Gene2vec

***** New: August 9th, 2019 update *****

We provided the evaluation script using target function proposed in the manuscript, as well as a gene2vec file in word2vec format.

Introduction

Gene2Vec is a distributed representation of genes based on co-expression. From a pure data-driven fashion, we trained a 200-dimension vector representation of all human genes, using gene co-expression patterns in 984 data sets from the GEO databases.

In this repository, we provided the relevent codes as well as pre-trained gene2vec files.

Installing Gene2vec

Requirements

Gene2vec relies on Python 3.6, TensorFlow 1.6+, gensim 3.4.0, numpy 1.14.0+, and matplotlib 2.1.2+

The Multicore t-SNE relies on a GitHub repository. Please follow their installation instruction.

Install

To download the codes, please do:

git clone https://github.com/jingcheng-du/Gene2vec.git
cd Gene2vec/
pip install -r requirements.txt

Usage

gene2vec.py

This script takes gene co-expression data as the input and output binary gene2vec file. It relies on gensim Word2vec module. The detailed parameters for Word2vec module can be seen in their document.

Please specify data directory, embedding output directory and data file ending pattern. You can direct run the script using the following command (txt means that the names of your data files end with "txt"):

python gene2vec.py data_directory output_directory txt

For co-expression data files, each line should have a pair of genes that is separated by a space. The format of the files should be like:

TOX4 ZNF146
TP53BP2 USP12
TP53BP2 YRDC

By default, the output will be in both in binary format and txt format. The binary format can be access using gensim module. For txt format, each line starts with the gene name and is followed by the vector. It can be opened by the text editor.

To speed up the training, please make sure you have a C compiler before installing gensim, see this article.

tsne_multi.py

We provide this script to map the high dimensional gene2vec to 2-D array. It relies on on a Parallel t-SNE implementation GitHub repository.

It takes the gene2vec txt file as the input and generates two separate files: a label file that each line is the gene name and a data file that each line has 2-D vector for the corresponding gene.

We implement a mutlipe-thread tsne training here. You can specify a set of iterations you'd like to train in parallel:

p = Pool(6)
p.map(TSNEWoker, ["100","5000","10000","20000","50000","100000"])

GGIPNN_Classification.py

GGIPNN stands for gene-gene interaction predictor neural network. Given the training, validation and testing datasets for gene-gene interaction, this script will a leverage multilayer perceptron (defined in GGIPNN.py) to predict gene-gene interaction and generate the AUC score.

Each dataset (training, validation and testing) should have two files:

1. a text file that each line has the names of the gene pair:

GPNMB BAP1
GPR34 CARD16
ELF5 TGFB2
LILRB2 NCR2

2. a label file that each line has either 1 (the corresponding gene pair has the gene-gene interaction) or 0 (doesn't have the gene-gene interaction):

0
0
1
1

You can specify whether to initialize the embedding layer using the pre-trained gene2vec or initialize the embedding layer randomly. You can also specify whether the embedding layer is trainable during the training process. Just to choose True or False here:

tf.flags.DEFINE_boolean("use_pre_trained_gene2vec", True, "use_pre_trained_gene2vec")
tf.flags.DEFINE_boolean("train_embedding", False, "train_embedding")

GTRxFigure.py

This script takes gene2vec tsne files (label and data) and tissue-specific genes expression files as the input and generates tissues-specific genes experssion maps.

The tissue-specific genes expression file should follow the following format. Each line starts with the gene name and followed by the z score:

ATRX	0.598962527411
TCOF1	-0.264690317639
NSRP1	0.716336803551
OPA1	0.223145913678
RHEB	0.978021549909
SEMA5B	-0.110593590242

Citation

If you use these codes in your publications, please cite this paper:

@article{2018gene2vec,
  title="Gene2vec: distributed representation of genes based on co-expression",
  journal="BMC Genomics",
  year="2019",
  month="Feb",
  day="04",
  volume="20",
  number="1",
  pages="82"author={Du, Jingcheng and Jia, Peilin and Dai, YuLin and Tao, Cui and Zhao, Zhongming and Zhi, Degui},
  issn="1471-2164",
  doi="10.1186/s12864-018-5370-x"
}