PARADE: a generative framework for enhanced cell-type specificity in rationally designed mRNAs

Matvei Khoroshkin, Arsenii Zinkevich, Elizaveta Aristova, Hassan Yousefi, et al., A generative framework for enhanced cell-type specificity in rationally designed mRNAs, biorXiv, 2024; doi: 10.1101/2024.12.31.630783v1

[Preprint]

Overview

mRNA delivery offers new opportunities for disease treatment by directing cells to produce therapeutic proteins. However, designing highly stable mRNAs with programmable cell type-specificity remains a challenge. Here, we present PARADE (Prediction And RAtional DEsign of mRNA UTRs), a generative AI framework for the design of untranslated RNA regions (UTRs) with tailored cell type-specific activity.

System Requirements

Hardware Requirements

The approach presented in PARADE requires only a regular PC with enough RAM to perform the operations defined by a user. To satisfy the minimal requirements and use the basic functions of PARADE (i.e. sequence activity prediction), only a computer with only about 4 GiB of RAM is necessary. However, if you would like to use PARADE generator, the generation results may depend on the number of sequences tested, therefore we recommend a computer with the following specs for optimal performance:

RAM: 16+ GiB
CPU: 8+ threads, 3.3+ GHz/core GPU: NVIDIA GeForce 3090/4090

OS Requirements

The development version was tested on Linux operating system:

OS: Ubuntu 20.04 LTS (GNU/Linux 5.15.0-116-generic x86_64)

The code, however, should be compatible with Windows, Mac, and Linux operating systems.

Usage

Setting up dependencies

To get going with PARADE, you should perform the following steps:

1. Install the latest version of conda package manager. We recommend using the latest version of Miniforge, however, Miniconda or Anaconda should also work fine.
2. Clone the current repository and cd into it.
3. Create a conda environment from YAML: conda env create -f environment.yml
4. The created environment can then be used in any convenient way, e.g. in JupyterLab, Visual Studio Code, PyCharm and other IDEs.

If you want to reproduce PARADE results (e.g. retrain the models), you also have to download data from the associated Zenodo repository and unpack it into the repository root directory.

Demo usage

Using PARADE predictive models

• Activity prediction for 5'UTR sequences [Tutorial]
• Activity prediction for 3'UTR sequences [Tutorial]
• Stability prediction for 3'UTR sequences [Tutorial]

Using PARADE generative models

• Generate UTRs with Diffusion [5'UTR, Notebook | 3'UTR, Notebook]
• Generate UTRs with genetic algorithm [5'UTR, Notebook | 3'UTR, Notebook]