Predicting immunological cross-reactivity: from genotype to antigenic phenotype

In general individual pathogen species are genetically diverse. This genetic variation often results in the immune system of infected hosts 'seeing' the infectious agent differently depending on the exact genotype of the strain of pathogen responsible for infection. The ability of different strains of pathogen to induce slightly different immune responses is referred to as antigenic variation. Antigenic variation is particularly prevalent among RNA viruses because these viruses lack proof-reading mechanisms and thereby incur unusually high rates of mutation when they replicate. The ability to quantitatively determine the extent to which two strains of pathogen are antigenically different is important for at least two reasons: 1) Only by quantifying antigenic differences can we start to understand how antigenic characteristics evolve; and 2) Many of the vaccines that are used to control viral pathogens have to be selected carefully so they match the strain of pathogen likely to cause infection. Studying antigenic variation in small viruses is sensible because almost the entire immune response is directed towards the outside of the virus coat, which by the standards of pathogens are reasonably simple structures encoded by just a small number of genes. If we can understand how variation in these genes results in antigenic variation it might be possible to predict antigenic similarities of these viruses from their genetic variation alone, which would greatly simplify and expedite studies of antigenic variation. In this proposal, we will develop computer programs that can predict antigenic characteristics from genetic variation in the genes coding for viral coat proteins. To do this, we will develop models of how an immune system might respond to different viral strains. Once we have a great deal of simulated data we can use them to 'train' computer programs to predict the antigenic characteristics of new strains.

An important goal of both epidemiological and viral evolutionary studies is to predict the antigenic similarity of different viral genotypes. The ability to easily determine antigenic similarity would greatly facilitate the empirical study of the evolution of antigenic novelty, informing us about when and how fast we can expect viruses to exhibit antigenic change. In this proposal I lay out a research program that aims to provide tools that will enable prediction of the antigenic similarity of different strains of FMDV from their capsid gene sequences alone. In stage 1 we will develop simplified 'in silico' models of immune reactions that simulate a polyclonal antibody response to different viral strains as represented by complete amino acid sequences of their capsid proteins. This immune model will exploit the substantial amount that is known about the structure and distribution of epitopes across the FMDV capsid, and will use as input existing capsid genotypes and additional strains predicted to derive from them. This immune model will enable the reactivity of the polyclonal response to one viral strain to be measured against another, thereby allowing pairwise antigenic similarity of different viral strains to be predicted. In stage 2 we will use these simulated data sets containing full-length capsid genes, and matrices containing estimates of their antigenic similarity to develop bioinformatic algorithms that will be able to predict the antigenic similarity of new pairs of capsid sequences for which antigenic data is lacking. We propose to try two different approaches: artificial neural networks, and kernel based machine learning methods. The performance of these algorithms can be assessed using simulated data, and pre-existing empirical data.