HydRA: Deep-learning models for predicting RNA-binding capacity from protein interaction association context and protein sequence.
Please cite: Jin et al., Molecular Cell, 2023
SONAR3.0
HydRA-seq
HydRA
The pre-computed HydRA scores for most human proteins are available in the supplementary Table S2 of the paper above. Also attached here.
- Installation
- Prepare the input data
- RBP prediction
- Occlusion Map analysis
- Train a new HydRA model
- Train and evaluate a new HydRA model
- StrucGNNs
- Practical Notes
HydRA is tested to work under Python 3.8. Run the following code in your terminal window to install HydRA. It was recommended to build a new conda environment and do the HydRA installation to avoid messing up your current python environment. An installation tutorial for conda could be found here.
## Create a conda environment for HydRA, and activate the environment
conda create -n "HydRA" python=3.8.12 -y
conda activate HydRA
## Install the dependency packages for HydRA
pip3 install pandas numpy protobuf==3.20.3 networkx==2.6.3 scikit-learn==0.22.1 tensorflow==2.6.0 keras==2.6.0 matplotlib==3.4.3 setuptools Pygments packaging protein-bert
## Install HydRA
pip3 install hydra-rbp --no-deps
## Alternative command: python3 -m pip install --index-url https://test.pypi.org/simple/ --no-deps --upgrade hydra-rbp
Download ProteinBERT-RBP model (i.e. ProteinBERT_TrainWithWholeProteinSet_defaultSetting_ModelFile.pkl) from the assets of Releases v0.1.21.37 in this repository here. (If you are Yeo Lab TSCC user, skip this step.)
The installation of Tensorflow on Mac OS (m1/m2) chips are different from that in Linux system (a good tutorial can be found here)
## Create a conda environment for HydRA, and activate the environment
conda create -n "HydRA" python=3.9 -y
conda activate HydRA
## Install tensorflow on Mac m1/m2
conda install -c apple tensorflow-deps -y
## Install scikit-learn 0.22 on Mac m1/m2, not feasible it.
## Install the dependency packages for HydRA
pip3 install pandas numpy protobuf==3.20.3 networkx==2.6.3 scikit-learn matplotlib setuptools Pygments packaging protein-bert
python -m pip install tensorflow-macos==2.7.0
## Install HydRA
pip3 install hydra-rbp --no-deps
## Alternative command: python3 -m pip install --index-url https://test.pypi.org/simple/ --no-deps --upgrade hydra-rbp
Download ProteinBERT-RBP model (i.e. ProteinBERT_TrainWithWholeProteinSet_defaultSetting_ModelFile.pkl) from the assets of Releases v0.1.21.37 in this repository here. (If you are Yeo Lab TSCC user, skip this step.)
The sequence should be stored in FASTA format and one protein sequence per file. The suffix .fasta is required in the filenames. Create a folder under your working directory, and put all the sequences of the query proteins in this folder.
The PPI and functional association information should be stored as edge lists using the tab-separated values (.tsv) format. In the tab-separated edge list files, each protein-protein interaction (PPI) or functional association (PPA) is expressed as a row with the proteins shown in the example below. The PPI and PPA information are stored separately in two .tsv files (shown as PPA_edgelist.tsv and PPI_edgelist.tsv in the example below).
P06241 P40337
Q99728 P38398
P09012 O14920
Q16665 P40337
Q09472 P04637
P22681 Q02750
Q00987 P04637
Q13546 P42345
P62993 P22681
Q01094 Q6P2Q9
Q86UP2 Q14145
Q6R327 P42345
P45983 Q9Y3T9
An example of the input data directory is shown below.
.
├── PIA_edgelist
│ ├── PPA_edgelist.tsv
│ └── PPI_edgelist.tsv
└── sequences
├── P06241.fasta
├── P09012.fasta
├── Q02750.fasta
├── Q13546.fasta
├── Q6P2Q9.fasta
├── Q86UP2.fasta
└── Q9Y3T9.fasta
The annotation file needs to be formatted in CSV format as shown below, with each row indicating the coordinates, type, name of the annotated region, and the coloring scheme (optional) for the annotations.
Run the following command by replacing the parts in the upper case, such as the location of your folder that contains the fasta files of your query proteins /PATH/TO/FASTA_FILE_FOLDER, the location where you want to store the output files /PATH/TO/OUTPUT_FOLDER, the location of the ProteinBERT-RBP model file you downloaded previously /PATH/TO/PROTEINBERT-RBP-MODEL, and a customized name for this prediction MAKE_A_NAME_FOR_THE_PREDICTION (optional). For predictions with PPI/PPA data, the file path for the PPI and PPA data is also needed. For Yeo lab TSCC user, you could use the /tscc/nfs/home/wjin/projects/RBP_pred/RBP_identification/HydRa2.0/data/ProteinBERT/ProteinBERT_TrainWithWholeProteinSet_defaultSetting_ModelFile.pkl on TSCC for /PATH/TO/PROTEINBERT-RBP-MODEL instead.
(1) Predicting protein sequences only.
HydRa2_predict --seq_dir /PATH/TO/FASTA_FILE_FOLDER \
--proteinBERT_modelfile /PATH/TO/PROTEINBERT-RBP-MODEL \
--outdir /PATH/TO/OUTPUT_FOLDER \
-n MAKE_A_NAME_FOR_THE_PREDICTION \
--no-PIA --no-PPA
(2) Predicting with protein-protein interaction (PPI), protein-protein association(PPA) data and protein sequences.
HydRa2_predict --seq_dir /PATH/TO/FASTA_FILES \
--proteinBERT_modelfile /PATH/TO/PROTEINBERT-RBP-MODEL \
--outdir /PATH/TO/OUTPUT_FOLDER \
-n MAKE_A_NAME_FOR_THIS_PREDICTION \
--PPI_edgelist PATH/TO/PPI_EDGE_LIST \
--PPA_edgelist PATH/TO/PPA_EDGE_LIST
If --PPI_edgelist and --PPA_edgelist are not provided, the default human PPI (from Mentha (downloaded on Jan 8, 2018) and BioPlex2.0) and PPA (from STRING_v10.5, while experimentally identified records are excluded) data will be used.
The final prediction scores are in the *_HydRa_predictions.csv file in the output folder.
Run the following command by replacing the parts in the upper case, such as the location of your folder that contains the fasta files of your query proteins /PATH/TO/FASTA_FILE_FOLDER, the location where you want to store the output files /PATH/TO/OUTPUT_FOLDER, the location of the ProteinBERT-RBP model file you downloaded previously /PATH/TO/PROTEINBERT-RBP-MODEL, and a customized name for this prediction MAKE_A_NAME_FOR_THE_PREDICTION (optional). To plot the annotation track (optional), showing the distribution of specific types of features (e.g., protein domain, disordered region, etc., along the protein sequence), the file path for the protein region annotation file /PATH/TO/PROTEIN_REGION_ANNOATION_FILE (see more details here) is also needed. For Yeo lab TSCC user, you could also use the /tscc/nfs/home/wjin/projects/RBP_pred/RBP_identification/HydRa2.0/data/ProteinBERT/ProteinBERT_TrainWithWholeProteinSet_defaultSetting_ModelFile.pkl on TSCC for /PATH/TO/PROTEINBERT-RBP-MODEL instead.
occlusion_map3 -s /PATH/TO/FASTA_FILE_FOLDER \
--out_dir /PATH/TO/OUTPUT_FOLDER \
--proteinBERT_modelfile /PATH/TO/PROTEINBERT-RBP-MODEL \
-n MAKE_A_NAME_FOR_THE_PREDICTION \
--annotation_file /PATH/TO/PROTEIN_REGION_ANNOATION_FILE \
--draw_ensemble_only
In the output folder, you will find the occlusion map plots (named as *_OcclusionMap1D.png or *_OcclusionMap1D.pdf) like the one below:
The amino acid resolution occlusion scores are scored in the files named as *_Occlusion_score_matrix_full_aac_addZscoresProtLenWiseFib_pvalues.xls in the output folder.
Run the following command by replacing the parts in the upper case, such as the location of your folder that contains the fasta files of your query proteins /PATH/TO/FASTA_FILE_FOLDER, the location where you want to store the output files /PATH/TO/OUTPUT_FOLDER, the location of the ProteinBERT-RBP model file /PATH/TO/PROTEINBERT-ORIGINAL-MODEL that you can download from here (Yeo lab TSCC user can use the file /tscc/nfs/home/wjin/projects/RBP_pred/RBP_identification/HydRa2.0/data/ProteinBERT/default.pkl on TSCC for this argument directly), and a customized name for this prediction MAKE_A_NAME_FOR_THE_PREDICTION (optional). Users can also specify the proteins for training by --train_list /PATH/TO/GENE_IDs_FOR_TRAINING where /PATH/TO/GENE_IDs_FOR_TRAINING could be a text file with one gene ID/name per row. To train the HydRA model with PPI/PPA data, the file path for the PPI and PPA data is also needed. Use the --RBP_list option to define the positive samples in your training set.
HydRa_train2 --model_name Model_Name --model_outdir /PATH/TO/OUTPUT_FOLDER -s /PATH/TO/FASTA_FILE_FOLDER \
--train_list /PATH/TO/GENE_IDs_FOR_TRAINING --PPI_edgelist /PATH/TO/PPI_EDGE_LIST --PPA_edgelist /PATH/TO/PPA_EDGE_LIST \
--RBP_list /YOUR/RBP/LIST --ProteinBERT_pretrainedBeforeFinetune_model /PATH/TO/PROTEINBERT-ORIGINAL-MODEL \
--new-pretrain
Run the following command by replacing the parts in the upper case, such as the location of your folder that contains the fasta files of your query proteins /PATH/TO/FASTA_FILE_FOLDER, the location where you want to store the output files /PATH/TO/OUTPUT_FOLDER, the location of the ProteinBERT-RBP model file /PATH/TO/PROTEINBERT-ORIGINAL-MODEL that you can download from here (Yeo lab TSCC user can use the file /tscc/nfs/home/wjin/projects/RBP_pred/RBP_identification/HydRa2.0/data/ProteinBERT/default.pkl on TSCC for this argument directly), and a customized name for this prediction MAKE_A_NAME_FOR_THE_PREDICTION (optional). User can also specify the proteins for training and evaluation by --train_list and --test_list where /PATH/TO/GENE_IDs_FOR_TRAINING and /PATH/TO/GENE_IDs_FOR_TEST could be text files with one gene ID/name per row. To train HydRA model with PPI/PPA data, the file path for the PPI and PPA data is also needed. Use the --RBP_list option to define the positive samples in your training and test set.
HydRa2_train_eval --model_name Model_Name --model_outdir /PATH/TO/OUTPUT_FOLDER -s /PATH/TO/FASTA_FILE_FOLDER \
--train_list /PATH/TO/GENE_IDs_FOR_TRAINING --test_list /PATH/TO/GENE_IDs_FOR_TEST \
--PPI_edgelist /PATH/TO/PPI_EDGE_LIST --PPA_edgelist /PATH/TO/PPA_EDGE_LIST \
--RBP_list /YOUR/RBP/LIST --ProteinBERT_pretrainedBeforeFinetune_model /PATH/TO/PROTEINBERT-ORIGINAL-MODEL \
--new-pretrain
More details and new updates on StrucGNNs for RBP predictions are located at: https://github.com/Wenhao-Jin/strucGNN_RBP
HydRA has incorporated the predictions from HydRA-seq and SONAR3.0 via statistical integration and has higher overall sensitivity, specificity, and precision than HydRA-seq and SONAR3.0. But in practice, whether to combine/intersect the positive prediction results from HydRA, HydRA-seq, and SONAR3.0 or “HydRA only” depends on the task at hand. For instance, if our goal is to create a comprehensive and exhaustive list of RBPs, meaning that we only care about sensitivity, we can combine the positive predictions from all three classifiers. However, if we aim to create a highly reliable list of RBPs, prioritizing precision and specificity, we can take the intersection of the positive predictions from the three classifiers. Nevertheless, combining positive predictions will often decrease specificity and precision, whereas taking the intersection will negatively impact sensitivity. Thus, by default, we assume we need a list of predictions with high specificity, precision, and sensitivity simultaneously, which is achieved by using the HydRA-only strategy.