Statistical Challenges and Considerations in Algorithm Development for Predictive Markers
Carrie Aldrich, Ventana Medical Systems, Inc.  Ping Jiang, Halozyme Therapeutics  Sihem Khelifa, Ventana Medical Systems, Inc.  *Jie Pu, Ventana Medical Systems, Inc.  James Ranger-Moore, Ventana Medical Systems, Inc.  Wilson Wu, Halozyme Therapeutics  Junming Zhu, Ventana Medical Systems, Inc. 

Keywords: Algorithm Development, Predictive Marker, Companion Diagnostic Test

Predictive biomarkers that can identify which patients are likely to benefit more from a new drug than a control intervention play an important role in precision medicine. One key step for co-development of a new drug and a companion diagnostic test is defining an appropriate scoring algorithm that can best capture the essentiality of the biomarker. Given the functional role of predictive biomarkers, one approach for evaluating candidate scoring methods is to test the interaction between treatment assignment and marker value. For markers with continuous measurement, a threshold (cut-off) that can be used to dichotomize the measure of the marker is also needed in order to inform treatment decisions. The presentation describes statistical challenges and considerations using a real-case example of developing the scoring algorithm for a recently developed companion diagnostic test. Ideally, a scoring method evaluation should be conducted in a randomized control trial, where each patient’s biomarker is assessed prospectively. However, practical limitations often necessitate retrospective algorithm development work on archived specimens collected from clinical studies that were generally underpowered either for testing any interaction effect or for evaluating a range of candidate cut-offs. Furthermore, because of the retrospective nature of such effort, not all patients may have specimens available/adequate for staining. This will lead to further sample size reduction and potential bias issues among the remaining samples. Proposed solutions and other statistical considerations such as influencing factors for cut-off selection (e.g., marker distribution, marker prevalence, and study endpoints) are discussed. Lastly, by describing the step-by-step evolution of the pathological assessments that have been attempted during the whole process, the presentation also demonstrates how biostatisticians can not only contribute to the statistical evaluation of candidate scoring algorithm(s), but also can mathematically derive/construct new ways to measure a predictive marker through close collaboration with subject-matter experts.