DeepSEED: Deep flanking sequences engineering for efficient promoter design

The code for official implementation of "Deep flanking sequences engineering for efficient promoter design" This codebase provides:

The conditional generative adversarial networks (GANs) that generate promoter sequences when giving conditional motif sequences. (the folder "Generator")
The predictor that could evaluate promoter activities based on densenet-lstm model. (the folder "Predictor")
The optimizer that could design promoter sequences with high activities when giving conditional motif sequences. (the folder "Optimizer")

Introduction
Environment Setup
Design Promoter Sequence
Bibtex

Introduction

We introduced an AI-aided promoter design framework, DeepSEED, that employs both experts’ knowledge and deep learning together to efficiently design synthetic promoters of diverse desirable functions. DeepSEED incorporates the user-defined cis-regulatory sequences as ‘seed’ and generates flanking sequences that match the ‘seed’. We showed that DeepSEED could automatically capture a variety of weak patterns like k-mer frequencies and DNA shape features from active promoters in the training set, and efficiently optimize the flanking sequences in synthetic promoters to better match these properties.

We validated the effectiveness of this framework for diverse synthetic promoter design tasks in both prokaryotic and eukaryotic cells. DeepSEED successfully designed constitutive, IPTG-inducible, and doxycycline(Dox)-inducible promoters with significant performance improvements, suggesting DeepSEED as an efficient AI-aided flanking sequence optimization approach for promoter design that may greatly benefit synthetic biology applications.

Environment Setup

Env Requirements:

MAC OS, Linux or Windows.
Python 3.7+.
PyTorch 1.7.1
CUDA 11.2 if you need train deep learning model with gpu.

Steps of using DeepSEED:

Install Python ref to Download Python

Install DeepSEED in virtualenv to keep your environment clean:

on macOS/Linux

pip install virtualenv  
# or 
pip install -i https://pypi.tuna.tsinghua.edu.cn/simple virtualenv 

virtualenv --python=python3 DeepSEED
cd DeepSEED
source ./bin/activate

Optional: After use, shutdown virtual environment with

deactivate

on Windows (open cmd with administrator's permission)

pip install virtualenv  
# or 
pip install -i https://pypi.tuna.tsinghua.edu.cn/simple virtualenv 
virtualenv --python=python3 DeepSEED
cd DeepSEED
.\Scripts\activate.bat

Optional: Shutdown virtual environment with

.\Scripts\deactivate.bat

Install Git, this step is optional if you does not install DeepSEED by git clone. Clone the source codes with git.
```
git clone https://github.com/WangLabTHU/deepseed.git
```
or, download the source codes and extract files and put it in the virtual environment directory you defined.

after 2/3, the directory of DeepSEED should have the following structure:

DeepSEED
    deepseed
        |-------
        |
        |-------data
                |...
        |-------Generator
                |...
        |-------Predictor
                |...
        |-------Optimizer
                |...
        |-------
        |...
    |...

After the extraction, download all dependencies with the following commend.

cd deepseed (or cd deepseed-main)
pip install -r requirements.txt

To boost download speeds, you could setup pip mirror such as

pip install -i https://pypi.tuna.tsinghua.edu.cn/simple -r requirements.txt  
# or
pip install -i https://mirrors.aliyun.com/pypi/simple  -r requirements.txt

It is recommended to use the official website to install pytorch.

# CPU version
pip install torch==1.7.1+cpu torchvision==0.8.2+cpu -f https://download.pytorch.org/whl/torch_stable.html
# GPU version
pip install torch==1.7.1 torchvision==0.8.2 -f https://download.pytorch.org/whl/torch_stable.html

The typical install time on a desktop computer depends on your download speed. Suppose your computer has 1MB/s when downloading python package, it will take about 1 hour to install all the necessary packages.

Run deepseed with python and enjoy it with following steps:

Design Promoter Sequence

We take design of 3-lacO IPTG-inducible promoters in E. coli as an example, to illustrate how to train the DeepSEED model and design the promoter sequences

1. Training the conditional GANs (Expected run time on GeForce GTX 1080Ti: 10 hours)

```
cd Generator
python cGAN_training.py
```

2. Training the predictor (Expected run time on GeForce GTX 1080Ti: 0.5 hour)

```
cd ../Predictor
python predictor_training.py
```

3. Design the promoter sequences with optimizer (Expected run time on GeForce GTX 1080Ti: 24 hours)

```
cd ../Optimizer
python deepseed_optimizer.py
```

4. Check the synthetic promoter sequences!

```
vi ./results/ecoli_3_laco.txt
```

Bibtex

Please consider citing our paper in your publications if the project helps your research. BibTeX reference is as follows.

@article{,
  title={Deep flanking sequences engineering for efficient promoter design},
  author={Pengcheng Zhang, Haochen Wang, Hanwen Xu, Lei Wei, Zhirui Hu, Xiaowo Wang
},
  journal={biorkiv},
  year={2023}
}

License

For academic use, this project is licensed under the MIT License - see the LICENSE file for details For commercial use, please contact the authors.

Reference

[1] Isola, Phillip, et al. "Image-to-image translation with conditional adversarial networks." Proceedings of the IEEE conference on computer vision and pattern recognition. 2017.

Parts of the code have been modified from the Pix2pix, which could be found at https://github.com/phillipi/pix2pix.

WangLabTHU/deepseed