TerminatorJ/DeepLocRNA

This is the counterpart of DeepLoc, but in RNA level

DeepLocRNA

Introduction

DeeplocRNA is a deep neural network that predicts the RNA localization, enabling the prediction across 4 RNA types (mRNA, miRNA, lncRNA, snoRNA) and different species (Human and Mouse).

Environment preperation

Please make sure anaconda is installed in your local machine, create a new working environment to run DeepLocRNA

conda create -n DeepLocRNA python=3.8

Enter the new created environment

source activate DeepLocRNA

To run the model, you should download DeepLocRNA via git or pypi

From git

pip install git+https://github.com/TerminatorJ/DeepLocRNA.git

or from pypi

pip install DeepLocRNA

Install two dependenies

pip install tensorflow==2.4.1

pip install typing-extensions==4.7.1

Train the model

if you want to train the model youself, please follow the following steps

Step 1: Preparing your FASTA file

Input format

DeepLocRNA works on FASTA files, e.g.

>test1
ACTGCCGTATCGTAGCTAGCTAGTGATCGTAGCTACGTAGCTAGCTAGCTACGATCGTAGTCAGTCGTAGTACGTCA
>test2
ACACACATGAGCGATGTAGTCGATGATGCATCGACGATCGATCGAGCTACGTAGCATCGATCGATGCATCGACGTAG

One can aldo use our prepared dataset to train the model as below

Step 2: Download this repository to your local machine

wget https://github.com/TerminatorJ/DeepLocRNA/archive/refs/heads/main.zip

Then compress the zip file

unzip main.zip

Step 3: Save encoded data

cd ./DeepLocRNA-main/DeepLocRNA

python ./fine_tuning_deeprbploc_allRNA.py --dataset ./data/allRNA/allRNA_all_human_data_seq_mergedm3locall2_deduplicated2_filtermilnc.fasta  

Afterwards, there will be "*_X.npy" in the "./DeepLocRNA/data/allRNA/allRNA_all_human_data_seq_mergedm3locall2_deduplicated2_filtermilnc" folder.

In order to do multiple RNA prediction, we will generate tags for all RNA species

python ./fine_tuning_deeprbploc_allRNA.py --dataset ./data/allRNA/allRNA_all_human_data_seq_mergedm3locall2_deduplicated2_filtermilnc.fasta --RNA_tag

Afterwards, there will be "*_X_tag.npy" in the "./DeepLocRNA/data/allRNA/allRNA_all_human_data_seq_mergedm3locall2_deduplicated2_filtermilnc" folder.

Step 4: Training the model

To train the model locally, RBP pre-trained model should be loaded first.

pip install ../parnet-develop

We provide two options to train the model

First, you can use standard training strategy, using single GPU to train the model. It is worth note that the training is bonded with 5-folds as default, which will repeat 5 times to go through the data.

python ./fine_tuning_deeprbploc_allRNA.py --dataset ./data/allRNA/allRNA_all_human_data_seq_mergedm3locall2_deduplicated2_filtermilnc.fasta --load_data --gpu_num 1 --species human --batch_size 8 --flatten_tag  --gradient_clip --loss_type BCE  --jobnum 001 --species Human

Arguments for training

Long	Description
--dataset	Path to the input fasta file, which is a mixture of different RNA types. Formatted according to the input format.
--load_data	loading the saved data, you should add this argument after generating X.npy and X_tag.npy.
--gpu_num	The number of gpus you want to use while training the model.
--species	The species you want to predict
--batch_size	The batch size while training the model for each step.
--flatten_tag	Add this tag to enable multi-RNA training.
--gradient_clip	whether using gradient clip to make the training process stable.
--loss_type	The loss function that used to train the model. Default: BCE.
--jobnum	The identified number that represents the run of each training job.

Model Prediction.

There should prepare your input file to ".fasta"(#Input-format) format

python ./DeepLocRNA/fine_tuning_deeprbploc_allRNA_prediction.py --fasta ./example.fasta --rna_types mRNA --species Human

Alternatively, you can also use our online webserver if you only have a couple sequences to be predicted ()

IG scores calculation

python fine_tuning_deeprbploc_allRNA_prediction.py --fasta ./example.fasta --rna_types mRNA --species Human --plot True

If you wish to get the precise nucleotide contribution, please choose "--plot" as True, and define the configure file yourself as "att_config.csv" before input in the input frame.

attribution config file

starts,ends
10,100
50,100

Where 10 to 100 is the interval that you want to get the attribution scores.

python fine_tuning_deeprbploc_allRNA_prediction.py --fasta ./example.fasta --rna_types lncRNA --att_config att_config.csv --species Human --plot True

Arguments for prediction

Long	Description
--fasta	Path to the input fasta file, formatted according to the input format.
--rna_types	The type of RNA you want to predict
--species	The species you want to predict
--att_config	Path to the customized position of a specific sequence as .csv. formatted according to att config format
--plot	Plot the attribution figures, this is mandatory if you want to visualize the explaination plots. Default: True