Many of the early successes in vaccine development were against infections caused by such viruses (smallpox, measles, mumps, rubella and yellow fever) with little or no known variability in their antigens, for which natural infection provides lifelong immunity against all known strains of the species. For the vaccines widely available by the 1970s, with the exception of polio, antigenic variation was not an important consideration.

Many recent vaccine development projects have faced the problem of antigenic variability. Some recently developed vaccines—notably, vaccines against the encapsulated nasopharyngeal bacteria, Haemophilus influenzae, Streptococcus pneumoniae, Neisseria meningitidis—have approached the problem of antigenic variability by multivalent formulations that raise immunity individually targeted against one or more of the most clinically significant antigenic (capsular) types. Such approaches have been largely successful, but in the case of S. pneumoniae, they have led to the selection of vaccine escape variants, namely strains carrying capsular types that are not included in the vaccine (Lipsitch 1999; Kyaw et al. 2006). Many of the remaining challenges for vaccine design—malaria, HIV/AIDS, trypanosomiasis, gonorrhoea and others—are against pathogens for which antigenic variation is common both within infected individuals and in the pathogen population as a whole, and for which natural immunity is imperfect and/or temporary. Indeed, for many of these pathogens, antigenic variability is so pervasive that no single protective antigen has been identified (as it was for the encapsulated bacteria), widening the scope of the challenge.

It is tempting to speculate that vaccine development is most difficult for pathogens that provoke the least effective immune responses following natural infection, and that these pathogens, in turn, are disproportionately prone to significant antigenic variability. This hypothesis suggests a practical benefit in understanding the mechanisms underlying antigenic variation, in order to both design vaccines that may overcome the problem, and predict how pathogen populations will respond to the selective pressure imposed by vaccines that target particular antigenic types. Antigenic variability is also of considerable interest from the perspective of basic population biology. Understanding the forces that maintain diversity in natural populations is a central question of both ecology and population genetics, and pathogen populations form excellent model systems for investigating these principles both theoretically and empirically. As this review will argue, antigenic variability takes widely different forms in different pathogen populations. The idea that host immune pressure leads to antigenic diversity is almost a truism, but has not been refined to the point that it can explain why different infectious agents show such different patterns of antigenic variability. This article is intended to encourage further studies towards the development of a more unified theory of antigenic variation, one that can explain this variety of patterns while making testable predictions about the biological mechanisms needed to maintain these patterns.

In §1, we describe five patterns of antigenic variability observed in different pathogens. In §2, we survey each of these patterns and their possible biological underpinnings in a single pathogen species. In §3, we consider some of the general principles that have been proposed to explain antigenic diversity, including mathematical models of the effects of diversifying selection by host immunity. In §4, we briefly consider non-immune mechanisms that may lead to antigenic diversity. In §5, we discuss the strengths and weaknesses of existing explanations for antigenic variability and lay out a set of theoretical, experimental and practical (vaccine development) questions that, we believe, would merit further study towards a more general understanding of the problem.

In this review, we define antigenic variability as the existence of two or more strains within a pathogen species for which the immune response provoked by one strain is more effective against the strain itself (the homologous strain) than against other (heterologous) strains. The concepts of species and strain within microbial populations are controversial because sex, in particular outcrossing sex, is possible yet optional for many important microbes. This liberated approach to sex (Levin & Bergstrom 2000) complicates the definition of species because the biological species concept, so useful for obligately sexual organisms, is not as clearly defined for those in which sex is optional and not tied to reproduction. On the other hand, the possibility of sex within nearly all microbial populations complicates the classical notion of a strain, which is based on clonal descent. We do not dispute the complexity of these arguments, but, in general, will take the notions of species and strain as if they were unproblematic.

Antigenic variability takes many forms. Here, we consider four pathogens that exemplify widely varying patterns of antigenic variability: the bacteria, S. pneumoniae and N. meningitidis; and the viruses, influenza A and HIV-1. We also consider infectious agents that rarely exhibit antigenic variability.

Pattern 1: little variation within a single host, but extensive population-wide variation that is consistent in space and time (S. pneumoniae).

Pattern 2: little variation within a single host, but extensive population-wide variation that can change in space and time (N. meningitidis).

Pattern 3: little variation within a single host, and little variation in the global population at a single time, but rapid variation over time on a scale of years (influenza A virus).

Pattern 4: extensive replacement of dominant types over time within a single host, with extensive and growing standing diversity in the global population (human immunodeficiency virus, type 1).

Pattern 5: since little or no antigenic variability is known, naturally acquired immunity and vaccine-induced immunity are universal or nearly so.

These patterns are neither fully exhaustive nor mutually exclusive; in particular, we do not consider here the pattern of within-host variation by high-frequency genetic changes (cassette expression or phase variation) used by pathogens including Plasmodium falciparum, trypanosomes and Neisseria gonorrheae (Criss et al. 2005; Dzikowski et al. 2006; Taylor & Rudenko 2006).

For each of the first four patterns of antigenic diversity, we consider here some relevant details of a particular pathogen that, we believe, exemplifies the pattern. General epidemiological details of these four infectious agents are given in table 1.

Table 1 Epidemiologic characteristics of four ‘typical’ pathogens with different patterns of antigenic diversity.

In a given location, the frequency distribution of particular serotypes or serogroups (groups of closely related serotypes) in invasive disease has remained nearly constant over several decades (figure 1b; Babl et al. 2001; Konradsen & Kaltoft 2002 and references therein), with the exception that serotype 1 has declined in frequency since the mid-twentieth century in several places, but has then rebounded in Europe while remaining low in the US. Unfortunately, similar data are not available for nasopharyngeal carriage, which should represent the population of pneumococci under selection for transmission. The top serogroups recovered from nasopharyngeal carriage (Bogaert et al. 2004a) or invasive or mucosal disease (Hausdorff et al. 2000, 2005) are also geographically consistent, although not identical, across continents. Despite this consistency, the total importance of the top serogroups varies geographically, with the most limited diversity in North America; part of this difference, however, probably reflects variation in sampling (Hausdorff et al. 2005).

Observing that different antigenic types (serotypes or serogroups) have roughly similar frequencies across time and space suggests the possibility that some strong selective force is maintaining these frequencies. Evidence for the strength of this selective force comes from a comparative study of antibiotic resistance and serotype composition across eight regions in the United States (McCormick et al. 2003). In the US, as elsewhere, resistance to antibiotics has been strongly concentrated in certain serotypes; for example, 42–84% of serotype 19A isolates were penicillin resistant, compared with only 0–6.8% of serotype 4 isolates (ranges are for the eight regions). Overall, penicillin resistance varied significantly between US regions, from 14% of pneumococci in California to 36% in Georgia and Tennessee; analyses by McCormick et al. (2003) attributed these differences to regional variation in selection for antibiotic resistance. One might have expected that this would lead to enrichment of the resistance-prone serotypes in the high-resistance regions compared with the low-resistance regions, i.e. the serotype composition would be changed by associated linkage selection for antibiotic resistance. This was not the case; high-resistance regions showed greater proportions resistant within each serotype, but no trend towards having more of the high-resistance serotype (McCormick et al. 2003). An interpretation of this finding is that whatever selection was maintaining, serotype composition in the populations of these states was stronger than the selection imposed by linkage of serotype to antibiotic resistance. Conversely, when strong selection to change serotype composition has been applied, by widespread use of the pneumococcal conjugate vaccine in the US, serotype composition of the pneumococcal population has changed and with it has brought down the rates of resistance (since most of the resistant strains are of serotypes included in the vaccine; Kyaw et al. 2006). Thus, at least among invasive disease isolates in the US, selection on serotype can change resistance patterns more readily than selection for resistance changes serotype composition.

As with most forms of antigenic variability, the existence of diverse serotypes of pneumococcus has been attributed to diversifying selection by host immune responses. In a population of n pneumococcal serotypes, with many hosts immune to many of the types, the n+1th serotype will have an advantage, as long as it is seen by host immune systems as being distinct from existing serotypes, because there will be more susceptible hosts in the population for this serotype. As it increases in frequency, its advantage will wane (opening the door for more diversification), while if it declines in frequency, the host population will be less exposed over time, resulting in less immunity and thereby restoring its advantage. This is the classic scenario of ‘negative, frequency-dependent’ diversifying selection on antigens by host immunity.

This scenario, while plausible, has important limitations in the pneumococcal case. First, in its simple form, it predicts that diversity should be high everywhere and at all times, but probably growing over time and not necessarily geographically coherent. Second, we and others who have looked into this failed to find evidence that naturally acquired immunity to S. pneumoniae is mainly serotype specific. Invasive disease incidence from various serotypes peaks around 1 year of age and falls in parallel, suggesting a single type of immunity acquired as children age, rather than the acquisition of multiple antigenic specificities, which should appear at different ages. Moreover, these declines begin before anticapsular antibody rises measurably in children, suggesting that anticapsular antibodies are not the mechanism of reduction in disease (Lipsitch et al. 2005). Similar patterns have been observed in nasopharyngeal carriage (Hogberg et al. 2007), and associations between anticapsular antibody and protection from pneumococcal carriage have been weak in epidemiologic (Goldblatt et al. 2005) and experimental human studies (McCool et al. 2002). In mice, acquired immunity following exposure to one serotype is equally effective against that type and other types, suggesting that it is mainly directed against one or more non-capsular antigens (which are probably more conserved across pneumococci; Malley et al. 2005; Trzcinski et al. 2005), and acquired immunity to carriage in mice is dependent on CD4+T cells, but not on antibody (Trzcinski et al. 2005; van Rossum et al. 2005), thereby excluding anticapsular antibody as a mechanism.

These findings cast doubt on the classical immune diversifying selection hypothesis as an explanation for pneumococcal capsular diversity. Other possible mechanisms of diversifying selection on pneumococcal capsules are under investigation, including the possibility that different capsular types are preferentially successful in different types of hosts, perhaps determined by host genetics, age (Bogaert et al. 2004a) or other factors.

Figure 1 Antigenic diversity, pattern 1: constant high level of standing between-host diversity consistent over space and time; little within-host diversity, exemplified by Streptococcus pneumoniae. (a) Schematic of this pattern. Triangles, individual hosts within (more ...)

Despite these similarities, the contribution of particular meningococcal strains to disease shows great variation in both space and time, in contrast to pneumococcus (figure 2a,b). Geographical variation is observed whether meningococci are classified by serogroup (figure 2b; Stephens 2007) or by clonal complex—a designation defined by the similarity of nucleotide sequence type (ST) of housekeeping genes, which are thought to be selectively neutral (Rosenstein et al. 1999; Yazdankhah et al. 2004). Temporal variation is also common. The incidence of disease caused by serogroup Y bacteria in the US has increased significantly from 2% of endemic disease in 1989–1991 to 39% for the period 1996–2001, whereas this serogroup is an infrequent cause of disease in Europe (Rosenstein et al. 1999; Raghunathan et al. 2004; Yazdankhah & Caugant 2004). Serogroup A has declined in incidence to very low levels in the US and Europe since World War II, but has been responsible for significant outbreaks in African countries and remains a major cause of meningitis there (Stephens 2007). It should be noted that while patterns of strains causing meningococcal disease show great geographical and temporal variation, the patterns among carriage isolates may be somewhat more homogeneous, at least within Europe (Caugant et al. 2007). The strong contrast drawn herein between patterns in the meningococcus and those in the pneumococcus should be seen as tentative until comparable samples of carriage isolates of the meningococcus can be assessed.

Several factors complicate the study of antigenic diversity in meningococcus. First, while the capsule (which defines meningococcal serogroup) is a vaccine antigen, its centrality as an antigen in the natural immune response is uncertain (as with the pneumococcus). Both encapsulated and acapsulate forms of the bacterium are common and each state is suggested to enable escape from immune recognition under different circumstances. While virtually all strains that cause disease are encapsulated (Dolan-Livengood et al. 2003; Yazdankhah & Caugant 2004; Yazdankhah et al. 2004), up to 30% of meningococcal isolates from carriers are non-serogroupable. It is suggested that the capsule may prevent immune recognition during systemic disease, increase survival in the environment and enhance transmission (Lipsitch & Moxon 1997; Dolan-Livengood et al. 2003). However, benefits of the acapsulate state have also been proposed, including evasion from a host immune response during carriage and facilitating adherence to the human nasopharyngeal epithelium (Hammerschmidt et al. 1996; Dolan-Livengood et al. 2003). Thus, any complete explanation of serogroup structure will require an explanation of the variable expression of the capsule observed in the species. Further knowledge of the regulation and role of the capsule could shed light on the transmission and pathogenic processes of N. meningitidis.

The value of the capsule in defining biologically distinct meningococcal types is also open to question. Clonal types defined by housekeeping genotype (or previously by protein electrophoretic type) seem to correlate better with many biological properties than capsular serogroup does, and meningococcal strains as defined by molecular methods have been observed to change serogroup (Swartley et al. 1997). Assessed by either serogroup or ST, temporal and geographical variations are common (Rosenstein et al. 1999; Yazdankhah et al. 2004). Other antigens may also be protective (Pizza et al. 2000; Urwin et al. 2004), raising questions as in the pneumococcal case about the uniqueness of the capsule as an important antigen for natural immunity.

Diversity is greater among carried meningococcal strains than among those causing disease, which represent a small number of hypervirulent lineages (Caugant et al. 1988; Yazdankhah et al. 2004). As colonization with the majority of strains of N. meningitidis produces no symptoms, the disease state is unimportant for much of its spread (Jolley et al. 2000). The low frequency with which virulent strains are isolated from carriers could be due to inherent low transmissibility of these strains. Alternatively, these strains could be highly transmissible but result in colonization of short duration (Caugant et al. 2007). As disease is thought to occur between 1 and 14 days after acquisition (Stephens 2007), it is plausible that either clearance or disease occurs soon after acquisition of a virulent strain and that this would explain the low point prevalence of such strains in all carriage studies. If the rate of transmission was significantly higher for some of the ‘hyperinvasive’ strains when compared to avirulent strains, it may be that the rate of disease per acquisition is very similar for the two types (Jolley et al. 2000). To discern which of these possibilities is correct, it will be essential to conduct large cohort studies over many months to learn what the lengths of carriage for different strains are. It may be the case that each strain has a characteristic carriage length or that the duration of carriage falls into two categories: one for avirulent strains and another for virulent ones.

A theory of strain structure proposed by Gupta and colleagues suggests that immune selection can organize polymorphic antigens into non-overlapping strains within a pathogen population (Gupta et al. 1996; Gupta & Maiden 2001). The allelic diversity of two antigenic outer membrane proteins (OMP) of N. meningitidis is suggested to be structured in such a manner. As an adaptive immune response is specific to a particular antigenic variant, such responses will limit the transmission of pathogens that share the same antigen or that cross-react with it. If the immune response is sufficiently strong, then strains that do not share alleles at antigenic loci will be selected as they do not interfere with each others' transmission. Thus, pathogen strains with unique combinations of antigenic alleles will be produced. Gupta suggests that the porA gene of N. meningitidis, which encodes a highly diverse OMP, shows signs of being structured in such a way (Gupta et al. 1996); further studies on porA and fetA, another gene encoding an OMP, are described in §5.

Figure 2 Antigenic diversity, pattern 2: substantial between-host diversity, changing over time and by geographic region; little within-host diversity, exemplified by Neisseria meningitidis. (a) Schematic as in figure 1a. (b) Global serogroup distribution of (more ...)

The major immune response that is thought to be important in preventing or clearing influenza A infection is an antibody response directed against the haemagluttinin (HA) protein (Zambon 1999). Phylogenetic studies of the HA nucleotide sequence within a single HA subtype (H3) have shown a pattern of slow genetic change over time, with considerable genetic change over a time span of decades, but rather limited diversity present in the global population at a given time (figure 3a,b; Fitch et al. 1991).

Recent studies have complicated this picture somewhat. Smith and colleagues have shown that the variation in HA nucleotide sequence does not map continuously into variation in antigenic similarity; rather, every few years, a new ‘antigenic cluster’ arises that differs significantly from previous strains, in that antibodies cross-react poorly between antigenic clusters. Within an antigenic cluster, however, significant amounts of genetic variation can occur without much apparent effect on the antigenic properties of the HA molecule. Thus, the rather smooth pattern of genetic variation leads to a more punctuated pattern of antigenic variation (Smith et al. 2004).

In contrast to the bacteria considered above, the key role of the immune system in selecting for antigenic variation in influenza is difficult to dispute. Strains isolated years apart provoke immune responses that are poorly cross-protective (Smith et al. 2004). Genetic changes in the HA are concentrated in the regions of the molecule that are most often the targets of antibodies (Plotkin & Dushoff 2003). Despite this clear evidence of selection for antigenic change, recent analyses suggest that significant antibody escape mutations are relatively rare, rarely reaching measurable levels within an infected host or becoming noticeable even within a season (Holmes et al. 2005; Wolf et al. 2006). This puzzling finding is related to a more general puzzle: why does antigenic diversification in influenza proceed in such an orderly fashion, with limited diversity at one time, rather than a great explosion of diversity in the population? Two proposed answers to this issue are discussed in §3.

Figure 3 Antigenic diversity, pattern 3: little diversity within-host or within the host population at a given time point, but rapid change over time, exemplified by influenza A virus. (a) Schematic as in figure 1a. (b) Phylogeny of haemagluttinin variants present (more ...)

The long infectious period of HIV-1 means that evolution occurs both within and among hosts. Thus, the large discrepancy in the degree of positive selection observed between the intra-host and the population-level phylogenetic trees of HIV-1 is surprising. The within-host tree is marked by a clear temporal structure showing continual adaptation to the host immune system and high rates of lineage extinction resulting in little genetic diversity at any time point (Grenfell et al. 2004). This is similar to the phylogenetic structure of influenza A on a population level (Grenfell et al. 2004). However, spatial dynamics appears to dominate the between-host HIV structure which shows multiple coexisting strains and weaker positive selection (Grenfell et al. 2004).

The remarkable genetic variation of HIV-1 within a host is a significant issue in terms of drug resistance, vaccine design and host immune response. This is due to a combination of factors: high mutation rate; a short generation time of approximately 2.6 days; production of approximately 10¹⁰ new virus particles per day; and frequent recombination (Sharp et al. 1995; Perelson et al. 1996). Strong immune selection pressure from neutralizing antibodies and cytotoxic T lymphocytes (CTL) drives the fixation of escape mutants that allow it to evade the host immune system (Ogg et al. 1998; Richman et al. 2003). Why then is such intense immune-driven selection not simply reflected on a population level? Rambaut et al. (2004) have proposed four explanations. Firstly, transmission is associated with a significant population bottleneck; over 99% of diversity in gag and env genes is lost upon sexual or vertical transmission (Edwards et al. 2006). Such a reduction in diversity affects the ability of HIV-1 to adapt on a population level, as advantageous mutations which arise in a new host have a reduced chance of transmission. Secondly, host behaviour can impact the action of selection at a population level. As HIV is largely a sexually transmitted disease, variation in the rate of partner exchange will also cause genetic drift (Grassly et al. 1999). By chance, viruses with advantageous mutations may be present in a person with a low rate of partner exchange and may not be transmitted for this reason. Thirdly, some advantageous mutations, for example CTL escape mutants, may not arise until late in infection after which point little transmission occurs (Goulder et al. 1997). Finally, CTL escape mutants may be transmitted to individuals with different HLA alleles from the donor. In such instances, escape mutants that were advantageous in the previous host may incur a high fitness cost in the current host without providing any benefit in terms of immune evasion. Studies of both humans and macaques found that CTL escape mutants that incur viral fitness costs reverted quickly upon transmission to individuals lacking the donor's restricting HLA alleles (Friedrich et al. 2004; Leslie et al. 2004), and that the rate of reversion reflects the magnitude of the viral fitness cost (Fernandez et al. 2005; Kent et al. 2005; Li et al. 2007). Therefore, HIV-1 evolution on a population level is driven by neutral demographic and spatial dynamics and both immune-mediated positive selection (Asquith et al. 2006) and purifying selection for replication ability. This balance has been elegantly demonstrated by a detailed longitudinal study of seven patients, in whom almost half of all substitutions observed were associated with escape from CTL responses or reversion of presumed escape mutations from prior hosts (Li et al. 2007). Moreover, the sites at which changes occurred to escape CTL responses were those that are most variable (and thus presumably least functionally constrained) in the global HIV-1 population (Allen et al. 2005).

An effective vaccine against HIV-1 will probably depend on careful selection of which epitopes should be targeted, taking into account selection pressure by a range of HLA alleles, associated fitness costs of escape mutants and the fact that the genetic diversity of HIV will change during the long process of vaccine development. Strains belonging to different subtypes can differ by up to 35% in their envelope protein and within-subtype diversity can reach 20% (Gaschen et al. 2002). In comparison, roughly 10% sequence divergence in the antibody-binding HA1 region of influenza HA typically exists between antigenic clusters, and this is more than enough divergence to compromise vaccine efficacy (Smith et al. 2004; Koelle et al. 2006). The most obvious method of creating a broadly effective vaccine is to include a variety of epitopes from different strains in a single formulation. Alternatively, artificial sequences, namely consensus or ancestral sequences, could provide the highest levels of cross-protection possible by any single sequence (Gaschen et al. 2002; Nickle et al. 2003). Appropriate construction of such sequences may require unbiased sequence datasets of each prominent subtype, as well as a quantitative understanding of the above individual- and population-level processes.

Figure 4 Antigenic diversity, pattern 4: substantial within-host diversity over time; little diversity at any single time point, and ongoing positive selection; global population does not reflect this selection but is highly diverse, exemplified by HIV-1. (a) (more ...)

Figure 5 Antigenic diversity, pattern 4: antigenic homogeneity, within and between hosts, over time. Schematic representation as in figure1a.

In the foregoing section, five patterns of antigenic diversity were described, and each was accompanied by a more or less plausible explanation relating the pattern observed to diversifying selection by host immune systems. In the ideal case, we would have more than just a good story for each pathogen and its pattern: we would have a list of characteristics which, once measured for a particular pathogen, would accurately predict the pattern of diversity observed. We have not reached this ideal—a sort of grand unified theory of antigenic diversity. However, there is a growing body of literature describing mathematical models of antigenic diversity, which taken together provides some testable hypotheses whose explanatory power for particular pathogens can be assessed. In this section, we briefly summarize some of the findings of this theoretical literature and suggest some of the relevant comparisons.

A general challenge in modelling the spread of multiple antigenic variants of a pathogen is that as the number of strains considered increases, there is an exponential increase in the number of equations required to track the population. Various elegant approaches have been employed to reduce the complexity of the problem, including assumptions of symmetry (such that all strains are identical and affect all other strains equally; Abu-Raddad & Ferguson 2004), status-based models (which track only the present ability of hosts to respond to strains, rather than their full past history of infection; Gog & Grenfell 2002; Gog & Swinton 2002), polarized immunity (in which infection confers total immunity on a proportion of hosts, rather than partial immunity on all hosts; Gog & Grenfell 2002; Abu-Raddad & Ferguson 2004) and the assumption that immunity reduces one's infectiousness to others, rather than one's risk of becoming infected (Gog & Grenfell 2002; Abu-Raddad & Ferguson 2004).

Having made varying combinations of simplifying assumptions, general models of pathogen diversity make the following predictions (summarized in table 2).

• If strains can be considered to exist in a ‘strain space’, in which distance corresponds to antigenic dissimilarity, the process of appearance and spread will tend to lead to clustering in strain space, because novel types similar to the most prevalent current types will constantly be generated, but types nearby an existing prevalent type will be at a disadvantage due to strong cross-immunity with the existing type. Therefore, highly prevalent types can exist only at some distance from one another, in strain space and/or time (Gog & Grenfell 2002; Gomes et al. 2002). This effect can, under certain conditions, maintain strain structure—collections of strains with non-overlapping sets of alleles at antigenic sites—despite recombination, because recombinants falling ‘between’ non-overlapping sets will be at a disadvantage (Gupta et al. 1996, 1998).
• The group of strains present in a population may be stable or changing over time. Rapid strain turnover is promoted by higher transmission (Gomes et al. 2002), weaker cross-immunity (Gog & Grenfell 2002; Gomes et al. 2002) and short-lived infections (Gupta et al. 1998; Gog & Grenfell 2002). In the case of multilocus genotypes that determine strain structure, Gupta (Gupta et al. 1998) finds a more complex relationship with cross-immunity, in which turnover (either cyclic or chaotic) is greatest at intermediate levels of cross-immunity, with lower levels giving unstructured stable coexistence of many types and high levels of cross-immunity promoting coexistence of non-overlapping strain types.
• Total pathogen prevalence is increased for pathogens having higher transmissibility and those with longer duration or less cross-immunity (Gupta & Galvani 1999; Abu-Raddad & Ferguson 2004).
• The standing diversity at any given time is greater for more transmissible pathogens and weaker cross-immunity (Abu-Raddad & Ferguson 2004) and longer infectious durations (Gog & Grenfell 2002; Abu-Raddad & Ferguson 2004).

Table 2 Predicted relationships between transmission characteristics of multi-strain pathogens (rows) and epidemiological features of strain diversity (columns), based on several theoretical studies.

Reassuringly, these predictions are largely qualitatively consistent across models despite differences in their simplifying assumptions, and in some cases are also consistent with previous simpler models (Dietz 1979; Gupta et al. 1994; Lipsitch 1997).

How well do these theoretical principles explain the observed patterns described in the previous section? Perhaps, the best match between theory and observation is for the case of influenza (the problem that motivated many of the theoretical studies): its short duration of infectiousness predicts limited diversity at one time, rapid turnover and low prevalence at any given time, which are qualitatively accurate. A problem with naive models is that diversity is generated too fast. Invoking either transient strain-transcending cross-immunity that does not depend on the distance between strains (Ferguson et al. 2003) or the existence of neutral variation that limits the rate of generation of antigenically novel variants (Koelle et al. 2006) can reduce this rate, thereby slowing the rate of strain turnover to realistic levels. In the general models, increasing the degree of antigenic distance-dependent cross-immunity (i.e. making the immunity extend to more distant strains) has a similar effect in reducing the rate at which escape variants are generated, but the details and other effects of these assumptions are different from those invoked in the influenza models (Ferguson et al. 2003; Koelle et al. 2006).

Some of the other patterns can perhaps be understood in the same framework, but the concordance is spotty. For brevity, we consider only S. pneumoniae; other comparisons are left as an exercise for the reader. S. pneumoniae differs from influenza mainly in the temporal stability of its antigenic patterns. If our hypothesis of strongly species-specific (not strain-specific) immunity is correct, then the models would predict slower turnover; the duration of pneumococcal carriage, while still short, is several-fold longer than infectiousness with influenza (Smith et al. 1993; Sleeman et al. 2006; Hogberg et al. 2007), also predicting slower strain turnover. Pneumococcal point prevalence is considerably higher than that of influenza, perhaps reflecting longer carriage duration and weaker overall immune responses, despite the fact that these immune responses are equally strong across different strains. These findings are concordant with the theoretical predictions just discussed (table 2). On the other hand, strong cross-immunity should lead to limited diversity at one time, yet in fact many serotypes coexist in the population.

This example suggests two considerations: first, that the qualitative predictions summarized in table 2 may be inadequate for the comparison of different pathogens, since two pathogens may differ in multiple respects, with countervailing predicted effects on some epidemiologic pattern (e.g. rate of turnover), making it impossible to predict which of the two pathogens will have a faster rate of turnover. Second, when considering individual pathogens, other factors not considered in general models may make an important contribution to the pattern of antigenic diversity. For example, immunity to pneumococcus appears to build up very gradually with exposure, and to take the form of shorter duration of carriage and reduced susceptibility to disease. While the immunity seems to be directed against conserved determinants and thus is effective against most or all serotypes, it may be more effective against some serotypes than against others (Hausdorff et al. 2005; Hogberg et al. 2007). More biological understanding may be required before general models can be adequately calibrated to particular pathogens.

While much antigenic variability is undoubtedly generated by interactions between a pathogen and the immune system, it should be remembered that probably no antigen produced by a pathogen has provocation of a specific immune response (or susceptibility to that response) as its evolutionary raison d'etre. Antigens are components of infectious agents with diverse functions of nutrient acquisition, interaction with host cells, etc. As a result, any process that selects for genetic or phenotypic variability—in ecological language, any temporal or spatial variation in the pathogen's habitat—may directly or indirectly select for antigenic variability. Examples of such mechanisms include variation between hosts due to genetics (Moxon et al. 2006), behaviour, competing flora (Bogaert et al. 2004b; Regev-Yochay et al. 2004) or other factors that may selectively favour different variants for reasons other than immune escape. Cell-to-cell or microanatomical variation within a host (Kang & Blaser 2006) may differentially favour particular genotypes (Briles et al. 2005); likewise, there may be selection for variation within a host to maintain a colonizing or infecting population within the host while also producing progeny that can colonize other hosts (Boles et al. 2004; Briles et al. 2005). Such hypotheses, which, to date, have been mainly invoked to explain the evolution of highly mutable ‘contingency’ loci in bacteria (Moxon et al. 2006), are serious contenders to explain some of the antigenic variation in pathogens for which immune selection has not been demonstrated as the cause.

The open questions raised by the viral examples above and those raised by the bacterial examples are rather different. We consider these in turn.

Important antigens have been identified for many important viruses, including those considered here. In such cases, the major theoretical challenge is to understand what constrains variation in the viruses—why measles and smallpox fail to escape immunity, while HIV does so within a host and influenza does so in host populations over time. It is probably not coincidental that the greatest within-host variation occurs in HIV-1 compared to these other viruses, since HIV-1 has by far the longest infection period. However, even this relationship is complex (Nowak & May 2000): short duration may be a cause of limited variability (if there is no time for variants to arise); it may be a consequence of limited variability (without immune escape, clearance occurs); and, as noted above, short duration may reduce diversity at the epidemiologic level as well (Gog & Grenfell 2002; Abu-Raddad & Ferguson 2004).

From a practical perspective, a major challenge of viral antigenic diversity is the design of vaccines. In the extreme case of HIV, in which immunodominant responses are strongly dependent on the host's HLA type (Bihl et al. 2006), it remains to be seen whether a vaccine can be developed that elicits strongly protective responses in a diverse population of hosts (Walker & Korber 2001). Similar considerations apply in the case of hepatitis C virus and other viruses for which vaccines remain elusive. In the case of influenza, an open challenge is to develop a vaccine that can provide ‘species-wide’ protection, not only against the antigenic drift variants that occur over time but also across influenza A subtypes, so that a vaccine could be usable when a new subtype appears (potentially causing a pandemic) without complete redesign (Doherty et al. 2006).

For both the pneumococcus and the meningococcus—and for other bacterial and eukaryotic pathogens—a central question is: what is (are) the key antigen(s) that trigger protective immunity? It will be instructive to revisit the broad distinctions made herein once the critical antigens for naturally acquired immunity are identified. From a scientific perspective, there are a number of promising approaches to identifying such antigens. These can be broadly divided into population-genetic approaches and experimental approaches. Theoretical (population-genetic) analyses predict that antigens should have several properties. First, they should be under positive selection, detectable by the pattern and rate of nucleotide substitution (Volkman et al. 2007) or perhaps even by the pattern of codon usage (Plotkin & Dushoff 2003; Plotkin et al. 2004). Second, they should be in linkage disequilibrium with one another (Gupta et al. 1996) despite sexual recombination in the pathogen genome. Identifying antigens by these characteristics and then assessing their pattern risks circularity (patterns of diversity assessed in antigens identified by their pattern of diversity), but certainly there is much to be learned by assessing patterns in known antigens, then using those patterns to ‘fish’ for candidates for additional unknown antigens. We have focused on spatio-temporal patterns in this review, but phylogenetic approaches (recently reviewed by Grenfell et al. 2004) use an additional dimension in the data and have been profitably applied to influenza (Wolf et al. 2006) among other pathogens.

In general, further comparisons between population genetic patterns of putative antigens and those of non-antigens will aid in understanding the selective forces maintaining variation. It is notable that the pattern of stable polymorphism in pneumococcal capsules described previously is observed in a species where strain types defined by ‘housekeeping’ genes fit well with a neutral model of evolution (Fraser et al. 2005). Similarly, comparison of HA evolution in influenza A against that of the whole genome has highlighted the importance of recombination and other processes that differentiate antigen evolution from ‘background’ evolution (Holmes et al. 2005). Use of these approaches will often depend on unbiased sampling of strains (Fraser et al. 2005; Holmes et al. 2005) and appropriate typing systems that gather data on targets of interest (Jolley et al. 2007).

For both bacteria and viruses, much work remains to understand how hosts confronted with a pathogen displaying multiple antigens ‘choose’ which antigens to respond to, which arms of the immune system will be involved and which of these responses will be protective—the general phenomenon of immunodominance. From the pathogen side, it would be of interest to understand, in the presence of a host response to multiple antigens, how much, if any, advantage is garnered by a variant that escapes some but not all of these responses. If we think of a genotype-fitness mapping in antigenic space, does a broad immune response always lead to a locally flat fitness landscape (because escaping from one response has no effect on fitness in the presence of other responses)? Does a very narrow immune response create a simple adaptive landscape in which any escape from the response is favoured (unless the escape mutation destroys function)? Does a response of intermediate breadth create a landscape of the form observed for influenza A, in which many changes are neutral (perhaps because they do not sufficiently reduce immune pressure; Smith et al. 2004; Koelle et al. 2006), but the right combination of changes can produce significant escape?

Understanding the answers to such questions would be valuable for the design of vaccines, especially (as is now almost universal) the design of subunit vaccines that contain a limited number of antigens. Until such answers are available, a novel approach for the design of subunit vaccines is that of ‘reverse vaccinology’ (Rappuoli 2000). Reverse vaccinology uses bioinformatic techniques to find vaccine candidates that are likely to be surface expressed (hence antibody-accessible), tests these candidates for immunogenicity and protective ability in a high-throughput way and suggests a combination of antigens for a multi-subunit vaccine (Rappuoli 2000; Mora et al. 2006). More recently, the concept has been updated to account for within-species variation in gene content (Mora et al. 2006). An analysis of inter-strain diversity of group B streptococcus (GBS) found that only 80% of genes are shared by all strains compared (Tettelin et al. 2005), while a study of Escherichia coli found that just 40% of a combined set of proteins were common to all strains analysed (Welch et al. 2002). Such diversity in gene content within pathogen species compounds the difficulty of producing an effective vaccine in the face of frequent and extensive antigenic polymorphism. To develop a vaccine against GBS, the genome sequences of eight GBS isolates were analysed as above and tested for immunogenicity. Four antigens were found to elicit protective immunity in mice and their combination conferred protection against a large number of strains (Maione et al. 2005). The novel aspect of the study is that none of the proteins could be classified as universal. Three of the proteins were absent in several strains tested and the fourth showed limited surface accessibility in some of these. Use of multiple genome sequences in such a way has been termed ‘pan-genomic reverse vaccinology’ in reference to the concept of a ‘pan-genome’ that is shared by members of a microbial species by descent and horizontal gene transfer, but which is not wholly possessed by any one strain (Mora et al. 2006).

In a variant on this approach, Urwin et al. (2004) have studied variation in OMP of N. meningitidis. Consistent with the predictions of Gupta et al. (1996), these antigenic and polymorphic proteins (PorA and FetA) show a relatively strong strain structure, with associations between particular alleles that are more than expected by clonal descent alone. Urwin et al. (2004) suggest that such structuring can greatly simplify vaccine design. As both PorA and FetA confer protection against reinfection with a homologous strain, careful selection of variants for inclusion in a vaccine could provide broad protection. Urwin et al. (2004) suggest that up to 90% of meningococcal strains could be covered by a limited number of antigen combinations. In principle, such a formulation is likely to be resistant against the generation of escape mutants as such mutants can only emerge and spread if they undergo simultaneous changes at both loci. It remains to be seen how successful such strategies will be in these and other pathogens, but these latter studies raise the exciting prospect that theoretical predictions about strain structuring may be rigorously tested in natural populations, and the results put to practical application in vaccine design.

Diversifying selection on both hosts and pathogens clearly plays a large part in their ongoing coevolution. Existing theoretical models provide a framework for understanding some of the observed patterns of antigenic diversification, but many questions remain—from how to identify antigens to why the patterns of diversity vary so widely between pathogens and how our evolving understanding of these patterns can be best used to design effective vaccines against the antigenically variable pathogens that continue to elude us.

This work was supported by the US National Institutes of Health cooperative agreement 5U01GM076497 (Models of Infectious Disease Agent Study) and grant 1R01 AI48935 to M.L.