Analysis and prediction of protective continuous B-cell epitopes on pathogen proteins

The applicability of peptides for the generation of preventive vaccines, therapeutics and diagnostics is an actively investigated field. Although historically disfavored the application of peptides in vaccine design is currently experiencing a renaissance[1]. While the focus is often on T-cell responses especially the generation of B-cell responses is of relevance against certain pathogens such as HIV to prevent initial infection [2]. It is thus not surprising that since the early days of computational biology scientists have attempted to predict the relevance of protein domains and peptides in several areas of application. Initial hallmarks of the field are represented, among many others, by work of Hopp and Woods [3, 4]. During the following and more recent years various methods and problems concerning the prediction of continuous B-cell epitopes have been proposed [5–12]. Recently the usability of amino acid scales for the prediction of B-cell epitopes has been profoundly questioned [13] and common standards regarding the validation of epitope predictions have been discussed [14].

Generally, B-cell antigenicity predictions should probably be understood as a measure of the likelihood to develop antibodies against a particular determinant or part of a surface, rather than another. In addition, most proposed classifiers of continuous epitopes are ultimately a composite of accessibility and charge-interaction potential prediction with a strong focus on delivering a few experimentally applicable peptides rather than an overall complete probability distribution for raising antibodies. In addition, continuous epitopes make up an undefined but presumably small part of the complete "epitope space" of an antigen. This even so when assuming distinct epitopes rather than a continuous surface and accepting dominant continuous elements of structural epitopes as continuous epitopes.

In this study we extend previous work by investigating a subgroup of continuous B-cell epitopes, namely protective continuous epitopes. This aspect has to the best of our knowledge not been systematically tackled so far.

From a biological point of view several principles should govern the availability and evolutionary behavior of protective amino-acid sites on proteins. One of the questions we ask is whether these principles or constraints lead to signals which can be used for predicting or rather detecting candidate epitopes. We assume that an antibody (and hence it's epitope) is protective because the function of the antigen (target protein) is inhibited and the activity of the organism is thus reduced. Or alternatively because immunological processes are activated leading to the destruction of the organism. The prior (protectivity class I) might most likely be expected in adhesion molecules or pathogenicity factors like matrix degrading proteases and toxins.

The second category (protectivity class II) would primarily refer to downstream events of antibody induced complement activation such as pore formation and opsonization leading to phagocytosis. It can be considered likely that the two mechanisms would often lead to differently characterized epitopes. Consequently as different functional constraints can be expected separate strategies for detecting them may be required. Basically any protein of high expression and density on the surface with at least one good epitope can be the target of class II protectivity. We define a "good" epitope as a surface area of high interaction potential, shape complementarity to the basic layout of an antibody [15] and dissimilar to self. Predicting that class therefore also requires to assess (or estimate) the density of a protein on the cell during pathogenesis, optimally experimentally or by inference from related organisms. A practically applicable continuous epitope could then be any exposed, possibly evolutionarily highly variable loop with an amenable antigenicity and solvent-accessibility profile. Class II protectivity may often be comparably straightforward to predict as soon as a target protein has been identified because selection of high scoring B-cell epitope scores often seems to be relatively straightforward and selection routines primarily falter in the domain of suboptimal scores as has been indicated by Sollner and Mayer [9]. However, immunologically sub-dominant or even cryptic B-cell epitopes can be of special interest regarding protectivity and inter-strain cross-reactivity [16] and are sometimes consistently immuno-silent during natural infection [17, 18]. As a consequence we focus on conserved and therefore presumably functionally constrained epitopes without post-translational modifications which still exhibit amenable antigenicity scores. Regarding the previous definition these epitopes may often fall into class I of protective continuous epitopes.

Such principles are primarily valid for pathogens already adapted to the host. Organisms in the process to adapt to new hosts or receptors can undergo significant alterations in their antigenic structure as has been demonstrated for SARS virus [19–21] and HIV [22], respectively.

We speculate that by means of functional constraints centers of biological activity can exert conserving pressure on closely associated potential epitopes while less relevant regions can be more variable. It may also be viable to suggest that conservation of posttranslational modification patterns may be different when comparing highly variable exposed loops and sites of functional relevance as modifications can play a major role in the masking of protective epitopes [23, 24] but may often be undesired near functional centers.

As Class II protectivity is to a certain degree already approached by standard B-cell epitope prediction (the maximization of antigenicity) and on the other hand depends on a bioinformatically more elusive factor (expression levels of pathogen protein) we see reason to focus on the prediction of conserved, functionally constrained epitopes (class I protectivity).

This work assesses in how far correlation between antigenicity, variability, post-translational modifications and protectivity/functional relevance can be put to use in a predictive model without the availability of 3D data. To compensate for the lack of 3D data multiple alignments of selected proteins are harnessed to derive information regarding the conservation of post-translational modification motifs as well as sequence variability per-se. We believe that understanding evolutionary "movement" of pathogen proteins allows insights into the importance of potential epitopes and we interpret such importance as indicator of protectivity.

This work describes the generation of a novel B-cell classifier based on the logistic regression of amino acid parameters derived from principal component analysis of a commonly used parameter database. We extend the classical application of antigenicity classifiers from the prediction of continuous B-cell epitopes to the prediction of protectivity by introducing measured variability and predicted modification patterns into the concept.

We could show that validation on the data-set by Blythe and Flower revealed close to random performance for both the gold standard ABCpred as well as for PCA19, although our method exhibited marginally better performance in comparison. This result is clearly relativated as we could show far better performance on a different dataset when considering only potentially relevant domains.

In detail, to assess the benefit of excluding presumably irrelevant domains from the prediction of protective continuous determinants a manual selection of protein regions was created based on standard bioinformatical tools. For this selection we used a curated set of 57 proteins with known protective or otherwise functionally altering, continuous epitopes. On this compilation protectivity prediction using PCA19 in combination with variability and modification likelihood performed significantly better after domain-accessibility filtering as measured by AROC values, while without filtering performance was comparable (although again slightly better) to antigenicity validation results on the Blythe et.al validation-set. Even when disregading eight proteins which contaminate the validation process due to similarity with the PCA19 training data the difference persists and is even enhanced.

Another primary observation is the gross variance in AROC values observed for various proteins. Approximately 65% of sequences exhibited AROC values < = 0.35 or > = 0.65, indicating a pattern substantially different from random noise. It may be wise to investigate the performance on a per-protein or per-epitope basis and to register how many percent of known, distinct sites were found with high reliability. That may help to avoid an averaging effect leading to the underestimation of the predictive power of classifiers. If a method performs well on every second protein or on certain epitopes that may be sufficient for many practical applications in the life-sciences field. This is especially so where high throughput is involved. Interestingly no correlation between the functional category of a protein or pathogen class (bacterial or viral) and the AROC could be established, indicating a more complex situation than just two types of epitope each associated with a distinct functional class.

By determining the number of theoretically required synthetic peptides to achieve satisfying protectivity in a vaccine or mAb approach we conclude that up to six peptides would be required to achieve success in every second protein. Such success naturally requires the existence of protective, largely conserved and possibly unmodified epitopes.

We also want to point out that the selected validation set presumably represents a combination of differently well mapped proteins. In addition, each of these can contain one or several protective B-cell epitopes of both class I and II. As we tried to detect only one of these in the set, neglecting the other, validation is skewed against our method. It may also be good to remember that calculated AROC values are averaged over entire proteins, not distinct epitopes.

The application of machine-learning procedures potentially combines the prediction of different epitope classes and allows an estimation of the information content in the data. A Random Forest model achieved 83% class separation in an equilibrated model, pointing towards the potential to significantly improve upon the single-score method.

Although the selection of pathogen proteins relevant for the stimulation of protective immune-responses can be enhanced by bioinformatics that is a topic distinct from the prediction of likely protective epitopes on these antigens [30, 31]. By combining in-silico ranking of likely protective targets and likely protective peptides from these targets completely automatized screening for applicable peptides is possible. Although methods do exist now for both aspects of in silico protectivity screening caution is still necessary. For example, the method published by Doytchinova and Flower predicts 95% of the proteins in our dataset as protective antigens (or at least antigens). On the other hand 65% of the proteome of Staphylococcus aureus COL (1727 of 2618 proteins) are also predicted to be relevant antigens which although possible seems to be a very high number suggesting a higher false positive rate than approximated by validation procedures in the publication. On the other hand all proteins are ranked by a score, so an order of predicted relevance is available, in our view essential for practical use. Unfortunately 701 proteins would be selected to cover all four proteins for which S. aureus protectivity data is available in the IEDB and which have close homologues also in strain COL (fibronectin-binding protein A and B, enterotoxin B and enterotoxin type A), yet not all may be required for a protective immune response and others may contain unmapped or discontinuous epitopes [32]. These findings indicate that several predictive methods and experimental data should be combined when selecting candidates for the generation of protective immune responses.

Our work solely focuses on the prediction of candidate peptides after such a selection has taken place, independent whether by means of experimental data, literature mining or purely bioinformatical/biological considerations.

We can show that prediction of protective or at least functionally relevant continuous B-cell epitopes can be efficiently done for approximately 50% of 57 analyzed proteins of pathogen origin. By minimizing sequence variability and probability of post-translational modification it can also be assumed that selected peptides are particularly suited for vaccine or monoclonal antibody generation. Exclusion of rationally selected domains strongly enhances the prediction of protective sites, indicating the relevance of a filtering step to restrict to immunologically likely accessible regions. Furthermore, analysis of correlation between variability, conservation of modification profiles and predicted antigenicity shows different and opposed categories of correlation thus indicating the existence of distinct epitope types. At this point it is difficult to verify or falsify our basic assumption of the two-class nature of protective epitopes. On the other hand the high percentage of epitopes with either positive or negative feature correlation indicates the existence of at least two types. Also, decision tree based models significantly outperform the single score-model pointing towards more complex relationships as well as possibly several distinct epitope signatures.

Future paths may lie in the detailed unraveling of parameter-associations and the utilization of more sophisticated classification methods such as regression trees for the assessment of biological relevance and protectivity in the selection of peptides for biotechnological application, as indicated by initial Random Forest classification.