DREAM

DREAM (DNA cis-Regulatory Elements with controllable Activity design platforM) is an efficient, scalable and explainable computational framework to design CREs from scratch.

Table of Contents

• Installation
• Dataset Preparation
• DREAM Training
• CRE Optimization

Installation

Prerequisites

• Python 3.10.13
• TensorFlow 2.13
• scikit-learn
• h5py
• pysam
• deap 1.4.1

Install Dependencies

conda install -r requirements.txt

Clone Repository

git clone https://github.com/cisGrammar/DREAM.git

DNA Sequence Encoding and Dataset Preparation

The structure of DREAM's enhancer activity prediction framework(SENet):

The script encode_sequences.py encodes DNA sequences into one-hot format and saves them as an HDF5 dataset.

Usage

Command Line Arguments

• --REFERENCE: Path to the reference genome in FASTA format (default: '/data/reference/melanogaster/dm3.fa').
• --output_h5: Path to save the output HDF5 file (default: 'alldata.h5').
• --traindata_df: Path to the training dataset CSV file (default: 'train_dataset.csv').
• --valdata_df: Path to the validation dataset CSV file (default: 'val_dataset.csv').
• --testdata_df: Path to the test dataset CSV file (default: 'test_dataset.csv').

Example Usage

To encode DNA sequences and prepare datasets, use the following command:

python dataset/dna_data.py --REFERENCE '/data/reference/melanogaster/dm3.fa' --output_h5 'alldata.h5' --traindata_df 'train_dataset.csv' --valdata_df 'val_dataset.csv' --testdata_df 'test_dataset.csv'

DREAM Training

The script train.py TensorFlow script for training an SE-ResNet model on enhancer activity dataset(e.g., STARR-seq dataset, MPRA dataset).

Usage

Command Line Arguments

The script train.py supports the following command line arguments:

• --cuda_devices: Specifies CUDA_VISIBLE_DEVICES setting (default: '0').
• --data_path: Path to the HDF5 data file containing training, validation, and test datasets (default: 'alldata.h5').
• --batch_size: Batch size for training (default: 512).
• --epochs: Number of epochs for training (default: 100).
• --patience: Patience for early stopping (default: 10).
• --checkpoint_path: Prefix for the model checkpoint files (default: 'checkpoint').

Example Usage

To train the SE-ResNet model, use the following command:

python train.py --cuda_devices '0' --data_path 'path/to/your/data.h5' --batch_size 512 --epochs 100 --patience 10 --checkpoint_path 'your_checkpoint_prefix'

CRE Optimization

The GA/ga.py is the implementation of an evolutionary algorithm using the DEAP framework to optimize the CRE activity based on the CRE activtiy prediction module's output.

Usage

Command-line Arguments

The GA/ga.py script supports the following command-line arguments:

• --sequence_length: Length of the DNA sequences to generate (default: 249).
• --nucleotide_frequency: Frequencies of nucleotides [A, C, G, T] (default: [0.25, 0.25, 0.25, 0.25]).
• --seed: Random seed for TensorFlow and NumPy (default: 12345).
• --cuda_devices: CUDA visible devices (default: "0").
• --best_model_checkpoint: Path to the best model checkpoint for fitness prediction (default: "checkpoint_keras/").
• --indpb: Probability of mutating each nucleotide in an individual (default: 0.025).
• --population_size: Initial population size (default: 100000).
• --NGEN: Number of generations to evolve (default: 90).
• --output_file: Path to save the evolution fitness dataframe CSV (default: "evolution_fits_df.csv").

Notice

Trained models are available for download via Baidu Netdisk. We recommend using these pre-trained models for enhancer/silencer element design and optimization. Download link: https://pan.baidu.com/s/1m4fEcxwm8GnmhtfKJeyD3g Access code: yywq

Example

python GA/ga.py --sequence_length 249 --nucleotide_frequency 0.25 0.25 0.25 0.25 --seed 12345  --best_model_checkpoint "/path/to/best_model/" --indpb 0.025 --n 1000 --NGEN 10 --output_file "/path/to/output.csv"