BiomarkerDiscoveryToolkit

An R package for biomarker discovery.

Background

A biomarker is defined as "a measurable indicator of some biological state or condition" (Wikipedia). In medicine, important application areas for biomarkers include disease diagnosis, prognosis of disease progression, prediction of therapy response, and pharmacodynamic assessments. In drug discovery and development, biomarkers that can predict drug efficacy in patients are immensely valuable, and may even be required for patient selection for clinical trials. Earlier in the drug discovery process, identifying biomarkers for drug response in vitro (cell lines) or in vivo (animal models) can help to elucidate mechanisms of action or generate hypotheses for patient biomarkers.

Nomenclature

In a bioinformatic context, a biomarker is a (molecular) feature that can be used to predict a biomedical property of interest (response) - either by itself or in combination with other biomarkers as part of a biomarker signature. Biomarker discovery thus means evaluating available features for associations with the response; either one by one (univariate analysis) or in combinations (multivariate analysis).

Examples of responses include:

• Genetic dependency from CRISPR/RNAi experiments
• Drug sensitivity from in vitro or in vivo experiments
• Patient response in a clinical trial

Examples of features include:

• Cell line or tumour lineage/tissue of origin
• Omics data:
  • Gene expression
  • Mutations
  • Copy numbers
  • Protein expression

Using the package

To use the in-development package, clone the Git repository to a local directory. Then run this in an R session (requires package devtools):

wd <- getwd() # save current working directory
setwd("/path/to/BiomarkerDiscoveryToolkit")
devtools::load_all()
setwd(wd) # restore current working directory

Current functionality

Drug dose responses (curve_fit.R)

• Dose-response curve fitting (using the dr4pl package)
• AUC calculation
• Visualization of dose-response curves (incl. IC50 and AUC annotations)

Univariate analyses (univariate.R)

1. Correlation-based analysis (correlations.R)

• For continuous response and continuous features
• Convenient calculation of correlation coefficients for many features
• Generation of null distributions by permutation
• Significance estimates
• Visualization of distributions and top hits (scatter plots)

2. Categorical analysis

• For categorical response and continuous features
• Compare feature values grouped by response using statistical tests
• Visualization of top hits (box plots)

3. Analysis of mutation data

• For continuous response and binary features
• Compare response values grouped by features using statistical tests
• Visualization of top hits (box plots)

4. Survival analysis

• Responses are surival times or similar (right-censored) time-to-event data
• Assess features using Cox proportional-hazards regression models
• Visualization of top hits (Kaplan-Meier curves)

Enrichment analyses (enrichment.R)

• Gene set enrichment analyses (GSEA; using the clusterProfiler package and others)
• Gene Ontology, Reactome and MSigDB gene sets
• Visualization using dot plots (with indication of directionality)

Multivariate analysis

• Random forest models (using the randomForest package)
• Trained using cross-validation and backwards feature elimination (powered by the caret package)
• Visualization of prediction performance and feature importance

Planned functionality

• R package to be submitted to CRAN or Bioconductor