martinspendl / discoveryscience24-paper

Accompanying code to the DS24 paper

Discovery Science 2024 // Transcription-decay decomposition loss using self-supervised CoxPH

This is code to reproduce key results and figures from the article: Latent embedding based on a transcription-decay decomposition of mRNA dynamics using self-supervised CoxPH.

Autoencoders are trained using different loss function.

We compare latent space representations of the models by evaluating their performance on three downstream tasks.

Instalation

The code was tested on Ubuntu 20.04.4 LTS and MacOS 13.1. With python version 3.10.8.

Follow these steps to prepare the environment:

• Clone the repository

git clone https://github.com/MartinSpendl/DiscoveryScience24-paper.git
cd DiscoveryScience24-paper

• Install the required packages

# using pip in a virtual environment
pip install -r requirements.txt

# using Conda
conda create --name <env_name> --file requirements.txt
conda activate <env_name>

Data

Data used for the analysis is publically accessible. Download file in the data/raw folder.

The Cancer Genome Atlas datasets

TCGA datasets from UCSC Xena portal: https://xenabrowser.net/datapages/

For clustering, load Illumina gene expressions and Phenotype data:

• TCGA-BRCA
• TCGA-KIRP
• TCGA-LGG

For survival, load Illumina gene expressions and Survival data:

• TCGA-KIRC
• TCGA-LUAD
• TCGA-LGG

METABRIC project data

Download multi-omic data from the cBioPortal.

Cancer Cell Line Encyclopedia (CCLE)

Download gene expression data and drug screening data from the Genomics of Drug Sensitivity in Cancer.

Genesets

L1000 geneset from the GEO project GSE92742 is already provided in the genesets folder.

Reproduce results

Firstly, run all the scipts in the /scipts directory:

cd scripts

python scripts/model-training-5-CV-CCLE-METABRIC.py --hyper-parameter-optimization

python scripts/model-training-5-CV-TCGA.py --hyper-parameter-optimization

python scripts/model-training-clustering.py --hyper-parameter-optimization

Note that due to hyper-parameter optimization, the training can take from several days to weeks if CUDA is not enabled.

Secondly, run notebooks in the /notebooks directory.

Figures from the notebooks are stored in the /figures directory.