RiboNN: A deep learning model to predict translation efficiency from mRNA sequence

For more information, please see our RiboNN paper.

• System requirements:

  This code has been tested on a system with 4 CPUs, 16 Gb RAM, and 1 NViDIA 10A GPU, with Ubuntu 20.04 as the OS (with CUDA Toolkit 11.3 installed). The required softwares are listed in environment.yml.

• To install project requirements:

  sudo apt install make
  
  # install mamba (https://github.com/conda-forge/miniforge) into "miniforge3/" in the home directory. 
  curl -L -O "https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-$(uname)-$(uname -m).sh"
  bash Miniforge3-$(uname)-$(uname -m).sh -b 
  
  # clone the repo
  git clone https://github.com/Sanofi-GitHub/RiboNN.git && cd RiboNN
  
  # install the RiboNN environment
  make install
  
  # activate the riboNN environment
  ~/miniforge3/bin/mamba shell init 
  source ~/.bashrc
  mamba activate RiboNN
  
  # create a directory for storing pretrained model weights
  mkdir models

  Note: Depending on your network speed, it may take a few minutes to install the required packages.

• To download pre-trained model weights for transfer learning or making predictions: Please download weights.zip from https://zenodo.org/records/15375573, unzip it, and put its contents in the models directory.

• To train the RiboNN model from scratch:

  1. Put the training data in a tab-separated text file in the "data" folder, which already contain an example training data file. The tab-separated text file should have columns named "tx_id" (unique transcript IDs), "utr5_sequence", "cds_sequence" (including start and stop codons), and "utr3_sequence". Alternatively, the file may have columns named "tx_id", "tx_sequence" (full transcript seuquences containing 5'UTR, CDS, and 3'UTR), "utr5_size" (lengths of the 5'UTRs), and "cds_size" (lengths of the CDSs).
  2. Edit the path to the training data ("tx_info_path") and other hyperparameters defined in the config/conf.yml file.
  3. Edit the code below line 18 of src/main.py to control how the model will be trained.
  4. Run make train at the terminal to start the training process.

• To do transfer learning (using pretrained multi-task models saved in the models directory):

  1. Put the training data in a tab-separated text file in the "data" folder, which already contain an example training data file. The tab-separated text file should have columns named "tx_id" (unique transcript IDs), "utr5_sequence", "cds_sequence" (including start and stop codons), and "utr3_sequence". Alternatively, the file may have columns named "tx_id", "tx_sequence" (full transcript seuquences containing 5'UTR, CDS, and 3'UTR), "utr5_size" (lengths of the 5'UTRs), and "cds_size" (lengths of the CDSs).
  2. Edit the path to the training data ("tx_info_path") and other hyperparameters defined in the config/conf.yml file.
  3. Edit the code below line 126 of src/main.py to control how the model will be trained.
  4. Run make transfer_learning at the terminal to start the training process.

• To make predictions after downloading the model weights (state_dict) into the models directory:

  1. Please create a tab-separated text file with columns named "tx_id" (unique transcript IDs), "utr5_sequence", "cds_sequence" (including start and stop codons), and "utr3_sequence". Alternatively, the file may have columns named "tx_id", "tx_sequence" (full transcript seuquences containing 5'UTR, CDS, and 3'UTR), "utr5_size" (lengths of the 5'UTRs), and "cds_size" (lengths of the CDSs).
  2. Save the text file as "prediction_input.txt" in the "data" folder. An example input file can be found in the "data" folder.
  3. (Optional) Edit the code below line 173 of src/main.py to control how the model will be used for prediction.
  4. Run make predict at the terminal.
  5. The predictions will be automatically written to a tab-separated file named "prediction_output.txt" in the "results" folder. Pre-existing files with the same name will be overwritten.

  Note: Input transcripts with 5'UTRs longer than 1,381 nt or combined CDS and 3'UTR sizes larger than 11,937 nt will be excluded in the output.