Research paper
A statistical approach to determining responses to individual peptides from pooled-peptide ELISpot data

https://doi.org/10.1016/j.jim.2016.05.006Get rights and content

Highlights

  • ELISpot assays of peptide pools enable the estimation of responses to individual peptides.

  • There was good agreement between estimates and confirmatory assays in an application to HIV-1 immunogenicity.

  • The estimates can inform substantial reductions in the number confirmatory assays.

  • We provide an easy-to-use free online tool for obtaining the individual-peptide estimates.

  • The method enables more efficient use of information from pooled ELISpot assay data.

Abstract

To investigate in detail the effect of infection or vaccination on the human immune system, ELISpot assays are used to simultaneously test the immune response to a large number of peptides of interest. Scientists commonly use “peptide pools”, where, instead of an individual peptide, a test well contains a group of peptides. Since the response from a well may be due to any or many of the peptides in the pool, pooled assays usually need to be followed by confirmatory assays of a number of individual peptides. We present a statistical method that enables estimation of individual peptide responses from pool responses using the Expectation Maximization (EM) algorithm for “incomplete data”. We demonstrate the accuracy and precision of these estimates in simulation studies of ELISpot plates with 90 pools of 6 or 7 peptides arranged in three dimensions and three Mock wells for the estimation of background. In analysis of real pooled data from 6 subjects in a HIV-1 vaccine trial, where 199 peptides were arranged in 80 pools if size 9 or 10, our estimates were in very good agreement with the results from individual-peptide confirmatory assays. Compared to the classical approach, we could identify almost all the same peptides with high or moderate response, with less than half the number of confirmatory tests. Our method facilitates efficient use of the information available in pooled ELISpot data to avoid or reduce the need for confirmatory testing. We provide an easy-to-use free online application for implementing the method, where on uploading two spreadsheets with the pool design and pool responses, the user obtains the estimates of the individual peptide responses.

Introduction

The ELISpot assay is a commonly used tool for studying immune responses to specific antigens by counting the number of single-cell responses in a sample of peripheral blood mononuclear cells (PBMC). The PBMC together with the studied antigen(s) are placed in wells on the assay plate and the responding cells form spots on the surface of the well. The total number of spots in a well represents the number of cells responding to the antigen (Anthony and Lehmann, 2003) and thus provides a direct estimate of the magnitude of the immune response to the antigen. Standard ELISpot plates consist of 96 wells, so that a single assay can generate data on the immune response to a large variety of antigens. When there are many potential antigens and only a small fraction of these are expected to elicit a response, investigators typically use peptide pools to improve the efficiency of the assay. Fig. 1 provides a simple illustration of a two-dimensional “matrix” pool design where 81 peptides of interest are arranged in 18 pools of 9 peptides each, with each peptide being allocated to two pools. The responses from the pools will typically not enable unique identification of peptides (as in Fig. 1(a)) and the assay proceeds with a confirmation step where some or all of the peptides in responding pools are tested individually. In the example in Fig. 1(b), peptides 31, 33, 49 and 51 would need to be tested individually so that using a pooled design reduces the testing to 22 wells (18 + 4) instead of 81, a significant reduction in effort and time. Higher-dimensional designs can provide better discrimination but even with a four-dimensional matrix, a confirmation step will usually be required.

Where there are no a priori peptides of interest, peptide pools are used by laboratory scientists to help separate responding peptides from nonresponding peptides without estimating the strength of the individual responses, since the observed data are the accumulated spot counts over several potentially responding peptides. In the first stage of the experiment, each well contains a pool (i.e. group) of peptides, so that if there is no response from the well, then all peptides in the pool can be eliminated from further investigation. Examination of the remaining pools rarely enables the elimination of all but a few peptides as in the simple examples above, so it is assumed that the responses observed might be due to any of the peptides and these are assayed individually at the second stage. Such an approach was used in a study of the cross-reaction of human influenza and avian flu (Lee et al., 2008). Clearly the efficiency of this method relies on the pool sizes and the combinations of peptides in the various pools (Roederer and Koup, 2003). The focus of our work is to use the pooled assay responses to estimate the strength of the response to each individual peptide. These estimates can be used to characterise the immune response or to help inform more targeted testing of individual peptides at the second stage. This can result in significant gains in efficiency, especially in applications where only a small proportion of peptides stimulate a response.

In this paper, we show that the use of peptide pool responses is not limited to identifying a subset of peptides for confirmatory testing, but these pool responses can be used to achieve estimates of the strength of a subject's response to the unique individual peptides in the pools. Since the data available are the combined responses to the pool of peptides in each well, but interest is focused on the response to an individual peptide, we can express this as an incomplete data problem (Dempster et al., 1977). The estimates of interest can then be obtained using the Expectation Maximization (EM) algorithm, which under some reasonable assumptions reduces to the solution of simple matrix equations. In Section 2, we describe the data that motivated this work, present our statistical method in detail and describe the simulation set-up we used to investigate its performance. In Section 3 we present the results of the simulation study and illustrate the method with an application to IFN-γ ELISpot assays of the immune responses of 6 participants in a HIV-1 vaccine trial for whom 199 peptides were assayed in 80 pools of size 9 or 10. We also compare our results to the individual-peptide confirmatory assays conducted for the same subjects. In Section 4 we discuss the method and outline the potential for further work in this area.

Section snippets

Materials and methods

This work was motivated by data from IFN-γ ELISpot assays conducted during a double-blind, randomised block, phase I HIV-1 vaccine trial (Borthwick et al., 2014) where three candidate vaccines (pSG2.HIVconsv DNA, ChAdV63. HIVconsv and MVA.HIVconsv) were administered to healthy, uninfected, low-risk adults. After a small safety arm, three groups of 10 volunteers were randomised to heterologous vaccine regimens or placebo in a ratio 8:2. For the assessment of vaccine immunogenicity, investigators

Statistical methods

The measure of response in an ELISpot assay, spot-forming-units (SFU), represents the number of cells that release the cytokine in response to the stimulating peptide(s). We model this underlying response of an individual to a specific peptide i,(i = 1…n), as a Poisson random variable, with intensity parameter βi which we refer to as the “peptide effect”. Let bij denote the (unobserved) response to peptide i in pool j. Then assuming independence of the individual responses within the same pool,

Simulation results

Table 1 displays the results of our simulations. For each of the β values drawn from the Gamma distributions representing low and moderate intensities, we summarise the results for data simulated from a Poisson process with and without overdispersion. The average estimates of the underlying intensities from the 200 simulations are presented with the empirical 95% confidence intervals, and the average bias and percentage bias are also reported. The results indicate very good accuracy for all

Discussion

We have shown that it is possible, in realistic situations, not only to separate potentially responding peptides from non-responding peptides using peptide pools, but to achieve reliable estimates of the strength of each response. We illustrated how the unobserved responses of individual peptides in peptide pools can be expressed as an incomplete data problem. We have derived the equations for estimating these individual effects using the EM algorithm and shown how these require just simple

Acknowledgements

This work was supported by the European and Developing Countries Clinical Trials Partnership (EDCTP), Grants CT.2006.33111.002 and SP.2011.41304.022, with co-funding from the Swedish International Development Cooperation Agency (SIDA). The work was also jointly funded by the UK Medical Research Council and the UK Department for International Development (DFID) under the MRC/DFID Concordat agreements (G0701669 and G1001757).

References (22)

  • A. Fiore-Gartland et al.

    Pooled-peptide epitope mapping strategies are efficient and highly sensitive: an evaluation of methods for identifying human T cell epitope specificities in large-scale HIV vaccine efficacy trials

    PLoS One

    (2016)
  • View full text