Bayesian modeling for analyzing heterogeneous response in preclinical mouse tumor models

In anticancer research, tumor growth measured in mouse models is important for assessing treatment efficacy for a treatment to progress to human clinical trials. Statistical analysis of time-to-event tumor volume data is complex because of heterogeneity in response and welfare-related data loss. Traditional statistical methods of testing the mean difference between groups are not robust because they assume common responses across a population. Heterogeneity in response is also seen in the clinic, and consequently, the assessment of the treatment considers the diversity through classification of the individual's response using the RECIST (Response Evaluation Criteria in Solid Tumors). To provide a comparable and translatable assessment of in vivo tumor response, we developed a statistical method called INSPECT (IN vivo reSPonsE Classification of Tumors) for analyzing heterogeneous responses through Bayesian modeling. This method can classify individual tumor behaviors into the categories of nonresponder, modest responder, stable responder, and regressing responder. Using both published and simulated data, we show that INSPECT methodology is more accurate and sensitive than existing methods with respect to balancing false-negative and false-positive rates. A case study demonstrates the value of INSPECT in drug projects for supporting the translation of drug efficacy from the preclinical phase into clinical trials. We also provide a package, "INSPECTumours," that launches a web interface to enable users to conduct the analysis and generate reports.