Use of Companion Diagnostics (CoDx) in Clinical Trials

Organizers: Sudhakar Rao (Janssen)
Chair: Sudhakar Rao (Janssen), Xiang Li (Janssen)
Vice Chair: Fabio Rigat (Janssen)

Songbai Wang (Janssen)
Xiang Li (Janssen)
Fabio Rigat (Janssen)

Title: A bivariate Bayesian framework for simultaneous evaluation of two candidate companion diagnostic assays in a new drug clinical trial
Speakers: Songbai Wang (Janssen)

Companion diagnostic tests play an important role in precision medicine. With the advancement of new technologies, multiple companion diagnostic tests can be rapidly developed in multiple platforms and use different samples to select patients for new treatments. Analytically validated assays must be clinically evaluated before they can be implemented in patient management. The status quo design for validating candidate assays is to employ one candidate assay to select patients for new drug clinical trial and then further evaluate the 2nd candidate assay in a bridging study. We propose a new enrollment strategy that employs two assays to select patients. We then develop a bivariate Bayesian approach that enables the totality of data to be used in evaluating whether these assays can be used independently or in a composite procedure in selecting right patients for new treatment. We demonstrate through simulations that when proper priors are available, the Bayesian approach is superior to classical methods in terms of statistical power.  At the end of this talk we will briefly discuss multiplicity issue when two candidate assays are simultaneously evaluated. This is a joint work with Richard M. Simon (R Simon Consulting).

Title: Calibration of the Use of Multiple Companion Diagnostics with Machine Learning based on Regularization
Speakers: Xiang Li (Janssen)

With increasing number of molecularly targeted therapy, the use of biomarkers and the development of corresponding companion diagnostic (CDx) are becoming more and more important in selection of patients for targeted drugs. Multiple companion diagnostics that identify the same or similar biomarkers are also used in a single clinical trial for patient enrolment, given that (1) if any of the CDx are positive; (2) acknowledging the limitations of each CDx; (3) patients could still be enrolled if one of the CDx is not available. However, clinical trial is usually only powered for the entire study population with any positive CDx, which may not reach desired power when analyzing each individual CDx. Furthermore, there is potential for dilution of apparent effectiveness of the drug if one CDx performs poorly.
We propose to obtain the estimate of treatment effect for every single CDx via borrowing information from other CDx, especially when there is a lack of other source of information that could be leveraged to increase the power. The commonly used approach for borrowing information may not perform well when there is heterogeneity between different CDx. To alleviate the problem, we develop a machine learning approach based on regularization with a tuning parameter to determine the degree of similarity between different CDx. The optimal tuning parameter is selected by minimizing the approximated mean squared errors via bootstrap sampling method, and the corresponding estimate of treatment effect is reported as the final estimate. One extension of our proposed approach is to use the selected optimal tuning parameter in an empirical Bayes fashion with commensurate prior instead of assuming a prior distribution in a full Bayesian approach. Comprehensive simulations were conducted to assess the model performance. The result shows increased power using both the proposed approach and the commensurate prior approach. Furthermore, the proposed approach had a better mean squared error than the Bayesian approach with commensurate prior and approach the estimates from the conventional approach without borrowing (e.g. maximum likelihood estimate) at a much faster rate when there is greater heterogenicity. This is a joint work with Hong Tian (Janssen) and Kevin Liu (Janssen).

Title: Is it worth it? Predicting success of clinical programs with and without biomarkers
Speaker: Fabio Rigat (Janssen)

Much is being written about the role of biomarkers in clinical development, from the need to rely on robust assays to the potential for updating clinical response criteria based on well understood markers. Here we focus in on the evaluation of the probability of success of clinical programs – defined as alternative sequences of clinical trials – using pre-specified outputs of exploratory and confirmatory biomarker data analyses for decision making using standard design and success criteria. Specifically, we illustrate common circumstances where investment in early confirmation of exploratory results via seamless biomarker cohorts determines a substantial increase in the probability of success. We relate these illustrative results to the underlying assay validation issues and published clinical trial data.