Skip to content
Using Deep Learning techniques to enhance orthology calls
Python R
Branch: master
Clone or download
mateuspatricio Measuring accuracy.
Here we have the plots for the model accuracy.
Latest commit 07f9f23 Sep 6, 2019
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
plots Measuring accuracy. Sep 6, 2019
.gitignore git should ignore __pycache__ Aug 23, 2019
.travis.yml Run both flake8 and pylint Aug 23, 2019
README.md README changes. Aug 29, 2019
access_data_rest.py Add files via upload Aug 23, 2019
create_genome_maps.py Add files via upload Jul 24, 2019
dist_matrix Add files via upload Jul 24, 2019
finalize_dataset.py
ftpg.py Add files via upload Aug 14, 2019
get_data.py Add files via upload Jul 24, 2019
model.py Add files via upload Aug 12, 2019
neighbor_genes.py Add files via upload Jul 24, 2019
pfam_folder_pred.py Add files via upload Aug 12, 2019
pfam_matrix.py Add files via upload Aug 30, 2019
pfam_parser.py Add files via upload Aug 29, 2019
prediction.py Add files via upload Jul 24, 2019
prediction_pfam.py
prepare_synteny_matrix.py
process_data.py Add files via upload Jul 24, 2019
process_negative.py Add files via upload Aug 12, 2019
read_data.py Add files via upload Jul 24, 2019
read_get_gene_seq.py
requirements.txt Create requirements.txt Jul 24, 2019
save_data.py Add files via upload Jul 24, 2019
select_data.py Update select_data.py Aug 14, 2019
selector_sp.py Rename selector.py to selector_sp.py Aug 14, 2019
sp_names Add files via upload Jul 24, 2019
species_tree.tree
threads.py
train.py Add files via upload Aug 12, 2019
tree_data.py Add files via upload Jul 24, 2019

README.md

The aim of this project is to use Deep Neural-Nets to predict homology type between a given pair of genes.

Requirements:

  1. A machine with atleast 8GB of RAM (although 16-32GB is recommended. It depends on the no. of homology databases that you are willing to use in the preparation of the dataset), a graphic card for training the deep neural nets. A single GPU machine would suffice. The model can be trained on CPU as well but will be a lot faster if trained on a GPU.
  2. A stable Internet Connection.
  3. A native/virtual python environment. Install the dependencies from requirements.txt using :
    pip install -r requirements.txt

Data Preparation:

The model uses a synteny matrix and some other factors derived from the species tree to make predictions.
Download The Required Files:

In order to download all the needed files run: python ftpg.py

It will download the gtf, cds and pep files.

This scripts writes the links of all the required gtf and cds files to gtf-link.txt and seq_link.txt. You can use your own script to download the files or just enter y when prompted for permission to download the files. It will automatically download all the files and store them in designated folder. You can manually download each files by pasting link from the files in the browser.

So as to prepare data we will need the following files:
1.All the gtf files to get the start and end locations of the genes and find their neighboring genes. This is used to create the synteny matrix which helps to see the conserved synteny among the genes.
2. All the cds files in FAST-A format. They are required but are not mandatory, if the files are not provided the sequences are directly accessed from the REST API but the process can be slow :( . It's better to have all the cds files.
3. All the pep files in FAST-A format. They are required to get the protein sequences to run the pfam scan on.

Homology Databases:
Additionally, you will need to download the homologies of your choice. They can be found at: ftp://ftp.ensembl.org/pub/current_tsv/ensembl-compara/homologies/

e.g.: ftp://ftp.ensembl.org/pub/current_tsv/ensembl-compara/homologies/homo_sapiens/Compara.97.protein_default.homologies.tsv.gz

All the databases have the same name so you have to rename the files with their respective speicies names. From Compara.97.protein_default.homologies.tsv.gz to species_name.tsv.gz

Homology Databases:
Additionally, you will need to download the homologies of your choice. They can be found at: ftp://ftp.ensembl.org/pub/current_tsv/ensembl-compara/homologies/

e.g.: ftp://ftp.ensembl.org/pub/current_tsv/ensembl-compara/homologies/homo_sapiens/Compara.97.protein_default.homologies.tsv.gz

All the databases have the same name so you have to rename the files with their respective speicies names. From Compara.97.protein_default.homologies.tsv.gz to species_name.tsv.gz All homology files must go into a directory called: data_homology/

Create Genome Maps:
The purpose is to create maps of all the genes present in the gtf files with respect to their chromosomes, a map of all the genes belonging to the same chromosome in the given species, a map of all the genes in the given species, a map of all the species whose data has been successfully read.
To create genome maps run this command:
python create_genome_maps.py:
Note: A precomputed Genome Maps can be downloaded from this link.

Select the Records from each homology database:
This step will select the specified no. of records from each of the homology databases on the basis of distant species,GOC score,homology type etc.
To select the data run:
python select_data.py number_of_records_to_be_selected_from_each_file.

Find the Neighbor Genes of the Selected Records:
This step will find the neighbor genes of all the selected records from the homology databases and write it to the processed directory.
Run:
python neighbor_genes.py

Create Synteny Matrix and Write Protein Sequences:
This step will create the synteny matrices and write the protein sequences on a FAST-A file named protein_seq_positive.fa.You will have to run the hmmer-scan on this file to get the PFAm domains.
To create the synteny matrices:
python prepare_synteny_matrix.py number_of_threads.
Note:The script uses multiprocessing to prepare synteny matrices. Since each thread has it own copy of all the resources its better to run with higher number of threads on a high-RAM machine.

Process The Negative Dataset:
Negative samples are a non-homologous pair of genes. You can get the negative samples from here.Download one of your choice :).
Process the negative set by using:
python process_negative.py negative_database_file_name number_of_threads<br>

Run the HMMER scan on the Protein Sequences:
The idea is that homologous genes will have overlapping domains. Run the hmmer scan on the protein_seq_positve.fa and pro_seq_negative.fa with the -domtblout option.

Parse the PFAm domain files:
This file will parse the hmmer scan database. You will have to parse both the positive samples and the negative sample database.
To parse, run:
python pfam_parser.py domtblout_file_name_positive domtblout_file_name_negative

Create PFAM matrices:
This step will create the pfam matrices. This might take some time...
Run:
python pfam_matrix.py negative_samples_50K.txt

Finalize the Dataset:
This step combines everything and finalizes the dataset by reading the processed factors and extracting some basic features from the species tree. Run:
python finalize_dataset.py name_of_negative_database_you_earlier_processed

IF YOU DID EVERYTHING RIGHT YOU SHOULD SEE A FILE NAMED dataset IN THE SAME DIRECTORY.

Train the Model:

This is where it gets interesting. You are gonna train your own model architecture or you can use the one given in the model.py script. You can directly change the model as well in the model.py.
To train the model, run:
python train.py model_name negative_start_composition negative_end_composition epochs learning_rate learning_rate_decay no_of_samples batch_size

Predictions:

To make predictions you need to have the prediction files in a pre-defined format like this file. All the fields have to tab seperated and in the same order.

Get the prediction files:
Create a new directory with any name of your choice in the code directory and paste all the files on which you want to make the predictions inside it.
Run:
python pfam_folder_pred.py directory_name
This will read all the files on the directory and write all the protein sequences on a FAST-A file named prediction_directory_name.fa. Run hmmer scan on that file.

Parse the PFAM file
This file parses the PFAM file and creates some maps.
Run:
python pfam_db_parser.py domtblout_file_name

Now you have all the resources required to Make predictions on the required file.
Copy the file you want to make predictions on from the directory that you created earlier to the code directory and Run:
python prediction_pfam.py file_name_with_extension model_name no_of_models number_of_threads start end name domtblout_file_name weights.
where:
start end:the locations from which you want to make predictions in the file.
name:Since you can run multiple predcitions at the same time this serves as a unique identifier for the temporary files being created.
domtblout_file_name:It is the name of the file that you get after running the hmmer scan on the fast-a files.
weights:e if you want to give equal weight to all the models, w followed by the same number of float literals as number of models parameter to assign specific weights.

You can’t perform that action at this time.