GDSCTools for mining pharmacogenomic interactions in cancer

1 Introduction

Cancers occur due to genetic alterations in cells accumulated through the lifespan of an individual. Cancers are genetically heterogeneous and as a consequence patients with similar diagnoses may vary in their response to the same therapy. The path towards precision cancer medicine requires the identification of specific biomarkers, such as genetic alterations, allowing effective patient selection strategies for therapy. Large-scale pharmacological screens such as the Genomics of Drug Sensitivity in Cancer (GDSC) (Garnett et al., 2012) and Cancer Cell Line Encyclopaedia projects (Barretina et al., 2012) have been used to identify potential new treatments and to explore biomarkers of drug sensitivity in cancer cells. In particular, the GDSC project releases database resources periodically (www.cancerRxgene.org) (Yang et al., 2013). A recent installment of this resource (version 17) includes cancer-driven alterations identified in 11 289 tumors from 29 tissues across 1001 molecularly annotated human cancer cell lines, and cell line sensitivity data for 265 anti-cancer compounds. A systematic identification of clinically-relevant markers of drug response uncovered numerous alterations that sensitize to anti-cancer drugs (Iorio et al., 2016).

Here, we present GDSCTools, a Python library that allows users to perform pharmacogenomic analyses as those presented in (Iorio et al., 2016). Our software complements an existing tool (Smirnov et al., 2016) by giving access to the full GDSC dataset and providing a powerful platform for statistical analyses and data mining through visualization tools.

2 Data formats and data wrangling tools

The GDSC database provides large-scale genomics and drug sensitivity datasets. The drug sensitivity dataset contains dose-response curves (e.g. cell viability for 5–9 drug concentrations) which can be used to derive drug sensitivity indicators (Garnett et al., 2012; Vis et al., 2016), such as the half-maximal inhibitory concentration (${\text{IC}}_{50}$) or the area under the curve (AUC) (Fig. 1A). In GDSCTools, logged ${\text{IC}}_{50}$ indicators are encoded as a N_c × N_d matrix, where N_c is the number of cell lines labeled with their COSMIC identifier (http://cancer.sanger.ac.uk/cosmic) and N_d is the number of drugs. For a given drug, we denote with Y_d the vector of logged ${\text{IC}}_{50}$s across the N_c cell lines. The genomic feature dataset $\mathbf{X}$ is also encoded as a N_c × N_f matrix, where N_f is the number of genomic features. In addition to a subset of the data files available in GDSCTools (version 17 only), users can also retrieve additional datasets online (e.g. methylation data, copy number variants, etc.) Database-like queries can be used to extract and use specific features (e.g. only gene amplifications or deletions). These database-like functionalities are part of the OmniBEM builder (Supplementary Material).

3 Data analysis tools

Using GDSCTools, genomic features can be investigated as possible predictors of differential drug sensitivity across screened cell lines. The statistical interaction Y_d ∼ X between drug response and genomic features can be tested within a sample population of cell lines from the same cancer type with a t-test. However, to account for possible confounding factors (including the tissue of origin, when performing pan-cancer analyses) a more versatile analysis of variance (ANOVA) is implemented. In this model, the variability observed in Y_d is first explained using the tissue covariate, subsequently using additional factors (e.g. microsatellite instability denoted by MSI), and finally by each of the genomic features in X (one model per feature). This can be mathematically expressed as Y_d ∼ C(tissue) + C(MSI) + … + feature, where the $C()$ operator indicates a categorical variable. An ANOVA test is performed for each combination of drug and genomic feature (Fig. 1B). Outcomes of this large number of tests (N_d × N_f) are corrected for multiple hypothesis testing using Bonferroni or Benjamini-Hochberg corrections. To account for P-value inflations due to differences in sample sizes, the effect sizes of the tested statistical interactions (computed with the Cohen and Glass models) are also included (Fig. 1C).

Unlike the ANOVA analysis that is performed on a one drug/one feature basis, linear regression models assume that drug response can be expressed as a linear combination of the status of a set of genomic features. GDSCTools includes three linear regression methods: (i) Ridge, based on an L2 penalty term, which limits the size of the coefficient vector; (ii) Lasso, based on an L1 penalty term, which imposes sparsity among the coefficients (i.e. makes the fitted model more interpretable) and (iii) Elastic Net, a compromise between Ridge and Lasso techniques with a mix penalty between L1 and L2 norms (see Supplementary Material for details). These three methods require the optimization of an α parameter (importance of L1 and L2 penalties) and a ρ parameter (mix ratio between L1 and L2 penalties; ElasticNet case only). This is performed via a cross validation to avoid over-fitting. The best model is determined using as objective function the Pearson correlation between predicted and actual drug responses on the training set. The final regressor weights are outputted as shown in Figure 1D. Significance of the final selected models is computed against.

4 Implementation and future directions

GDSCTools is available on http://github.com/CancerRxGene/gdsctools. It is fully documented on http://gdsctools.readthedocs.io. Pre-compiled versions of the library are available on https://bioconda.github.io/. GDSCTools can be used via standalone applications to analyse a user defined set of drugs (and genomic features) and assemble the results in an HTML report. We also provide solutions based on the Snakemake framework (Köster and Rahmann, 2012) to parallelize the analysis on distributed cluster farm architectures such as LSF or SLURM (Supplementary Material). Besides analysis of pharmacogenomic datasets, GDSCTools can provide the framework for discovering new biomarkers through integration/mining of novel and heterogeneous datasets, including pharmacological, RNA interference or increasingly available genetic screens (e.g. CRISPR), alternative drug response metrics (e.g. AUC) or implementing new analytical tools. The augmentation of genomic features with information obtained from online web services (Cokelaer et al., 2013) like pathway enrichment [e.g. via OmniPath (Turei et al., 2016)] will further extend functionality and usefulness of GDSCTools.

Conflict of Interest: none declared.

Supplementary Material