DeepFRI
Deep functional residue identification 
Citing
@article {Gligorijevic2019,
author = {Gligorijevic, Vladimir and Renfrew, P. Douglas and Kosciolek, Tomasz and Leman,
Julia Koehler and Cho, Kyunghyun and Vatanen, Tommi and Berenberg, Daniel
and Taylor, Bryn and Fisk, Ian M. and Xavier, Ramnik J. and Knight, Rob and Bonneau, Richard},
title = {Structure-Based Function Prediction using Graph Convolutional Networks},
year = {2019},
doi = {10.1101/786236},
publisher = {Cold Spring Harbor Laboratory},
URL = {https://www.biorxiv.org/content/early/2019/10/04/786236},
journal = {bioRxiv}
}
Dependencies
DeepFRI is tested to work under Python 3.7.
The required dependencies for DeepFRI are TensorFlow, Biopython and scikit-learn. To install all dependencies run:
pip install .
Protein function prediction
To predict protein functions use predict.py script with the following options:
seqstr, Protein sequence as a stringcmapstr, Name of a file storing a protein contact map and sequence in*.npzfile format (with the following numpy array variables:C_alpha,seqres. Seeexamples/pdb_cmaps/)pdbstr, Name of a PDB file (cleaned)pdb_dirstr, Directory with cleaned PDB files (seeexamples/pdb_files/)cmap_csvstr, Filename of the catalogue (in*.csvfile format) containg mapping between protein names and directory with*.npzfiles (seeexamples/catalogue_pdb_chains.csv)fasta_fnstr, Fasta filename (seeexamples/pdb_chains.fasta)model_configstr, JSON file with model filenames (seetrained_models/)ontstr, Ontology (mf- Molecular Function,bp- Biological Process,cc- Cellular Component,ec- Enzyme Commission)output_fn_prefixstr, Output filename (sampe prefix for predictions/saliency will be used)verbosebool, Whether or not to print function prediction resultssaliencybool, Whether or not to compute class activaton maps (outputs a*.jsonfile)
Generated files (see examples/outputs/):
output_fn_prefix_MF_predictions.csvPredictions in the*.csvfile format with columns: Protein, GO-term/EC-number, Score, GO-term/EC-number nameoutput_fn_prefix_MF_pred_scores.jsonPredictions in the*.jsonfile with keys:pdb_chains,Y_hat,goterms,gonamesoutput_fn_prefix_MF_saliency_maps.jsonJSON file storing a dictionary of saliency maps for each predicted function of every protein
DeepFRI offers 6 possible options for predicting functions. See examples below.
Option 1: predicting functions of a protein from its contact map
Example: predicting MF-GO terms for Parvalbumin alpha protein using its sequence and contact map (PDB: 1S3P):
>> python predict.py --cmap ./examples/pdb_cmaps/1S3P-A.npz -ont mf --verbose
Output:
Protein GO-term/EC-number Score GO-term/EC-number name
query_prot GO:0005509 0.99824 calcium ion binding
Option 2: predicting functions of a protein from its sequence
Example: predicting MF-GO terms for Parvalbumin alpha protein using its sequence (PDB: 1S3P):
>> python predict.py --seq 'SMTDLLSAEDIKKAIGAFTAADSFDHKKFFQMVGLKKKSADDVKKVFHILDKDKDGFIDEDELGSILKGFSSDARDLSAKETKTLMAAGDKDGDGKIGVEEFSTLVAES' -ont mf --verbose
Output:
Protein GO-term/EC-number Score GO-term/EC-number name
query_prot GO:0005509 0.99769 calcium ion binding
Option 3: predicting functions of proteins from a fasta file
>> python predict.py --fasta_fn examples/pdb_chains.fasta -ont mf -v
Output:
Protein GO-term/EC-number Score GO-term/EC-number name
1S3P-A GO:0005509 0.99769 calcium ion binding
2J9H-A GO:0004364 0.46937 glutathione transferase activity
2J9H-A GO:0016765 0.19910 transferase activity, transferring alkyl or aryl
(other than methyl) groups
2J9H-A GO:0097367 0.10537 carbohydrate derivative binding
2PE5-B GO:0003677 0.53502 DNA binding
2W83-E GO:0032550 0.99260 purine ribonucleoside binding
2W83-E GO:0001883 0.99242 purine nucleoside binding
2W83-E GO:0005525 0.99231 GTP binding
2W83-E GO:0019001 0.99222 guanyl nucleotide binding
2W83-E GO:0032561 0.99194 guanyl ribonucleotide binding
2W83-E GO:0032549 0.99149 ribonucleoside binding
2W83-E GO:0001882 0.99135 nucleoside binding
2W83-E GO:0017076 0.98687 purine nucleotide binding
2W83-E GO:0032555 0.98641 purine ribonucleotide binding
2W83-E GO:0035639 0.98611 purine ribonucleoside triphosphate binding
2W83-E GO:0032553 0.98573 ribonucleotide binding
2W83-E GO:0097367 0.98168 carbohydrate derivative binding
2W83-E GO:0003924 0.52355 GTPase activity
2W83-E GO:0016817 0.36863 hydrolase activity, acting on acid anhydrides
2W83-E GO:0016818 0.36683 hydrolase activity, acting on acid anhydrides, in phosphorus-containing anhydrides
2W83-E GO:0017111 0.35465 nucleoside-triphosphatase activity
2W83-E GO:0016462 0.35303 pyrophosphatase activity
Option 4: predicting functions of proteins from contact map catalogue
>> python predict.py --cmap_csv examples/catalogue_pdb_chains.csv -ont mf -v
Output:
Protein GO-term/EC-number Score GO-term/EC-number name
1S3P-A GO:0005509 0.99824 calcium ion binding
2J9H-A GO:0004364 0.84826 glutathione transferase activity
2J9H-A GO:0016765 0.82014 transferase activity, transferring alkyl or aryl
(other than methyl) groups
2PE5-B GO:0003677 0.89086 DNA binding
2PE5-B GO:0017111 0.12892 nucleoside-triphosphatase activity
2PE5-B GO:0004386 0.12847 helicase activity
2PE5-B GO:0032553 0.12091 ribonucleotide binding
2PE5-B GO:0097367 0.11961 carbohydrate derivative binding
2PE5-B GO:0016887 0.11331 ATPase activity
2W83-E GO:0097367 0.97069 carbohydrate derivative binding
2W83-E GO:0019001 0.96842 guanyl nucleotide binding
2W83-E GO:0017076 0.96737 purine nucleotide binding
2W83-E GO:0001882 0.96473 nucleoside binding
2W83-E GO:0035639 0.96439 purine ribonucleoside triphosphate binding
2W83-E GO:0032555 0.96294 purine ribonucleotide binding
2W83-E GO:0016818 0.96181 hydrolase activity, acting on acid anhydrides, in phosphorus-containing anhydrides
2W83-E GO:0032550 0.96142 purine ribonucleoside binding
2W83-E GO:0016817 0.96082 hydrolase activity, acting on acid anhydrides
2W83-E GO:0016462 0.95998 pyrophosphatase activity
2W83-E GO:0032553 0.95935 ribonucleotide binding
2W83-E GO:0032561 0.95930 guanyl ribonucleotide binding
2W83-E GO:0032549 0.95877 ribonucleoside binding
2W83-E GO:0003924 0.95453 GTPase activity
2W83-E GO:0001883 0.95271 purine nucleoside binding
2W83-E GO:0005525 0.94635 GTP binding
2W83-E GO:0017111 0.93942 nucleoside-triphosphatase activity
2W83-E GO:0044877 0.64519 protein-containing complex binding
2W83-E GO:0001664 0.31413 G protein-coupled receptor binding
2W83-E GO:0005102 0.20078 signaling receptor binding
Option 5: predicting functions of a protein from a PDB file
>> python predict.py -pdb ./examples/pdb_files/1S3P-A.pdb -ont mf -v
Output:
Protein GO-term/EC-number Score GO-term/EC-number name
query_prot GO:0005509 0.99824 calcium ion binding
Option 6: predicting functions of a protein from a directory with PDB files
>> python predict.py --pdb_dir ./examples/pdb_files -ont mf --saliency --use_backprop
Output:
See files in: examples/outputs/
Training DeepFRI
To train DeepFRI run the following command from the project directory:
>> python train_DeepFRI.py -h
or to launch jobs run the following script:
>> ./run_train_DeepFRI.sh
Output
Generated files:
model_name_prefix_ont_model.hdf5trained model with architecture and weights saved in HDF5 formatmodel_name_prefix_ont_pred_scores.pcklpickle file with predicted GO term/EC number scores for test proteinsmodel_name_prefix_ont_model_params.jsonJSON file with metadata (GO terms/names, architecture params, etc.)
See examples of pre-trained models (*.hdf5) and model params (*.json) in: trained_models/.
Functional residue identification
To visualize class activation (saliency) maps use viz_gradCAM.py script with the following options:
saliency_fnstr, JSON filename with saliency maps generated bypredict.pyscript (see Option 6 above)list_allbool, list all proteins and their predicted GO terms with corresponding class activation (saliency) mapsprotein_idstr, protein (PDB chain), saliency maps of which are to be visualized for each predicted functiongo_idstr, GO term, saliency maps of which are to be visualizedgo_namestr, GO name, saliency maps of which are to be visualized
Generated files:
saliency_fig_PDB-chain_GOterm.pngclass activation (saliency) map profile over sequence (see fig below, right)pymol_viz.pypymol script for mapping salient residues onto 3D structure (pymol output is shown in fig below, left)
Example:
>>> python viz_gradCAM.py -i ./examples/outputs/DeepFRI_MF_saliency_maps.json -p 1S3P-A -go GO:0005509
Output:
Data
Data (train and validation) used for training DeepFRI model are provided as TensorFlow-specific TFRecord files and they can be downloaded from:
| PDB | SWISS-MODEL |
|---|---|
| Gene Ontology(19GB) | Gene Ontology(165GB) |
| Enzyme Commission(13GB) | Enzyme Commission(117GB) |
Pretrained models
Pretrained models can be downloaded from:
- Models (use these models if you run DeepFRI on GPU)
- Newest Models (use these models if you run DeepFRI on CPU)
Uncompress tar.gz file into the DeepFRI directory (tar xvzf trained_models.tar.gz -C /path/to/DeepFRI).
18 Aug 19, 2022
14 Nov 10, 2022
153 Dec 14, 2022
114 Dec 20, 2022
4 Nov 28, 2022
27 Dec 13, 2022
53 Dec 04, 2022
8 Aug 17, 2022
180 Jan 05, 2023
1 Nov 09, 2021
32 Dec 25, 2022
50 Oct 28, 2022
21 Aug 31, 2022
3 Jun 17, 2022
2 Nov 29, 2022
212 Dec 25, 2022
69 Dec 12, 2022
3 Aug 19, 2022
318 Dec 14, 2022
160 Jan 04, 2023