Sequential Compression of Gene Expression Data Across Latent Space Dimensions

Gregory Way and Casey Greene 2018

University of Pennsylvania

The repository stores data and data processing modules to sequentially compress gene expression data.

Named after the mechanical device developed by Alan Turing and other cryptologists in World War II to decipher secret messages sent by Enigma machines, BioBombe represents an approach used to decipher hidden messages embedded in gene expression data. We use the BioBombe approach to study different biological representations learned across compression algorithms and various latent dimensionalities.

In this repository, we compress three different gene expression data sets (TCGA, GTEx, and TARGET) across 28 different latent dimensions (k) using five different algorithms (PCA, ICA, NMF, DAE, and VAE). We evaluate each algorithm and dimension using a variety of metrics. Our goal is to construct reproducible gene expression signatures with unsupervised learning.

Links to access data and archived results can be found here: https://greenelab.github.io/BioBombe/

Citation

Compressing gene expression data using multiple latent space dimensionalities learns complementary biological representations Way, G.P., Zietz, M., Rubinetti, V., Himmelstein, D.S., Greene, C.S. Genome Biology (2020) doi:10.1186/s13059-020-02021-3

Approach

Our approach is outlined below:

BioBombe Training Implementation

Our model implementation is described below.

Analysis Modules

To reproduce the results and figures of the analysis, the modules should be run in order.

Name	Description
0.expression-download	Download and process gene expression data to run through pipeline
1.initial-k-sweep	Determine a set of optimal hyperparameters for Tybalt and ADAGE models across a representative range of k dimensions
2.sequential-compression	Train various algorithms to compress gene expression data across a large range of k dimensions
3.build-hetnets	Download, process, and integrate various curated gene sets into a single heterogeneous network
4.analyze-components	Visualize the reconstruction and sample correlation results of the sequential compression analysis
5.analyze-stability	Determine how stable compression solutions are between and across algorithms, and across dimensions
6.biobombe-projection	Apply BioBombe matrix interpretation analysis and overrepresentation analyses to assign biological knowledge to compression features
7.analyze-coverage	Determine the coverage, or proportion, of enriched gene sets in compressed latent space features for all models and ensembles of models
8.gtex-interpret	Interpret compressed features in the GTEX data
9.tcga-classify	Input compressed features from TCGA data into supervised machine learning classifiers to detect pathway aberration
10.gene-expression-signatures	Identify gene expression signatures for sample sex in GTEx and TCGA data, and MYCN amplification in TARGET data

Algorithms

See 2.sequential-compression for more details.

Computational Environment

All processing and analysis scripts were performed using the conda environment specified in environment.yml. To build and activate this environment run:

# conda version 4.5.0
conda env create --force --file environment.yml

conda activate biobombe