Controlling the Reproducibility of AC50 Estimation during Compound Profiling through Bayesian β-Expectation Tolerance Intervals

During drug discovery, compounds/biologics are screened against biological targets of interest to find drug candidates with the most desirable activity profile. The compounds are tested at multiple concentrations to understand the dose-response relationship, often summarized as AC50 values and used directly in ranking compounds. Differences between compound repeats are inevitable because of experimental noise and/or systematic error; however, it is often desired to detect the latter when it occurs. To address this, the β-expectation tolerance interval is proposed in this article. Besides the classical acceptance criteria on assay performance, based on control compounds (e.g., quality control samples), this metric permits us to compare new estimates against historical estimates of the same study compound. It provides a measure that detects whether observed differences are likely due to systematic error. The challenge here is that limited information is available to build such compound-specific acceptance limits. To this end, we propose the use of Bayesian β-expectation tolerance intervals to validate agreement between replicate potency estimates for individual study compounds. This approach allows the variability of the compound-testing process to be estimated from reference compounds within the assay and used as prior knowledge in the computation of compound-specific intervals as from the first repeat of the compound and then continuously updated as more information is acquired with subsequent repeats. A repeat is then flagged when it is not within limits. Unlike a fixed threshold such as 0.5log, which is often used in practice, this approach identifies unexpected deviations on each compound repeat given the observed variability of the assay.