Cancer in silico drug discovery: a systems biology tool for identifying candidate drugs to target specific molecular tumor subtypes

Abstract

Large-scale cancer datasets such as The Cancer Genome Atlas (TCGA) allow researchers to profile tumors based on a wide range of clinical and molecular characteristics. Subsequently, TCGA-derived gene expression profiles can be analyzed with the Connectivity Map (CMap) to find candidate drugs to target tumors with specific clinical phenotypes or molecular characteristics. This represents a powerful computational approach for candidate drug identification, but due to the complexity of TCGA and technology differences between CMap and TCGA experiments, such analyses are challenging to conduct and reproduce. We present Cancer in silico Drug Discovery (CiDD; scheet.org/software), a computational drug discovery platform that addresses these challenges. CiDD integrates data from TCGA, CMap, and Cancer Cell Line Encyclopedia (CCLE) to perform computational drug discovery experiments, generating hypotheses for the following three general problems: (i) determining whether specific clinical phenotypes or molecular characteristics are associated with unique gene expression signatures; (ii) finding candidate drugs to repress these expression signatures; and (iii) identifying cell lines that resemble the tumors being studied for subsequent in vitro experiments. The primary input to CiDD is a clinical or molecular characteristic. The output is a biologically annotated list of candidate drugs and a list of cell lines for in vitro experimentation. We applied CiDD to identify candidate drugs to treat colorectal cancers harboring mutations in BRAF. CiDD identified EGFR and proteasome inhibitors, while proposing five cell lines for in vitro testing. CiDD facilitates phenotype-driven, systematic drug discovery based on clinical and molecular data from TCGA.

Figure 1

A CiDD analysis produces a list of candidate drugs to treat tumors with the molecular or clinicopathological phenotype of interest and a list of cell lines that are representative of the phenotype of interest.

Figure 2

A CiDD workflow shows the 5 main steps of an analysis with their data set dependencies. Input to this workflow includes point mutations (such as BRAF V600E) or other molecular and clinical phenotypes of interest paired with a cancer type (e.g., CRC). The primary output includes a candidate drug list that has been annotated with drug databases and a list of cell lines for subsequent experimentation.

Figure 3

CiDD-generated heat map and clustering of BRAF V600E mutated CRCs based on TCGA Illumina GA RNA sequencing data. Differentially expressed genes comparing BRAF V600E and BRAF wildtype samples were identified using the Limma package in R and required to have a Benjamini Hochberg adjusted p-value <= 0.05 and a minimum log fold change >= 2. Hierarchical clustering of the samples and genes were performed using hclust with a “pearson” distance measure in R. The BRAF V600E gene expression signature is represented with the vertical colored bar on the right side of the figure, where red represents down-regulated genes and blue up-regulated genes. BRAF V600E mutant samples all reside within 2 sample clusters of the heatmap, which suggests that the BRAF V600E signature captures the gene expression response of BRAF V600E mutations.