Understanding Memory B Cell Selection

For more information:ali.ma@wecistanche.com

ABSTRACT

The mammalian adaptive immune system has evolved over millions of years to become an incredibly effective defense against foreign antigens. The adaptive immune system's humoral response creates plasma B cells and memory B cells, each with its own immunological objectives. The affinity maturation process is widely viewed as a heuristic to solve the global optimization problem of finding B cells with high affinity to the antigen. However, memory B cells appear to be purposely selected earlier in the affinity maturation process and have a lower affinity. We propose that this memory B cell selection process may be an approximate solution to two optimization problems: optimizing for affinity to similar antigens in the future despite mutations or other minor differences and optimizing to warm start the generation of plasma B cells in the future. We use simulations to provide evidence for our hypotheses, taking into account data showing that certain B cell mutations are more likely than others. Our findings are consistent with memory B cells having a high affinity to mutated antigens but do not provide strong evidence that memory B cells will be more useful than selected naive B cells for seeding the secondary germinal centers.

Click to Cistanche NZ for memory

1 Introduction

The immune system is an effective threat mitigation system that deploys several learned identification algorithms. While the innate immune system is adept at identifying foreign invaders, or antigens, it must engage the adaptive immune system to create a more massive and specific response. A core aspect of the adaptive immune system’s humoral response is training two types of B cells through a process called affinity maturation (AM): plasma B cells which generate antibodies to identify the current antigen, and memory B cells which are used in subsequent immune responses to identify similar antigens in the future. The AM process is highly unusual, in that a specific region of DNA within participating B cells is mutated to generate offspring which are selected to have a higher affinity to the antigen in question. The preservation of DNA sequences is usually of utmost importance in most cells, but the region of the genome which defines the shape of the B cell receptor must be rapidly modified for the B cell receptor to have a chance of becoming better at recognizing the antigen of interest (Meyer-Hermann et al., 2012). These mutations are responsible for the B cells’ incredible ability to recognize practically any antigen that they are presented, making the mammalian adaptive immune system one of the most effective learning identification systems in the natural world.

Stephen Lindsly, Maya Gupta b, Cooper Stansbury, Indika Rajapakse

a Department of Computational Medicine and Bioinformatics, University of Michigan, Ann Arbor, United States

b Google Research, Mountain View, CA, United States

c Department of Mathematics, University of Michigan, Ann Arbor, United States

In this paper, we consider whether the plasma B cell and memory B cell generation processes can be interpreted as trying to satisfy specific objectives, and if so, can we state these objectives precisely? We borrow standard ideas from machine learning, where it is common to first specify an ideal mathematical object to be optimized (such as minimizing the expected error rate of a learned identification system), then propose heuristic algorithms that approximately optimize that mathematical objective. Similarly, we hypothesize that due to evolutionary pressures, the AM processes act like heuristics that approximately optimize for idealized immunological objectives. We hypothesize what those evolutionarily-adaptive immunological objectives might be, then compare how well different B cells satisfy these objectives when faced with adversarially-mutated antigens via simulations. These findings lead us to propose new hypotheses about the implicit objectives of the immune system's training of naive B cells to become memory B cells.

First, in Section 2, we review how naive B cells are recruited and trained to become plasma B cells and present a hypothesis for the objectives of this training mathematically. In Section 3, we consider the mathematical optimization objectives of the more enigmatic training process that leads to the generation of memory B cells. We define plasma B cell training or training plasma B cells as the process of generating plasma B cells from naive B cells during AM. Similarly, we define memory B cell training or training memory B cells as the process of generating memory B cells from naive B cells during AM. We test our hypotheses via simulations in Section 4 and conclude with a discussion of open questions in Section 5.

2. Plasma B cell training

We review the AM process that trains naive B cells to become plasma B cells, then consider what mathematical criteria the plasma B cell training may have evolved to optimize.

2.1. Affifinity maturation of plasma B cells

AM begins by recruiting naive B cells with some initial affinity to the antigen to secondary lymphoid organs. These naive B cells, along with T follicular helper (TFH) cells and follicular dendritic cells, concentrate into temporary structures known as germinal centers (Fig. 1A) (Meyer-Hermann et al., 2012; Tas et al., 2016). Germinal centers (GCs) ensure that these cells are nearby, thus facilitating rapid mutation and evaluation of B cells receptor sequences. During AM, B cells are evaluated by TFH cells for their affinity to the antigen through the length of interaction between them, based on antigen presentation by B cells (Mesin et al., 2016; Murphy and Weaver, 2016). If the initial affinity of a B cell is high, it receives a chemical signal from the TFH cell to move to a separate area of the GC and proliferate. While the B cell is proliferating, a specific section of the genome called the hypervariable region is exposed to an enzyme, activation-induced cytidine deaminase (AID) (Muramatsu et al., 2000; Bannard and Cyster, 2017). AID can deaminate cytosine creating uracil, a nucleotide that is not normally found in DNA. The operation that repairs these changes is error-prone, leading to mutations in the DNA sequence (Martin et al., 2015).

The process of deamination and mutations during repair is referred to as somatic hypermutation (SHM). After proliferating, the B cells return to the area of the GC containing TFH cells and are reevaluated for their affifinity towards the antigen. This iterative process of proliferation, mutation, and affinity evaluation continues until the B cells have a sufficiently high affifinity to the antigen. At this point, the B cells differentiate into plasma B cells and begin to produce antibodies which allow for the immune system to eradicate the antigen.

2.2. Affinity maturation as an Algorithm

We model AM in Algorithm 1, which we use to simulate AM in our experiments. We simplify a few known or uncertain issues about AM, detailed in SubSection 2.3.

Each naive B cell receptor sequence is generated randomly by a combinatorial mix of its V, D, J, and C gene segments, as well as through junctional diversity between these segments (Sompayrac, 2019). Naive B cells span at least 100 million possibilities (Sompayrac, 2019). The naive B cells recruited to a germinal center are cells that already have some promising affifinity s to the antigen a. SHM then mutates nucleotides in a region of the DNA sequence approximately 100 base pairs (bp) long which contributes to the definition of the B cell receptor structure (Martin et al., 2015). Mutations in the hypervariable region are on the scale of 106 times more likely than mutations outside of this region (Martin et al., 2015). Mutations can be swaps, insertions, and deletions in a categorical space modeled as fA; T; C; G;£g100, where £ connotes a deletion. It should be noted that although the combinatorial space of possible mutations is large, many specific mutations immediately lead to apoptosis. AM optimization is parallelized and distributed over G germinal centers, which algorithmically can be thought of as G different parallel processors. We model the different germinal centers as working independently, though there may be biochemical signaling between them. In practice, an organism trains for multiple independent antigens simultaneously, but for simplicity, we consider one antigen at a time.

2.3. Known simplifications of Algorithm 1

We note that Algorithm 1 simplifies a few known characteristics of AM. We believe these simplifications are minor and that they do not affect the major conclusions of this work.

We model the algorithm as T discrete iterations but in practice, AM is a continuous process that is partly time-limited because of antigen decay and external pressures. However, to our knowledge, there is not an exact limit to the number of divisions that can occur or an exact timeline that must be met during AM. Furthermore, there is a chance that the immune system does not find a solution fast enough, causing the host to die. While we recognize that this occurs in the natural system, it is not the focus of our simulation. Therefore, we use a fixed number of iterations as an approximation of the time limits the real immune system faces. In addition, we identify the highest affinity B cells at the end of our simulation as plasma B cells for simplicity, but plasma B cells are not selected simultaneously at the end of AM and may not have the absolute highest affinity.

Before SHM begins, the initial B cell population may have undergone undirected proliferation, which means the initial random sample may be better modeled as random clusters of B cells. Algorithm 1 allows the germinal centers to grow without bounds, though the die-off rated will tend to limit the population size in the germinal centers. In practice, the size of germinal centers is also bounded by physical volume constraints and biochemical resource constraints, We use a constant die-off rate. but there is some evidence die-off probability decreases as affinity increases (Anderson et al, 2009). We include a simplified version of this behavior by biasing the death probabilities of B cells in the germinal center towards lower affinity cells.

T cells measure nearby B cells for their affinity, so there is only some probability that a specific B cell will have its affinity measured, and that probability a B cell's affinity gets measured is thought to be a function of the spatial organization (which is indirectly affected by affinity)and direct affinity. The affinity and spatial proximity of a particular B cell influence its likelihood to be selected and induced to proliferate. We do not consider the spatial organization between TaH cells and B cells explicitly. The strength of the proliferation signal is generally proportional to the affinity of the Bcell, such that a high-affinity Bcell is more likely to proliferate many times before returning for another iteration of affinity evaluation(Mesin et al,2016; De Silva and Klein, 2015). Similar to the increased likelihood of death for a low-affinity B cell, we bias the selection of B cells for proliferation towards higher affinity B cells.

Algorithm 1 may oversimplify AM in other ways as well that we are not aware of, or that are not yet known.

2.4. Plasma B cells are created to optimize antigen affifinity given a limited time

AM is a process that has long been framed as the immune system acting as a global optimization algorithm trying to find a B cell that best identifies a given antigen through SHM (Theodosopoulos and Theodosopoulos, 2002). That is, AM acts as if it were a heuristic to solve, where a is a given antigen, b is a B cell from the set B of all possible B cells, and s is the affinity function that models the quality of the lock-and-key physical and biochemical interaction of b and a. Note that B is a very large categorical space defined by the variable-length DNA sequence that encodes the B cell receptor.

However, objective (1) does not recognize the fact that a plasma B cell does not need to be a perfect match to the antigen. In fact, there appears to be a sufficient affifinity s such that once an affinity of s is reached, the B cell is induced to differentiate into a plasma B cell. Further, the immune system is under time pressure to produce such sufficiently high-affifinity plasma B cells as fast as possible.

Therefore, we propose that a more realistic model of what the plasma B cell generation process is optimizing should also depend on the sufficient affinity t>0， and the given set 多。 of initial naive B cells in the germinal center. To capture the time pressure, we model probabilistic mutations to B cells in the germinal center at each discrete time iteration. Given a naive B cell b E， let M(b)∈ be a new random B cell produced by a single random mutation of b. Let M*(b)= M(M(..(M(b))..)) denote the random B cell generated after Krandom mutations of b so that the random B cell M(b)can be any of the B cells reachable by Kmutations M of the initial B cell b, with the corresponding probabilities dependent on the sum of the likelihood of the different mutations paths that could produce M* (b)starting from b.

Then we hypothesize the plasma B cell selection process is a heuristic evolved to minimize the number of mutations K∈needed so that on average K random mutations will produce at least one B cell in the germinal center with sufficient affinity t to the antigen:

where EI.] is the standard expectation operator (average) concerning the random variable's possible outcomes weighted by their probabilities. This objective is consistent with our Algorithm 1.

Clearly, the immune system is not a sentient entity that explicitly formulates the criteria(2)and subsequently identifies a heuristic to optimize it. Rather, we hypothesize that evolutionary pressures have been selected for a plasma B cell generation process that better optimizes(2).

3. Memory B cell training

Similar to plasma B-cells, memory B cells are created within the germinal center, but there are key differences in the generation of these two cell types to achieve their respective objectives(Mesin et al,2016; Weisel and Shlomchik,2017; Suan et al,2017). We first review how memory B cells are created, and then consider what criteria they are optimized for, analogous to our criteria(2)for plasma B cells

3.1.Background on memory B cells

While the name memory B cell may invoke the idea that a memory B cell is a long-term copy of a plasma B cell, the truth is more complicated. Memory B cells do not undergo the entire AM process as plasma B cells do. In fact, memory B cells are characterized by their relatively low affinity compared to plasma B cells, and low SHM load (number of mutations gathered from SHM)(Suan et al,2017). This implies that while memory B cells have initially high affinity to an antigen relative to the naive B cell repertoire, they do not undergo AM to the extent of plasma B cells, and tend to have lower affinity to the current antigen than plasma B cells.

The gene BACH2 plays an important role in the development of memory B cells within the germinal center, and in their eventual differentiation(Suan et al,2017: Shinnakasu et al,2016). BACH2 is inversely correlated with the help a B cell receives from Te cells, and the resulting weak interactions with T cells allow for BACH2 expression to remain high. Critically,

the relationship between BACH2 expression and Tp cells allows for some help from T-cells for cell survival within the germinal center but prevents the B cell precursor from entering the area where it would proliferate and mutate via SHM. This leads to three subsets of B cells within the germinal center:(1) high-affinity B cells which are selected for by T cells to proliferate and mutate via SHM, and eventually lead to plasma B cell differentiation,(2) moderate affinity B cells(low compared to plasma B cell precursors, high compared to the average naive B cell)whose selection by Tw cells is tempered by BACH2, leading to memory B cells, and (3) low-affinity B cells which receive little or no help from Tw cells leading to apoptosis(Mesin et al,2020; Taylor et al,2015).

Memory B cells are similar to naive B cells in terms of their transcriptional profiles, which enables them to circulate freely within the organism and to monitor for future instances of antigens. Despite these similarities, they exhibit over-expression of anti-apoptotic genes which allows for the memory B cell to live for extraordinarily long periods of time and therefore the ability to recognize antigens in the future (Suan et al.2017).

3.2. What are memory B cells optimized to do?

If the immune system's objective were rote memorization of the highest affinity B cell receptors to the current antigen a, we might expect the memory B cell receptor repertoire to be nearly identical to the plasma B cells receptor repertoire, but they are not. One might alternatively expect AM to take advantage of the luxury of time it has before it needs the memory B cells to mutate more so that the memory B cells could have even higher affinity s(b, a) to the antigen than the plasma B cells, further optimizing (1). That also does not appear to be the case. While both those options should be biologically feasible, the immune system does something radically different to create memory B cells: it selects memory B cells earlier in AM than plasma B cells, and thus the memory B cells on average have lower affinity s(b.a)than plasma B cells.

To explain why the memory B cells are so poorly-fit to the cur-rent antigen a, we propose two hypotheses for the objective function that memory B cells may be heuristically trying to optimize. Our two hypotheses follow from the dual role of memory B cells (Dogan et aL,2009; Weisel and Shlomchik, 2017). First, when future incarnations of the antigen attack, the memory B cells are used as-is to differentiate into plasma B cells and eradicate the mutated antigen. Second, memory B cells are used to warm start AM's secondary training of plasma B cells. In fact, recent evidence shows that a large portion of the plasma B cells in the secondary response are memory B cells from the first response, and some memory B cells are also used to seed the new germinal centers to optimize secondary response plasma B cells(Mesin et al,2020).

3.3. Training for affinity to a mutated antigen

We propose that the key issue for memory B cells is that the future instance of the antigen they must mitigate is almost certainly a mutation àof the original antigen a. At the time the memory B cells are created, the future mutated antigen a is unknown, but we can characterize it as a randomly mutated antigen A. In this paper, we use the standard probability notation that a capital letter denotes a random variable, and its corresponding lower-case letter denotes the realization of that random variable. For example, if you roll a six-sided die, the random value Xe {1,2..6} refers to the die roll before you look at it because at that point you only know the probability of its six values, but once you see the die roll, it is a deterministic value xE {1,2....,6}.

Let A be a random antigen drawn from some conditional probability distribution PA~ja that depends on the current antigen a, and models the probability of possible future mutations to a and the probability that such a mutation is presented to the host organism within the lifespan of memory B cells selected during the immune response to a. If only a single memory B cell were required for a secondary response, a logical generalization of (1) would be to select a memory B cell that will, on average, have high affifinity to the random mutated antigen A.

If all mutations of the antigen a are equally likely and the affifinity score s was a nice linear function, then the solution to (3) might be the same as (1). But we expect there to be substantive asymmetry in the probability of different antigen mutations, so we expect the solution to (3) will be different than the solution to (1). This is the same principle as in the famous Wayne Gretzky quote about hockey, I skate to where the puck is going to be, not where it has been

Fig. 2 illustrates the criterion in (4), showing that diversity in the memory B cells helps cover the space of probable mutations of the original antigen a. The criterion (4) assumes a fixed choice of N, but if you could also optimize (4) for N, then you would always prefer a larger number N of memory B cells. However, there is also downward pressure on N due to the physical resources needed to store and maintain those cells, and time pressure before the antigen decays away. We emphasize again that we are not hypothesizing that the memory B cells themselves try to optimize (4), rather than evolutionary pressures would have preferred memory B cell generation processes that optimized(4).

3.4. Optimizing for warm starting training for a mutated antigen

We do not mean to suggest that the memory B cells directly optimize(5), but rather that evolutionary pressures might have preferred memory B cell selection processes that better optimize (5). Goal (4) and goal (5) will probably have different optimal solutions depending on the probability of different mutations of the antigen and the B cells, though the same heuristic memory B cell selection process might do pretty well at both objectives. It is not yet known how important memory B cells are to the secondary response plasma B cell training in germinal centers, yet some evidence shows that secondary response germinal centers are comprised of more naive B cells than one might expect (Mesin et al., 2020).

3.5. Why are memory B cells not copies of plasma B cells? 

Both (4) and (5) appear to require knowledge of the probability distribution PA~ of different mutations the antigen may undergo, and the probability distribution PMK ðbÞ of a mutated B cell after K mutations. However, for many symmetric choices of PA~ and PMK ðbÞ, the exact distributions might not matter much: the immune system could cheaply achieve a good approximate solution to (4) and (5) by simply making the memory B cells copies of the plasma B cells; there does not appear to be any biochemical restriction preventing exact replication. However, memory B cells do in fact appear to be selected differently than the plasma B cells. We present two hypotheses as to why.

Our first hypothesis was already introduced in Section 3.3 and Fig.2: because the immune system gets to select a set of memory B cells in (4), it pays to have more diversity in the memory B cells than one gets by copying the plasma B cells. Plasma B cells tend to be less diverse because they are trained to optimize (1), which even with multiple local minima in the shape of sðb; aÞ will limit their diversity. Memory B cells are more diverse than plasma B cells because they are selected earlier in the maturation process. We believe this diversity is important to optimize affifinity to the mutated antigen as per (5) because the true affifinity function s is a highly nonlinear function of the amino acid sequences of a B cell b and antigen a that arise from complex biochemical properties and physical lock-and-key structures (Carneiro and Stewart, 1994; Kilambi and Gray, 2017; Ambrosetti et al., 2020)

Our second hypothesis is that the warm start objective (5) for the secondary germinal centers is not well-optimized by a copy of the plasma B cell set because there is evidence that the probabilities PMjb of the random mutations of the B cells in SHM are asymmetric: certain mutations of B cells are much more likely than others. That makes some B cells a more flexible starting point for warm-starting than the original plasma B cells, which may have trouble mutating to match the new antigen. Evidence for asymmetric PMjb is that many researchers have noted AID preferential targeting of specific motifs (Martin et al., 2015; Stavnezer, 2011; Chen and MacCarthy, 2017; Keim et al., 2013). As mutations would, by definition, change the specific sequence that AID targets, it is reasonable to infer that the first mutation of this location is easier than future ones. Once the sequence is changed, AID is less likely to target this location. Overall, the preferential targeting of AID would make it harder for this region to mutate further or reverse back to the original sequence.

This asymmetry in the probability of moving around the space of all B cells via mutations during AM creates a disconnect between the plasma B cell objective (1) and the objective of being a good warm start solution to future plasma B cell training as per (5). Specifically, a plasma B cell might have made many difficult-to reverse mutations to optimize (1) for the current antigen a. In contrast, the chosen memory B cells appear to be under-optimized for fitting the current antigen a, but we hypothesize they can more easily mutate in a secondary response AM to better fit the random future antigen A~. Overall, we note that how well the objectives (1), (4), and (5) align depends on the symmetry of PA~; PMjb, and the nonlinearity of s.

4. Simulations 

We use the AM algorithm (Algorithm 1) to model how plasma B cells and memory B cells are trained and show through two simulations that the simulated memory B cells are better than the simulated plasma B cells at optimizing our hypothesized objectives (4) and (5), thus providing evidence that these objectives are biologically reasonable. We fifirst demonstrate the mechanics of affifinity maturation in simulated primary immune responses, then compare different potential initial conditions for simulated secondary immune responses.

These simulations do not account for all of the real-world issues at play, such as that a viral mutation must not harm the virus's functionality, and the issues described in 2.3. Despite these limitations, we argue these simulations capture many of the key issues needed to illustrate that our hypothesized objectives are consistent with the difference in plasma and memory B cell training. Complete code for our simulations will be made available upon request.

4.1. Simulation set-up

Our simulations follow Algorithm 1 for the AM process. We initialize a naive B cell repertoire (10,000 cells)with B cell receptors that are represented by a random sequence of 10-50 amino acids, where each amino acid is drawn uniformly over the space of 61 non-stop codons(creating a non-uniform distribution over the amino acids). We simulate the antigens as sequences derived from known antigenic sequences of chicken ovalbumin, bovine milk, and wheat (Honma et al. 1996; Liu and Sathe, 2018). Each antigenic sequence is 17 amino acids long for consistency. We simulate the affinity metric s between a B cell and the antigen using the standard local align MATLAB function, which finds the optimal alignment between two sequences using the BIOSUM50 matrix and returns a score reflecting how similar two sequences are in this alignment Henilkoff and Henikoff,1992. We use this score as a measure of affinity between the B cell receptor and the antigen for simplicity, but note that it is only a rough approximation of the more complex structural compatibility between the B cell receptor and an antiphon.

Mutations of the B cell during AM are modeled in the codon space, where codons of the B cell receptor are replaced with one of the 61 codon possibilities. While SHM mutates B cells on a single nucleotide level, working in the codon space prevents the added complication of filtering out nonsensical B cell receptor sequences. In addition to the replacement of codons, codons in B cell receptor sequences can be inserted or deleted. Insertions and deletions are less likely to occur than swapping for another codon, based on rates of each type of mutation observed in humans (Gibbs et al., 2003; 1000 Genomes Project Consortium et al., 2015; Zia and Moses, 2011). The codons defining each B cell receptor are chosen to mutate at random, but we simulate codons that contain C cytosines to be C þ 1 time more likely to be mutated than codons without cytosine, reflecting biological biases to nucleotide sequence motifs (Martin et al., 2015; Stavnezer, 2011; Chen and MacCarthy, 2017; Keim et al., 2013). We also impose a transition bias between codons, making some swaps more likely than others based on a mutability matrix derived from the BLOSUM50 matrix (Henikoff and Henikoff, 1992; Veerassamy et al., 2003).

For our initial simulation of a primary adaptive immune response to an antigen (primary response), we randomly select 50 naive B cells from the B cell repertoire from the top 1,000 of the 10,000 naive B cell repertoire in terms of affifinity to said antigen. This reflects the recruitment of naive B cells with some partial affifinity by TFH cells to germinal centers (Mesin et al., 2016; Tas et al., 2016). These 50 ‘founder’ B cells are then duplicated 20 times to form a germinal center population of 1,000 cells, reflecting the growth period of germinal center formation (Amitai et al., 2017). For each iteration of the simulation, 50 B cells are selected for proliferation and 50 B cells are selected for removal. Higher affifinity B cells have a higher selection probability for proliferation, while lower affifinity B cells have a higher selection probability for removal. The B cells selected for proliferation are duplicated and mutated, replacing all B cells selected during this iteration. This process imitates apoptosis of low affifinity B cells from lack of TFHcell help and proliferation of B cells with high affifinity after being selected by TFHcells. These mutations have the possibility to increase, decrease, or have no effect on the affifinity of the B cell receptors. We also impose constraints on which mutations can occur on a particular iteration, simulating a negative selection process due to damaging or potentially dangerous mutations. This entire process repeats over 100 iterations.

As a basis for both Simulations 1 and 2, we extract 50 cells during the first half of our primary response simulation as the simulated memory B cells The cells are randomly chosen from the top 25th percentile of germinal center B cells. Another 50 cells are selected at the very end of the primary response to represent plasma B cells, where these cells exhibit the highest 50 affinity scores to the antigen. This reflects a slight deviation from Algorithm 1, as we do not know the threshold t a priori. We establish t for the secondary response based on the affinity of the plasma cells chosen at the end of the primary response. As expected, the simulated memory B cells have an overall lower affinity to the antigen compared to the plasma B cells, but higher than the initial set of naive B cells.

Our simulated mutations of the antigen for the secondary response are derived from a uniform random swap of any of the codons for any other(including possibly itself, ie. a no-op), which creates a non-uniform distribution over the amino acids as some amino acids are coded for by multiple codons. Mutations of the antigen can also include insertions or deletions of codons, with equal probability to any codon.

4.2. Simulation 1:Affinity to a mutated antigen

A secondary infection could involve an antigen that has been mutated or an antigen that is similar to a related pathogen. We simulate the changes in the antigen from the primary to secondary infection by causing adversarial mutations to our antigenic sequence. First, we generate 1,000 uniformly random mutations of antigen a's sequence. The random mutations may be a swap of any amino acid to any other, an insertion of any amino acid, or a deletion of amino acid. Of those candidate mutations, we keep the one that has the lowest average affinity (worst case)to the set of 50 plasma B cells from the primary response, to reflect that a potentially dangerous secondary infection would likely be from a more challenging mutation. We repeat this random process a total of K times to produce an antigen with K mutations. We take that adversarial antigen to be the worst-case realization of P; from the candidate mutations.

We then test the different B cell populations by the hypothesized goal of(4).Fig.3 shows the average affinity between increasingly mutated antigens(chicken ovalbumin, bovine milk, and wheat)with mutations K=1....10 for different cell populations at the end of the primary response averaged over 100 independent runs(Honma et al,1996; Liu and Sathe,2018). Fig 3 shows that our simulated plasma B cells are the best choice to maximize(4) for a small number of adversarial mutations, our simulated memory B cells are the best choice between~2-6 mutations, and naive B cells are best after many mutations(7+). Our simulations are too simplified for the specific transition points to be meaningful, but we argue they do provide strong evidence that the plasma B cells are likely not optimal for identifying substantially-mutated anti-gens. This may further suggest that there is some region in the mutation space where memory B cells are more useful than plasma B cells or naive B cells as an initial condition for a secondary response.

We hypothesize that the fact that plasma B cells do not always have the highest affinity to mutated antigens is driven by the greater diversity of the memory B cells and naive B cells. While the mutations occur in DNA space, the relevant diversity is in the resulting nonlinear physical and biochemical space that defines the affinity to the antigen. Approximately measuring the B cell diversity in each population using the pairwise BLOSUM similarities in each set shows substantial diversity differences, with the plasma B cells having an average within-set BLOSUM similarity of 46, the memory B cells having a much lower average within-set BLOSUM similarity of ~27, and the naive B cells having even lower within-set BLOSUM similarity of ~9(from a representative primary response).

4.3. Simulation 2: Mutations needed for secondary response plasma B cell training

In this simulation, we compare how well the three types of B cells perform as seeds for the secondary response germinal center. We mimic a secondary response training of a new set of plasma B cells optimized for high affinity to a mutated antigen (described in Simulation 1). We initialize the secondary response plasma B cell optimization with one of three choices:(i)the N=50 plasma B cells generated in the primary responsibility for the original antigen a,(ii)the N=50 memory B cells generated in the primary responsibility for the original antigen a, or (i)N=50 naive B cells. We populate the germinal center in an identical way to the primary response, using the three sets of 50 B cells as our new founder cells. As in the primary response, the naive B cells are random but chosen to have some initial affinity to the now-mutated antigen to simulate recruitment to the germinal center. Each of these three secondary response germinal centers undergo AM in an identical fashion to the primary response.

Fig.4 are representative examples of multiple secondary response simulations, given the same primary response and the same K=1.....10 mutations on the bovine milk antigen(Liu and Sathe,2018). Specifically, it shows the average affinity of the 0.05% highest affinity cells over 100 iterations in secondary responses.Fig.4 highlights that for just one or two antigen mutations, plasma B cells tend to have the highest initial affinity to the mutated antigen. For three or more adversarial mutations, naive and memory B cells are better initial conditions for the secondary response than the primary response plasma B cells, and reach an affinity of t in fewer iterations (Figs.4 and 5).Fig.4 also conveys how the recruited secondary response naive B cells tend to have a higher initial affinity to the mutated antigen than either the primary response plasma B cell or memory B cells once the antigen has been sufficiently mutated. This was expected, but we were surprised at how few mutations it took for the naive B cells to have the highest affinity at the first iteration of the secondary response. That is. for relatively few codon mutations in the new antigen, we find empirical evidence that seeding a secondary response with primarily B cells from the naive repertoire is an advantageous strategy. We were not able to simulate a region in mutation space where the memory B cells from the primary response consistently had the highest affinity at the first iteration of the secondary response, likely due to considerable variation between simulated secondary responses.

Similarly, Fig. 5 shows the average number of iterations for plasma, memory, and naive B cells to reach t overall secondary response simulations for all three antigens. The iteration number, where the threshold t is reached, is averaged over all secondary response simulations for 50 independent simulations of the primary responsibility for each antigen. Again, our simulations did not show a region of mutation space where memory B cells were consistently the best warm start conditions. We found this to persist across several refinements of our simulation, suggesting this finding might not be an artifact of too coarse a simulation. We hypothesize that this is due to the selection of naive B cells for the secondary germinal center being biased to have some initial affinity to the mutated antigen (as described in the primary response simulation). That is, naive B cells in Fig. 4 are selected for initial affinity to a, while memory and plasma B cells derived from naive B cells are selected for having some initial affinity to the original antigen a(Fig.3). We initially suspected that our affinity bias for the selection of naive cells was too strong, but both naive and memory B cells are recruited simultaneously in real secondary germinal centers to have the best chance of creating new plasma B cells. In fact, recent experimental evidence suggests that secondary response germinal centers are comprised of more naive B cells than previously thought (Mesin et al,2020). It is also possible that there is an antigen mutation regime for which the memory B cells are indeed more effective than naive B cells for warm starting, but that our simulations were not realistic enough to capture it.

We note that the secondary response simulations often take more than 100 iterations to reach the affinity threshold t from their corresponding primary simulation. When calculating the average number of iterations to reach t in Fg.5, we set these cases to the maximum iteration number of 100. We hypothesize that this phenomenon occurs because the adversarial mutations of d may create a more difficult problem for the simulated germinal center to solve. That is, the mutations may be specifically biased towards potential weak areas in the mutation probability space as a result of the antigen's own survival objectives.

5. Conclusions and open questions

We hypothesized that the dual role of memory B cells can be captured by two objectives(4)and (5) and that these objectives would not be as well-optimized by copying plasma B cells that are trained for(1), due to their over-fitting the original antigen. Our simulations, while limited, provide strong evidence that plasma B cells would not optimize(4)or(5)once the antigen underwent sufficient adversarial mutations. We believe this sub-optimality of plasma B cells against mutated antigens provides a role for the different selection mechanisms used for memory B cells.

Our simulations show a limited range of antigen mutations over which our simulated memory B cells may be optimal; for substantial mutations, we show random naive B cells can work even better. Biologically, this conclusion is supported by recent findings on secondary adaptive immune responses, where naive B cells make up the vast majority of B cells in secondary germinal centers (Mesin et al,2020). Instead, memory B cells are often used to directly differentiate into plasma B cells, rather than undergoing further somatic hypermutation (Mesin et al, 2020). Plasma B cells are a one-time solution and are highly overfit to the current anti-gen of interest. Memory B cells provide a more approximate solution to the current antigen, which is kept within the body to recognize future antigens with similar characteristics. If a future antigen is so different from what has been previously encountered by the immune system that no memory B cells can identify it, a new solution is formed from scratch using naive B cells.

Memory B cells play two roles, both differentiating into plasma B cells and re-initiating germinal centers, but these roles may be played by distinct subpopulations(Dogan et al。2009; Weisel and Shlomchik, 2017). These distinct subpopulations of memory B-cells might have resulted from distinct AM processes, or changes in the AM process over the AM time span that we have not explicitly modeled in our Algorithm 1(Weisel et al,2016). Thus our two memory B cell objectives in (4)and(5)may apply to independent memory B cell populations. Here we investigated a simplified model, where memory B cells were considered a unified group. However, we were not able to show via simulations a regime in which the memory B cells were clearly better than secondary naive B cells for re-initiating the secondary germinal response. Our results suggest this re-initialization task might be a weaker or rarer role of the memory B cells. These results align with recent experimental evidence that similarly noted memory B cells were less prevalent in secondary germinal centers than previously assumed (Mesin et al., 2020). However, even if memory B cells are not always needed for the secondary response, it might be that in some cases they are very important for warm starting, which might still exert evolutionary pressure on their selection process. The evolutionary pressures on memory B cell selection in nature are not known but may be elucidated through the integration of computational simulations and biological experiments. Actively monitoring the affifinity of B cells during affifinity maturation, as well as detecting when and why GC B cells become memory B cells, may assist with the development of more accurate and complex models in the future.