Study participant QA013 was initially infected with a HIV clade D virus in 1995. At just over one year post initial infection (pii), a clade A variant was identified that had not been detected at the previous visit a few months earlier, indicating superinfection (SI) occurred in the intervening period: approximately 264–385 dpii (Chohan et al., 2005). Both viruses persisted in this individual for at least 6 years, and our lab previously isolated the bnAb QA013.2 at 2282 dpii (just over 6 years pii), along with several less potent nAbs that also contributed to plasma breadth. (Williams et al., 2018). To investigate the functional development of this V3/glycan-specific bnAb, we performed deep sequencing of the subject’s antibody repertoire using four longitudinal peripheral blood mononuclear cell (PBMC) samples spanning time points from pre-HIV infection (D-379) to 765 dpii (D765) (Figure 1A). With these deep sequencing data, computational inference methods designed for antibody lineage analyses were used to infer the probable naive B cell receptors (BCRs) that seeded the heavy and light clonal families of QA013.2. We then performed ancestral sequence reconstruction using Bayesian phylogenetics to infer the most probable routes of development from naïve BCR sequence to mature sequence (Figure 1B, Figure 1—figure supplements 1–4; Dhar et al., 2020; Ralph and Matsen, 2016). Antibodies were constructed based on the most statistically probable routes of development for the QA013.2 bnAb heavy and light chains, which involved eight variable heavy chain (VH) intermediate sequences and six variable light chain (VL) intermediate sequences between respective naïve BCR and mature sequences (Figure 1C, Supplementary file 1) Of note, several of the inferred intermediate sequences shared ≥98% nucleotide identity with sequences sampled from the longitudinal B cell repertoire of QA013, and in some cases, there were identical sequences observed (Figure 1C, Supplementary file 2).

The heavy chain clonal family of QA013.2 included a total of 42 members and this clonal family ranked among the top 10 largest clonal family clusters that were sampled in this subject’s immune repertoire. These clonally related sequences were derived from an inferred naïve BCR that used VH3-7*01, VD1-1*01, and VJ5*02 genes and had a 21-residue CDRH3 loop. The VH lineage matured over the course of six years to accumulate 20% somatic hypermutation (SHM) – which corresponded to a total of 38 amino acid mutations across the variable region in the mature QA013.2 bnAb (Figure 1C). The QA013.2 light chain clonal family included >1000 members likely due to overclustering, but it did not rank among the top 10 largest light chain clonal families that were sampled in the repertoire (also due to overclustering, see Materials and methods). The inferred naive light chain BCR used germline genes VL1-40*01 and VJ2*01 and the lineage acquired fewer mutations over the course of maturation, with a total of 12% SHM present in the mature lambda light chain (Figure 1C).

From the deep sequencing data, which was informed by samples up to 765 dpii, we observed that the inferred antibody lineages acquired limited SHM between the inferred naive BCRs and the latest inferred intermediates (Int6_VH and Int5_VL). More than 70% of the total SHM (84 of 117 nucleotide mutations across VH and VL regions combined) present in QA013.2 was not detected in the sequence data through 765 dpii, suggesting that either we did not sample sequences containing these mutations, or that these mutations accumulated between 765 dpii and the time when QA103.2 was isolated at 2282 dpii. Thus, we can only reliably infer the lineage through the early stages of development; the later stages of the bnAb’s evolutionary pathway are beyond the scope of the computationally inferred lineage (Figure 1B,C). Despite this lack of resolution, the lineage and longitudinal development that we were able to decipher offer a means to dissect the determinants of antibody binding, potency, and breadth for QA013.2. Here, we defined neutralization breadth as the mediation of cross-clade heterologous neutralization of Tier 2 HIV pseudoviruses as measured by the TZM-bl assay.

To determine whether key residues mediating breadth were represented in the inferred intermediate sequences defined through 765 dpii, we selected these intermediates (Int6_VH and Int5_VL; Figure 1), as well as inferred naive heavy and light chain BCR sequences and paired them together in various combinations (Figure 2A) to test their capacity to bind and neutralize HIV relative to the mature QA013.2 bnAb. The resulting monoclonal antibodies (mAbs) were tested for binding to autologous and heterologous Envelope protein (Env) by biolayer interferometry (BLI). The inferred naive BCR had weak but detectable binding to the Env-gp120 from a clade A autologous superinfecting virus variant from 765 dpii; there was even weaker binding to heterologous clade A BG505.SOSIP.664 trimer (Figure 2B,C). Although these data were insufficient for calculating reliable binding kinetics, they suggest that the naïve BCR of the QA013.2 bnAb lineage is capable of recognizing the superinfecting virus. The inferred naive BCR did not detectably bind to the clade D virus from initial infection (Figure 2—figure supplement 1).

By contrast, the mature QA013.2 bnAb exhibited strong binding kinetics to both the autologous SI gp120 and heterologous BG505 Env trimer, with equilibrium dissociation constants of 94.9 nM and 3.2 nM, respectively. A mAb consisting of the mature heavy chain paired with the inferred naive light chain (mat_VH 0_VL) bound to autologous SI gp120; however, the binding affinity was much weaker relative to the mature bnAb and not suitable for calculating binding kinetics. This same mAb bound to heterologous BG505 Env trimer (K_D = 89.6 nM) with an affinity that was approximately 30-fold less than that of the mature bnAb (Figure 2B,C). The converse mAb pairing (0_VH mat_VL) did not result in any binding capacity to either tested antigen.

Intermediate mAb Int6_VH Int5_VL, with 6% total SHM, represents the final inferred intermediates from the deep sequencing data, and they were detected in the sampled NGS sequences at 99–100% nucleotide identity, further validating the sequence of these inferred intermediates (Supplementary file 2). We could not detect binding of the intermediate mAb to the gp120 of the autologous SI virus, but there was clear evidence of binding to heterologous BG505 Env trimer (K_D = 513 nM) (Figure 2C), demonstrating HIV Env specificity. However, the binding affinity to heterologous trimer was more than 100-fold lower when compared to the mature bnAb. Together these studies demonstrate that VH is important for facilitating initial HIV recognition and binding. This finding is supported by data collected for both autologous Env-gp120 and heterologous native-like trimer, despite the native-like Env trimer yielding tighter binding affinities and dissociation constants for all tested mAbs. While early affinity maturation of the QA013.2 VH lineage led to measurable binding to heterologous Env antigen (Figure 2C), higher affinity binding required subsequent mutations, including acquisition of 30 non-synonymous mutations across the variable region between Int6_VH and mature_VH. (Figures 1B,C and 2B,C).

We tested the same lineage mAbs for neutralization using a subset of HIV Env pseudoviruses that were neutralized by the mature QA013.2 bnAb, including the same clade A autologous Env variant from 765 dpii used to generate gp120 for binding studies (Williams et al., 2018). Neutralization was not observed for the inferred naive BCR (0_VH0_VL) (Figure 2D). While binding was detected between the intermediate mAb (Int6_VHInt5_VL) and the heterologous BG505 Env trimer, it did not neutralize the same virus (Figure 2C,D). Thus, the inferred naive and intermediate Int6_VHInt5_VL mAbs represent early members of the lineage with the capacity to bind HIV Env trimer that have not yet acquired the ability to neutralize either autologous or heterologous variants. While the inferred Int6_VHInt5_VL intermediate was validated by near-perfect sequence matches in the sampled NGS sequences (Supplementary file 2), we cannot rule out the possibility that another lineage member with neutralization capacity was also present by 765 dpii, because deep sequencing antibody variable regions from blood PBMC samples often results in incomplete sampling (Horns et al., 2019; Laserson et al., 2014). Indeed, we observed low sequence read depth across all QA013 PBMC sample timepoints and antibody isotypes (Table 1).

When the mature VH was paired with the inferred naive light chain (mat_VH0_VL), autologous and cross-clade Tier two heterologous neutralization was achieved, suggesting that the approximate timeframe in which neutralization-mediating mutations were acquired in VH likely occurred between 765 dpii (resolution of computationally inferred lineage) and 2282 dpii (isolation of mature bnAb QA013.2). The inferred naive heavy chain paired with the mature light chain (0_VHmat_VL) did not result in any detectable neutralization of the virus panel (Figure 2D), suggesting that affinity maturation of the heavy chain is critical for imparting both binding and neutralization. When Int5_VL was paired with the mature heavy chain (mat_VH Int5_VL), there was an average fourfold increase in neutralization potency for all viruses tested compared to mat_VH0_VL (Figure 2D).

To better identify which residues contribute to QA013.2’s ability to bind and neutralize Env, we performed single particle cryo-electron microscopy (cryo-EM) on the QA013.2 Fab and BG505.SOSIP.664 trimer complex (Figure 3A,B). The known structure of BG505.SOSIP.664 trimer (PDB ID: 5aco) was fit into the resulting 4.15 Å resolution EM density map along with a homology model for the variable domain of QA013.2 Fab (Figure 3A,B). The map density corresponding to the Fab variable domain was well resolved, with prominent loops and separation of strands in beta sheet regions. These features facilitated high confidence docking of the Fab homology model into the EM map. Subsequent real-space refinement of the BG505.SOSIP.664 + QA013.2 Fab model against the map enabled a good fit of secondary structural elements within the EM density (Figure 3—source data 1). The complex structure of QA013.2 Fab bound to BG505.SOSIP.664 delineated antigen contacts characteristic of other V3/glycan-specific bnAbs, including conserved glycans as well as the linear GDIR motif (Barnes et al., 2018; Mouquet et al., 2012; Pejchal et al., 2011).

From the cryo-EM structure, it is apparent that QA013.2 interacts not only with the N332 glycan as previously reported (Williams et al., 2018), but also with the glycan at site N301, and the linear GDIR motif at the base of the V3 loop (Figure 3C–E). Key regions of the antibody paratope include the CDRH1 and CDRH2 which are nearest the N301 glycan, as well as CDRH3 and CDRL3, which are in close proximity to the N332 glycan. We observe that the CDR loops of the heavy chain contain many polar and charged residues, which likely facilitate hydrogen bonding interaction with the Env glycans and the charged residues in the GDIR motif (Figure 3—figure supplement 1A). The CDRH3 loop is distinctively long (21 residues) and appears to tilt away from the rest of the heavy chain toward the light chain, generating a cleft between CDRH2 and CDRH3 and exposing an extended interface that has been observed for other V3/glycan-specific bnAbs including 10–1074 and PGT121 (Figure 3—figure supplement 1B; Mouquet et al., 2012). Seven key residues in the Fab CDRH3 loop (L103 – I109) lie adjacent to the conserved GDIR motif. Based on our sequencing data of the sampled repertoire of this subject, these amino acids (excluding Y105 and D107) were likely introduced during affinity maturation of the heavy chain between 765 dpii and 2282 dpii because they were not detected by NGS through 765 dpii (Figure 1C), suggesting a potential role for these residues in facilitating antibody breadth and/or potency.

With paratope data in hand, we had the opportunity to show which regions of QA013.2 were playing a role in its functional activity by dissecting the regional and individual residue contributions interpreted through the structure. We began by examining the functional importance of VH mutations surrounding the CDR loops that were present in the mature antibody but not in the ≤765 dpii inferred intermediates. We synthesized six additional variations of QA013.2 VH, illustrated in Figure 4A, that we based on Int6_VH as the template and included all CDR-localized and CDR-adjacent FWR residues present in the mature QA013.2 VH (Supplementary file 3). All VH chimeras were paired with the mature light chain (mat_VL) and tested for neutralization capacity (Figure 4A). Mutations in individual CDRs, including adjacent FWR mutations, were not sufficient to achieve potent neutralization comparable to the mature antibody when paired with the mature light chain. We did observe weak autologous neutralization when mutations surrounding FWRH3-CDRH3 were added to Int6_VH. Mutations spanning both FWRH1-CDRH1 and FWRH3-CDRH3 of the mature VH together were required to obtain heterologous neutralization, consistent with the cryo-EM map of key antigen contacts (Figures 3 and 4A). Interestingly, the determinants of autologous and heterologous neutralization in this antibody lineage were different, in which autologous virus neutralization required less SHM and was facilitated by residues spanning CDRH2-FWRH3-CDRH3 (Figure 4A). To determine the importance of the framework residues and examine loss of function, we reverted all framework mutations in the mature VH back to the inferred naive BCR residues (∆FWR). This reversion resulted in no detectable binding or neutralization activity, demonstrating that residues in the heavy chain FWRs are essential for QA013.2 binding and neutralization, as has been noted for other HIV bnAbs (Figure 4A,B; Klein et al., 2013).

In a complementary approach, we used the mature VH as a template and reverted all of the mutations in FWRH1 alone (∆FWRH1), as well as those that spanned FWRH1-CDRH1 (∆CDR1_VH) and CDRH3 (∆CDR3_VH) back to the inferred naïve VH and as above, paired these variants with the mature light chain (Figure 4A). We observed identical increases in the half maximal inhibitory concentration (IC₅₀) neutralization of autologous superinfecting virus for both the ∆FWRH1 and the ∆CDRH1 chimeric antibodies, with a 89-fold increase in IC₅₀ neutralization relative to the mature bnAb (Figure 4A). We also observed a dramatic reduction in heterologous virus neutralization for these same VH chimeric antibodies (∆FWRH1 or ∆CDRH1). Specifically, we observed 733-fold and 250-fold increases in the IC₅₀ for ∆FWRH1 and ∆CDRH1 chimeric antibodies respectively, corresponding to decreased neutralization of clade C virus QC406.F3 (Figure 4A). Reversion of the six residues spanning FWRH1-CDRH1 also led to weakened binding to heterologous BG505.SOSIP.664 trimer (K_D = 40.6 nM) relative to the mature bnAb (Figure 4A,B). When the CDRH3 was reverted back to naïve (∆CDR3_VH), binding and neutralization of both autologous and heterologous virus was maintained, but there was an impact on potency across all viruses tested; although this antibody could still neutralize a heterologous clade C variant, it did not neutralize SF162 at the highest antibody concentration tested (Figures 3 and 4A,B). These data suggest that the CDRH3 does contribute to the overall breadth and potency of QA013.2, while mutations in FWRH1-CDRH1 are critical for achieving complete cross-clade Tier two neutralization breadth.

To further interrogate the above hypothesis, we generated single point mutations in the mature VH at residues of interest in FWRH1-CDRH1 and CDRH3 that reverted them to naïve (Figure 5A,B). These heavy chain point mutants were also paired with the mature light chain and tested for binding to Env and neutralization. We found that four of the six residues spanning FWRH1-CDRH1 individually contributed to the mature bnAb’s ability to neutralize autologous and heterologous virus. Residues E26, G28, and I24 were most critical, with average increase in IC₅₀ of 191-, 27-, and 25-fold, respectively, across all pseudoviruses tested when these residues were reverted to naïve (Figure 5C,D). While reversion of individual CDRH3 residues did not abolish neutralization, which correlates with the chimeric antibody results (Figure 4), one residue in particular (D106) demonstrated a much higher IC₅₀ value when mutated back to naive. D106 is adjacent to the GDIR motif in the structure. Reversion of this residue to naive did not inhibit clade A autologous virus neutralization, but did substantially impact heterologous neutralization, with an average 40-fold increase in IC₅₀ across all tested pseudoviruses (Figure 5E). However, the D106V point mutant only resulted in a modest decrease in binding to heterologous BG505.SOSIP.664 trimer relative to the mature bnAb (Figure 5F). These data suggest that while D106 is not critical for Env binding, this residue does contribute to heterologous neutralization, likely by interacting directly with the linear GDIR motif. Taken together, it is clear that although the antibody paratope of QA013.2 includes the CDRH3, where interaction with the N332 glycan and GDIR motif occurs, the QA013.2 paratope is also comprised of critical residues in FWRH1 and CDRH1.

Although less impactful than the heavy chain, our aforementioned data suggested that light chain development indeed contributes to QA013.2 neutralization potency (Figure 2). To identify which regions of the light chain were essential for these contributions, we reverted regions of SHM within Int5_VL back to naive (Int5∆FWR2-FWR3_VL and Int5∆CDR3_VL) and paired these with the mature heavy chain (Figure 6A). Reversion of the residues spanning FWRL2-FWRL3 abrogated antibody binding to BG505.SOSIP.664 heterologous Env (K_D = 1.3 µM) relative to the mature bnAb by more than 400-fold. This effect was less pronounced when residues in the CDRL3 were reverted to naïve (Figure 6A,B). These light chain variants also demonstrated reduced neutralization potency relative to the mature bnAb. Reversion of residues present in FWRL2-FWRL3 resulted in an average 19-fold increase in IC₅₀ (range 6–42 fold) and reversion of residues in CDRL3 caused a 14-fold increase in average IC₅₀ (range 1–41 fold) across all pseudoviruses tested when compared to the mature bnAb IC₅₀ values (Figure 6A). Together, these data demonstrate that residues in FWRL2-FWRL3 and CDRL3, which are situated along the cleft between heavy and light chains and near the N332 glycan, contribute to the mature bnAb’s binding affinity and neutralization potency, with residues in FWRL2-FWRL3 appearing to have a larger influence on bnAb binding. Notably, the residues spanning FWRL2-FWRL3 do not interact directly with the N332 glycan on Env. However, the residues in FWRL3 may influence the conformation of residues in the adjacent CDRL3 loop, which interacts with the N332 glycan. Furthermore, the residues in FWRL2 interact with FWRH4 of the Fab heavy chain (residues 117–120) which follows the CDRH3 loop (Figure 6C). Attempts to identify the light chain residues in FWRL2-FWRL3 responsible for mediating this interaction were inconclusive, as reversion of the residues located in FWRL2 and CDRL2-FWRL3, separately, did not influence the IC₅₀ values of the tested pseudoviruses, indicating that the cumulative effect of these mutations was beneficial (Figure 6A).

We used mutational antigenic profiling to interrogate the effect of all functionally tolerated single amino acid mutations in BG505.W6.C2 T332N Env on viral escape when under selective pressure from the mature QA013.2 bnAb (Dingens et al., 2017; Figure 7). In addition to the more traditional signatures of Env escape from V3/glycan-specific bnAbs, including mutations that disrupt the N301 and N332 potential N-linked glycosylation sites (PNGS) or alter the GDIR motif (Dingens et al., 2019), we also observed viral escape via mutations in the V1 loop (Figure 7A,B). In addition to the two PNGS, enrichment of viral escape mutations was strongest at sites D325, R327, and H330; validation of these data using pseudoviruses with individual mutations in TZM-bl neutralization assays showed that mutations to these sites can result in >32-fold change in IC₅₀ relative to the wild-type BG505.T332N (Figure 7B,C). However, not all mutations to these sites resulted in viral escape. Similar to 10–1074 (Dingens et al., 2019), while D325E was enriched in mutational antigenic profiling and had a >32-fold effect in the TZM-bl assay, D325N/S mutants were not enriched and had negligible effects when tested individually via TZM-bl neutralization assay.

Viral escape in the V1 loop occurred at two sites that are adjacent in the BG505 structure; D141 and R151 (HXB2 numbering), both of which caused a modest ~3-fold change in IC₅₀ relative to the wildtype virus (Figure 7B,C). While the V1 loop of Env does not appear to be in direct contact with QA013.2 Fab in the complex structure, the backbone Cα atoms of the Fab CDRH3 loop and Env V1 loop are within 8 Å of each other (Figure 7D), suggesting there could be amino acid side-chain interactions between the two loops that can affect neutralization.

Interestingly, while the glycan at N332 is essential for bnAb neutralization and disrupting this motif results in viral escape, we observed that it is not required to achieve antibody binding – as demonstrated by the mature bnAb binding to the gp120s from the autologous clade D and BG505.W6.C1 viruses, both of which lack the PNGS at site 332 (Figure 7—figure supplement 1). It is likely that this binding, while measurable, is too weak to confer neutralization, confirming the importance of the N332 glycan.

The QA013.2 bnAb has previously been shown to potently neutralize tier two viruses from clades A, B, and C, as well as several tier three viruses from clades B and C, but it did not neutralize clade D or A/D recombinant viruses (Williams et al., 2018). This study demonstrated the importance of the glycan at N332 for QA013.2-mediated neutralization, and this finding was corroborated in the cryo-EM structure, functional mutagenesis, and viral escape data that we present here (Figures 3–5 and 7). The structural data also suggest that the N301 glycan is playing a major role in facilitating antibody neutralization, based on its location adjacent to FWRH1-CDRH1 residues in the antibody VH region that we show are critical for bnAb breadth (Figures 3–5). Moreover, the antigenic profiling data highlight this glycan site as a key region facilitating viral escape in the V3 loop (Figure 7). However, the role of the glycan at N301 had not been examined directly in prior studies. To confirm the dependence of QA013.2 bnAb on the glycan at N301, we generated a N to K mutation at position 301 in BG505.W6.C2 T332N, which is the virus used for antigenic profiling, as K301 was identified as a major escape mutation (Figure 7). Elimination of the glycan at N301 abolished all neutralization activity despite the presence of the N332 glycan in BG505.W6.C2 T332N (Figure 8). These data demonstrate that both N301 and N332 glycans in combination are required to achieve bnAb neutralization.

We also tested the mature bnAb against an expanded panel of clade A and clade D pseudoviruses to better define the specificity of this bnAb with regard to the clade D initial infecting virus and clade A superinfecting virus. Similar to the findings with autologous virus, QA013.2 neutralized clade A but not clade D viruses; interestingly, the bnAb also did not neutralize D/A recombinant variants. Clade A virus neutralization was not universal, notably those viruses missing either glycan or the GDIR motif were not neutralized, which aligns with the predictions of key antigen contacts identified in the cryo-EM structure of QA013.2 bound to native-like Env trimer (Figure 3). The mature bnAb is only capable of neutralizing viruses that contain N301, N332, and the linear GDIR motif (Figure 8).

Interestingly, there are some viruses with all three of the V3 features that are not neutralized, which implies that there are other key residues involved. The antigenic profiling data demonstrate that charged residues in the V1 loop are also important for neutralization by QA013.2 (Figure 7). Indeed, we noted that all viruses that are neutralized by QA013.2 contain a (negative charge/branched hydrophobic/positive charge) motif in the final 10 residues of the V1 loop, except MK184.W0.D1, which is a clade C/D recombinant with clade D V1-V5 sequence (Wu et al., 2006). The four viruses that contained the main antigenic contacts in Env (N301, N332, and GDIR) but were not neutralized by the mature bnAb were either subtype D in variable loops 1 and 2 (D/A recombinants; Wu et al., 2006), or they lacked this [negative charge/branched hydrophobic/positive charge] in V1. Overall, the neutralization results from this expanded panel demonstrates that the mature bnAb can only neutralize viruses from clades A, B, and C, not clade D (Figure 8), as suggested by studies using a smaller virus panel (Williams et al., 2018) and that V1 influences the antibody’s function.

Peripheral blood mononuclear cell (PBMC) samples were collected between 1994–1997 from female subject QA013 at various times pre- and post-HIV infection. QA013 was enrolled in a prospective cohort of HIV negative, high-risk women in Mombasa, Kenya and was 30 years old at the time of her enrollment in 1993 (Martin et al., 1998). During the study, QA013 seroconverted following initial infection with a clade D virus (1995), and was subsequently superinfected with a clade A virus between 265 and 385 days post initial infection (Chohan et al., 2005). Study participants were treated according to Kenyan National Guidelines and did not receive any antiretroviral treatment during the time period in which the samples used in this study were collected. Antiretroviral therapy was offered as of March 2004, with support from the President’s Emergency Plan for AIDS Relief. This study was approved by members of the ethical review committees at the University of Nairobi, the Fred Hutchinson Cancer Research Center, and the University of Washington. Study participants provided written informed consent prior to enrollment.

For antibody production: HEK 293 F cells (RRID:CVCL_6642; originally derived from female human embryonic kidney cells) were obtained from Invitrogen (Thermo Fisher Scientific, Waltham, MA, catalog #R790-07) and grown at 37°C in Freestyle 293 Expression Medium (Thermo Fisher Scientific, catalog #12338002) in baffle-bottomed flasks rotating at 135 rpm. These cells were tested for mycoplasma and found to have no contamination. We verified the identity of these cells through short tandem repeat (STR) DNA profiling and found that they are 92% identical to HEK 293-FT cells (RRID:CVCL_6911), which are derived from the parent HEK 293-F cells (RRID:CVCL_6642) based on 15 markers.

For pseudovirus production: 293 T cells (RRID:CVCL_1926; a transformed cell line originally derived from female human embryonic kidney cells) were obtained from ATCC (Manassas, VA, catalog #CRL-3216) and grown at 37°C in DMEM media with added fetal bovine serum (10%), penicillin (10,000 units/mL), streptomycin 10,000 μg/mL, and Amphotericin B (250 ng/mL) (all from Thermo Fisher Scientific). These cells were tested for mycoplasma and found to have no contamination. We verified the identity of these cells through short tandem repeat (STR) DNA profiling and found that they are 91% identical to Anjou 65 cells (RRID:CVCL_3645), which are derived from the parent HEK 293T/17 cells (RRID:CVCL_1926) based on 8 markers.

For neutralization assays: TZM-bl cells (RRID:CVCL_B478; originally derived from female cancerous human cervical tissue) were obtained from NIH AIDS Reagent Program (Germantown, MD, catalog #ARP-8129) and grown at 37°C in DMEM media with added fetal bovine serum (10%), penicillin (10,000 units/mL), streptomycin 10,000 μg/mL, and Amphotericin B (250 ng/mL) (all from Thermo Fisher Scientific). These cells were tested for mycoplasma and found to have no contamination. We verified the identity of these cells through short tandem repeat (STR) DNA profiling and found that they are 93% identical to HeLa cells (RRID:CVCL_0030) based on 15 markers.

Four longitudinal PBMC samples from 379 days pre-HIV infection (D-379), D-351, 385 days post initial infection (D385), and D765 were thawed in a 37°C water bath, diluted 10-fold in RPMI, and pelleted at 300 x g for 10 min. Cell pellets were washed with 1X phosphate-buffered saline (PBS), counted using trypan blue solution, centrifuged, and then total RNA was extracted using the AllPrep DNA/RNA Mini Kit (Qiagen, Germantown, MD) according to the manufacturer’s protocol. Isolated RNA was stored at −80 °C.

Antibody sequencing was performed as previously described (Simonich et al., 2019; Vigdorovich et al., 2016) using the SMARTer RACE 5’/3’ Kit (Takara Bio USA Inc, Mountainview, CA). Briefly, RACE-ready cDNA synthesis was performed using primers specific for IgM, IgG, IgK, and IgL as previously described (Simonich et al., 2019). We used molecular barcoding to eliminate errors introduced during library preparation and sequencing, which facilitated accurate measurement of genetic diversity and improved ancestral sequence reconstruction. Template switch adaptor primers that included unique molecular identifiers were used in the cDNA synthesis step: SmartNNNa 5’ AAGCAGUGGTAUCAACGCAGAGUNNNNUNNNNUNNNNUCTTrGrGrGrGrG 3’ (Turchaninova et al., 2016), where ‘rG’ indicates ribonucleoside guanosine bases. Following synthesis, cDNA was diluted in 10 mM Tricine-EDTA according to the manufacturer’s recommended protocol. First-round PCR of 14 cycles was performed using Q5 High-Fidelity Master Mix (New England BioLabs, Ipswich, MA) and nested gene-specific primers at 20 µM. First-round amplicons were used as templates for second-round PCR amplification (six cycles), which added MiSeq adaptors. Second-round products were gel purified and indexed using Nextera XT P5 and P7 indices (Illumina, San Diego, CA). Gel-purified indexed libraries were quantified using the KAPA library quantification kit (Kapa Biosystems, Wilmington, MA) using an Applied Biosystems 7500 real-time PCR machine. Libraries were denatured and loaded onto an Illumina MiSeq using 600-cycle V3 cartridges, according to the manufacturer’s instructions.

Following next-generation sequencing, sequence reads were pre-processed into amplicons using FLASH, primers were trimmed using cutadapt, and the FASTX-toolkit was used to remove sequence reads containing low-confidence base calls (N’s) as previously described (Simonich et al., 2019; Vigdorovich et al., 2016). Sequence reads from each timepoint were merged together into an aggregate dataset that was deduplicated and annotated using partis (https://github.com/psathyrella/partis) with default settings. Sequences that were out of frame or contained internal stop codons were removed, while singletons, or sequences that were observed only once in the sampled repertoire, were included in an attempt to retain undersampled or rare sequences. Sequencing statistics for this study can be found in Table 1.

Clonal family clustering was performed on both individual timepoint and aggregated datasets using both partis seeded and unseeded clustering methods (Ralph and Matsen, 2016). The seeded method extracts all sequences that are clonally related to a particular ‘seed’ sequence of interest. The ‘seeds’ here refer to the previously-isolated QA013.2 VH and VL sequences (Williams et al., 2018). Unseeded clustering, on the other hand, clusters all sequences in the repertoire into clonal families. Since this is more computationally demanding, some of the very large datasets were subsampled down to 50–150 k sequences. Three random subsamples were analyzed and compared in order to measure statistical uncertainties. No subsampling was done for the seeded analyses.

Because the light chain has much lower diversity than the heavy chain, and because B cells proliferate after successful heavy chain rearrangement but before light chain rearrangement, light chain clustering without heavy/light chain pairing information inevitably incorrectly clusters together many different light chain families. This overclustering artificially inflates the number of sequences in the clonal family, resulting in most inferred light chain families containing sequences stemming from several different light chain rearrangements.

The light chain family resulting from clonal family clustering was very large (1158 members), and because this apparent family was likely composed of sequences from several different light chain rearrangements, each of which are likely confined to separate lineages, a pruning procedure was performed based on phylogenetic relatedness to the QA013.2 antibody chain using FastTree 2 (Price et al., 2010). This resulted in the light chain clonal family being reduced to the most relevant 100 sequences (https://github.com/matsengrp/cft/blob/master/bin/prune.py). The VH clonal family did not undergo any such pruning as it contained fewer than 100 sequences (note that the tenfold different in size between VH and VL clonal families is itself strong evidence for overclustering).

The heavy and pruned light chain clonal families were then analyzed with linearham (https://github.com/matsengrp/linearham), a phylogenetic hidden Markov model that accounts for both VDJ recombination and SHM when inferring ancestral sequences including the inferred naive B cell receptor (BCR) (Dhar et al., 2020). Partis software also infers the unmutated common ancestor, or naive sequence; however, the version of partis used here assumed a balanced, star-like clonal family tree architecture, and the alternative naive sequences it returns are not based on a rigorous probabilistic model (current versions of partis no longer assume star-like phylogenies). Linearham is thus more accurate for trees, and also returns proper uncertainty estimates, because it samples naïve sequences from their posterior distribution and accounts for uncertainty in the naive rearrangement process.

As part of the linearham package, RevBayes (Höhna et al., 2016) was run using an unrooted tree model with the general time-reversible (GTR) substitution model. Each RevBayes run had customized settings to ensure that likelihood and posterior estimated sample sizes were >100 for the heavy and light chain lineages. For both VH and VL lineage inference, MCMC iterations were 100,000 with a thinning frequency of 100 iterations and 10,000 burn-in samples. Linearham generates summary graphics following phylogenetic inference that display only the edges that satisfy the given posterior probability threshold, and only the nodes that contact edges above the threshold. These graphics detail the relative confidence in unique inferred sequences and amino acid substitutions (Figure 1 and Figure 1—figure supplements 2 and 3). Inferred intermediate sequences present in the most probable lineage paths were selected for study (see Figure 1).

To determine whether the computationally inferred naive BCR and intermediate ancestor sequences for VH and VL lineages were observed in the sampled NGS sequences, we ran the following script: https://github.com/matsengrp/cft/blob/master/bin/blast.py. A local BLAST database was created for the two seeded clonal families and queried for sampled NGS sequences that had high nucleotide sequence identity to each of the VH and VL lineage members using the ‘blastn’ command (Biopython package). E-value of 0.001 was used; other settings were default. Blastn ‘hits’ for each lineage member were sorted according to their percent nucleotide identity and alignment length. Percent nucleotide identity of the top blastn ‘hit’ for each lineage member is listed in Figure 1.

Antibody heavy and light chain sequences were synthesized as ‘fragmentGENES’ by Genewiz (GENEWIZ, South Plainfield, NJ) or gene blocks by Integrated DNA Technologies (IDT, Coralville, IA) and cloned into IgG1 and IgL vectors as previously described (Williams et al., 2015a). Human embryonic kidney (HEK) 293 F cells (RRID:CVCL_6642; Thermo Fisher Scientific, Waltham, MA, catalog #R790-07) were transfected with equal amounts of heavy and light chain plasmid DNA using the Freestyle Max transfection system (Thermo Fischer Scientific, catalog #16447100) according to the manufacturer’s instructions. Transfections proceeded at 37°C with 8% CO₂ in baffled-bottom flasks, rotating at 135 rpm and were harvested after 3–6 days. Supernatants were collected following centrifugation of cells at 1900 rpm for 10 min. Antibody supernatants were passed three times over columns packed with immobilized Protein G resin (Thermo Fisher, catalog #20397) as previously described (Scherer et al., 2014).

Antibody chimeras were synthesized as fragmentGENES as shown in Figure 4A, while VH mutagenesis (Figure 5) was performed using phosphorylated primers (25 µM) that introduced the mutation of interest into the variable heavy chain, using mature QA013.2 VH DNA in the IgG1 plasmid as a template (2.5 ng/µL). PCR-mediated mutagenesis was performed using Phusion Hi Fidelity polymerase (Thermo Fisher Scientific, catalog #F-530) with the following thermocycler settings: denature at 98°C for 1 min, 30 cycles of 98°C for 30 s, annealing at 55–70°C for 30 s, extension at 72°C for 5 min, and final extension at 72°C for 10 min. PCR products were digested with DpnI (New England BioLabs, catalog # R0176) overnight at 37°C and purified using Qiagen PCR Purification kit (Qiagen, catalog #28104), with products eluted in ddH₂0. Mutants were ligated at room temperature for at least two hours prior to transformation into One Shot TOP10 chemically competent cells (Thermo Fisher Scientific, catalog #C404003). Mutant antibody chains were sequence-confirmed prior to transfection and antibody purification as described above.

HIV pseudovirus (PSV) was produced in HEK 293 T cells (ATCC, Manassas, VA, CRL-3216) and used in TZM-bl neutralization assays as previously described (Goo et al., 2012). Briefly, monoclonal antibodies were incubated with PSV at 37°C for 1 hr and TZM-bl cells (obtained through the NIH AIDS Reagent Program, Germantown, MD, catalog #8129 from Dr. John C. Kappes, and Dr. Xiaoyun Wu) were added at 10,000 cells/well and the assay was harvested after 48 hr at 37°C by adding Gal-Screen substrate (Thermo Fisher Scientific, catalog #T1028) as previously described (Simonich et al., 2019). IC₅₀ values represent the half maximal inhibitory concentration in ug mL⁻¹ at which 50% of the virus was neutralized. All neutralization data are the average of at least two independent experiments, each performed in technical duplicate (see Figure 2—source data 1; Figure 4—source data 1; Figure 5—source data 1; Figure 6—source data 1; Figure 7—source data 1; Figure 8—source data 1). IC₅₀ values that deviated from the technical replicates by >10-fold were excluded as outliers from the average IC₅₀ calculation.

QA013 monoclonal antibody binding was measured using biolayer interferometry on an Octet RED instrument (Forte Bio, Fremont, CA). All antibodies were diluted to 10 µg mL⁻¹ in a 0.22 µm filtered binding buffer solution composed of 1X PBS, 0.1% bovine serum albumin, 0.005% Tween-20, and 0.02% sodium azide. Antibodies were immobilized onto anti-human IgG Fc capture biosensors (Sartorius, Göttingen, Germany, catalog #18–5063) and dipped into analyte solution of several different antigens as detailed in the text (Figures 2 and 4–6): autologous gp120 monomers from the initial infecting virus Env (clade D) at 2 µM, the superinfecting virus variant from 765 days post initial infection (clade A) at 2 µM, heterologous BG505.SOSIP.664 Env trimer containing the N332 glycan (clade A) at 500 nM, or wildtype BG505.W6.C1 gp120 monomer at 2 µM. Antigens were diluted to their respective starting concentrations in the same buffer as above and a series of six, twofold dilutions were tested as analyte in solution at a shake speed of 600 rpm at 23°C. The kinetics of mAb binding were measured as follows: association was monitored for 180 s, dissociation was monitored for 180 s, and regeneration was performed in 10 mM Glycine HCl (pH 1.5, GE Healthcare, Chicago, IL, BR-1003–54). Binding-affinity constants (K_D; on-rate, k_on; off-rate, k_dis) were calculated using ForteBio’s Data Analysis Software 7.0. Responses (nanometer shift) were calculated using data that were double reference subtracted using reference wells and non-specific binding of biosensor to analyte. Data were processed by Savitzky-Golay filtering, prior to fitting using a 1:1 model of binding kinetics. Prism v9.0 (GraphPad, San Diego, CA) was used to plot both the processed data (following double reference subtraction) and fit data (after global analysis) of each antibody + antigen combination. Each biolayer interferometry experiment presented in this study was performed a minimum of two independent times, each time using freshly diluted antibodies in binding buffer (see Figures 2 and 4–6 – BLI Source Data 1 and 2). Both QA013 gp120 monomers were synthesized as custom proteins through Cambridge Biologics (Brookline, MA). BG505.SOSIP.664 trimer was expressed and purified as previously described (Verkerke et al., 2016) and the BG505.W6.C1 gp120 protein was ordered through Immune-Tech (Immune Technology Corp., New York, NY, IT-001–176 p).

Purified BG505.SOSIP.664 trimers were mixed with threefold molar excess of QA013.2 Fab (fragment antigen binding) in 1X PBS at pH 7.4. The mixture was diluted with PBS to approximately 0.04 mg/mL final protein concentration and incubated for 45 min on ice. Lacey grids with a thin continuous carbon film (400 mesh) (Electron Microscopy Sciences, Hatfield, PA) were glow discharged (negative charge) under vacuum using 20–25 mA current for 30 s. A 3.0 µL aliquot of sample was applied to these grids at 4°C and 100% humidity, blotted for 3.5 s and immediately plunge frozen in liquid ethane using a Vitrobot Mark IV specimen preparation unit (FEI Company, Hillsboro, OR).

Vitrified grids were imaged using an FEI Titan Krios operating at 300keV equipped with a Gatan K2 summit direct detector and a post-specimen energy filter. Micrographs were collected at ×105,000 magnification with a pixel size of 1.35Åper pixel in counting mode. Each image received a dose rate of ~8e⁻/pixel/s with 200 ms exposure per frame and 50 frames per image collected. Estimated defocus ranged from 1.7 to 3.5 µm. Data were collected in two separate sessions with a total of 6475 micrographs using the automated data collection software Leginon (Carragher et al., 2000).

All data processing steps were performed within the Relion (version 3.0.7) software setup (Scheres, 2012). Frame alignment and dose-weighting were performed using MotionCor2 (Zheng et al., 2017), and CTF estimation was done using CTFFIND4 (Rohou and Grigorieff, 2015). Further processing was continued with Relion. A total of 1,082,742 particles were picked using the LOG-based automatic particle picking routine. The particles were initially extracted at 4x binning with a pixel size of 10.8 Å/pixel. The binned particle stack was subjected to two rounds of unsupervised 2D classification, out of which classes that clearly showed SOSIPs with three Fabs bound were selected. A total of 113,470 particles from these classes were used for 3D refinement. The previously published negative stain reconstruction of BG505.SOSIP.664 complexed to QA013.2 Fab (EMD-7471) was used as an initial model (Williams et al., 2018). The initial model was low-pass filtered to 60 Å and refinement was performed on the full particle stack with C3 symmetry imposed. Subsequent map sharpening and post processing in Relion resulted in a 4.186 Å structure using the ‘gold-standard’ FSC cutoff of 0.143. The unbinned particle stack was further subjected to CTF-refinement at per micrograph level which improved the map to its final resolution of 4.15 Å (Figure 3—figure supplement 2).

The atomic model of BG505.SOSIP.664 trimer (PDB ID: 5ACO) (Lee et al., 2015) was rigid body fitted into the 4.18 Å resolution map using UCSF Chimera (Pettersen et al., 2004). No further modification to the SOSIP model was performed as the atomic model agreed well with the map. In the cryo-EM map, the variable region of the Fab domain was well-resolved with separation in beta-sheets observable, but the constant domain of the Fab was only partially resolved. Thus, a QA013.2 Fab model consisting of only its variable heavy and light chains was generated using the SWISS model software (Waterhouse et al., 2018, p.). Five different models for QA013.2 Fab variable domain model were computationally generated using available PDB structure templates. All five models were comparable when superposed onto one another, except for the orientation of the variable CDRH3 loop. The Fab model generated by homology modeling with the crystal structure of an infant antibody ADI 19425 (PDB ID:6apc) (sequence identity of 71.18% and 0.97 coverage) was selected as the final model for further analysis.

The cryo-EM map was sharpened with BG505.SOSIP.664 atomic model (PDB ID:5ACO) input using the Autosharpen program (Terwilliger et al., 2018) in the Phenix software suite (Liebschner et al., 2019), in order to enhance features in the Fab density. The QA013.2 Fab model was then computationally docked into the map using ‘Dock In Map’ program in the Phenix software suite (Liebschner et al., 2019). The SOSIP+Fab model was then refined using Realspace Refinement in the Phenix package using reference model restraints from a high-resolution atomic model of BG505.SOSIP.664 (PDB ID:6mu6) (Liebschner et al., 2019). The data collection, refinement parameters and final model statistics are summarized in Figure 3—source data 1, and the cryo-EM map and coordinates are included in Figure 3—source data 2.

Mutational antigenic profiling was performed as previously described (Dingens et al., 2017). Briefly, 0.5–1.0x10⁶ TZM-bl infectious units of independent BG505.W6.C2 T332N mutant proviral DNA libraries (Haddox et al., 2018) were incubated with the QA013.2 bnAb at ≥ IC₉₅ concentration for one hour. One million SupT1.CCR5 cells were then infected with these mutant libraries in the presence of 100 µg/mL DEAE-dextran (VWR, Radnor, Pennsylvania, catalog #97061–684). In parallel to QA013.2 antibody selection, each mutant virus library was also infected into 1x10⁶ SupT1.CCR5 cells without antibody selection to serve as the non-selected control. Non-integrated viral DNA was isolated from each pool of infected cells at 12 hr post infection using a Qiagen Miniprep kit. The env gene was amplified from each sample (selected and non-selected) using a barcoded subamplicon sequencing approach that generates seven tiling subamplicons across env as previously described (Haddox et al., 2018). These pools of amplicons were then deep sequenced on an Illumina HiSeq using 250 bp paired-end reads.

Following deep sequencing, differential selection values were calculated as described in Doud et al., 2017. This included calculating the enrichment of each amino acid at individual sites relative to the wild-type by comparing the mutation counts of each residue in the antibody selected sample to the mock treated sample. Percent infectivity and correlation values between replicates are provided as Figure 7—figure supplement 1. We took the median values across all experimental replicates and visualized the differential selection data on logoplots rendered by dms_tools2 using weblogo (Crooks et al., 2004). Throughout this manuscript, we focused only on positively enriched mutations.

To validate the mutational antigenic profiling data, TZM-bl neutralization assays were performed using BG505.T332N pseudoviruses bearing single point mutants representing residues that were enriched for viral escape at individual sites across Env. TZM-bl neutralization assays were performed as described above in at least two independent assays, each performed in technical duplicate (see Figure 7 – Neut Source Data 1). The fold change in IC₅₀ of each point mutant relative to BG505.T332N wild-type pseudovirus was calculated independently for each experiment and then averaged across all replicates.

Longitudinal antibody gene sequencing data of QA013’s B cell repertoire can be found at BioProject SRA, accession PRJNA674442. The EM map and atomic coordinates for QA013.2 complexed to BG505.SOSIP.664 are deposited under accession codes EMD-24195 and PDB 7N65.

The computationally inferred antibody variable region sequences, chimeras, and mutant antibody chains produced in this study have not been deposited into GenBank because computationally-generated sequences are not accepted, but they are available in Supplementary files 1 and 3. Mature QA013.2 antibody chain GenBank accession numbers are MH003558.1 (VH) and MH003564.1 (VL). GenBank accession numbers for HIV Env variants used in this study can be found in the Key Resources Table above.

The open-source software used in this study is publicly available on GitHub: partis (https://github.com/psathyrella/partis; Ralph, 2021a), pruning of clonal family sequences (https://github.com/matsengrp/cft/blob/master/bin/prune.py; Harkins, 2021a), linearham (https://github.com/matsengrp/linearham; Jasti, 2021), and Blast validation (https://github.com/matsengrp/cft/blob/master/bin/blast.py; Harkins, 2021b). The specific commands and run logs for the lineage analysis can be found on Zenodo. Open-source software to analyze mutational antigenic profiling datasets is available at https://jbloomlab.github.io/dms_tools2/ (Bloom, 2021).