DCell

A deep neural network simulating cell structure and function

Introduction

DCell is an application to provide an easy-to-use user interface and interpretable neural network structure for modeling cell structure and function.

Reference implementation is available here:

• http://d-cell.ucsd.edu/

Publication

Using deep learning to model the hierarchical structure and function of a cell. Jianzhu Ma, Michael Ku Yu, Samson Fong, Keiichiro Ono, Eric Sage, Barry Demchak, Roded Sharan & Trey Ideker. Nature Methods, 2018

Directory Structure:

• training/code: folder containing lua code for both neural network training and prediction.
• training/TrainData: Training and predicting data.
• training/Topology: Topology files for gene ontology.
• backend: python wrapper code to perform predictions.
• frontend: Javascript files to construct the web application server.
• data-builder: Source files and scripts for backend database.

Dependencies:

The code is based on Lua Torch running on a GPU linux system. See here for installation and basic tutorials.

Demo

Training cmd:

th Train_DCell.lua -train training_file -test testing_file -topo ontology_file -save model_file

Predicting cmd:

th Predict_Dcell.lua -load model_file -test testing_file -out result_file [-gindex  gene_index_file]

Examples of training/testing files are in TrainData/ and ontology files are in Topology/.

Topology file defines the structure of an ontology as:

• ROOT: term_name #genes
• GENES: gene1, gene2, ...
• TERMS: child_term1, child_term2

Output:

The model trained for each iteration will be saved in "-save model_file". The training program will produce a gene index mapping file saved in the same folder.

The predicting program will load both gene index file and trained model file and save the predictions in "-out result_file".

Data availability

To train the ontology on genetic interaction or growth using the gene ontology.

Please download the ontology at:

• http://geneontology.org/page/download-ontology

Genetic interaction and growth is at:

• http://thecellmap.org/costanzo2016

D-Cell predictions for Costanzo et al. 2010 dataset is at:

• http://chianti.ucsd.edu/~kono/ci/data/deep-cell/predictions/prediction.tar.gz

The running time on a standard Tesla K20 GPU takes <2 minutes for terms like "DNA repair", and 2-3 days for using the GO and ~7 millions training data.

User Documentation

Please visit our wiki.

---

© 2017-2018 UC, San Diego Trey Ideker Lab

Developed and Maintained by Keiichiro Ono (kono ucsd edu)