targets that user specify for prediction.

y_task1	y_task2	y_task3	y_task4	y_task5	y_task6
0/1	0/1	0/1	0/1	0/1	0/1

*_ProteinFiles.csv

Protein File	Node Num	Removed NaNs
EGFR_HUMAN	1207	0
PTC1_HUMAN	1444	0
DDX10_HUMAN	872	0
...	...	...
RBL1_HUMAN	1065	0

*_nodes_ProteinID.csv

If more than one protein is given to the model, this file keeps track of the residues that belong to each protein.

node_id	protein_id_flag	protein_id
0	0	EGFR_HUMAN
1	0	EGFR_HUMAN
2	0	EGFR_HUMAN
3	0	EGFR_HUMAN
4	0	EGFR_HUMAN
...	...	...
1207	1	E2F8_HUMAN
1208	1	E2F8_HUMAN
1209	1	E2F8_HUMAN
...	...	...

Amino Acid Residue (AA) feature vector:

feature ID	feature name	feature Description
0-19	feat_A, feat_B, ..., feat_Y	primary structure - one-hot-encoding binary vector of the 20 natural amino acid names
20-23	feat_PHI, feat_PSI, feat_TAU, feat_THETA	Dihedral angles, φ, ψ, τ, θ
24	feat_BBSASA	Back Bone Solvent Accessibility
25	feat_SCSASA	Side Chain Solvent Accessibility
26	feat_pLDDT	AlphaFold per-residue confidence metric
27-41	feat_DSSP_*	Secondary structure features, e.g. α-helix and β-sheet
42	feat_CentricDist	Euclidean distance of AA residue from the center of protein
43	feat_CentricCosineDist	Cosine distance of AA residue from the center of protein
44	feat_iPlus	AA info of a node before node i in protein sequence
45	feat_iMinus	AA info of a node after node i in protein sequence

📂 Output files

---

All output files are saved in this directory ./results/graph_*A/ . Specific sub-directories are created according to model parameters, so that user can keep track of different test cases.

When training NodeCoder

In a cross-validation setting, the performance scores are saved in a .csv file like Model_Performance_Metrics_Fold1.csv , for all folds. In addition to this, model state is also saved in CheckPoints sub-directory at certain epochs. The default checkpoint epoch is 50, as well as the epoch to save model performance, but user can change these with checkpoint_step and performance_step. At the end of training on each fold, the inference is performed by finding the optimum epoch and loading corresponding trained model at the optimum epoch. At the end of inference, an output file is saved in Prediction sub-directory that includes the predicted labels for all proteins in validation set. This can be useful for ranking proteins.

Once training NodeCoder is completed, the results are all saved in results folder with the following tree structure (the example here is for 8A cut-off distance, 5 folds for cross-validation and Ligand as prediction task):

results
└── graph_8A
    └── 9606
        └── 5FoldCV
            └── y_Ligand_HiddenLayers_38_28_18_8_50Epochs_LR0.01
                ├── CheckPoints
                ├── Model_Performance_Curves.jpg
                ├── Model_Performance_Metrics_Fold1.csv
                └── Prediction

Model parameters such as network structure, number of epochs and learning rate (LR) are reflected in the created subdirectory.

When predicting with NodeCoder

When running inference with trained NodeCoder, the prediction results are saved in a sub-directory with protein name. The prediction result is a csv file like KI3L1_HUMAN_prediction_3Tasks_results.csv , which is a dataframe that contains the target labels, predicted labels and prediction probability of the labels per node (AA residue) for all tasks of interest, {y1, y2, ..., yn}.

node_id	protein_id_flag	protein_id	Task 1 Target	Task 1 Prediction	Task 1 PredictionProb	...	Task n Target	Task n Prediction	Task n PredictionProb
0	0	E2F8_HUMAN	0/1	0/1	float		0/1	0/1	float
1	0	E2F8_HUMAN	0/1	0/1	float		0/1	0/1	float

Once inference is completed, the results are all saved in results folder, with protein ID as the name of the subdirectory. The following tree structure is an example for KI3L1_HUMAN as the protein of interest, 8A cut-off distance and 5 folds for cross-validation):

results
└── graph_8A
    └── 9606
        └── 5FoldCV
            ├── KI3L1_HUMAN
            │         ├── KI3L1_HUMAN_ProteinFiles.csv
            │         ├── KI3L1_HUMAN_edge_features.csv
            │         ├── KI3L1_HUMAN_edges.csv
            │         ├── KI3L1_HUMAN_features.csv
            │         ├── KI3L1_HUMAN_nodes_ProteinID.csv
            │         ├── KI3L1_HUMAN_prediction_1Tasks_results.csv
            │         └── KI3L1_HUMAN_target.csv
            └── ...

🤝 Collaborators

---

This project has been sponsored by Mitacs and Cyclica Inc..
The main contributors are:
Nasim Abdollahi, Ph.D., Post-doctoral Fellow at University of Toronto, Cyclica Inc.
Ali Madani, Ph.D., Machine Learning Director at Cyclica Inc.
Bo Wang, Ph.D., Canada CIFAR AI Chair at the Vector Institute, Professor at University of Toronto
Stephen MacKinnon, Ph.D., Chief Platform Officer at Cyclica Inc.

🔐 License

MIT Licence

📄 Citing this work

@article {2021,
	author = {Abdollahi, Nasim and Madani, Ali and Wang, Bo and MacKinnon,
	Stephen},
	title = {Residue characterization on AlphaFold2 protein structures using graph neural networks},
	year = {2021},
	doi = {},
	publisher = {NeurIPS},
	URL = {https://www.mlsb.io/papers_2021/MLSB2021_Residue_characterization_on_AlphaFold2.pdf},
	journal = {NeurIPS, MLSB}
}