Computational Genomics Final Project

This repository contains a final project for COMS 4761 -- Computational Genomics at Columbia University in the City of New York.

• GitHub repo: https://github.com/k3jph/coms4761.git
• Free software: MIT

Our project replicates aspects of the GRADIS for the supervised learning of gene regulatory networks based on graph distance profile of transcriptomics data. For this, we have replicated the basic analysis in R, and used methods other than SVM. In addition, to simplify some of the processing, we have exported the data from Excel to CSVs.

Features

All referenced files can be found in src, which is organized:

• data folder containing E. coli and S. cerevisiae inputs in the correct format for GRADIS use. Also includes the raw DREAM5 data for both organisms in raw and the pre-processing script used to transform to the proper input form/files.
• gnn folder containing the GRGNN implementation from https://github.com/juexinwang/GRGNN, as well as the output results used in our analysis. Includes a separate README file with quickstart instructions.
• R folder contains the majority of our GRADIS implementation scripts, as translated from the original MATLAB implementation here: https://github.com/MonaRazaghi/GRADIS.

Quickstart

Data Transformation

The GRADIS scripts expect three input files:

• Genes.csv with the list of gene names for the organism
• Network.csv listing labeled TF-gene interactions (where 1 indicates a positive interaction)
• Expression.txt with collected gene expression levels. Each row represents a sample.

Raw DREAM5 data for E. coli (Network 3) and S. cerevisiae (Network 4) is available in data\raw, as pulled from https://www.synapse.org/#!Synapse:syn2787211. To transform these files into the expected format for our GRADIS implementation (discussed below), navigate to the data directory, and run the following command:

python data_processing.py <network_id>

where network_id = 3 for E. coli and network_id = 4 for S. cerevisiae.

Usage

GNN

bash install.sh

to install the required software and libraries. Node2vec and DGCNN are included in software folder.

Unzip DREAM5 data

cd data/dream

unzip dreamdata.zip

cd ../../

(Optional): Preprocessing DREAM5 data

cd preprocessing

python Preprocessing_DREAM5.py 3

python Preprocessing_DREAM5.py 4

In this program, data3 means E.coli dataset, data4 means S. cerevisae dataset

Train and test E. coli with hop 1 and embedding, Type:

python Main_inductive_ensemble.py  --traindata-name data3 --testdata-name data3 --hop 1 --use-embedding

Train and test S. cerevisae with hop 1 and embedding, Type:

python Main_inductive_ensemble.py  --traindata-name data4 --testdata-name data4 --hop 1 --use-embedding

R

To test three different methods for differentiating between positive and negative, for E. coli use:

Rscript src/R/glmmer-ec.R

For S. cerevisiae, use:

Rscript src/R/glmmer-sc.R

Both scripts include a vanilla GLM, random forest, and naïve Bayes as classifiers. The code is built on top of the R library caret, so switching classifiers is trivial.

To identify and train negatives over the E. coli data, use:

Rscript src/R/gradis-neg-ec.R

To identify and train negatives over the S. cerevisiae data, use:

Rscript src/R/gradis-neg-sc.R

Both scripts use random forest by default, but also use caret, so switching is, again trivial. However, even rapidly training models will take multiple hours. Using random forest requires up to 24 hours on new Apple silicon.

For more information

• "Supervised learning of gene-regulatory networks based on graph distance profiles of transcriptomics data"
• "Inductive inference of gene regulatory network using supervised and semi-supervised graph neural networks"
• James P. Howard, II <jh@jameshoward.us>