Genomic Impact Transformer (GIT)

Introduction

This repository contains PyTorch implementation of GIT model (and its variants) in the following paper: Yifeng Tao, Chunhui Cai, William W. Cohen, Xinghua Lu. From genome to phenome: Predicting multiple cancer phenotypes based on somatic genomic alterations via the genomic impact transformer. Proceedings of the Pacific Symposium on Biocomputing (PSB). 2020.

The preprocessed TCGA dataset, and gene embeddings mentioned in the paper are also released below.

Prerequisites

The code runs on Python 2.7, and following packages are used:

PyTorch 0.1.12_2, pandas, pickle, numpy, random, argparse, os.

Data

All the required data can be downloaded to the directory:

cd genome-transformer
wget www.cs.cmu.edu/~yifengt/paper/2019a/data.tar.gz -O data.tar.gz
tar -zxvf data.tar.gz

TCGA dataset

The preprocessed SGA-DEG TCGA dataset is available at data/dataset.pkl, which contains SGAs, DEGs, cancer types, and barcodes of 4,468 cancer samples from TCGA:

data = pickle.load( open("data/dataset.pkl", "rb") )

Gene embeddings

Two types of gene embeddings are available:

Gene2Vec-pretrained gene embeddings: data/gene_emb_pretrain.npy;
Gene2Vec-pretrained + GIT-finetuned gene embeddings: data/gene_emb_finetune.npy.

You may extract the gene embeddings in Python:

gene_emb_matrix = numpy.load("data/gene_emb_finetune.npy")

It is a 19782 by 512 matrix, where the index of each row can be mapped to a gene name through data/idx2gene.txt.

Trained GIT model

The parameters of trained GIT model are stored at data/trained_model.pth.

Replicate experiments

Train GIT model

You can train the GIT from scratch and then evaluate its performance on test set:

python test_run.py

Evaluate GIT model

If you have already downloaded the trained parameters of GIT, you can directly evaluate its performance:

python test_run.py --train_model False

GIT variants

You may run more GIT variants, e.g., GIT-init, GIT-attn, GIT-can etc., by checking the code and comments of test_run.py, or:

python test_run.py --help

FAQ

Q: How to prepare my own input data and run experiments on the new data?
A: You can directly prepare the dataset.pkl pickle file. However, we also provide pieces of code (prepare_dataset.py) to faciliate you to convert more general data types (csv and txt) into the dataset.pkl.

First, you need to format your own data into three files: mydata/cancer_type.txt (each line a cancer type), deg.csv (each row a sample, each column an over/under-expressed gene), sga.txt (each row a sample name and mutated genes, comma separated).

Second, run the following code to convert and merge these files into a single mydata/dataset.pkl file:

python prepare_dataset.py --input_dir mydata --output_dir mydata

Third, run experiments on the new data:

python test_run.py --initializtion False --input_dir mydata --output_dir mydata

Citation

If you find the data or code from this repository helpful, please cite this paper:

@inproceedings{tao2020git,
  title = {From Genome to Phenome: Predicting Multiple Cancer Phenotypes based on Somatic Genomic Alterations via the Genomic Impact Transformer},
  author = {Tao, Yifeng  and 
    Cai, Chunhui  and 
    Cohen, William W.  and 
    Lu, Xinghua},
  booktitle={Pacific Symposium on Biocomputing},
  year = {2020},
}

coder-drinker/genome-transformer