epiG4NN: G-quadruplex prediction in live cells using epigenetic data

This repository contains the scripts for the epiG4NN data preparation pipeline and the code for epiG4NN model training.

• epiG4NN: G-quadruplex prediction in live cells using epigenetic data

  • Overview
  • Installation
  • G4 formation prediction with pre-trained models
  • Data preparation pipeline for new model training
  • Training of new models
  • Utils

Overview

epiG4NN is a deep convoluational neural network architectiure to predict formation of G-quadruplexes in cells. epiG4NN model uses the potential G4 sequence (PQS) padded to 1000 nucleotides for context, matched with the respective epigenetic data of choice.

You can use epiG4NN to predict G4 formation, given the sequence and the epigenetic mark data (H3K4me3, ATAC-seq, H3K27ac, H3K4me1, H3K36me3 and H3K9me3 are supported), or train your own model.

Installation

Clone the epiG4NN repository to a local folder. Use the file environment.yml to create a conda environment:

conda env create -f environment.yml
conda activate epig4nn

G4 formation prediction with pre-trained models

This repository contains models trained on various epigenetic marks in A549 cells (see preprint for more details) that can be used for predictions.

To make predictions, python script predict.py needs a .fa or fasta-style .txt file with 1000nt sequences centered at G4, and the respective normalized (0.0-1.0 range) epigenetic arrays in a .csv file. The epigenetic data table should be structured as follows: (n_rows, n_columns) = (n_samples, 1000). The first column should correspond to the sample IDs from the .fa file. See pqs_chr22.txt and h3k4me3_293T_chr22.csv in the data folder for an example. If the prediction is made using a sequence-only model (no epigenetic data), only the sequence file is required.

predict.py has the following arguments:

Argument	Default value	Description
--seqs	None	(required) input .fa file with 1000nt G4 sequences
--epi	None	(required) input .csv file with 1000nt epigenetic arrays
--model	None	(required) model to use: seq, h3k4me3, atac, h3k27ac, h3k4me1, h3k36me3, h3k9me3
--output	None	(required) output filename prefix; will be saved to output folder

Examples of usage:

#H3K4me3-based model prediction
python predict.py --seqs data/pqs_chr22.txt --epi data/h3k4me3_293T_chr22.csv --model h3k4me3 --output out_h3k4me3

#sequence-based model prediction
python predict.py --seqs data/pqs_chr22.txt --model seq --output out_seq

The output will be saved to the output folder with the specified prefix.

Data preparation pipeline for new model training

data_prep.sh contains the full data preparation pipeline for new model training. Prerequisites: human genome file, .bedgraph with experimental G4 scores, .bedgraph with epigenetic scores. A toy dataset for pipeline execution is stored in the data/tiny folder. Usage:

1. Download and unpack the human genome from the UCSC website into the data folder:

cd data/
wget --timestamping 'ftp://hgdownload.soe.ucsc.edu/goldenPath/hg19/bigZips/hg19.fa.gz' -O hg19.fa.gz
gunzip hg19.fa.gz
cd ..

2. Run the data preparation script with the toy G4 and epigenetic data in data/tiny. PQS sequences found in hg19 are used by default. If you are preprocessing your own data, edit lines 9-10 in data_prep.sh to point to your files.

bash data_prep.sh

This script will save the preprocessed .npy inputs into data/inputs_numpy/train_1000nt.

Training of new models

After generating the inputs with the data_prep.sh script, one can train either a bare sequence model (train_seq.py) or a model with epigenetic inputs (train_epi.py).

train_seq.py and train_epi.py have the following arguments:

Argument	Default value	Description
--model_name	None	(required) model name prefix
--inputs	None	(required) folder with inputs split by chromosome
--kernel_size1	11	kernel size 1 for convolution
--kernel_size2	11	kernel size 2 for convolution
--no_stacks	2	number of RB stacks (convolutional towers)
--dilation1	4	dilation rate 1
--dilation2	4	dilation rate 2
--batch	32	batch size for training
--const_lr	False	use constant learning rate
--q	4.08277148e-02	G4 score binarization threshold

Note that we use continuous G4 labels for training that are binarized using a predefined threshold q. If you are using your own data, this must be redefined. Examples of training:

#H3K4me3-based model training
python train_epi.py --inputs data/inputs_numpy/train_1000nt --model_name tiny_293T_h3k4me3

#sequence-based model training
python train_seq.py --inputs data/inputs_numpy/train_1000nt --model_name tiny_293T_seq

Utils

PQS_search.py file contains a python re implementation for PQS search in the human genome

Table of contents generated with markdown-toc