Improved drug-screening tests of candidate anti-cancer drugs in patient-derived xenografts through use of numerous measures of tumor growth determined in multiple independent laboratories

Background: Researchers screen candidate anti-cancer drugs for their ability to inhibit tumor growth in patient-derived xenografts (PDXs). Typically, a single laboratory will use a single measure of tumor growth. Purpose: An effective drug-screening test as one that correctly identifies whether a drug treatment inhibits or does not inhibit tumor growth. We document improvements in the experimental design and statistical analysis of drug-screening tests based on the criteria of sensitivity and specificity. Methods: We analyzed two published datasets. The response of each PDX model was known in advance. This information provided for statistical ground-truth classification. One dataset analyzed growth inhibition in the presence of one specific drug treatment for two PDX tumor models for numerous labs. A second dataset reported tumor growth of many PDX models in the presence of many drugs. A PDX model for which the treatment showed no tumor growth inhibition is referred to as Progressive Disease (PD). A PDX model for which the treatment showed complete tumor growth inhibition is referred to as Completely Responsive (CR). We created and analyzed four drug-screening tests, based on p-values for either a single-measure and single-lab, or p-values from meta-analysis and multiple-test correction. The outcome of each screening test was that either the drug treatment was effective or it was not. For both datasets, we computed median sensitivities and specificities by applying bootstrap resampling, and specification of a significance level. Results: Our results showed that drug screening tests utilizing p-values from meta-analysis of numerous labs, or multiple test correction, produced median sensitivities and specificities that were always at least as high as those for the Single-Measure, Single-Lab test. This result was true for all significance levels. The 95% confidence intervals were usually greater in length for the Single-Measure, Single-Lab screening test.