Position-dependent Differential Targeting of Somatic Hypermutation

Abstract

Somatic hypermutation (SHM) generates much of the antibody diversity necessary for affinity maturation and effective humoral immunity. The AID-induced DNA lesions and error-prone repair that underlie SHM are known to exhibit intrinsic biases when targeting the immunoglobulin (Ig) sequences. Computational models for SHM targeting often model the targeting probability of a nucleotide in a motif-based fashion, assuming that the same DNA motif is equally likely to be targeted regardless of its position along the Ig sequence. The validity of this assumption, however, has not been rigorously studied in vivo. Here, by analyzing a large collection of 956,157 human Ig sequences while controlling for the confounding influence of selection, we show that the likelihood of a DNA 5-mer motif being targeted by SHM is not the same at different positions in the same Ig sequence. We found position-dependent differential SHM targeting for about three quarters of the 38 and 269 unique motifs from more than half of the 292 and 1,912 motif-allele pairs analyzed using productive and non-productive Ig sequences respectively. The direction of the differential SHM targeting was largely conserved across individuals, with no allele-specific effect within an Ig heavy chain variable (IGHV) gene family, but was not consistent with general decay of SHM targeting with increasing distance from the transcription start site. However, SHM targeting did correlate positively with the mutability of the wider sequence neighborhood surrounding the motif. These findings provide insights and future directions for computational efforts towards modeling SHM.

Introduction

Effective humoral immunity is driven by B cells that can produce antibodies with high binding affinity towards their target antigens. Somatic hypermutation (SHM) generates the antibody diversity necessary for affinity maturation. During SHM, the enzyme AID deaminates a cytidine into uracil. The canonical and mutagenic versions of various repair pathways, such as base excision repair and mismatch repair, recognize and compete to resolve this lesion (1). As a result, point mutations are introduced at a rate of approximately 10⁻³ per base pair per division (2) to the immunoglobulin (Ig) variable region, not only at but also around the initial lesion site. B cells with mutations that increase their binding affinity are then positively selected for. While largely stochastic, there are biases intrinsic to SHM in terms of AID targeting preferences and error-prone repair. Computational models exist that attempt to capture such biases and predict which positions are more or less likely to mutate as a result of this SHM process. Understanding and modeling these biases are important for many aspects of computational analysis of B cell receptor (BCR) sequences, such as building phylogenetic models that take into account of such B cell-specific mutation biases (3), understanding the lineage development of broadly neutralizing anti-HIV antibodies that require extensive SHM (4), quantifying the selection pressure on Ig sequences (5), and simulating Ig sequences with realistic mutation profiles (6).

Computational efforts to model SHM have long focused on the immediate DNA context of the mutations. These models attempt to estimate the relative likelihood of SHM targeting based on the local microsequence context. Studies from before the high-throughput sequencing era identified classical hot or cold spots made up of specific DNA motifs. Hot spots canonically include 5’WRC/GYW3’ and 5’WA/TW3’ motifs, where W={A, T}, R={G, A}, Y={C, T}, and the position mutating is underlined (7, 8). Cold spots include motifs of the pattern 5’SYC/GRS3’, where S={C, G} (9). It was observed that hot spots tended to be targeted by AID for SHM more often, whereas cold spots tended to be targeted less. Beyond these categorical designations, a hierarchy of mutabilities exists that is highly dependent on the surrounding nucleotides (10, 11). More recent works have largely reaffirmed the overall hierarchy amongst these 3-mer motifs, but have also shown significant differences among classical hot and cold spots by extending this microsequence context-based approach into models involving 5-mer (12-14) or even 7-mer motifs (15, 16). One widely used 5-mer model (17-22) was developed by (12). This synonymous, 5-mer, functional (S5F) model was able to estimate SHM targeting without the confounding influence of selection by focusing on synonymous (silent) mutations that occurred at central positions of DNA 5-mer motifs on productively rearranged, or “functional”, sequences (12). An alternative approach to leverage information from both non-synonymous (replacement) and synonymous mutations relies on the use of non-productively rearranged, or “non-functional”, sequences to avoid the confounding influence of selection. This model is referred to as the replacement and synonymous, 5-mer, non-functional (RS5NF) model (13). Both models attempt to model the targeting preference of SHM by assigning to 5-mer motifs the relative probability of its central nucleotide being targeted given the composition of the adjacent four nucleotides (12, 13).

Beyond the microsequence context, it has been suggested that SHM targeting could vary due to positional effects within the Ig sequence (23). (24) reported an exponential decay in mutation frequency at the 3’end due to the distance from the transcription start site, predicting a fall to <1% of the maximum mutation frequency at ~2100 bp downstream of the 5’ boundary. (25) reported a sharp reduction in mutations separated by around 50 nucleotides for the nontranscribed strand, raising the possibility that AID might consistently skip a fixed distance during targeting. Past studies have also reported differences between the framework regions (FWRs) and the complementarity-determining regions (CDRs) beyond hot and cold spots, though they tended to be either low-throughput with analysis of fewer than 1,000 mutations from human heavy chain variable region sequences (8, 26) or did not fully account for selection (27). Generally, these proposed positional effects are not part of the microsequence context-based motif models for SHM.

A key assumption of the microsequence context-based models, regardless of whether they rely on 3-mer, 5-mer, or 7-mer motifs, is that the targeting probability of a motif is constant regardless of the position of the motif along the Ig sequence. However, the validity of this assumption has not been rigorously investigated. Here, we investigated whether the 5-mer microsequence context was sufficient to capture variations in mutability by examining the assumption that the probability of a 5-mer motif being targeting by SHM is the same regardless of its position along the Ig sequence. Specifically, we identified 5-mer DNA motifs that occur at two different positions in the same Ig heavy chain variable (IGHV) gene allele, and tested whether these motifs were targeted at the same rate. To control for selection, we either analyzed only motifs where mutation at the central nucleotide is always synonymous when using productively rearranged sequences, or performed the analysis using non-productively rearranged sequences. Based on close to one million BCR sequences, we found position-dependent differential SHM targeting for about three quarters of the unique motifs analyzed. The direction of the differential SHM targeting was largely conserved across individuals with no allele-specific effect within an IGHV family, and correlated positively with the mutability of the wider sequence neighborhood surrounding the motif.

Materials & Methods

Motif-allele pairs (MAPs)

A motif-allele pair (MAP) captures exactly two instances of the same DNA 5-mer motif in an IGHV allele with the central nucleotides located at position i (5’ site) and position j (3’ site), where i<j. If a motif appears m times in an IGHV allele, where m>2, this is captured by $\left(\begin{array}{c}\hfill m\hfill \\ \hfill 2\hfill \end{array}\right)$ MAPs. Under the S5F analysis, which used productively rearranged sequences, the MAPs analyzed were restricted to motifs whose central nucleotide can mutate synonymously only. Under the RS5NF analysis, which used non-productively rearranged sequences, the MAPs analyzed included motifs whose central nucleotide can mutate both synonymously and non-synonymously.

Reference IGHV alleles

An initial set of 372 reference IGHV alleles for Homo sapiens were downloaded from IMGT/GENE-DB release 201918-4 under the “F+ORF+in-frame P” specification (28). Only alleles labelled functional and non-partial were kept, leaving 218 alleles, each of which fully covered IMGT-numbered nucleotide positions 1 to 312. Additionally, one allele each from 6 pairs of alleles with identical nucleotide sequences was removed. These were: IGHV1-69D*01, IGHV3-23D*01, IGHV3-30-5*02, IGHV3-30-3*03, IGHV3-30-5*01, and IGHV2-70D*04, which were identical sequence-wise to, respectively, IGHV1-69*01, IGHV3-23*01, IGHV3-30*02, IGHV3-30*04, IGHV3-30*18, and IGHV2-70*04. The curated set of reference IGHV alleles contained 212 unique alleles.

Overview of datasets

Five previously reported BCR repertoire datasets comprising 30 human subjects were re-processed in a consistent manner.

The dataset from (20) originated from a study (PRJNA338795) examining the BCR repertoire of myasthenia gravis patients (20). From this study, we started with bulk-sequenced heavy chains from FACS-sorted CD19+CD27− naïve B cells and CD19+CD27+ memory B cells from 4 healthy controls. For analysis, we used only sequences from CD19+CD27+ memory B cells.

The dataset from (29) originated from a study (PRJNA429427) comparing IgD−CD27− double negative (DN) B cells and IgD−CD27+ non-DN B cells (29). From this study, we analyzed bulk-sequenced heavy chains from FACS-sorted CD19+IgD−CD27+ class-switched memory B cells from 3 healthy controls.

The dataset from (30) originated from a study (PRJNA300878) on the influence of inheritability on the BCR repertories of twins (30). From this study, we started with bulk-sequenced heavy chains from FACS-sorted CD20+CD27− naïve B cells and CD20+CD27+CD38low memory B cells from 10 nominally healthy subjects belonging to five monozygotic twin pairs. For analysis, we used only sequences from CD20+CD27+CD38lo memory B cells.

The dataset from (31) originated from a study (PRJNA349143) on the antibody response induced by flu vaccine (31), the samples collected in which were later re-sequenced (32). From the re-sequencing (32), we started with bulk-sequenced heavy chains from unsorted PBMCs from 3 healthy adults. For analysis, we used only class-switched sequences.

The dataset from (33) originated from one of the largest studies (PRJNA406949) to date examining the circulating BCR repertoire (33). From this study, we started with bulk-sequenced heavy chains from unsorted PBMCs from 10 healthy adults. For analysis, we used only class-switched sequences.

Re-processing of datasets

For the datasets from (20), (29), (30), and (31), we obtained sequences pre-processed via pRESTO as described in (20), (29), (30), and (32) respectively.

For the dataset from (33), we downloaded assembled reads made publicly available by (33) at https://github.com/briney/grp_paper (2019-08-27) and performed further pre-processing in pRESTO v0.5.11 (34). For each subject, there were 6 biological samples, each with 3 technical replicates. For each biological sample, duplicate reads in its technical replicates were collapsed into unique sequences. Biological samples for the same subject were then combined. Isotypes were assigned by local alignment of the 3’ end of each sequence to IgM- and IgG-specific internal constant region sequences (IgM: GGGCGGATGCACTCCC; IgG: GGCCCTTGGTGGARGC). Sequences with inconsistent isotype assignment and IgM or IgG primer alignment were removed.

For each dataset, initial germline V(D)J gene annotation was performed using IgBLAST v1.14.0 (35) with IMGT/GENE-DB release 201918-4 (28). Additional quality control was performed, requiring sequences to align exclusively to heavy chain V and J genes; have a minimum V segment coverage from nucleotide positions 1 to 312 under the IMGT unique numbering scheme (36); and have Ns at fewer than 10% of V segment positions. Productively rearranged sequences without a junction length that was a multiple of 3 were removed, where junction was defined to be from IMGT codon 104 encoding the conserved cysteine to codon 118 encoding phenylalanine or tryptophan. Individual genotypes were computationally inferred using TIgGER v0.3.1 (37) and used to finalize V(D)J annotations. To remove potential chimeric sequences, sequences containing more than 5 mismatches from the germline in any 10-bp window were identified using SHazaM v0.1.11 (38) and removed.

Productively rearranged sequences were then used for inferring B cell clones via hierarchical clustering (32) on a by-subject basis. Using Change-O v0.4.5 (38), Productively rearranged sequences were first partitioned based on common V and J gene annotations and junction region lengths. Within each partition, sequences whose junction regions were within a specified normalized Hamming distance from each other were clustered as clones. This distance threshold was determined by manual inspection in conjunction with kernel density estimates, in order to identify the local minimum between the two modes of the bimodal distance-to-nearest distribution representing, respectively, clonally related and unrelated sequences. The thresholds used for the datasets from (20), (29), (30), (31), and (33) were, respectively, 0.1, 0.12, 0.08, 0.1, and 0.05 units of normalized Hamming distance. Following clonal clustering, clonal consensus germline sequences were reconstructed for each clone. A single clonal representative sequence was then chosen for each clone. For the sorted datasets ((20), (29), and (30)), the most highly mutated memory B cell sequence, if any, was used. For the unsorted datasets ((31) and (33)), the most highly mutated class-switched sequence, if any, was used.

Non-productively re-arranged sequences which contained out-of-frame rearrangements in the junction region, and which contained no indel in the V or J segment or stop mutation (30), were aggregated across subjects and across datasets.

For the dataset from (33), nucleotide positions 1 through 50 were excluded from the analysis to account for the fact that the V gene primers annealed to FWR1 (33).

Statistical tests

For a given MAP, to test whether any observed difference in the mutation frequency of the central nucleotide of the 5-mer motif at the 5’ site and at the 3’ site was statistically significant, a binary sign test, or McNemar’s test, was performed (39). First, a contingency table was constructed, in which n₁₁, n₁₂, n₂₁, and n₂₂ denote the number of sequences in which, respectively, there was no mutation at either site; there was no mutation at the 5’ site but there was at the 3’ site; there was mutation at the 5’ site but there was not at the 3’ site; and there were mutations at both sites. A test was performed if there were at least 35 sequences containing the MAP (n₁₁+n₁₂+n₂₁+n₂₂ ≥ 35), and if there were at least 5 sequences in which there was mutation at one site but not at the other (n₁₂+n₂₁ ≥ 5). The test statistic, $\frac{n_{12}-n_{21}}{\sqrt{n_{12}+n_{21}}}$, follows a standard Normal distribution. A two-sided test was performed, and the direction of the observed difference was noted. Correction for multiple testing was performed by controlling for false discovery rate (40).

Under the meta-analysis, for a given MAP tested in more than one subject, a combined one-sided p-value was calculated for each direction using Stouffer’s method (41). The combined p-value of the direction with the lower value was doubled and controlled for false discovery rate.

Targeting models

An updated version of the S5F SHM targeting model was built with SHazaM v0.1.11 (38) using the most highly mutated, memory or class-switched, and productively rearranged sequences as clonal representatives (same as in “Re-processing of datasets”) from all subjects from all datasets. Default parameters, “minNumMutations=50” and “minNumSeqMutations=500”, were used for the “createSubstitutionMatrix” and “createMutabilityMatrix” functions, respectively.

Simulation

Simulated sequences were generated using the “shmulateSeq” function from SHazaM v0.1.11 (38). In each simulation run, for each observed sequence, the number of nucleotide mutations in the V segment was first calculated using the “observedMutations” function. An equal number of mutations were then introduced into the corresponding germline sequence with probabilities specified by an S5F SHM targeting model.

Motif Analysis

Enrichment for particular nucleotides at each position in the motifs from MAPs was assessed using pLogo v1.2.0 (https://plogo.uconn.edu/) (42). For comparing MAPs that showed higher mutation frequency at the 5’ site with those that showed lower mutation frequency at the 5’ site, the set of motifs from MAPs that showed significant differential SHM targeting was used as the background.

Correlation between observed mutation frequency and surrounding germline mutability

We calculated the correlation between the observed mutation frequency of the central nucleotide of the MAPs and the germline mutability of the neighborhoods of varying sizes surrounding the MAPs. The germline mutability of a neighborhood surrounding the central nucleotide of a given motif was calculated by averaging the individual mutabilities estimated by the updated S5F targeting model of the 5-mer motifs within that neighborhood. The size of the neighborhoods ranged from up to 28 5-mer motifs (30 nucleotides) in the 5’ direction and up to 43 5-mer motifs (45 nucleotides) in the 3’ direction. Depending on the positions of a MAP and the neighborhood size, if the neighborhood covered more nucleotides than there were in the germline allele, truncation was performed and only the 5-mer motifs present were considered. For each neighborhood size, the percentage of MAPs whose observed mutation frequency correlated positively with the surrounding germline mutability was calculated.

Results

Germline landscape of motif-allele pairs (MAPs)

To examine whether the local nucleotide context is sufficient to capture the relative likelihood of SHM targeting, we proposed to test whether two instances of the same DNA 5-mer motif at different positions in a single IGHV allele would be targeted at equal rates. As a first step, we surveyed the germline landscape for DNA 5-mer motifs that appear at more than one site along the same IGHV allele. We define a motif-allele pair (MAP) as two instances of the same 5-mer motif in an IGHV allele (Materials & Methods). We collected 212 unique, functional, and non-partial human germline IGHV reference alleles from IMGT/GENE-DB (28), and searched for MAPs along this curated set of germline IGHV alleles. In doing so, we divided the MAPs into two categories, depending on the types of mutations possible at the central nucleotide of a 5-mer motif. The first category of MAPs involved motifs whose central nucleotide only mutates synonymously. In total, we found 876 such MAPs involving 210 unique alleles and 44 unique motifs (Figure 1A; Supplemental Figure 1A). Alleles contained 4.2 MAPs on average, with most alleles harboring between 2 and 6 MAPs (Figure 1B). Around 93% of the MAPs occurred at classical SYC/GRS cold spots or neutral spots, with the remaining 1% and 5% of the MAPs occurring at classical WA/TW and WRC/GYW hot spots, respectively (Figure 1C). The median distance between two instances of the same motif was 39, 54, 51, and 9 nts for MAPs occurring at cold, neutral, WA/TW hot, and WRC/GYW hot spots, respectively (Figure 1C). The second category of MAPs involved motifs whose central nucleotide could also mutate non-synonymously. We found a total of 12,704 such MAPs involving 212 unique alleles and 389 unique motifs (Supplemental Figures 1B and 2A). Alleles contained 59.9 MAPs on average, with most alleles harboring between 50 and 65 MAPs (Supplemental Figure 2B). Within MAPs, there was a median distance of 85, 79, 69, and 89 nts between instances for those occurring at cold, neutral, WA/TW hot, and WRC/GYW hot spots, respectively (Supplemental Figure 2C). Overall, the widespread presence of MAPs across germline IGHV alleles supports a systematic investigation of the MAPs observed in high-throughput BCR repertoire sequencing data.

Figure 1.

Motif-allele pairs (MAPs) in which the central nucleotide of the motif can only mutate synonymously. (A) MAPs found in an example set of germline IGHV alleles. The list of MAPs in IGHV2-5*01 is enumerated in the box. (B) Distribution of the number of MAPs harbored in a germline IGHV allele. (C) Distribution of the nucleotide distance between the central nucleotides at the 5’ and 3’ sites of a MAP.

MAPs are targeted by SHM at different rates

To test whether MAPs are targeted for SHM at the same rate, we gathered curated bulk-sequenced BCR heavy chain sequences from 30 human subjects from five published studies (Materials & Methods). These sequences either came from FACS-sorted memory B cells or, if derived from unsorted PBMCs, were filtered to include only class-switched, and hence putatively antigen-experienced, sequences. In total, we collected 936,714 productively rearranged BCR sequences and19,443 non-productively rearranged BCR sequences (Table 1).

Table 1.

Heavy chain immunoglobulin sequences from bulk sequencing from five B cell receptor (BCR) repertoire datasets used for analysis

Dataset	Source	Isotype filter	Subject	Number of productively rearranged sequences a	Number of non- productively rearranged sequences	Reference & Accession
(20)	FACS-sorted CD19+CD27+ memory B cells b	-	HD07	4401	29	(20); PRJNA338795
			HD09	3240	16
			HD10	3974	21
			HD13	4249	29
(29)	FACS-sorted CD19+IgD−CD27+ memory B cells	-	HC024	5221	37	(29); PRJNA429427
			HC038	4404	11
			HC263	11199	55
(30)	FACS-sorted CD20+CD27+ CD38low memory B cells b	-	TW01A	1502	267	(30); PRJNA300878
			TW01B	1056	20
			TW02A	5950	887
			TW02B	4336	1094
			TW03A	3309	896
			TW03B	1866	402
			TW04A	5506	1348
			TW04B	6974	1991
			TW05A	3196	890
			TW05B	1315	266
(31)	Unsorted PBMCs	Class-switched (IgA, IgE, IgG)	hu420139	25700	718	(31-32); PRJNA349143
			420IV	35702	593
			PGP1	12825	249
(33)	Unsorted PBMCs	Class-switched (IgG)	316188	90724	836	(33); PRJNA406949
			326650	58993	406
			326651	215683	3261
			326713	102876	1243
			326737	30366	213
			326780	51748	540
			326797	95961	1119
			326907	33148	162
			327059	83602	1503
			D103	27688	341

^a:one clonal representative sequence per clone

^b:additional cell subsets used for genotyping and clonal inference but not for analysis (see Materials & Methods)

We first analyzed the SHM frequency of MAPs among the productively rearranged BCR sequences. To avoid the possibility that selection pressures could influence the observed mutation frequency at the different sites, we focused this analysis (hereafter the “S5F analysis”) on the set of MAPs where all mutations at the central nucleotide were synonymous. For each MAP, within each subject, we compared the mutation frequency at the two sites of the MAP (Figure 2A). Specifically, for a MAP observed in at least 35 sequences, in which at least 5 sequences saw mutation at one site but not at the other, we performed a McNemar’s test to determine whether the likelihoods of observing a mutation at one site but not at the other were significantly different (Materials & Methods). As one example, consider the MAP comprising the motif CTGAG occurring at IMGT-numbered nucleotide positions 57 and 282 along IGHV3-74*02 (Figure 2A, left). This MAP was observed in a total of 1,388 sequences in Subject 316188, and accumulated a mutation at one site but not the other in 74 of these sequences. Ten sequences contained mutation at position 57 but not at position 282 (frequency = 0.14), while 64 sequences contained mutation at position 282 but not at position 57 (frequency = 0.86). In other words, the 5’ site was significantly less likely than the 3’ site to mutate with a difference in likelihood of 0.73 (False Discovery Rate (FDR) = 1.7*10⁻⁹). As a second example, consider the MAP comprising ACCAG occurring at positions 105 and 237 along IGHV1-8*01 (Figure 2A, right). This MAP was observed in a total of 270 sequences in Subject HC263, and accumulated a mutation at one site but not the other in 53 sequences. Forty-four sequences contained a mutation at position 105 but not at position 237 (frequency = 0.83), while 9 sequences contained mutation at position 237 but not at position 105 (frequency = 0.17). In other words, the 5’ site was significantly more likely than the 3’ site to mutate with a difference in likelihood of 0.66 (FDR = 1.6*10⁻⁵). Using the abovementioned thresholds, there was sufficient data to test for 12-218 MAPs involving 5-57 unique alleles and 9-27 unique motifs across the 30 subjects (Table 2; Supplemental Figure 3A). In 28 of the 30 subjects, we found MAPs in which there was a statistically significant difference (FDR < 0.05) between the two sites in the likelihood of being targeted by SHM (Figure 2B). Amongst these MAPs, the median difference in the likelihood of being targeting by SHM between the two sites ranged from 0.46 to 1.00, with 1.00 being the scenario where all of the mutations were observed to accumulate at one site and never the other site (Figure 2B). Thus, there exist MAPs that are targeted by SHM at different rates.

Figure 2.

Position-dependent differential SHM targeting within MAPs. (A) Examples of MAPs being tested for differential SHM targeting in individual-subject S5F analysis and aggregate RS5NF analysis pooling all non-productively rearranged sequences. Gray shaded regions correspond to CDR1 and CDR2. (B) Difference in likelihood of mutating at one site but not at the other between the 5’ and 3’ sites of MAPs showing significant differential SHM targeting in each individual-subject S5F analysis. Each point represents a MAP and is colored by the classical categorization of the motif in the MAP. Numbers at the top indicate the number of MAPs found significant. (C) Percentage of MAPs showing significant differential SHM targeting in observed sequences (solid) versus sequences from 500 simulated datasets (hollow) based on the S5F model.

Table 2.

Number of motif-allele pairs (MAPs) detected, tested, and found significant for differential somatic hypermutation (SHM) targeting

Analysis	Dataset	Subject	MAPs Detected a	MAPs Tested b					Unique Alleles Tested	Unique Motifs Tested	MAPs Significant
				Total	Cold SYC/ GRS	Neutral	Hot WA/ TW	Hot WRC/ GYW
S5F	(20)	HD07	189	70	17	46	1	6	25	24	1
		HD09	180	57	13	37	1	6	22	20	0
		HD10	216	50	12	33	1	4	22	20	5
		HD13	214	81	17	54	2	8	26	27	0
	(29)	HC024	70	62	16	40	0	6	16	26	20
		HC038	79	69	18	46	2	3	21	20	10
		HC263	102	94	19	67	3	5	23	26	28
	(30)	TW01A	199	21	9	10	1	1	11	11	5
		TW01B	190	12	5	6	1	0	5	9	3
		TW02A	232	112	35	68	3	6	33	29	30
		TW02B	198	90	30	53	2	5	29	23	17
		TW03A	225	56	17	35	2	2	21	19	12
		TW03B	202	29	12	14	3	0	14	14	10
		TW04A	207	103	31	60	4	8	34	28	26
		TW04B	219	117	39	67	4	7	37	29	41
		TW05A	213	64	19	41	1	3	22	23	15
		TW05B	204	20	10	7	1	2	10	13	3
	(31)	hu420139	234	203	43	143	2	15	51	34	54
		420IV	251	218	53	149	3	13	57	37	61
		PGP1	245	179	42	123	2	12	50	32	21
	(33)	316188	209	80	24	53	3	0	37	22	52
		326650	176	69	22	45	2	0	33	19	47
		326651	202	81	27	51	3	0	37	22	63
		326713	246	96	29	64	3	0	43	22	65
		326737	210	64	20	41	3	0	32	19	33
		326780	222	75	21	52	2	0	37	20	45
		326797	207	74	24	48	2	0	35	20	51
		326907	202	57	16	39	2	0	31	17	30
		327059	208	81	27	51	3	0	39	20	56
		D103	253	77	23	51	3	0	37	19	35
RS5NF	All	All	6579	1912	180	368	396	968	68	269	813

^a:Number of MAPs present in at least one sequence

^b:Number of MAPs with sufficient data for testing (see Materials & Methods)

Next, we analyzed the SHM frequency of MAPs among the non-productively rearranged BCR sequences. As these sequences contained out-of-frame junctions and were presumably not subjected to selection (30), we were able to explore in this analysis (hereafter the “RS5NF” analysis) a larger number of MAPs, some of which gave rise to non-synonymous mutations when the central nucleotide mutated, and which were therefore inaccessible previously due to the restriction to MAPs giving rise to only synonymous mutations. Given the smaller number of non-productively rearranged sequences in each subject, we performed the analysis by combining all such sequences across subjects. There was sufficient data to test for 1,912 MAPs involving 68 unique alleles and 269 unique motifs (Table 2; Supplemental Figure 3B). In 813 MAPs, the likelihoods of one site mutating and other not mutating were found to be statistically different (FDR < 0.05), with the median difference between the two sites being 0.80 (Supplemental Figure 4). These 813 MAPs involved 199 unique motifs, including 19 SYC/GRS cold spots, 34 WA/TW hot spots, and 35 WRC/GYW hot spots.

Since SHM is a stochastic process, to test the possibility that the observed differences in the likelihood of one site of a MAP mutating and the other not mutating rose by chance, we stochastically simulated sequences to see whether the same extent of differences could be detected. We preserved the MAP abundance and the mutational load of the observed data by generating an equivalent number of sequences with underlying germline alleles and introducing an equivalent number of mutations per sequence as the observed data. The mutations were introduced according to targeting probabilities specified by an updated S5F model of SHM targeting that we constructed as part of this study. The original S5F model estimated the SHM targeting probabilities based on 806,860 synonymous mutations found in 446,027 sequences (12). We built an updated S5F model based on 4,074,051 synonymous mutations found in our curated set of 936,714 productively rearranged sequences (Supplemental Figure 5A). The two models correlate highly in terms of 5-mer motif mutability (Spearman correlation coefficient = 0.86) (Supplemental Figure 5B). In the updated S5F model, we continued to see a distinct hierarchy of 5-mer motif mutability (Supplemental Figure 5C) and a strong bias for nucleotide transitions over transversions (Supplemental Figure 5D), as observed in the original model (12). Upon analyzing the MAPs from 500 simulated datasets, we found on average 0.03%, 0.06%, 0.02%, 0.05%, and 0.03% MAPs per dataset with significant differential SHM targeting for, respectively, Subject HD10, HC263, TW04B, 420IV, and the RS5NF analysis pooling non-productively rearranged sequences, compared to significantly higher (Bonferroni empirical p < 0.05) percentages of 10.00%, 29.79%, 35.04%, 27.98%, and 42.52% in the observed data (Figure 2C). Hence, it was unlikely that the observed differential SHM targeting on MAPs was due to chance. Taken together, these results suggest that there is differential SHM targeting on the same DNA 5-mer motif occurring at different positions along the same IGHV allele.

Direction of differential SHM targeting consistent across subjects and across alleles within an IGHV family

To investigate whether the observed differential SHM targeting was likely to be an effect intrinsic to the MAPs, we examined the consistency in the direction of the differences observed across individual subjects. We reasoned that a lack of consistency in directionality would imply that the observed differences were due to factors extrinsic to the MAPs themselves. Of the 134 MAPs that tested significant in more than one individual-subject S5F analysis, 118 (88%; p = 1.1*10⁻²⁰, one-sided binomial test) showed consistency in directionality across all the subjects in which the MAP tested significant (Figure 3; Supplemental Figure 6). These 118 MAPs involved 24 unique motifs, including 6 SYC/GRS cold spots, 1 WA/TW hot spot, and 2 WRC/GYW hot spots. Thus, it is likely that two instances of the same motif experience differential SHM targeting because of some characteristic intrinsic to the MAP, such as the specific allele in which the MAP occurs, the positions of the motifs, or the broader sequence context surrounding the motifs.

Figure 3.

Direction of differential SHM targeting is largely conserved across individuals. A meta-analysis identified 190 MAPs with significant differential SHM targeting, for which the directions observed in individual-subject S5F analyses, aggregate S5F analysis pooling all productively rearranged sequences, and aggregate RS5NF analysis pooling all non-productively rearranged sequences are also shown. Each column identifies a MAP. The identities of the MAPs are detailed in Supplemental Figure 6.

In the S5F analysis, we analyzed exclusively MAPs that gave rise to only synonymous mutations in an effort to avoid being skewed by the selection bias exerted upon the productively rearranged sequences. To assess the possibility that the observed differential SHM targeting on these MAPs were still driven by selection, despite the fact that their central nucleotides could only mutate synonymously, we compared the direction of the differences observed in the productively rearranged sequences in the S5F analysis to that observed in the non-productively rearranged sequences – thought to be not subjected to selection (30) – in the RS5NF analysis. To facilitate comparison, we first synthesized the results from the individual-subject S5F analyses by performing a meta-analysis across subjects (Materials & Methods). Of the 292 unique MAPs involving 38 unique motifs that were tested using productively rearranged sequences in more than one subject and therefore included in the meta-analysis, 190 showed significant differential SHM targeting (FDR < 0.05). These 190 MAPs involved 29 unique motifs, including 6 cold SYC/GRS cold spots, 1 WA/TW hot spot, and 3 WRC/GYW hot spots. The meta-analysis result was consistent with the result from the RS5NF analysis pooling non-productively rearranged sequences from all subjects. Of the 39 MAPs that tested significant in both the meta-analysis and the RS5NF analysis, 34 (87%; p = 1.2*10⁻⁶, one-sided binomial test) agreed in terms of directionality (Figure 3; Supplemental Figure 6). Hence, it is unlikely that the differences observed in the MAPs in the S5F analysis were due to selection.

Next, to investigate whether there is an allele-specific effect to the observed differential SHM targeting on a motif, we investigated the consistency in directionality across different alleles. From the MAPs that tested significant in the meta-analysis, we examined the direction of differential SHM targeting on motifs occurring at the same IMGT-numbered nucleotide positions across different alleles. In all but four cases, the alleles of concern were all found within the same IGHV family. These involved 23, 6, 55, and 48 MAPs comprised of the same motifs occurring at the same positions in different alleles of, respectively, the IGHV1, IGHV2, IGHV3, and IGHV4 family (Figures 4A-D). All but three of these MAPs showed consistency in directionality across alleles within their respective IGHV family. One exception was the motif CTGAG at positions 57 and 282, which was targeted more frequently at the 5’ site in IGHV3-23*04, whereas less frequently in the other IGHV3 alleles (Figure 4C). The remaining two exceptions involved the motif GGGGG at positions 24 and 48, which was targeted more frequently at the 5’ site in IGHV3-48*02 and IGHV3-66*02, whereas less frequently in IGHV3-15*01, IGHV3-15*07, IGHV3-21*01, and IGHV3-23*01 (Figure 4C). The four cases in which the alleles of concern spanned different IGHV families involved 41 MAPs comprised of the same motifs occurring at the same positions in alleles across one or more of the IGHV1, IGHV3, IGHV5, and IGHV7 families. Of the four unique motifs involved in these 41 MAPs, CCTGG and CTGAA showed consistency in directionality across alleles from different IGHV families, while GTGCA and TCTGG displayed inconsistency (Figure 4E). Interestingly, GTGCA showed inconsistency amongst IGHV1 alleles, whereas TCTGG showed consistency amongst IGHV1 alleles but inconsistency between IGHV1 and IGHV3 alleles (Figure 4E). Overall, the direction of differential SHM targeting appears to be largely conserved across MAPs involving the same motif at the same positions in different alleles, at least within the same IGHV family.

Figure 4.

Direction of significant differential SHM targeting for MAPs found in multiple (A) IGHV1 alleles; (B) IGHV2 alleles; (C) IGHV3 alleles; (D) IGHV4 alleles; and (E) alleles across different IGHV families. Each point represents a MAP, with the x-axis indicating the allele in which the MAP is found and the y-axis indicating the 5-mer motif and the IMGT-numbered positions of the central nucleotides at the 5’ and 3’ sites.

Differential SHM targeting is correlated with surrounding mutability

To investigate the potential mechanisms underlying the differential SHM targeting observed in MAPs, we first examined characteristics intrinsic to the MAPs that showed significant differential SHM targeting in the meta-analysis (Figure 5A). We compared the motifs from the MAPs that showed higher mutation frequency at the 5’ site with those that showed lower mutation frequency at the 5’ site. We found enrichment for A at the first nucleotide and for T at the central nucleotide in the motifs from the former group of MAPs, and enrichment for C at the first nucleotide in the motifs from the latter group of MAPs (Bonferroni p < 0.05) (Materials & Methods; Supplemental Figures 7A-B). We found no significant difference in the abundancy of different types of motif (cold/neutral/hot), or of alleles from different IGHV families (Supplemental Figures 7C-D), with the Spearman correlation coefficients for abundancy rank being 1 and 0.9 respectively for the types of motif and the IGHV families. Overall, in terms of intrinsic characteristics, we found different patterns of nucleotide enrichment in MAPs showing differential SHM targeting.

Figure 5.

Potential factors related to the direction of (A) Position-dependent differential SHM targeting in 190 MAPs that tested significant in the meta-analysis. Gray shaded regions correspond to CDR1 and CDR2. (B) Fold change in mutation frequency at 5’ site over that at 3’ site versus distance between the 5’ and 3’ sites of a MAP, with the linear regression line shown in red. Relationship between the mutation frequency of each MAP and (C) the percentage of classical hot spots, or (D, E) the average mutability, in a neighborhood extending 16 (C, D) or 23 (E) 5-mer motifs in the 5’ direction and 27 5-mer motifs in the 3’ direction. Each line represents a MAP. MAPs for which the slope was undefined due to 5’ and 3’ sites having equal percentages of surrounding hot spots or average mutability were excluded. One-sided binomial tests were performed for the proportions of positive slopes being >0.5.

It has been suggested that SHM targeting decreases with distance from the transcription start site (24). Although we did not find that MAPs with an increased mutation frequency at the 5’ site were over-represented in our data (P > 0.05), the effect of distance may only be apparent when considering MAPs that are further apart. Thus, we examined whether there was an incremental relationship between the distance from the 5’ site of a MAP to its 3’ site, and the fold change in mutation frequency at the 5’ site over that at the 3’ site. We performed a linear regression and found that an x nt increase in the distance between the 5’ and 3’ sites of a MAP tended to correlate with a fold change in 5’ over 3’ site mutation frequency of 2^−0.00985x, or 0.99^x times (p = 1.3*10⁻¹⁰, t-test) (Figure 5B). The fact that 0.99^x is smaller than 1 for any positive x implies that MAPs whose motifs are farther apart tend to be targeted less frequently at the 5’ site compared to at the 3’ site, consistent with the observation in Figure 5A. In addition, the fact that 0.99^x decreases as x increases implies that there is a decremental relationship where a greater distance between the sites is associated with a smaller fold change in 5’ over 3’ site mutation frequency. Hence, we did not find evidence that the 5’ site of MAPs was targeted more frequently, or that greater distance between the 5’ and 3’ sites correlated with greater difference in SHM targeting.

The CDRs tend to have a higher density of classical hot spot motifs compared to the framework regions (FWRs) (p = 7.8*10⁻³⁷, one-sided paired Mann-Whitney test; Supplemental Figure 8A) and are known to carry higher mutational loads (43). We hypothesized that proximity to these regions might lead to higher SHM targeting that is not fully captured by the 5-mer motif context. Thus, we tested whether the MAP site closer to a CDR would have a higher mutation frequency. For each MAP, we compared the distance to the nearest CDR for the 5’ site and for the 3’ site, with the corresponding mutation frequency at each site. We did not find a significant relationship between the distance to the nearest CDR and the mutation frequency (p = 0.77, one-sided binomial test; Supplemental Figure 8B).

Even if not driven by proximity to a CDR, it is possible that increased mutability of the surrounding sequence leads to a higher mutation frequency by recruiting the SHM machinery to the local vicinity. Thus, we tested whether the more mutated site in each MAP tended to be found in a neighborhood with higher overall mutability, where mutability was either based crudely on the abundance of classical hot spots, or estimated more precisely by the updated S5F model. We first tested a neighborhood extending 16 5-mer motifs 5’ and 27 5-mer motifs 3’ of the motif. This neighborhood size is based on the finding that error-prone repair can introduce mutations extending at least 29 nt 5’ and 18 nt 3’ of the cytosine deaminated by AID (44). With respect to a mutation that is part of the mutational spread, the spread could extend 18nt 5’ and 29 nt 3’ of itself, or, equivalently, 16 5-mer motifs and 27 5-mer motifs in the 5’ and 3’ direction respectively. With this neighborhood size, we found a significant enrichment (p = 2.6*10⁻¹⁰, one-sided binomial test) amongst 72% of the MAPs for positive correlation between the mutation frequency in each MAP and the abundance of classical hot spots (either WA/TW or WRC/GYW or both), with no strand specificity (Figure 5C; Supplemental Figure 9). Similarly, we found a significant enrichment (p = 7.0*10⁻⁹, one-sided binomial test) amongst 71% of the MAPs for positive correlation with the average mutability of the neighborhood estimated by the S5F model (Figure 5D). In addition, we also investigated whether other neighborhood sizes could give a greater enrichment for MAPs with positive correlation between the observed differential SHM targeting pattern and S5F-based mutability of the neighborhood. We performed a grid search for a neighborhood size for which we saw the most MAPs with positive correlation (Materials & Methods). We found the optimal neighborhood size to be 23 5-mer motifs (25 nt) in the 5’ and 27 5-mer motifs (29 nt) in the 3’ direction. With this neighborhood size, we found a significant enrichment (p = 5.2*10⁻¹⁴, one-sided binomial test) amongst 77% of the MAPs for positive correlation (Figure 5E). Taken together, these results suggest that differences in the mutation frequency of the same motif at different positions are at least partly driven by the overall mutability of the wider neighborhood surrounding the motif.

Discussion

In this study, we examined a key assumption of the microsequence context-based SHM targeting model. Using curated BCR sequences from five published studies and under a MAP framework, we investigated whether it was equally likely for a DNA 5-mer motif to be targeted by SHM regardless of its position along the same IGHV allele. To avoid being skewed by the selection pressure exerted on B cells as part of affinity maturation, we analyzed either MAPs found on productively rearranged sequences in which the motif could only mutate synonymously, or MAPs found on non-productively rearranged sequences which presumably were not subject to selection.

We found statistically significant, position-dependent, differential SHM targeting for about two thirds of the 292 MAPs tested based on productively rearranged sequences and for about half of the 1,912 MAPs tested based on non-producitvely rearranged sequences. In both cases, the significant MAPs involved about three quarters of the unique motifs tested (29 out of 38 and 199 out of 269, respectively). Importantly, some of these MAPs involved classical WA/TW and WRC/GYW hot spots, suggesting that the “hotness” of a hot spot could also vary depending on its position. For a motif that occurs twice on the same allele, the difference in the likelihood that one site has a mutation while the other does not ranged from 0.092 to 1. The magnitude of difference correlated loosely and negatively with the number of sequences (Pearson correlation coefficients −0.43 and −0.40; Supplemental Figures 10A-B), suggesting either that the effect size decreased as the amount of data increases, or that more data was able to reveal more subtle differences. Either way, a difference that is at the minimum around 0.1, or 10%, should not be negligible. Indeed, when we performed the same analysis on simulated sequences into which the same number of mutations as observed were introduced according to SHM targeting probabilities specified by the current S5F model, we did not detect the same differential SHM targeting, suggesting that the observed effect was not due to chance. Hence, the next generation of motif-based SHM targeting models would be improved by taking into account of such position dependency.

Upon examining the direction of the observed differential SHM targeting, we found that it was largely consistent across individuals and across different IGHV alleles within the same gene family. Due to the small number of motifs analyzed that occurred at the same positions across IGHV alleles from different gene families (Figure 4E), it remains inconclusive whether there is a gene-specific effect to the differential SHM targeting observed. Furthermore, we found positive correlation with the mutability of the wider sequence neighborhood surrounding the motif. One mechanism that could underlie this correlation is spreading of mutations from the initial AID-induced lesion to the surrounding sequence neighborhood due to error-prone repair pathways (44). Under this model, 5-mers that are surrounded by AID hot spots are more likely to be in the range of mutation spreading events, resulting in a higher targeting probability. Consistent with this hypothesis, we found that the correlation between 5-mer motif targeting and the surrounding neighborhood was strongest when the neighborhood extended 25 nucleotides in the 5’ direction and 29 nucleotides in the 3’ direction from the central nucleotide of the motif. This is highly similar to the observation that SHM could spread mutations in a neighborhood extending 29 nucleotides in the 5’ direction and 18 nucleotides in the 3’ direction from the initial AID lesion site (44). Further studies are necessary in order to quantify the potential influence of mutational spread on position-dependent differences in SHM frequency. One possibility is to study SHM patterns in Ig sequences from antigen-experienced B cells that are deficient in MutSα-mediated mismatch repair (1). While knockout mice targeting MutSα have been reported (45-48), mutations in MutSα, a heterodimer made up of MSH2 and MSH6, are rare in human. However, high-throughput repertoire sequencing data from individuals with such mutations were recently reported by (49).

Our analysis focused on pairs of 5-mers that occurred at different positions along a germline IGHV allele. While we often visualized these MAPs as lines connecting two points, where the two points represented the two positions of occurrence (e.g., Figure 5A), these lines should not be interpreted as a trend. It is also important to point out that our conclusions do not require that a motif occurring more than twice along an allele follow a monotonically increasing or decreasing pattern amongst its multiple occurrences. Among motifs eligible for the S5F analysis, the maximum number of occurrences for a single motif within an allele was three. These motifs were captured by three separate MAPs: one for positions i and j; one for positions i and k; and one for positions j and k, where i<j<k. Within the 190 MAPs that tested significant for differential SHM targeting (Figure 5A), there were only14 cases in which the motif occurred three times along an allele (Supplemental Figure 10C). In 9 of these 14 cases, the mutation frequency of the motif did not monotonically increase or decrease from 5’ to 3’. Instead, for these cases, we found high correlation between the mutability of the local sequence neighborhood surrounding the motif and its mutation frequency (Supplemental Figure 10D), consistent with the broader results for MAPs.

A common view in the field holds that SHM targeting decreases with distance from transcription start (50-52). Although this finding is primarily attributed to a study by (24), (53) correctly points out that this study only established that SHM frequency exponentially decays downstream of the J segment, and did not investigate the V(D)J exon itself. The fact that we observed MAPs in the V segment with both positive and negative slopes of differential SHM targeting (Figure 5A), with many MAPs accumulating more mutations at the 3’ site, argues against a model where SHM frequency generally decreases with distance from transcription start. Thus, our results challenge the assumption that the exponential decay observed by (24) applies to the V(D)J exon.

Instead of modeling mutational spread around 5-mer motifs, an alternative possibility would be to expand the number of surrounding nucleotides in the original motifs and use 7-mer, 9-mer, or even 11-mer motifs. As the length of the motif increases, an increasing challenge that such an approach faces is that a possible motif that exists in theory may not exist in reality given the composition of the Ig germlines, thereby greatly reducing the number of motifs eligible for studying. In addition, it has been suggested that even extending the microsequence context to 7-mer motifs only accounts for 70%-80% of the variations in mutability (23). Indeed, based on their 7-mer motif model, (16) noted that in order to determine SHM hot spots, one also needs to consider the co-occurrence of nearby mutations, in addition to the 7-mer microsequence context.

While we found significant differential SHM targeting for MAPs involving alleles from all seven IGHV families, an inherent limitation of the allele-specific MAP framework is that it requires more data compared to a motif-based approach. Whereas a motif-based approach aggregates all sequences in which a motif is found in an allele-blind manner, the MAP framework requires sequences from a specific allele for the analysis of a particular MAP. Consequently, only 292 unique MAPs out of the theoretical set of 876 were tested in more than one subject in the S5F analysis; and only 1,912 unique MAPs out of the theoretical set of 12,704 were tested in the RS5NF analysis. Also, for subjects in the dataset from (33), we did not test for any MAP involving a WRC/GYW hot spot (Table 2), because the 5’ site of such MAPs all fell within the first 50 IMGT-numbered nucleotide positions (Supplemental Figure 1A), which were excluded from analysis as “burn-in” for the sequencing (Materials & Methods). In addition, while we used a version of the IMGT reference alleles that was most up-to-date at the time of analysis, and which contained computationally identified and experimentally validated novel alleles (54) from the datasets from (20), (30), and (31), we did not computationally infer novel alleles for the datasets from (29) and (33). Should novel polymorphisms exist in those two datasets, that could be a potential source of error as undetected polymorphisms would be mistaken for mutations. Nevertheless, the likelihood that a novel polymorphism happens to fall within one of the MAPs analyzed should be small. Indeed, it has been estimated that less than 0.1% of the mutations used for the S5F analysis should be affected by unknown polymorphism (12).

In summary, we conclude that the SHM targeting probability of a DNA 5-mer motif is not constant along the Ig sequence and depends on the position of the motif. This finding has significant implications for SHM analysis, as current analytical frameworks generally assume that mutation targeting can be accurately captured by a 5-mer sequence context or even more minimally (for example, the WRCY/RGYW hot spot). Incorporating positional-dependence will be an important challenge for future SHM targeting models, perhaps by explicitly modeling the initial AID-induced lesion, followed by mutational spread.

Key Points.

• Somatic hypermutation targeting for a DNA 5-mer motif can differ by position.

• Direction of differential targeting conserved across subjects and alleles.

• Differential targeting correlates with mutability of wider neighborhood of motifs.