The conceptual framework and the high potential of in vitro
drug activity-based GEMs for predicting patient response to chemotherapy has been laid out in previous studies.(6
) In this study we demonstrated two major breakthroughs in GEM-based prediction for patients’ chemotherapeutic responses: 1) concordant prediction performance of unaltered in vitro
-based GEMs on geographically and ethnically diverse cancer patient cohorts in three different cancer types, and 2) simultaneous use of multiple, parallel historical patient data sets for efficient GEM assessment. Together, we believe these provided a highly encouraging possibility in applying such in vitro
-based GEMs to guide patients with more effective chemotherapies.
Some recent GEMs initially reported with their superior prediction performance failed to perform well by an independent group’s validation.(11
) This may be due to the so-called selection bias
when such multi-gene predictors were trained both with modeling and applying microarray datasets of numerous candidate genes. Our current study has avoided this pitfall by independently testing multiple geographically- and ethnically-diverse patient sets and by consistently predicting both clinical tumor response and patient survival outcome of 477 patients across three different cancer types. Therefore, we believe the prediction performance of the in vitro
-based GEMs here will be highly likely realized in clinical practice even though the statistical power of some of these GEMs may appear to be a bit lower than those reported in other recent studies applied only to a single patient set at a time.(17
We found that the GEM score was independent of conventional clinical parameters and was always a significant predictor of tumor response to pharmacotherapy even when all other clinical variables were considered together, suggesting that the GEMs are not simply surrogates of standard tumor characteristic variables. Also, GEMs generated by various different training sets that contained distinct stages of bladder cancer successfully stratified patient survival with only a slight decrease in statistical significance whereas the accuracy to distinguish the ability to responders from non-responders was maintained (Supplementary Figure S2
). While the assumption of independence of relevant agents was necessary due to our GEM derivation from the single-agent drug activity data of NCI-60, our GEM scores correlated well with clinical responses and survival of diverse patient cohorts of the three cancer types treated with multi-agent pharmacotherapy in this study. This demonstrated that even though our combination GEMs could not capture synergistic drug effects, the main efficacy of combination regimens appeared to be reflected by simple additive effects of single drug GEMs. It is conceivable that future in vitro
work on doublet or higher-order drug combinations may help in modifying the algorithm to effectively incorporate drug synergies.(18
We, however, note that each GEM’s reported sensitivity, specificity, NPV, PPV values, or Kaplan-Meier curves here (derived from its ROC curve) showed its optimal performance on the applying particular data set. This does not yet represent validation of each drug GEM’s pre-set threshold to call a patient case + (responder) or − (non-responder), but rather shows proof-of-a-concept, illustrating that such GEM scores are informative in stratifying patients’ responses to the agent. In order to derive an exact cutoff criterion on a specific GEM assay for clinical use, a standard diagnosis assay platform and procedure should be developed for routine clinical practice, from which a fixed cutoff value can be defined for a target patient population, which, we believe, is quite feasible in the near future.
Generating GEMs using in vitro-based approaches does also have some theoretical and practical limitations. For example, it cannot be used to develop GEMs for agents which do not have any effect on cell lines in vitro. While this requirement is strictly embedded in the design of the approach, the composition of the cell panel may be tailored to the expected cellular and molecular target of the tested agent. For example, while using the NCI-60 panel may not result in an effective generation of a drug response profile for an anti-angiogenic agent that targets endothelial cells, carrying out such an experiment on an endothelial cell panel may provide the necessary data which can be used for GEM development. Similarly, for agents that target the immune system, panels composed of the appropriate cells may permit GEM development. As we have shown above, the critical requirement for the cell panel is to provide effective dose response information for the agent in question rather than be required to be composed of the same histological tumor types as the human tumors whose response to therapy is assessed. Hence, using an endothelial cell panel for an anti-angiogenic agent and generating GEMs for the use of such an agent in bladder cancer, for example, seems to be justified.