A formal comparison of different methods for establishing cut points to distinguish positive and negative samples in immunoassays
Introduction
Biotechnology derived therapeutics may induce an unwanted immune response resulting in the formation of anti-drug antibodies (ADA). As a consequence of the development of ADA efficacy and safety of the therapeutic protein could be impaired. For example, binding or neutralizing antibodies may affect pharmacokinetics or functionality of the therapeutic protein or even induce autoimmunity when the ADA cross-react with endogenous counterparts. In addition, unwanted immune responses may lead to allergic reactions. As a result, drug induced immune responses to a therapeutic protein are a major concern and need to be assessed during drug development.
Consequently there is a need to develop appropriate assays for the detection and characterization of ADA. In 2007, the European Medicines Agency (EMA) published a guideline that describes the general strategy for the development and validation of assays for immunogenicity assessment of biotechnology derived therapeutic proteins [1]. A multi-tiered approach for the testing of patient samples is recommended. In the first instance a screening assay is used for rapid identification of positive samples while subsequently an additional confirmatory assay is used to confirm the results of the screening assay. As a third step, a functional assay for assessment of the neutralizing capacity of antibodies is recommended. Screening, confirmatory and functional assays for detection and characterization of ADA need to be validated [2], [3].
A critical step during assay development and validation is the definition of an appropriate cut-off that can be used to distinguish between positive and negative samples in the screening assay. This initial assay needs to be as sensitive as possible to maximize the detection of true positive samples and should be designed to avoid classifying positive samples as negative. A proportion of false positive samples is acceptable as they can be identified by the following confirmatory assay while costs and time urge to take few samples to this second stage. This approach ensures that the assays will detect as many patients who have indeed developed antibodies.
A valid statistical approach needs to be elaborated to define a reliable cut-off value used in screening and confirmatory assays [4]. For defining an appropriate cut point usually control samples obtained from healthy subjects or untreated patients are used. Such a pool of control samples is in most cases of heterogeneous composition, containing sub-populations consisting of true negative samples as well as true and false positive samples. The portion of each sub population has impact on the final cut-off value if one assumes that indeed all samples are truly negative. For example, a high content of true positives in the sample population due to specific pre-existing antibodies used for calculating the cut-off would result in a high number of false negative evaluation of samples. Consequently it is crucial to use statistical methods that deal with potential (false) positive samples appropriately when determining a cut point. Different strategies to detect and characterize ADA's have been discussed in [4], [5] but no formal evaluation of the methods has yet been undertaken.
In this paper we evaluate a variety of established and less established methods for cut point determination. We will introduce the methods in Section 2 before we compare them thoroughly via simulation (Section 3). We conclude with an in-depth discussion and some future directions.
Section snippets
Methods to determine cut point
In this section we will describe various methods for determining cut points. Many of the methods are informed by the discussions in [4], although some adjustments have been made to enable automated cut point determination in the simulations to follow. Most importantly no outlier removal is incorporated prior to applying the various methods as different criteria will result in different cut points. Furthermore, the simulated data studied later do not contain outliers and hence such outlier
Comparison of methods
In this section we focus on comparing the different methods to establish cut points introduced in Section 2. We will first illustrate the cut points obtained by the different methods and where they differ on a real data example and then evaluate the methods formally in an extensive simulation study. All analyses and simulations were performed in R [14] Version 2.10.1. The mixture approaches used the gamlss.mx package [15] while the package msm [16] was used for generating truncated normal
Discussion
In this paper seven methods for cut point estimation were formally compared in terms of their ability to identify positive and negative samples. Due to the overlap between the distributions of positive and negative samples, it is inherent to methods determining cut points that the performance to identify a certain category can be increased only at the cost of decreasing the ability to correctly identify the other category, a relationship clearly seen in the simulation results. As these assays
Conflict of interest
J.-P. Lawo, M.J. Wolfsegger, J. Singer, P. Allacher, F. Horling are employees of Baxter.
References (18)
- Committee for medicinal products for human use, Guideline on immunogenicity assessment of biotechnology derived...
- et al.
Recommendations for the design and optimization of immunoassays used in the detection of host antibodies against biotechnology products
J. Immunol. Methods
(2004) - et al.
Recommendations for the design, optimization, and qualification of cell-based assays used for the detection of neutralizing antibody responses elicited to biological therapeutics
J. Immunol. Methods
(2007) - et al.
Recommendations for the validation of immunoassays used for detection of host antibodies against biotechnology products
J. Pharma. Biomed. Anal.
(2008) - et al.
Recommendations on risk-based strategies for detection and characterization of antibodies against biotechnology products
J. Immunol. Methods
(2008) - et al.
An analysis of variance test for normality (complete samples)
Biometrika
(1965) - et al.
Graphical Methods for Data Analysis
(1983) Subchronic toxicity studies
Falsification of propensity models by statistical tests and the goodness-of-fit paradox
Philos. Math.
(2007)