(T1430-01-03) Bioanalytical Cross Validation: A More Statistically Based Approach to Establish Differences between Bioanalytical Methods and Laboratories

Purpose: While it is understood that cross validations between bioanalytical methods or between laboratories is required, industry practice has not been consistent. Often, QCs and selected study samples are tested once with each method or once at each laboratory. While industry working practice historically uses the ISR acceptance criteria, where at least two thirds of the tested study samples are within ±30% of the mean for large molecules, this practice is not consistent either. As part of the newly finalized ICH M10 Bioanalytical Method Validation Guidance, a cross validation is required when 1) data are obtained from different fully validated methods within a study, 2) data are obtained within a study from different laboratories with the same bioanalytical method or 3) when data are obtained from different fully validated methods across studies that are going to be combined or compared to support special dosing regimens, or regulatory decisions regarding safety, efficacy, and labelling. While the ICH M10 does not set acceptance criteria for the assessment, it suggests that bias can be assessed using statistically based approaches. There have been instances in which cross validations have met acceptance criteria, however a systemic bias between methods and/or labs have led to an apparent PK difference, hence raising the question whether a correction factor needs to be applied due to directional systemic bias between methods and/or labs. In addition, an understanding of trend in analysis over time is needed when data from multiple studies may be compiled for Model Informed Drug Development extrapolation approaches.

Methods: Here we are presenting a recent cross validation between two clinical PK methods, which was performed following a different strategy, utilizing a more statistically based approach comprising of mixed model investigation of variance components, Bland-Altman analysis of agreement, and Deming regression analysis. In this work, incurred study samples were selected to represent the overall sample population in the clinical study to the point of testing. The samples were chosen from different study cohorts and subjects, to cover the study timeline, and to have their concentrations spanning for the PK profile (C_max, elimination and trough). Those samples along with the QCs were tested in three separate runs by originator and comparator, with each run performed on a different day. This more comprehensive analysis was used to better characterize the within-run and between-run variability of the methods and uncover any systemic bias between the methods.

Results: The statistical analysis revealed that: 1) the within-method run-to-run variability was noticeably higher for the originator method, but the level of variability was negligible relative to the sample-to-sample variability within-method; 2) the comparator method read samples 4% higher than the originator on average, with the 95% limits of agreement between 0.93 and 1.17; 3) in Deming regression, the 95% confidence interval of the intercept contains 0, and that of the slope contains 1. These indicated that the two PK methods are suitable to be used interchangeably in supporting study sample analysis.

Conclusion: In summary, this more comprehensive cross validation was conducted in alignment with the new ICH M10 Guidance and provided a more thorough understanding of performance comparison between originator and comparator, benefiting from consideration of run-to-run variance and the more statistically sound data analysis. The acceptance criteria were not pre-defined, but with the statistically based approaches described above, a convincing conclusion can be drawn that the two methods being cross validated can be used interchangeably. Data was also shared with clinical pharmacologists/modeler, who further confirmed that the differences identified in the statistical analysis have no impact on the PK endpoints. We therefore propose that the new approach can serve to better understand between assay and laboratory variability, and to investigate the potential for systemic and directional bias, which has the potential to confound data analysis when two methods or laboratories are used within a study.