Abstract
Antibody responses emerge from the competition of B cell lineages with different antigen receptors, each produced by the recombination of germline immunoglobulin genes. Which lineages win out can depend on subsequent somatic mutations that improve antigen binding, yet lineages using specific germline alleles can have higher affinity than others from the start or a higher propensity to adapt. How much do those germline-encoded advantages determine the outcome of B cell competition, potentially leading to predictable allele frequencies and sequence motifs in the response to the same antigen in different individuals? In simulations, we show that selection for receptors with germline-encoded specificity can lead to similar germline allele frequencies between individuals early in the response. As B cell lineages evolve, those early advantages are often overcome by lineages using different germline alleles in different individuals, leading to increasingly contingent patterns of germline allele usage over time. Consistent with simulations, mice experimentally infected with influenza virus have increasingly dissimilar germline allele frequencies and do not converge on similar CDR3 sequences or similar somatic mutations. These results suggest germline-encoded specificities might be selected to enable fast recognition of specific antigens early in the response, while diverse evolutionary routes to high affinity limit the predictability of responses to infection and vaccination in the long term.
Introduction
Antibodies owe their diversity and potency to evolution on two different timescales. B cell receptors, the precursors of secreted antibodies, are encoded by immunoglobulin genes diversified over hundreds of millions of years (Marchalonis et al., 1998; Flajnik, 2002; Das et al., 2008). Recombination of separate sets of genes encoding the receptor’s heavy and light chains, combined with insertions and deletions at the alleles’ junctions, produces a unique receptor in each maturing B cell (Hozumi and Tonegawa, 1976; Brack et al., 1978; Jackson et al., 2013). The result is a diverse repertoire of naive (antigen-inexperienced) B cells collectively capable of binding virtually any antigen. Once activated, naive B cells expand into lineages that compete with each other for access to antigen and can undergo selection for somatic mutations that improve binding (Eisen and Siskind, 1964; Jacob et al., 1991; Victora and Nussenzweig, 2012). How much the various lineages grow, what antigens and epitopes they target and how well they do so determine the ultimate specificity and potency of the antibody repertoire.
A central question in the study of adaptive immunity is how much these outcomes depend on the initial set of germline immunoglobulin genes versus the subsequent evolution of fully-formed B cell receptors. Affinity maturation can vastly improve binding (Liao et al., 2013; McCarthy et al., 2019), yet high affinity for particular epitopes can be “hardcoded” on individual germline alleles from the start. These germline-encoded specificities might arise as evolutionary spandrels — byproducts of immunoglobulin gene diversification — (Gould and Lewontin, 1979; Sangesland and Lingwood, 2021), but they could be subsequently selected. For instance, germline alleles with innately high affinity for bacterial antigens might arise from long-term selection in vertebrate populations to recognize commonly encountered pathogens and commensals via broad classes of epitopes shared by these organisms (Yeung et al., 2016; Collins and Jackson, 2018; Sangesland et al., 2020).
Three lines of evidence support the idea that specific germline immunoglobulin alleles are better than others at binding particular antigens. First, structural characterization of individual antibodies shows that some variable (V) alleles can bind specific epitopes through germline-encoded motifs in complementarity-determining regions (CDRs) 1 and 2 (e.g., West et al. 2012; Pappas et al. 2014; Yeung et al. 2016; Yuan et al. 2020; Voss et al. 2021). These antigen-binding regions are encoded solely by the receptor’s V allele, whereas CDR3 spans the junction of V alleles with joining (J) and, in the case of the heavy chain, diversity (D) alleles. While epistatic interactions with other alleles might be important, germline-encoded motifs in CDRs1-2 could make specific V alleles more likely than others to bind specific antigens. A second line of evidence comes from experiments with transgenic mouse strains that each have a single heavy-chain V allele but multiple alleles in the other sets. Strains with specific V alleles have higher antibody titers against specific antigens than the other strains (Sangesland et al., 2019, 2020). A third line of evidence comes from sequencing studies, which often show that specific alleles, allele combinations or CDR3 sequences are overrepresented in the response to particular epitopes (reviewed by Dunand and Wilson 2015 and Sangesland and Lingwood 2021). Overrepresentation of specific alleles is often interpreted as a consequence of, and as indirect evidence for, the kind of germline-encoded specificity revealed by structural analyses or transgenic mouse experiments. In addition to different propensities to encode receptors with high affinity, overrepresentation of particular alleles might also reflect different potentials for subsequent adaptation during affinity maturation.
Yet the degree to which specific germline alleles are consistently overrepresented in individuals exposed to the same antigen varies widely for reasons that are poorly understood. In some cases, only a few germline alleles are represented in the response (e.g., Crews et al. 1981; Cumano and Rajewsky 1985, 1986; Guthmiller et al. 2021, 2022), suggesting that germline-encoded specificities strongly predict the outcome of B cell competition and lead to highly similar repertoires. In other cases, most germline alleles are used (e.g. Di Niro et al. 2015; Kuraoka et al. 2016; Nielsen et al. 2020; Robbiani et al. 2020; Sakharkar et al. 2021), only some of which appear overrepresented with respect to controls, suggesting that initial advantages of B cells with specific germline alleles do not strongly predict the outcome of B cell evolution and competition. Those studies vary not only in the complexity of the antigen and the type of B cell studied, but also in the amount of time since the exposure and thus the extent of affinity maturation. How the overrepresentation of specific alleles and the degree of similarity between individuals changes during the course of the response has not been systematically investigated.
Using simulations and experiments, we show that stochasticity and contingency in B cell evolution and competition counterbalance initial selection for receptors with specific germline alleles. Germline alleles can give B cell lineages an advantage over others early in the response, but which lineages ultimately dominate also depends on factors that are largely random with respect to the choice of germline allele, such as the occurrence and timing of mutations in different lineages. In simulations, these factors tend to produce increasingly contingent patterns of allele usage during the course of the response. We find patterns consistent with those dynamics in the B cell response of mice experimentally infected with influenza virus. Specific heavy-chain V alleles are consistently abundant in the repertoires of different infected mice early on, but allele usage becomes less consistent over time as large mutated lineages come to dominate the repertoire. Those dominant lineages use different germline V alleles in different mice, and lineages sharing the same allele rarely evolve the same somatic mutations. These results suggest that germline-encoded specificities and those evolved later by affinity maturation are important at different phases of the response. Selection to reinforce germline-encoded specificities in the long-term evolution of jawed vertebrates might be driven by the fitness benefits of responding rapidly to commonly encountered pathogens. The lack of consistency in germline allele usage or specific mutations later in the response also suggests no pronounced differences in the adaptability of different immunoglobulin genes.
Results
We begin by asking what factors might affect germline allele frequencies in the B cell response, before turning to a mathematical model to understand how those factors interact. By response we mean the set of B cell populations that bind a specific epitope, antigen or pathogen: germinal center cells, memory cells and short- and long-lived plasma cells. Since all those populations descend from naive B cells, germline allele frequencies in the response depend partly on allele frequencies in the naive repertoire, which vary widely between alleles but tend to be positively correlated between individuals (although heritable variation exists; Glanville et al. 2011; Watson et al. 2017; Collins et al. 2020). Here, we focus on factors that cause some alleles to be over or underrepresented in the response relative to their baseline frequency in naive B cells. How much a germline allele increases or decreases in frequency depends on how many naive B cells using that allele are activated, and how much they divide inside or outside germinal centers, relative to naive cells using other alleles. Within specific cell types, germline allele frequencies also reflect how often cells using each allele differentiate into each cell type. We start by focusing on germline allele frequencies in the response as a whole, across cell types.
There are two non-mutually exclusive reasons why the total number and size of B cell lineages involved in the response might vary between germline immunoglobulin alleles (Figure 1). The first is if using particular germline alleles tends to give B cell lineages an advantage over others. This advantage could be a higher initial affinity, a greater capacity to evolve high affinity during affinity maturation, or both:
Germline-encoded affinity
Since affinity is a property of the entire recombined receptor, not of its individual constituent alleles, naive B cells using a particular germline allele have a distribution of possible affinities depending on the choice of alleles from the other sets in the heavy and light chains (including insertions and deletions at the alleles’ junctions). If individuals have similar sets of germline alleles at similar frequencies in the naive repertoire, and if the probabilities of different allele combinations are similar, then the affinity distribution for any given germline allele will be similar in different individuals. Yet different germline alleles might have different affinity distributions, leading to different fitness distributions for naive B cells using each allele. For instance, naive B cells using a specific heavy-chain V allele might bind the antigen well across all combinations with other alleles via CDRs1-2 (Figure 1, orange allele), while naive B cells using a different V allele may bind poorly across the board (Figure 1, purple allele) or have low affinity overall but high affinity in certain combinations (Figure 1, green allele).
Germline-encoded adaptability
Like initial affinity for a particular antigen, the potential for a B cell receptor to evolve higher affinity is a property of the entire receptor, not of individual germline alleles. Yet receptors using different alleles might have different propensities to adapt (Figure 1), for instance if they tend to have different rates of beneficial and deleterious mutations. Variation in mutability occurs because the enzymes responsible for mutating the B cell receptor target different nucleotide motifs at different rates (Rogozin and Kolchanov, 1992; Rogozin and Diaz, 2004; Yaari et al., 2013; Wei et al., 2015), so variation in the motif composition of germline immunoglobulin alleles can lead to differences in the frequency and distribution of mutations. Variation in the relative probabilities of beneficial and deleterious mutations arises from epistasis: mutations are more likely to change affinity or disrupt the receptor’s function in some backgrounds than in others (Boyer et al., 2016; Schulz et al., 2021).
A second reason why specific germline alleles might become over or underrepresented is the role of chance and contingency in B cell activation, evolution and competition. Contingency means that although these processes are not random (since they are shaped by selection for affinity), their precise outcome depends on the occurrence, order and timing of events that are largely unpredictable (Gould, 1989; Beatty and Carrera, 2011; Blount et al., 2018). Which lineages come to dominate the response, and which germline alleles they use, will be contingent on how those events play out. Several sources of stochasticity in B cell dynamics could lead to contingent germline allele frequencies:
Stochasticity in B cell activation and in the colonization of germinal centers
Because the number of naive B cells is finite and the probabilities of different VDJ combinations vary by orders of magnitude (Elhanati et al., 2015), rare germline allele combinations with high affinity may be present in the naive repertoires of some individuals but not others. Even if present in most individuals, low-frequency, high-affinity germline-allele combinations might, simply by chance, be recruited only in some of them (for instance, if none of the rare naive cells come near the site of the response in a given individual).
Since germinal centers have a limited size, lineages that happen to arrive first might prevent others from establishing in the germinal center (similar to species competing for access to a site; Chase 2003; Fukami 2015). Whether such “priority effect” does occur in germinal centers is unknown.
Stochasticity in the timing, order and effect of mutations
Which lineages ultimately evolve the highest affinity and outcompete the others is contingent on the precise timing, order and effect of mutations in each lineage. For instance, the timing and effect of mutations affect the outcome of clonal interference – when multiple affinity-increasing mutations within a lineage or in different lineages compete for fixation (Desai and Fisher, 2007). Due to epistasis, the same B cell lineage could end up with very different affinities by acquiring mutations in different orders (Starr and Thornton, 2016).
Demographic stochasticity and genetic drift
Demographic stochasticity and genetic drift might be important, especially early in the response when population sizes are small. Demographic stochasticity might tip the balance of competition between lineages, driving some to extinction purely by chance. Genetic drift might cause new mutations to be fixed within a lineage even if they are neutral or deleterious or become extinct even if they are beneficial. The loss of newly arisen beneficial mutations due to drift is important even in large populations.
While both contingency and germline-encoded advantages can cause germline alleles to increase in frequency relative to the naive repertoire, only the latter are expected to produce consistent deviations in the response of different individuals exposed to the same antigen (provided individuals have similar sets of germline alleles) (Figure 1). The correlation in frequency deviations between individuals can therefore be used to measure how strongly germline-encoded advantages shape the outcome of B cell competition and evolution.
Similarity in germline allele frequencies reflects a balance between contingency and germline-encoded advantages
To understand how germline-encoded advantages interact with chance and contingency to shape the B cell repertoire, we used a stochastic mathematical model to simulate B cell evolution and competition in germinal centers (Methods: “Model of B cell dynamics”; Table 1). Rather than making quantitative predictions based on realistic parameter values, our goal was to investigate the qualitative behavior of germline allele frequencies in the response and their deviations from the naive repertoire under different scenarios. The model focuses on the subset of the B cell response derived from germinal centers (the canonical sites of somatic hypermutation and B cell evolution), without considering extrafollicular B cell populations that expand outside of germinal centers (although reports of selection and somatic hypermutation in those populations suggest they might have similar dynamics; Di Niro et al. 2015; Elsner and Shlomchik 2020). To simulate selection for affinity, B cells are stochastically sampled to immigrate or divide based on their affinity relative to other cells in the naive repertoire (in the case of immigration) or in the germinal center (in the case of division). Dividing B cells then undergo affinity-changing mutations with some probability. To represent variation in germline-encoded affinity and adaptability, B cells using different heavy chain germline V alleles can have different naive affinity distributions. The variation within each distribution in turn represents the effects of stochasticity in VDJ recombination. The model also allows different germline V alleles to have different mutation rates, representing one aspect of variation in adaptability. We simulated 20 individuals, each with 15 germinal centers. We based these simulated individuals on mice for which we empirically estimated the set of heavy chain V alleles and their frequencies in the naive repertoire. These mice typically had about 75 heavy-chain V alleles (60-70 of which were typically shared between a pair of mice), and allele frequencies in the naive repertoire were strongly correlated between mice (Figure 3A-B).
If all germline alleles have the same naive affinity distribution and the same mutation rate, the model predicts that allele frequencies will be positively correlated between individuals early in the response but less so over time (Figure 2, left column; Figure 2–Figure Supplement 1). The positive correlation early on arises from the correlation in naive allele frequencies between individuals: Assuming identical affinity distributions between germline alleles, the alleles tend to arrive in germinal centers in the same frequencies in which they occur in the naive repertoire. The subsequent decrease in allele-frequency correlations reflects the increasing role of stochasticity and contingency. Over time, due to selection, each germinal center tends to become dominated by the lineage with the highest affinity (Figure 2, left column, top row). With no differences in affinity or adaptability between germline alleles, which lineages ultimately evolve the highest affinity is completely random with respect to the choice of germline allele, and so allele-frequency correlations between individuals decrease while deviations from the naive repertoire (measured as the ratio between experienced and naive frequencies) remain uncorrelated throughout the response (Figure 2, left column, bottom row).
If some germline alleles tend to encode receptors with higher affinity than others, the model predicts that both allele frequencies and their deviations from the naive repertoire will be positively correlated between individuals early in the response, but chance and contingency reduce this correlation over time (Figure 2, middle column; Figure 2–Figure Supplement 2). Without somatic hypermutation, both types of correlation remain high over time as germinal centers are consistently dominated by B cell lineages using high-affinity alleles (Figure 2–Figure Supplement 3). As the rate or the effect size of somatic hypermutation increases, so does the opportunity for B cell lineages using low-affinity alleles to overcome the initial advantage of those using high-affinity alleles (Figure 2–Figure Supplement 3). In the model, precisely which low-affinity alleles are used by lineages that do so is a matter of chance, since all B cells have the same probability of acquiring benefi-cial mutations irrespective of the germline V allele they use. As a result, both types of correlation between individuals decrease over time. In practice, due to different fitness landscapes between germline alleles, germline-encoded advantages might be preferentially overcome by lineages with specific alleles, allele combinations or heavy-light chain pairings.
Finally, when some germline alleles have a higher mutation rate than others, B cell lineages with high-mutation alleles are likely to dominate germinal centers in the long term due to their propensity to adapt (Figure 2–Figure Supplement 5), countering the tendency for allele frequencies to become less correlated over time and leading to a positive correlation in frequency deviations later in the response (Figure 2, right column; Figure 2–Figure Supplement 4).
Contingent allele frequencies in the mouse response to influenza infection despite evidence of germline-encoded advantages
We compared these simulated dynamics with the B cell response of C57BL/6 mice infected with influenza virus once or twice and sacrificed at different times points (8, 16, 24, 40 and 56 days after the primary infection, with the secondary infection on day 32; Materials and Methods: “Experimental infection of mice with an influenza A/H1N1 virus”). Because influenza viruses do not naturally infect mice, any germline-encoded specificities for influenza antigens are either evolutionary spandrels or the product of selection to recognize molecular patterns shared between influenza and pathogens that have historically infected mice. We used RNA sequencing to estimate the frequencies of germline alleles and the relative sizes of B cell lineages in each mouse. We focused on heavy-chain sequences sampled from the mediastinal lymph node because, consistent with previous work (Sealy et al., 2003), cell sorting data indicated that lymph node B cells were induced by the influenza infection (control mice had very few germinal center, plasma or memory cells in the mediastinal lymph node; Figure 3–Figure Supplement 1). Early in the mouse response to influenza, lymph node populations likely consist of extrafollicular plasma cells expanding outside of germinal centers, with germinal-center derived cells arriving later (Sealy et al., 2003) and persisting for as long as six months (Yewdell et al., 2021). Most germline V alleles observed in a mouse (across all tissues and cell types sampled) were represented in the influenza-induced lymph node populations, suggesting that most mouse V alleles can produce at least some receptors capable of binding influenza antigens (Figure 3). To compare the observed mouse responses with our simulations, we measured the correlation in germline V allele frequencies and in their deviations from the naive repertoire between pairs of infected mice (Materials and Methods: “Estimating correlations between mice”).
As expected, influenza infection led to competition and affinity maturation in mouse B cell lineages (Figure 3–Figure Supplement 2). Serum antibody titers against the infecting virus measured by ELISA rose about 1,000 fold between days 8 and 24 and remained high. In parallel to this rise in antibody titers, germinal center and plasma cell populations became increasingly dominated by a few lineages, suggesting that lineages varied in fitness due initial differences in affinity, differences acquired during the lineages’ subsequent evolution, or both. Lineages sampled at later time points had more high-frequency amino acid mutations within them (those present in 50% or more of the reads in a lineage). Those mutations include fixed mutations and those potentially rising to fixation via selection for affinity, and they are unlikely to have arisen from sequencing and amplification errors (which we estimate at 1.8 per thousand nucleotide bases; Materials and Methods: “B cell receptor sequencing”). These trends were visible in the lymph nodes of infected mice but not apparent in other tissues or in control mice (Figure 3–Figure Supplement 3), suggesting they were driven by the influenza infection. (Influenza-specific lineages may have been present in other tissues, but our data do not allow us to distinguish them from lineages elicited by other antigens.)
Plasma cells and germinal center cells were ultimately dominated by lineages using different germline V alleles in different mice, consistent with the role of contingency observed in our simulations. Early in the response, germline allele frequencies in those cell types were correlated between mice (Figure 4A, left panel). In both cell types, this initial similarity was likely partly due to the correlated germline frequencies in the naive repertoire (Figure 3B). In early plasma cells, it also re-flected the consistent overrepresentation of specific germline alleles, suggesting that those alleles contributed to higher affinity or adaptability than did others (Figure 4A, Figure 4B). For instance, in day-8 plasma cells, IGHV14-4*01 increased in frequency relative to the naive repertoire in all 6 mice with enough data, becoming the most common V allele in 4 mice and the second most common in the other 2 (Figure 4B). In contrast, at later time points for plasma cells (Figure 4–Figure Supplement 1) and throughout the response for germinal center cells (Figure 4–Figure Supplement 2), the most common V allele was usually different in different mice, and most germline alleles were overrepresented relative to the naive repertoire in some mice but not in others. These results suggest that while germline-encoded advantages may strongly shape the early B cell response, they do not predict B cell fitness in the long run.
To further test if the effect of germline-encoded advantages was strongest early in the response, we compared the observed patterns with a null model in which a lineage’s fitness is independent of which germline V allele it uses, mimicking the equivalent-alleles scenario in our simulations. We did so by keeping the observed distribution of lineage sizes (a proxy for lineage fitness) while randomly assigning each lineage’s germline V allele based on naive repertoire frequencies. Early in the plasma cell response, germline alleles were overrepresented much more consistently between mice than expected under this null model, suggesting that the early response was strongly shaped by the advantages associated with using those alleles (Figure 4–Figure Supplement 3). Later in the response, however, specific alleles were not overrepresented in different mice more often than expected if lineage fitness was independent of the germline V allele.
In contrast with germinal center and plasma cells, germline allele frequencies in memory cells remained similar between mice (Figure 4A) — and similar to naive allele frequencies within each mouse (Figure 3C) — throughout the response. Differences between memory cells and the other cell types could be expected if a higher fraction of memory cells is unrelated to the influenza infection (uninfected controls had more memory B cells than plasma or germinal center cells in their lymph nodes, although they had fewer lymph node memory cells than did infected mice; Figure 3– Figure Supplement 1). In addition, these differences between cell types might reflect the relationship between affinity and B cell differentiation. Since activated B cells with low affinity are more likely than others to exit germinal centers and differentiate into memory cells (Viant et al., 2020), dominant lineages with high affinity for influenza antigens might contribute less to the memory cell population than they do to the germinal center and plasma cell populations. Consistent with that possibility, the increasing dominance by a few large and mutated lineages seen in germinal center and plasma cells of infected mice was not evident in their memory cells (Figure 3–Figure Supplement 2). Without dominance by a few lineages and whatever germline alleles they happen to use, germline allele frequencies in memory B cells might not stray as far from naive repertoire frequencies as do germline allele frequencies in the other cell types. Consistent deviations from the naive repertoire still occur (Figure 4A, Figure 4–Figure Supplement 4), as would be expected if using specific germline alleles makes B cell activation more likely. Of the activated B cells using those germline alleles, those cells with lower affinity than the rest might then differentiate into memory B cells soon after activation. For instance, IGHV14-4*01, which was consistently overrep-resented relative to the naive repertoire in early plasma cells, was also overrepresented in the memory cells of 50% or more of the mice at every time point (Figure 4–Figure Supplement 5).
Germline V alleles consistently overrepresented early in the response have low predicted mutability in CDRs
Our simulations suggest that the consistent overrepresentation of specific germline alleles early in the response is more likely to reflect germline-encoded affinity than germline-encoded adaptability (Figure 2). With sequence data alone, we cannot determine if consistently overrepresented germline V alleles do generate receptors with especially high affinity for influenza antigens. We can, however, estimate potential differences in adaptability between germline alleles based on their sequences alone, using estimates of the propensity of different nucleotide motifs to undergo somatic hypermutation (although those estimates were derived from mouse light-chain rather than heavychain genes; Cui et al. 2016) (Methods: “Mutability analysis of germline alleles”).
We found no clear evidence that germline alleles with higher predicted mutability in the CDRs (which might give those alleles a higher rate of affinity-changing mutations) tended to increase in frequency relative to the naive repertoire (Figure 4–Figure Supplement 6). Neither did germline alleles with lower mutability in the structurally-important framework regions (FRs; where mutations are more likely to be deleterious than in CDRs) tend to increase in frequency. Instead, in day-8 plasma cells we found the opposite relationship: germline alleles tended to increase in frequency relative to the naive repertoire if they had high mutability in FRs and low mutability in CDRs. The consistently overrepresented and dominant allele IGHV14-4*01, for instance, is predicted to be one of the least mutable in CDRs 1 and 2, (Figure 4–Figure Supplement 6). Two other germline alleles consistently overrepresented in day-8 plasma cells, IGHV1-82*01 and IGHV1-69*01 (5 of 6 mice with enough data), have similarly low predicted mutability in CDR1, though not in CDR2. If those alleles do have a high propensity to bind influenza antigens, low mutability in CDRs 1 and 2 might reduce the chance that mutations disrupt this initial binding, potentially reinforcing the fitness advantage of B cells using those alleles.
B cell lineages sharing the same germline V allele rarely had mutations in common
While germinal center cells and plasma cells were increasingly dominated by large lineages with somatic mutations, the sheer number of mutations acquired by a B cell lineage did not predict its success. The biggest lineages in the influenza-induced B cell populations did not generally more mutations than smaller lineages (Figure 3–Figure Supplement 5). This observation is consistent with previous work showing that the number of mutations in the B cell receptor does not predict affinity or neutralization strength (Viant et al., 2020; Sakharkar et al., 2021; Neumeier et al., 2021). Thus, successful lineages might be those that acquire one or a few substitutions with large effects on affinity, instead of many substitutions with smaller effects.
We found no tendency for these mutations to be the same in B cell lineages using the same germline V allele. Most pairs of lineages with the same V allele had no high-frequency mutations in common (Figure 3–Figure Supplement 4). For specific cell types and specific V alleles, we found some instances of high-frequency mutations shared by multiple lineages. However, they were constrained to one or two mice, suggesting they might be an artifact of the incorrect partitioning of a single large lineage into several small ones. Overall, these results suggest that influenza infection does not strongly select the same mutations in B cell lineages with the same germline V alleles. Multiple ways to improve affinity might be possible for the same germline V allele, especially if epistatic interactions between the V segment and the other segments cause the same mutation to have different effects in different lineages.
Limited evidence of selection for specific CDR3 sequences
While binding can occur via the two CDRs solely encoded by the V segment, it often occurs via CDR3, which spans the junction between the segments. Thus, while selection for receptors with specific germline V alleles seems to have a limited effect, influenza antigens might also select for receptors with specific CDR3 sequences. To investigate this possibility, we computed the amino acid sequence and biochemical similarity of CDR3 sequences sampled from different mice and matched for the same length (Methods: “Measuring CDR3 sequence similarity”). On average, length-matched CDR3 sequences from the influenza-induced populations of different mice were no more similar than sequences sampled from their naive repertoires (Figure 3–Figure Supplement 6). This result suggests that influenza infection in mice does not strongly select for B cell receptors with particular CDR3 sequences. While individuals exposed to the same pathogen are more likely to share specific CDR3 sequences compared with healthy individuals (Ortega et al., 2021), our results suggest those convergent CDR3s do not make up a large fraction of the response.
Finally, influenza antigens might select for combinations of specific germline alleles and CDR3 sequences (Jackson et al., 2014; Harshbarger et al., 2021). To investigate this possibility, we computed the similarity of CDR3 sequences from different mice matched both for the same length and the same germline heavy-chain V allele. Length- and allele-matched CDR3 sequences from the lymph node populations of different mice were not, overall, more similar than length- and allele-matched sequences from the naive repertoire (Figure 3–Figure Supplement 6). This result suggests that, overall, binding influenza antigens with a specific germline V allele does not require specific CDR3 sequences, even if specific combinations of germline alleles and CDR3 sequences can be found in the response of different individuals (Jackson et al., 2014). In day-56 plasma cells, we found higher similarity between length- and allele-matched CDR3 sequences than in the naive repertoire, driven by two clusters of sequences with different lengths (each using a small set of V alleles; Figure 3–Figure Supplement 6, Figure 3–Figure Supplement 7). Collectively, these CDR3 sequences occurred at a very low frequency throughout the response but made up a significant proportion of plasma cell sequences on day 56. Why these combinations of specific CDR3 sequences and specific alleles might have been selected only late into the secondary response is unclear.
Discussion
How much effective B cell responses depend on particular germline immunoglobulin genes versus their subsequent evolution by affinity maturation has important consequences for adaptive immune evolution and vaccination strategies but remains understudied. In simulations, we find that initial germline-encoded advantages are mostly overcome by B cell lineages using different germline alleles in different individuals. These contingent outcomes arise because the growth of B cell lineages also depends on factors that are largely unpredictable, including the timing, order and effect of mutations in different lineages, genetic drift, demographic stochasticity and stochasticity in VDJ recombination. Our simulations and experiments suggest that the effects of contingency increase over time: The longer B cell lineages evolve, the more opportunity there is for differences to accumulate as those processes play out. Like evolution in general (Gould, 1989; Blount et al., 2018; Xie et al., 2021; Park et al., 2022), the evolution of B cell repertoires in different individuals experiencing identical primary infections might become decreasingly predictable at the genetic level over time. Yet, as is often the case in other systems (Lässig et al., 2017), the resulting phenotype is remarkably predictable: potent antibodies reliably emerge in most individuals, suggesting there are many different ways to achieve high affinity against the same pathogen.
A testable prediction suggested by those results is that germline allele usage might diverge between people following repeated exposures (such as sequential influenza infections or vaccines) or over time during chronic infections (such as HIV). This prediction depends on the extent to which responses to repeated or prolonged infections rely on the reactivation of preexisting memory cells and their reentry into germinal centers (Li et al., 2012; Andrews et al., 2015; Mesin et al., 2020; Turner et al., 2020; Hoehn et al., 2021). Divergence in germline allele frequencies might be small if the response to each exposure is dominated by lineages newly recruited from the naive repertoire. In contrast, successive bouts of evolution by recalled B cell lineages might increase the chance that they overcome germline-encoded advantages. Vaccine strategies focused on the recruitment of specific alleles (McGuire et al., 2014; Jardine et al., 2016; Bonsignori et al., 2017; Lin et al., 2020) might be hindered in their immediate goal by contingent patterns of allele usage in different people, especially if the strategy involves multiple immunizations or immunizations in people with extensive immune memory.
In addition to varying over time, the similarity of the induced B cell repertoire might also vary with the complexity of the antigen, since antigens encoding multiple epitopes present more potential specificities. Although certain amino acid motifs can make some germline alleles highly polyreactive (Hwang et al., 2014; Shiroishi et al., 2018), individual alleles might be unlikely to have a consistent advantage over others across all epitopes in an antigen or all antigens in a pathogen. Thus, germline-encoded specificities might be more apparent in B cells specific for a single epitope than in the set of all B cells binding the antigen or pathogen. Some previous observations are consistent with this hypothesis. For instance, the response to haptens (simple antigens with few potential epitopes) tends to be dominated by one or a few alleles (Cumano and Rajewsky, 1985, 1986), while the response to complex antigens can use many (Kuraoka et al., 2016). Only a handful of germline alleles are represented in monoclonal antibodies specific for narrowly defined sites on influenza hemagglutinin (Guthmiller et al., 2021, 2022), while tens of alleles are present in monoclonal antibodies that bind different sites on the major domains of the SARS-CoV-2 spike protein (Robbiani et al., 2020; Sakharkar et al., 2021). Many germline V alleles are found in monoclonal antibodies against the IsdB protein of Staphylococcus aureus, but antibodies targeting each particular epitope tend to use only one or two of them (Yeung et al., 2016).
Understanding this variation further requires overcoming limitations of our analyses. Similarity between individuals might decay even faster in genetically diverse outbred populations than in the inbred mice we used in the experiments. In simulations, we varied the strength of contingency by varying the frequency and the effect size of somatic mutations relative to the variation in affinity from VDJ recombination alone. Although genetic drift, demographic stochasticity and priority effects in the colonization of germinal centers were present in our model, we did not systematically explore their impacts. Understanding the importance of those processes might require longitudinal data to resolve the timing of cell arrivals in germinal centers and the lineages’ population dynamics early in the response. Complementing our sequence analysis, affinity measurements could be used to estimate the affinity distributions of naive B cells using different germline alleles, compare variation within and between those distributions, and directly test if alleles with higher-affinity distributions tend to be used by B cell lineages with high growth rates. Affinity measurements could also be used to study germline allele usage in pathogen-specific B cell lineages outside of lymph nodes, which we could not identify with sequence data alone.
Finally, if germline-encoded specificities are most consequential early in the response, longterm selection to “pre-adapt” germline genes might be linked to the benefits of responding rapidly to commonly encountered pathogens. Mathematical models suggest that maintaining innate defenses against a particular pathogen becomes more advantageous the more frequently the pathogen is encountered (Mayer et al., 2016). Germline alleles specific to common pathogens or pathogenic motifs might be selected, effectively hardcoding innate defenses into the adaptive immune system (Collins and Jackson, 2018). A reliable supply of receptors against common enemies might be especially important in small and short-lived organisms, which can more quickly die of infection and have fewer naive B cells with which to cover the vast space of possible pathogens (Collins and Jackson, 2018). Reinforcing germline-encoded specificities might also be especially useful when the opportunity for adaptation is limited, as might be the case for pathogens that induce extrafollicular responses without extensive B cell evolution (although affinity maturation can occur outside of germinal centers; Di Niro et al. 2015; Elsner and Shlomchik 2020). Understanding what conditions favor similar versus contingent allele usage in the antibody repertoire may thus shed light on the long-term evolution of immunoglobulin genes.
Materials and Methods
Model of B cell dynamics
We modeled B cell evolution and competition in germinal centers using stochastic simulations based on a Gillespie algorithm. There are three types of independent events in the model: immigration of individual B cells into germinal centers, cell division and death. The total rate of events λ is given by where the terms on the right-hand side correspond to the rate of each kind of event (mutation is associated with cell division and is therefore not an independent event). The algorithm consists of drawing the time to the next event by sampling from an exponential distribution with rate λ. Once an event has occurred, we a make a second draw to determine its type. The probability for each type of event in this second draw is proportional to the corresponding event-specific rate (e.g., the probability that the next event is a cell division is λdivision/λ). After determining the event type, we update event rates and draw the time to the next event, and so on until a maximum time tmax is reached. For each germinal center, we record the number of cells in each B cell lineage (and the V alleles used by the lineages) at the end of day 1 and then every 5 days starting on day 5.
Immigration of B cells into germinal centers is restricted to an initial period with duration timm. Parameter Itotal controls the expected number of lineages that enter each germinal center during that time (each recruited B cell is the founder of an individual lineage). Given those parameters, we let λimmigration be a linearly decreasing function over time reaching 0 at timm, with intercept and slope chosen such that Itotal lineages are expected to enter each GC by that point (λimmigration then remains at 0 until the end of the simulation).
Once an immigration occurs, we randomly sample a single immigrant from a newly generated recruitment pool of 1,000 naive cells whose V alleles are drawn with replacement from the naive repertoire. For each member of the recruitment pool, we sample an affinity value based on the naive affinity distribution associated with its V allele. By default, all alleles have the same normal affinity distribution with mean and standard deviation equal to 1 (we sample from the associated truncated distribution to avoid negative values). Depending on the scenario, naive B cells using specific V alleles may have a different distribution with mean 1 + s and the same standard deviation. The probability that each cell in the recruitment pool is chosen as the new immigrant is then proportional to its affinity.
The rate of cell divisions depends on the total number of cells inside the germinal center, N, and on the rate of cell division for each individual cell, μ(N):
To represent competition for antigen, μ(N) decreases with N so that it equals a fixed per-cell death rate δ when the population is at carrying capacity (N = K):
Once a division event occurs, we randomly sample a B cell to divide. The probability that each is B cell is chosen is proportional to its affinity. Each dividing B cell has some probability of having a mutation that changes affinity by a normally distributed amount with mean 0 and standard deviation β (affinity is set to 0 if the mutation produces a negative value).
Finally, with a fixed per-cell death rate, the population-level death rate is given simply by
When simulating alleles with higher naive affinity or higher mutation rates than others, we chose the set of 5 alleles present in all mice with average naive frequency of 2-3% (the typical median frequency in the naive repertoire).
Experimental infection of mice with an influenza A/H1N1 virus
We infected 40 8-week-old female C57BL/6 mice weighing 20-22g (8 for each time point) intranasally with 0.5 LD50 of a mouse-adapted pandemic H1N1 strain (A/Netherlands/602/2009) in a total of 30 μL of PBS under full anesthesia. In addition, two controls for each time point were given PBS only. All mouse experiments were approved by The University of Chicago Institutional Animal Care and Use Committee (IACUC protocol 71981).
Tissue processing, cell sorting and nucleic acid extraction
We prepared single cell suspensions from the mediastinal lymph node, spleen and both femurs harvested at the indicated time points. B cells were first enriched from the splenocyte suspension by MACS (magnetic activated cell sorting) using the Pan B cell Isolation Kit (Miltenyi Biotec), followed by staining for FACS (fluorescence activated cell sorting). The lymph node and bone marrow cells were directly stained for FACS. Antibodies used for sorting were anti-B220 (clone RA3-6B2, Biolegend), IgD (clone 11-26c.2a, Biolegend), anti-CD4 (clone RM4-5, Biolegend), anti-CD8 (clone 53-6.7, Biolegend), anti-CD38 (clone 90, Biolegend), anti-CD95 (clone Jo-2; BD Biosciences), anti-CD138 (clone 281-2, Biolegend), anti-F4/80 (clone BM8, Biolegend), anti-GL7 (clone GL7, BD Biosciences), anti-Sca-1 (clone D7, Biolegend), and anti-TER-119 (clone TER-119, Biolegend). Antibody stainings were preceded by adding Fc block (anti-CD16/CD32; clone 2.4G2, BD Biosciences). For sorting, the cells were first gated on size and granularity (forward and side scatter, respectively) to exclude debris, followed by doublet exclusion. We sorted naive (IgD+B220+), plasma (IgD-Sca-1hiCD138hi), memory (IgD-B220+CD95-CD38hi) and germinal center (IgD-B220+CD95+ CD38loGL-7+) cells after excluding cells expressing CD4, CD8, TER-199 or F4/80 (to exclude T cells, erythroid cells and macrophages). After spinning down cells and removing the PBS supernatant, we extracted DNA and RNA from the cell pellets using the AllPrep DNA/RNA Mini Kit (Qiagen), according to the manufacturer’s protocol. All samples were kept frozen until sequenced.
B cell receptor sequencing
We generated immunoglobulin heavy chain (IGH) DNA libraries from complementary DNA generated from 10-500 ng of total RNA using Superscript III (Invitrogen) reverse transcriptase and random hexamer primers. For PCR amplifications, we used multiplexed primers targeting the mouse framework region 1 (FR1) of IGHV in combination with isotype-specific primers targeting constant region exon 1 of IgA, IgD, IgE, IgG, or IgM (Table 2). We performed separate PCR reactions for each isotype to avoid formation of inter-isotype chimeric products. We barcoded each sample with 8-mer primer-encoded sequences on both ends of the amplicons and performed PCR amplification in two steps. First, we generated amplicons using primers with the partial Illumina adapter, the sample-specific barcode and the locus-specific sequence. In the second step, we performed another PCR to complete the Illumina adapter sequence and to ensure final products were not amplified to saturation. We purified pooled products by agarose gel electrophoresis and extraction. We used a 600 cycle v3 kit to sequence products using an Illumina MiSeq instrument.
We estimated the rate at which errors were introduced during amplification and sequencing by comparing the sequenced reads with the reference sequence for the corresponding isotype. Because the constant region does not undergo somatic hypermutation, we counted each mismatch between the end of the J gene and the beginning of the conserved region primer as an error introduced by sequencing and amplification. Based on 187,500 errors found out of 104,092,368 bases analyzed, we estimated the error rate to be 1.80 mutations per thousand bases (95% binomial CI 1.79-1.81).
ELISA
We coated 96-well ELISA plates (Thermo Fisher Scientific) overnight at 4°C with eight hemagglutination units (HAU) of virus in carbonate buffer. We used horseradish peroxidase (HRP)-conjugated goat anti-mouse IgG antibody (Southern Biotech) to detect binding of serum antibodies, followed by development with Super Aquablue ELISA substrate (eBiosciences). We measured absorbance at 405 nm on a microplate spectrophotometer (Bio-Rad). We analyzed serum samples starting at a top dilution of 1:20 (PBS controls and day 8 animals) or 1:1000 (all other samples), followed by 2-fold dilutions in 8 (PBS controls and day 8 animals) or 16 steps. We determined the end titer as the last dilution point with an OD value of > 2x the blank average OD value for each respective plate.
Estimating the frequencies of V alleles and B cell lineages
We used partis v0.15.0 to partition sequences into lineages and identify the germline alleles used by each lineage’s naive ancestor (accounting for variation in the set of germline alleles present in each mouse; Ralph and Matsen 2016a,b, 2019). We used the fraction of reads corresponding to each allele as a proxy for the frequency of that allele in each B cell population. To reduce the error in frequency estimates, we excluded B cell populations with fewer than 100 reads in a mouse. Since we did not barcode individual cells or RNA molecules during sequencing, the number of reads with a particular sequence reflects not only the number of B cells with that sequence but also their transcription levels. However, we found similar results using the number of unique sequences to estimate the abundance of each lineage or allele (i.e., counting multiple identical reads from the same mouse, tissue, cell type and isotype only once; Figure 4–Figure Supplement 7).
We measured the size of each lineage in each lymph-node B cell population as the number of reads from that lineage in that population (as opposed to the number of reads in the lineage across all cell types and tissues). B cell lineages were mostly confined to a single tissue and usually dominated by a single cell type (Figure 3–Figure Supplement 8; note that the partitioning of sequences into lineages was agnostic to cell type and tissue).
While we initially considered Dump-IgD+B220+ cells as naive cells, we noticed that many sequences obtained from them were extensively mutated relative to their inferred germline genes and were also inferred to be part of large clonal expansions. To exclude reads originating from non-naive B cells sorted as IgD+B220+, we considered a read as likely coming from a naive cell if it met all of the following criteria: 1) it came from IgD+B220+ samples; 2) its isotype was IgM or IgD; 3) it belonged to a clone that had a single unique sequence across its reads (and the reads all came from IgD+B220+ samples), and 4) that sequence had at most two nucleotide mutations in the V gene region. To compute naive frequencies, we pooled sequences meeting those criteria across all tissues. When computing experienced-to-naive frequency ratios, we adjusted the frequencies of germline alleles that were sampled in an experienced B cell population but not in naive B cells, since those alleles must have been present in naive B cells even though they were not sampled. In those cases, we imputed a single sequence to the allele in the naive repertoire then recalculated naive allele frequencies accordingly. When computing frequency deviations from the naive repertoire, we excluded mice with fewer than 100 naive reads even if the corresponding experienced population had more than 100 reads.
To test if our results were robust to uncertainty in the identification of naive B cells in our data, we alternatively estimated V allele frequencies from naive B cells (CD138-CD19+IgD++IgM+CD23++ CD21+PI-) sampled by Greiff et al. (2017) from the spleen of healthy C57BL/6 mice. For these data, we processed raw paired-end reads using presto v.0.6.2. (Vander Heiden et al., 2014), then used partis v0.15.0 to identify germline V alleles for a random sample of 20,000 sequences per mouse. V allele frequencies (measured by the fraction of total reads assigned to each gene in each mouse) were positively correlated between this independent dataset and the designated naive populations from our data (mean Spearman correlation coefficient between pairs of mice from each dataset = 0.68, interquartile range 0.60 − 0.77). We repeated the analysis of pairwise frequency-deviation correlations over time after replacing naive frequencies in our mice with the average frequency of each gene in the Greiff et al. (2017) dataset, preserving the number of reads in each mouse. When calculating the average allele frequencies in the alternative data set, we artificially assigned a single read to alleles present in our mice but absent from the alternative data set (since genes present in the experienced cells cannot be entirely missing from the naive repertoire) and re-normalized frequencies so they would sum to 1. Frequency deviations calculated based on this alternative data set were similar to those estimated using our own data (Figure 4–Figure Supplement 8).
Estimating correlations between mice
We used Pearson’s correlation coefficient to measure the correlation between mice in germline allele frequencies and their deviations from the naive repertoire. Pearson’s coefficient was better able to discriminate between different scenarios than Spearman’s coefficient, which measures the correlation in frequency deviation ranks instead of using the actual values. In simulations, using Spearman’s coefficient leads to a positive correlation in frequency deviations between individuals even in the scenario where all alleles are functionally equivalent (Figure 2–Figure Supplement 6). This pattern is driven by the exclusion from the response of alleles with very low naive frequencies (which tend to be the same alleles in different individuals), as those alleles are unlikely to be represented in lineages that successfully establish in germinal centers (repeating the simulations assuming all alleles have identical frequencies in the naive repertoire eliminates this pattern; Figure 2–Figure Supplement 6).
Identifying overrepresented germline alleles
To determine which germline alleles were consistently overrepresented in experienced B cell populations relative to the naive repertoire, we compared the frequency deviations for each germline allele (separately for each type of B cell) with the distribution expected if alleles were sampled based on naive frequencies alone (maintaining the observed the number of sequences in each mouse). For each germline allele, we then counted the number of mice with stronger deviations from the naive repertoire than expected under this null distribution (using a 95% bootstrap confi-dence interval).
Mutability analysis of germline alleles
To estimate the mutability of mouse germline V alleles, we used the RS5NF mutability scores estimated by Cui et al. 2016 using non-functional mouse kappa light-chain sequences and implemented in R package shazam. These scores describe the relative mutability of all possible 5-nucleotide motifs. We estimated the mutability of each framework region (FR) and complementarity-determining region (CDR) as the average score across motifs in the region. We then calculated an average for all FRs weighted by the length of each FR, and similarly for CDRs. We used igblast v1.14.0 to identify the FRs and CDRs of each germline V allele sequence identified by partis.
Measuring CDR3 sequence similarity
We compared the amino acid sequence similarity and biochemical similarity of pairs of CDR3 sequences sampled from different mice and matched either for length alone or both for length and V allele. To limit the number of comparisons, we proceeded as follows. For each pair of mice, we chose one mouse and sampled 500 sequences of the same cell type. For each sequence length represented in this sample, we paired sequences from the first sample with randomly chosen sequences of the same length from the second mouse. If matching sequences both for length and V allele, we did this second sampling separately for each combination of V allele and sequence length present in the first sample. This procedure matches sequences while preserving the length distribution (or the joint distribution of length and V alleles) in the first sample.
We measured amino acid sequence similarity as the proportion of sites with the same amino acid in both sequences. Following previous work (Hershberg and Shlomchik, 2006; Saini and Hershberg, 2015), we measured biochemical similarity as the proportion of sites in which the amino acids of both sequences belonged to the same category in the classification by (Chothia et al., 1998): hydrophobic (F, L, I, M, V, C, W), hydrophilic (Q, R, N, K, D, E) or neutral (S, P, T, A, Y, H, G).
Data and code availability
Code for the analyses is available at http://github.com/cobeylab/v_gene_selection. Data, intermediate files and results are available on Zenodo (https://doi.org/10.5281/zenodo.7080191). Raw fastq files are available on SRA [Accession number pending].
Competing interests
C.T.S. has consulted for Alvea / Telis Bioscience Inc. on the design of universal influenza vaccines. The other authors report no competing interests.
Supplementary information
Acknowledgments
Lauren McGough and Christopher Russo made helpful comments and suggestions to a draft of the manuscript. This project has been funded in part with Federal funds from the National Institute of Allergy and Infectious Diseases, National Institutes of Health, Department of Health and Human Services under grant DP2 AI117921, CEIRS Contract No. HHSN272201400005C and CEIRR contract 75N93021C00015 to S.C., grants R01AI127877, R01AI130398, U19AI057229 to S.D.B and grants U19AI082724, U19AI109946, U19AI057266, CEIRS contract HHSN272201400005C, CEIRR contract 75N93019R00028 and CIVIC contract 75N93019C00051 to P.C.W. The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIAID or the National Institutes of Health. The project was also partly supported by a Complex Systems Scholar Award from the James S. McDonnell Foundation awarded to S.C. M.C.V. was partly supported by a William Rainey Harper Dissertation Fellowship awarded by the University of Chicago. This work was completed in part with resources provided by the University of Chicago Research Computing Center.