Genomics of Drug Sensitivity in Cancer

Why do some drugs have multiple Drug IDs?

The drug data presented at cancerrxgene.org is the result of a combined screening effort at the Massachusetts General Hospital (Boston, USA) and the Wellcome Sanger Institute (Cambridge, UK). Some drugs have been screened at both sites, resulting in two Drug IDs for the same drug. Drug ID values below 1000 indicate drugs that have been screened at MGH, while Drug IDs greater than 1000 were screened at Sanger.

Some associations will be overlapping between the pan-cancer and cancer-specific, whereas some will be unique to a specific cancer type. The results from these analyses can be queried separately using the drop-down menus on each results page.

Cancer-specific analyses were performed for 18 cancer types where n > 15 cell lines were available.

What is the difference between a pan-cancer and cancer-specific ANOVA analysis?

We have performed two types of analyses on our drug sensitivity and genomic datasets. A pan-cancer analysis correlates drug sensitivity for cell lines from all cancer types with genomic features identified from the analysis of patient tumours across multiple different cancer types. A cancer-specific analysis correlates drug sensitivity for cancer cell lines of the specified cancer type with genomic features identified in patients from the designated cancer type. The cancer types are those described by The Cancer Genome Atlas (TCGA).

Some associations will be overlapping between the pan-cancer and cancer-specific, whereas some will be unique to a specific cancer type. The results from these analyses can be queried separately using the drop-down menus on each results page.

Cancer-specific analyses were performed for 18 cancer types where n > 15 cell lines were available.

How do I know if an association is statistically significant? What threshold should I use?

The ANOVA calculates a p-value to determine the significance of each drug interaction. A Benjamini-Hochberg multiple testing correction was applied to the resulting p-values. A p-value threshold of <10-3 and a false discovery rate threshold equal to 25% were used to call significant associations across all the performed analyses.

In addition to the statistical significance, we advise you to investigate the effect size for each association and to prioritise genomic features associated with a large effect on drug sensitivity. This can be visualised using the scatter plot function.

What genetic features are included in the ANOVA analysis for associations with drug response?

The ANOVA analysis currently correlates coding mutations and regions of recurrent copy number alteration with drug sensitivity data (IC50 values). These molecular alterations have been identified from the analysis of >11,000 patient tumours and subsequently mapped onto the cell lines for identifying molecular markers of drug sensitivity and resistance.

Why is drug sensitivity correlated with recurrent aberrant copy number alterations (CNA)? How do I know which genes are included in each segment?

We identified regions of focal recurrent copy number alteration (amplification and homozygous deletion) from patient tumours, and these regions were then used to correlate with drug sensitivity observed in cancer cell lines. The pan-cancer set of segments was derived from the entire collection of cancer specific segments. CNA may contain many genes and in many cases the cancer gene(s) operative within the segment is unclear. For this reason, we report the association between a CNA and a specific compound, and any known cancer genes included within a CNA. Some segments do not contain known cancer genes. The complete list of genes included in each CNA is available on the relevant webpages, or in the download section of the website.

How are the drug targets and pathway defined?

The target refers to the nominal therapeutic target(s) of a compound. In many, if not all, instances compounds have additional targets not listed here.

The target pathway/process has been manually curated based on our current understanding of cancer biology, therapeutic application and the biological processes in disease.

How have the cancer types been defined? What do the acronyms mean?

The cancer types have been defined using The Cancer Genome Atlas disease studies (https://tcga-data.nci.nih.gov/datareports/codeTablesReport.htm).

TCGA Label	Definition
ACC	Adrenocortical carcinoma
ALL	Acute lymphoblastic leukemia
BLCA	Bladder Urothelial Carcinoma
BRCA	Breast invasive carcinoma
CESC	Cervical squamous cell carcinoma and endocervical adenocarcinoma
CLL	Chronic Lymphocytic Leukemia
COREAD	Colon adenocarcinoma and Rectum adenocarcinoma
DLBC	Lymphoid Neoplasm Diffuse Large B-cell Lymphoma
ESCA	Esophageal carcinoma
GBM	Glioblastoma multiforme
HNSC	Head and Neck squamous cell carcinoma
KIRC	Kidney renal clear cell carcinoma
LAML	Acute Myeloid Leukemia
LCML	Chronic Myelogenous Leukemia
LGG	Brain Lower Grade Glioma
LIHC	Liver hepatocellular carcinoma
LUAD	Lung adenocarcinoma
LUSC	Lung squamous cell carcinoma
MB	Medulloblastoma
MESO	Mesothelioma
MM	Multiple Myeloma
NB	Neuroblastoma
OV	Ovarian serous cystadenocarcinoma
PAAD	Pancreatic adenocarcinoma
PRAD	Prostate adenocarcinoma
SCLC	Small Cell Lung Cancer
SKCM	Skin Cutaneous Melanoma
STAD	Stomach adenocarcinoma
THCA	Thyroid carcinoma
UCEC	Uterine Corpus Endometrial Carcinoma

How have the IC50 and AUC values been calculated? What are the RMSE values?

The IC50 and AUC are estimated using a non-linear mixed effect model fitting a sigmoid curve to the experimental dose response data. We are screening a large collection of cell lines with a diverse selection of anti-cancer drugs and consequently observe a wide-range of drug responses in our experimental data. The curve-fitting algorithm readily models sensitivity to a drug that falls within the range of experimental screening concentrations. In many instances however, a significant proportion of cell lines will be only partial responsive or resistant to a given drug within the range of experimental screening concentrations. For completeness we have reported these IC-values but they should be interpreted carefully because the confidence in the reported value is reduced. For certain downstream analyses it may be appropriate to restrict the IC50 value to the reported maximum screening concentration, or use an alternative output such as AUC.

The model error can be estimated by the root mean squared error (RMSE) which is indicative of how far the model approximation deviates from the measured observations. Where the RMSE is > 0.3 the fit is considered poor and the results have been filtered from further analysis.

The screening concentration for each compound can be found in the download page.

Why have some IC50/AUC values changed from previous releases of the GDSC dataset?

IC50 and AUC estimates for a given cell line and drug may differ slightly to those previously published on the GDSC website. This is because the estimation model has changed even though the underlying data may be the same.

In this data release dose response curves have been fitted using a multilevel fixed effects model. Previously a Bayesian MCMC model was used that utilized the data for a single cell line/drug experiment to estimate the response. The multilevel model estimates the response for a cell line/drug combination using the available response data across all cell lines and drugs as input. One consequence of this model is that dose response estimates for a single experiment may differ depending on the context of the overall dataset.

Are details for the dose-response curve fitting published?

For more details see:

Multilevel models improve precision and speed of IC50 estimates - Vis DJ et al., Pharmacogenomics, 2016.

http://www.ncbi.nlm.nih.gov/pubmed/27180993

What is the data sharing policy?

Wellcome is committed to data sharing wherever possible. This maximises the value and scientific impact of the data, and ensures transparency and equity in exploitation of the opportunities created. The Institute recognises the need for researchers to be appropriately credited for their scientific contribution and investment in data generation. It is therefore expected that all researchers both honour agreements in line with Fort Lauderdale's data sharing principles and appropriately acknowledge the contributions of others.

For more information about the Institute's data policies, please refer to our legal http://www.sanger.ac.uk/legal/index.html page, which has sections on data sharing and use.

Alternatively you can download the PDF version of the Data Sharing Policy.

Users have a non-exclusive, non-transferable right to use data files for internal proprietary research and educational purposes, including target, biomarker and drug discovery. Excluded from this licence are use of the data (in whole or any significant part) for resale either alone or in combination with additional data/product offerings, or for provision of commercial services.