Gregory Way and Casey Greene

A transcriptome can describe the total state of a tumor at a snapshot in time. In this repository, we use cancer transcriptomes from The Cancer Genome Atlas PanCancerAtlas project to interrogate gene expression states induced by deleterious mutations and copy number alterations.

The code in this repository is flexible and can build a Pan-Cancer classifier for any combination of genes and cancer-types using gene expression, mutation, and copy number data. In this repository, we provide examples for building classifiers to detect aberration in TP53 and Ras signalling.

The Ras signalling pathway is a major player in cancer development and treatment resistance. We observed that nearly 60% of all tumors in TCGA have mutations or copy number alterations in at least one of 38 core pathway genes (Sanchez-Vega et al. 2018).

We applied our approach to detect Ras pathway activation using KRAS, HRAS, and NRAS gain of function mutations and copy number gains to define our gold standard Ras hyperactivation events. We train a supervised classifier to detect when a tumor has activated Ras.

For more details about the approach, see our paper published in Cell Reports. The paper should be cited as:

Way, GP, Sanchez-Vega, F, La, K, Armenia, J, Chatila, WK, Luna, A, Sander, A, Cherniack, AD, Mina, M, Ciriello, G, Schultz, N., The Cancer Genome Atlas Research Network, Sanchez, Y, Greene, CS. 2018. Machine Learning Detects Pan-cancer Ras Pathway Activation in The Cancer Genome Atlas. Cell Reports 23(1):172-180.e3 doi:10.1016/j.celrep.2018.03.046

We have previously described the ability of a machine learning classifier to detect an NF1 inactivation signature using Glioblastoma data (Way et al. 2016). There, we applied an ensemble of logistic regression classifiers to the problem, but the solutions were unstable and overfit. To address these issues, we posited that we could leverage data from diverse cancer types to build a pancancer NF1 classifier. We also hypothesized that a Ras classifier would be able to detect tumors with NF1 inactivation since NF1 directly inhibits RAS activity.

We are also interested in building a classifier to detect TP53 inactivation. TP53 is the most highly mutated gene in cancer and regulates several important oncogenic processes such as apoptosis and DNA damage response (DDR). We include a pipeline to build and evaluate a machine learning TP53 classifier. See tp53_analysis.sh for more details.

The description for this analysis can be viewed in the following publication:

Knijnenburg, TA, Wang, L, Zimmermann, MT, Chambwe, N, Gao, GF, Cherniack AD, Fan, H, Shen, H, Way, GP, Greene, CS, Liu, Y, Akbani, R, Feng, B, Donehower, LA, Miller, C, Shen, Y, Karimi, M, Chen, H, Kim, P, Jia, P, Shinbrot, E, Zhang, S, Liu, J, Hu, H, Bailey, MH, Yau, C, Wolf, D, Zhao, Z, Weinstein, J, Li, L, Ding, L, Mills, GB, Laird, PW, Wheeler, DA, Shmulevich, I, The Cancer Genome Atlas Research Network, Monnat Jr, RJ, Xiao, Y, Wang, C. 2018. Genomic and Molecular Landscape of DNA Damage Repair Deficiency across The Cancer Genome Atlas. Cell Reports 23(1):239-254.e3 doi:10.1016/j.celrep.2018.03.076

All data was released by the TCGA PanCancerAtlas project. The compendium of papers is described here. Supplementary data from these papers can be downloaded from the NCI. The specific data used in the analyses presented here are archived on Zenodo Gene expression) and copy number data can be accessed here.

See scripts/initialize/download_data.sh for more details.

Also note that the definitions for all TCGA cancer-type acronyms is stored in data/tcga_dictionary.tsv.

The repository must be cloned onto local machine before analyses can proceed.

# Make sure git-lfs (https://git-lfs.github.com/) is installed before cloning
# If not, run `git lfs install`
git clone git@github.com:greenelab/pancancer.git

cd pancancer

We provide two distinct example pipelines for predicting TP53 and NF1/RAS loss of function.

1. TP53 loss of function (see tp53_analysis.sh)
2. Ras signaling hyperactivation (see ras_analysis.sh)

For custom analyses, use the scripts/pancancer_classifier.py script with command line arguments.

python scripts/pancancer_classifier.py ...