Utilizing Machine Learning to Improve Neutralization Potency of an HIV‑1 Antibody Targeting the gp41 N‑Heptad Repeat

Abstract

The N-heptad repeat (NHR) of the HIV-1 gp41 prehairpin intermediate (PHI) is an attractive potential vaccine target with high sequence conservation across diverse strains. However, despite the potency of NHR-targeting peptides and clinical efficacy of the NHR-targeting entry inhibitor enfuvirtide, no potently neutralizing NHR-directed monoclonal antibodies (mAbs) nor antisera have been identified or elicited to date. The lack of potent NHR-binding mAbs both dampens enthusiasm for vaccine development efforts at this target and presents a barrier to performing passive immunization experiments with NHR-targeting antibodies. To address this challenge, we previously developed an improved variant of the NHR-directed mAb D5, called D5_AR, which is capable of neutralizing diverse tier-2 viruses. Building on that work, here we present the 2.7Å-crystal structure of D5_AR bound to NHR mimetic peptide IQN17. We then utilize protein language models and supervised machine learning to generate small (n < 100) libraries of D5_AR variants that are subsequently screened for improved neutralization potency. We identify a variant with 5-fold improved neutralization potency, D5_FI, which is the most potent NHR-directed monoclonal antibody characterized to date and exhibits broad neutralization of tier-2 and −3 pseudoviruses as well as replicating R5 and X4 challenge strains. Additionally, our work highlights the ability of protein language models to efficiently identify improved mAb variants from relatively small libraries.

Introduction

Since the discovery of HIV-1 as the causative agent of AIDS more than 40 years ago, scientists have tirelessly pursued the development of an effective prophylactic HIV vaccine. Unfortunately, these efforts have yet to bear fruit, with every major HIV vaccine efficacy trial ending in failure. Most recently, the failure of the Mosaico (HVTN 706) and PREPVacc trials in 2023 extinguished the hope that an HIV vaccine would be available this decade. ,

The challenges to the development of an HIV-1 vaccine are numerous, − among the greatest being HIV-1’s high mutation rate and resulting enormous genetic diversity. In particular, Env, the only HIV-1 surface protein and the main target for HIV-1 neutralizing antibodies, is incredibly sequence-diverse among HIV-1 strains. Much of the HIV vaccine field is focused on the identification and characterization of broadly neutralizing monoclonal antibodies (bnAbs) capable of neutralizing viruses from divergent strains and clades, with the subsequent goal of developing vaccines to elicit antibodies similar to those bnAbs. , The vast majority of bnAbs identified to date target the prefusion native state of Env, before the conformational changes have begun that enable HIV-1 viral fusion. , Consequently, nearly all HIV-1 vaccine development efforts thus far have focused on vaccine antigens mimicking prefusion Env.

Work by our group and others has instead focused on antibody discovery and vaccine approaches targeting the highly conserved fusion intermediate form of Env. During HIV-1 membrane fusion, the gp120 subunit of Env first binds to CD4 and coreceptors, triggering a conformational change in which the N-terminal hydrophobic fusion peptide of gp41 inserts into the host cell membrane. The highly conserved N- and C-heptad repeats (NHR and CHR, respectively) of gp41 then span the intermembrane distance for a period of minutes , until the NHR and CHR snap together to form the energetically favorable trimer-of-hairpins, physically pulling the two membranes closer together to drive membrane fusion. , This transient conformation in which the NHR and CHR extend between the viral and host cell membranes is known as the prehairpin intermediate (PHI) and has been proposed as a potential vaccine target.

The PHI, and in particular the NHR region of the PHI, possesses several attributes favorable for a potential vaccine target. First, the NHR is highly conserved, with amino acid residues comprising the gp41 hydrophobic pocket ≥95% sequence conserved among strains in the Los Alamos National Laboratories sequence repository. Indeed, a trimeric D-peptide fusion inhibitor targeting the gp41 pocket demonstrated a high barrier to resistance development in viral passaging experiments and as monotherapy was capable of preventing viral rebound in chronically infected rhesus macaques after discontinuation of combination antiretroviral therapy. ,− Second, the NHR can be neutralized by antibodies with lower levels of somatic hypermutation than those targeting epitopes on prefusion Env, suggesting it may be possible for germline antibodies to bind to NHR-based antigens in contrast to germline-inferred versions of most bnAbs which cannot bind native Env. These results are promising for potential NHR-based vaccines, as binding of germline antibodies to immunogens is a necessary first step toward further in vivo affinity maturation. Finally, a peptide fusion inhibitor targeting the NHR, enfuvirtide, received FDA approval in 2003, validating the NHR as a safe and effective clinical target in humans. ,

Nonetheless, potently neutralizing NHR-targeting antibodies have thus far not been identified. While various versions of the three-stranded coiled coil formed by the NHR have been created and used as vaccine candidates in animals, − the neutralization potencies of these antisera, as well as those of first generation anti-NHR monoclonal antibodies (mAbs), − were modest and mostly limited to HIV-1 isolates that are highly sensitive to antibody-mediated neutralization (referred to as tier-1 viruses). However, two recent developments support continued exploration of the NHR as a potential vaccine target. The first was the demonstration that the neutralization potency of NHR-targeting antibodies is improved ∼5000 fold by the presence of the high-affinity Fc receptor FcγRI on the surface of host cells, with preservation of this potentiation effect in polyclonal antisera elicited by NHR-based immunogens. The second was the engineering of D5_AR, a higher potency variant of the first characterized NHR-targeting antibody D5, which established that targeting the NHR could protect against tier-2 HIV-1 strains and, in the presence of FcγRI, potently neutralize a broad array of tier-2 and SHIV challenge strains (IC₅₀ < 0.1 μg/mL).

Building on those advancements highlighting the promise of the NHR for vaccine design, we report the crystal structure of D5_AR bound to NHR-mimetic peptide IQN17 at 2.7Å resolution to determine the structural basis for its increased potency over the parent antibody D5, including identifying three new side chain interactions within the IQN17 binding paratope. We subsequently utilize a combination of protein language models and supervised machine learning to engineer an improved NHR-targeting antibody, D5_FI, with 5-fold improved neutralization potency. We then demonstrate that D5_FI is the most potent NHR-directed monoclonal antibody characterized to date and broadly neutralizes tier-2 and −3 pseudoviruses as well as replicating R5 and X4 challenge strains.

Results

Structure of D5_AR Bound to IQN17 at 2.7 Å Resolution

To better understand the structural basis for HIV-1 neutralization of D5_AR at the NHR, we solved the crystal structure of the antibody fragment in complex with a gp41 NHR antigen (Table ). We used the recombinantly expressed fragment antigen-binding (Fab) region of D5_AR and the synthetic IQN17 peptide, a 45-residue coiled-coil NHR trimer, to assemble the antibody–antigen complex. We purified the complex to homogeneity and obtained crystals that diffracted to 2.7 Å resolution. We found that our crystal was twinned and applied the twin law (h, –h–k,l) throughout refinement. With space group H 3, the crystal contains one heavy chain and one light chain of the D5_AR Fab, and one IQN17 protomer per asymmetric unit.

1. Data Collection and Refinement Statistics for D5_AR:IQN17 Complex.

protein	D5_AR: IQN17
PDB ID	8VWE
Data Collection
wavelength	0.97946
resolution range (Å)	41.48–2.2
space group	H3
a, b, c (Å)
α, β, γ	159.97 159.97 67.97
90 90 120
Twin Law	(h, –h–k, l)
total reflections	205628
unique reflections	32,536
multiplicity	19.8
completeness (%)	92
mean I/sigma	29.2
R meas	0.069
R pim	0.015
CC1/2	0.980
CC*	0.995
Refinement
reflections used in refinement	32,569
reflections used for R-free	2003
R work	0.205
R free	0.210
number of non-hydrogen atoms	3380
macromolecules	3012
ligands	320
solvent	48
protein residues	462
RMS(bonds) (Å)	0.004
RMS(angles) (°)	0.865
Ramachandran favored (%)	98.01
Ramachandran allowed (%)	1.99
Ramachandran outliers (%)	0
rotamer outliers (%)	0
clashscore	7.53
average B-factors	75

The structure of the complex revealed three IQN17 protomers are arrayed into a three-stranded, parallel coiled-coil (Figure A), consistent with previously determined NHR-mimetic structures. ,,, The antibody-binding epitope mapped to the gp41 hydrophobic pocket formed by two adjacent IQN17 protomers. Each IQN17 trimer has three hydrophobic pockets, to which three D5_AR Fab molecules bind simultaneously. The D5_AR paratope consists of the complementarity-determining regions (CDRs) of the heavy and light chains, encompassing solvent-accessible surface areas of 586.5 and 264.1 Ų, respectively. The trimeric structure of the D5_AR:IQN17 complex suggests that D5_AR binds to the PHI conformation of gp41 without distorting the NHR coiled coil and without causing steric clashes between Fabs. As with D5, D5_AR engages the gp41 hydrophobic pocket primarily with the heavy chain CDR2 and CDR3 loops, with additional stabilizing, nonpocket contacts made by light chain CDR3.

1.

Structural basis of improved efficacy of NHR-targeting antibody D5_AR compared to D5_WT. (A) Crystal structure of D5_AR Fab (heavy chain in teal, light chain in light blue) bound to IQN17 peptide (white), shown as if looking parallel to the viral and cell membranes with the viral membrane below and the cell membrane above (top) and as if looking perpendicular from the cell membrane toward the viral membrane (bottom). Three asymmetric units are shown, each containing one helical IQN17 peptide and one copy each of the heavy and light chain Fab fragments. (B) Space filling model of a single D5_AR Fab bound to the IQN17 trimer, with binding epitope (residues within 4 Å of any IQN17 residue) highlighted in light blue (top). Amino acid sequences within the D5_WT and D5_AR heavy chain CDR2 and CDR3 loops that form the binding interface, shown with amino acid substitutions in D5_AR highlighted in teal (middle). D5_AR binding interface shown head-on, with interacting residues labeled in teal and amino acid substitutions present in D5_AR as compared to D5_WT shown as sticks (bottom). (C) Comparison between D5_WT bound to 5-Helix from 2CMR PDB (left) and D5_AR bound to IQN17 trimer from the structure reported in this work, 8VWE PDB (right). In the former, 5-Helix is shown as a space-filling model in gray and D5_WT HC (dark green) and LC (light green) are shown as ribbons with residues as sticks. For D5_AR, the IQN17 trimer is shown in space-filling gray and D5_AR HC (teal) and LC (light blue) are shown as ribbons with residues as sticks. Residues that differ from D5_WT to D5_AR are shown in pink. (D) Three new interactions between D5_AR and the NHR hydrophobic pocket present in the D5_AR:IQN17 structure that are absent in the D5_WT:5-Helix structure: (top) new salt bridge between Ser50 of D5_AR HC (teal) and Gln39 of IQN17 (light gray), (middle) new hydrophobic (van der Waals) interaction between Leu54 of D5_AR HC (teal) and Leu29 of IQN17 (light gray), and (bottom) new hydrophobic (van der Waals) interaction between Ala59 of D5_AR HC (teal) and Gln39 of IQN17 (light gray). Bond distances (dotted lines) are shown with measurements in yellow.

Comparison of NHR-Bound D5_AR to D5

Eight amino acid substitutions in the heavy chain CDR2 and CDR3 loops differentiate D5 and D5_AR (Figure B). We compared our D5_AR:IQN17 structure with that of D5 solved in complex with NHR mimetic 5-Helix (PDB: 2CMR). The latter structure only has one NHR pocket available for binding, while D5_AR was solved here in complex with IQN17, with three Fabs binding three NHR pockets. Despite that difference, we can derive mechanistic insights by comparing the NHR-bound structures of each Fab (Figure C). In particular, the overall binding mode of the two antibodies is quite similar, with most engagement with the NHR made by the Fab heavy chain CDR2 and CDR3 loops, and fewer contributions from the light chain CDR3 loop. Indeed, in quantitative terms, the backbone RMSD between D5 and D5_AR Fabs is 0.489, with even more conformational similarity of the CDR loops involved in binding (backbone RMSD between heavy chain CDR2 loops, 0.272 and between heavy chain CDR3 loops, 0.264). In both cases, the hydrophobic phenylalanine at position 55 of the Fab heavy chain fits closely within the NHR hydrophobic pocket. Of the amino acid substitutions between D5_AR and D5, three D5_AR residues in the heavy chain CDR2 loop are within the paratope making direct contact with IQN17: Ser50, Leu54, and Ala59. Each of these may contribute to the increase in binding affinity of D5_AR over D5: Ser50 forms a new salt bridge with Gln29 in IQN17 that was not present in the D5:5-Helix structure (Figure D, top); Leu54 forms a new van der Waals interaction with IQN17 Leu29 (Figure D, middle); Ala59 forms a new van der Waals interaction with a side chain carbon of IQN17 Gln29 (Figure D, bottom).

The D5_AR substitutions in the heavy chain CDR3 loop are not making direct contact with the hydrophobic pocket, so they may have more indirect effects on the neutralization activity of D5_AR. Taken together, these observations advance our mechanistic understanding of antibody-mediated neutralization at the NHR hydrophobic pocket and lay the groundwork for further antibody engineering to improve the potency of NHR-directed mAbs.

Protein Language Models Efficiently Propose Single-Residue Substitutions that Improve Neutralization Potency Greater than 3-Fold

Protein language models trained on large, diverse protein sequence data sets have been increasingly used as an efficient way to engineer proteins. Recently, our group demonstrated that protein language models can efficiently improve antibody binding affinity and neutralization potency using solely primary amino acid sequence and without access to information regarding ligand, structure, binding site, or other molecular details. Here, we used this ensemble of 6 language models to predict a small set of single amino acid substitutions based on the heavy chain (HC) and light chain (LC) sequences of D5_AR Fab, queried separately (Figure A). A total of 24 substitutions, 14 D5_AR HC variants and 10 D5_AR LC variants, were recommended by the protein language model ensemble. The proposed substitutions spanned the HC CDR2, HC CDR3, LC CDR1, and framework regions. We expressed each of these variants as a full-length IgG and tested them for neutralization potency compared to D5_AR IgG in TZM-bl cells with HXB2 pseudovirus (Figure B, and Supporting Table 1), yielding six improved variants (IC₅₀ > 1.5-fold lower than D5_AR). Strikingly, three of the six single amino acid substitutions that improved potency are located in the D5_AR framework regions.

2.

Protein language models and supervised machine learning identify D5_AR variants with improved neutralization potency. (A) Schematic of workflow used to identify language-model proposed variants of the variable domain of the D5_AR heavy chain (VH) and light chain (VL). VH and VL were independently queried to identify highly probable single amino acid substitutions, which were then expressed and tested for neutralization potencies. (B) IC₅₀ fold improvement of language-model proposed single amino acid substitutions to D5_AR (n = 24), with neutralization measured in a TZM-bl cells against HXB2 pseudovirus. The sequences, IC₅₀ values, and standard deviation are listed in Supporting Table 1 for all 24 variants. D5_AR is shown in teal, substitutions in framework regions are shown in lavender, and substitutions in CDR loops are shown in purple. Horizontal dotted line represents neutralization potency of D5_AR. (C) IC₅₀ fold improvement in neutralization potency as compared to D5_AR of 27 variants made by combining multiple beneficial substitutions from panel B, as measured in TZM-bl cells against HXB2 pseudovirus. Sequences, IC₅₀ values, and standard deviation are listed in Supporting Table 2. D5_AR is shown in teal, D5_CS in magenta, and the horizontal dotted line represents neutralization potency of D5_AR. (D) Schematic of workflow used to generate proposed amino acid variants based on a supervised learning model trained on N = 41 VH/VL-IC₅₀ pairs. The supervised learning model was used to propose single amino acid substitutions in both the original D5_AR VH and VL and the VH and VL from the best combined variant, D5_CS, identified in panel 2C. The top 40 scoring variants from both antibodies were expressed and tested for neutralization potency in TZM-bl cells with HXB2 pseudovirus. (E) Neutralization IC₅₀ values of D5_AR and variants proposed by the supervised learning model tested with HXB2 pseudovirus in TZM-bl cells (n = 71). Values are shown on a log₁₀ scale normalized to D5_AR (teal), excluding three variants for which the IC₅₀ could not be determined due to poor curve fit of the neutralization curves. D5_AR (teal), and D5_FI (purple), the most potent variant, are shown in color while all other variants are shown in greyscale. Sequences and IC₅₀ values for each variant are shown in Supporting Table 3. (F) Neutralization curves of D5_AR and D5_FI. For each antibody, six 5-fold dilutions were tested, with error bars showing the SEM of two technical replicates. Neutralization curves for each antibody are nonlinear fit of least-squares regression ([Inhibitor] vs response (three parameters)) in Prism with bottom constrained to 0. Percent infection is normalized to wells on the same plate with virus only and no inhibitor (100%) and cells only (0%).

Combinations of Single Beneficial Substitutions Yield Combined Variants with Improved Potency

An important step in directed evolution campaigns is to test combinations of beneficial mutations to evaluate for additive or synergistic effects that could yield even more potent combined variants. Given the large combinatorial search space of the six improved variants across the two chains of the antibody, we refined the set of candidate substitutions for combination by considering only those that were identified by at least two of the six models used in the protein language model ensemble. In addition, we chose to include equivocal (neutralization affinity 0.8–1.2 fold that of parent) substitutions D27Y HC and K111Q HC to investigate whether there may be synergy between equivocal and beneficial substitutions. We tested combinations of single amino acid substitutions that individually improved neutralization of D5_AR (Figure C, and Supporting Table 2), yielding several combined variants with neutralization potencies similar to that of the best singly substituted variants (up to 3.3 fold). As combinatorial testing of single beneficial substitutions did not yield additive or synergistic benefits, we set out to employ additional methods for identification of improved variants.

Supervised Machine Learning Methods Yield the Highest-Potency D5_AR Variant

To attempt to identify variants with further improved potency, we first curated a data set including the IC₅₀ fold-change mutagenesis data collected to date and utilized the paired sequence-IC₅₀ fold changes to train a supervised machine learning model over language model embeddings to learn pseudovirus neutralization potency fold-change from wild-type D5_AR as a surrogate fitness function (Figure D). This data set included 41 sequence-IC₅₀ pairs from all combined variants (n = 27) and the single substitutions suggested by 2 or more language models in the ensemble (n = 14, data for the ten single residue substitutions suggested by only a single language model had not yet been collected). This supervised model then recommended candidate variants of both D5_AR and the highest-potency variant in the training set (D5_CS), of which the top 20 sequences from each library were selected using both greedy and upper confidence bound acquisition schemes, yielding a total of 74 unique VHs and VLs for testing (some variants were selected by both the greedy and upper confidence-bound acquisition methods). We then tested these variants for neutralization potency in the TZM-bl assay against HXB2 pseudovirus (Figure E,F and Supporting Table 3).

From this 74-variant library, we identified the most potent NHR-targeting antibody described to date, D5_FI, which neutralized HXB2 7.4-fold more potently than D5_AR. This design is comprised of four heavy chain substitutions, three which had been proposed by the protein language model ensemble (D27F, E74T, and S108Y) and the fourth which was informed by the supervised model (T57F). The T57F substitution proposed by the supervised model is a threonine to phenylalanine substitution at a residue contacting the IQN17 hydrophobic pocket in the D5_AR crystal structure, which may increase hydrophobic interactions as a mechanism for improved neutralization potency. The other substitutions do not make direct contact with IQN17 and may be subtly changing the overall conformation or orientation of the binding loops. T57F had not been tested previously as a single substitution, so we compared the neutralization potency of D5_AR with the single T57F substitution to D5_FI, demonstrating that D5_FI is 3-fold more potent than D5_AR T57F and that fuller improvement in neutralization potency requires T57F in the context of the other D5_FI substitutions (Supporting Figure 1).

D5_FI is More Potent than D5_AR in All Tier-2 and Tier-3 Pseudoviruses Tested

To compare the potency of D5_FI to D5_AR against a broader panel of viruses, we tested both antibodies against a panel of tier-2 and tier-3 pseudoviruses from diverse clades (tier-2:246F3, 25710, BJOX, CE0217, CE1176, CH119, CNE55, TRO.11, X1632; tier-3:33.7, PVO.4, Figure A, Figure B, and Supporting Table 4). D5_FI more potently neutralized all 11 viruses tested, with an average IC₅₀ fold improvement of 5-fold over D5_AR.

3.

D5_FI neutralizes tier-2 and −3 pseudoviruses more potently than D5_AR and exhibits potentiated neutralization in TZM-bl/FcγRI cells. (A) Neutralization IC₅₀ values of D5_AR and D5_FI in a panel of 11 tier-2 and tier-3 HIV pseudoviruses tested in the TZM-bl neutralization assay, shown on a log₁₀ scale. IC₅₀ in μg/mL for each antibody-virus pair is displayed above each column and is the average of two technical replicates with error bars = SEM. Tier-3 viruses are indicated with an asterisk in the figure legend. Neutralization data for each virus was analyzed using unpaired t tests, with significance indicated (p < 0.05 = *, p < 0.01 = **). A two-way ANOVA was used to analyze differences in neutralization of D5_AR and D5_FI across the viral panel, with p = 0.0023. (B) Potency-breadth curves of IC₅₀ (dotted line) and IC₈₀ (solid line) titers for both D5_AR (teal) and D5_FI (purple), with y-axis representing percent coverage of the 11 virus panel shown in panel A. (C) Neutralization potency of D5_FI tested with replicating HIV-1 strains NSN-FX (blue) and CH040 (green) in TZM-bl cells (solid line, filled points), and FcγRI-expressing TZM-bl cells (dotted line, open points). Error bars are SEM of two technical replicates, curves are nonlinear fit of least-squares regression ([Inhibitor] vs response (three parameters)), graphed in Prism without constraints. Percent infection is normalized to wells on the same plate with virus only and no inhibitor (100%) and cells only (0%). (D) Neutralization potency of D5_FI (purple) and D5_FI LALAPG (pink) in TZM-bl cells (solid, filled points) and TZM-bl/FcγRI cells (dotted line, open points) with HXB2 pseudovirus. Error bars are SEM of two technical replicates, curves are nonlinear fit of least-squares regression ([Inhibitor] vs response (three parameters)), graphed in Prism with bottom constrained to 0. Percent infection is normalized to wells on the same plate with virus only and no inhibitor (100%) and cells only (0%).

D5_FI Neutralizes Replicating HIV-1 Strains and is Potentiated by FcγRI in an Fc-Binding Dependent Manner

Finally, we then assessed the neutralization potency of D5_FI against two replicating HIV-1 challenge strains that are commonly used in humanized mouse models: NSN-FX (an X4 strain) and CH040 (an R5 strain). We tested D5_FI with each of these viruses in TZM-bl cells as well as FcγRI-expressing TZM-bl cells to assess the extent of FcγRI-mediated potentiation with replicating HIV-1. We found that D5_FI exhibits potentiation of ∼10³ for both viruses (Figure C), which was expected given the similar extent of potentiation observed for both D5 and D5_AR. In TZM-bl cells, D5_FI neutralized NSN-FX four times more potently than D5_AR (Supporting Figure 2). In addition, to verify that FcγRI-mediated potentiation observed is due to binding of the Fc region to the FcγRI receptors, we find that a version of D5_FI in which three amino acid substitutions ablate binding to FcγRI (D5_FI LALAPG) does not exhibit potentiation in the tier-1B pseudovirus HXB2 (Figure D).

Discussion

We used machine-learning approaches to engineer an improved neutralizing antibody targeting the highly conserved transiently exposed NHR domain of HIV-1 gp41 and found it to be the most potent NHR-targeting antibody identified to date. This improved antibody also broadly neutralizes diverse HIV strains, including the least neutralization-sensitive viruses (tier-3) and replicating challenge strains.

Our work provides two contributions to the field of HIV-1 vaccine development. First, the structure of D5_AR bound to NHR-mimetic IQN17 adds to the relatively small repertoire (D5, HK20, and Fab 8066 , ) of NHR-targeting neutralizing antibodies bound to their targets, only one of which, Fab 8066, had been solved in complex with a trimeric N-peptide mimetic. The D5_AR:IQN17 structure presented here demonstrates that D5_AR Fab can occupy each of the three NHR hydrophobic pockets simultaneously without steric clash between Fabs or structural perturbation of the NHR coiled-coil. Furthermore, comparing the structure of D5_AR to that of D5 reveals a structural basis for the improved potency of D5_AR, including a new salt bridge and new hydrophobic contacts formed at the antibody–antigen interface.

A second contribution of this work is the development and characterization of the most potent NHR-targeting antibody characterized to date. Previously, both NHR-targeting antibodies and anti-NHR antisera derived from immunization were found to be weakly neutralizing. − In recent work, we have shown both that an improved NHR-targeting antibody could neutralize tier-2 viruses from diverse clades and that the presence of the high-affinity Fc receptor FcγRI improved the neutralization potency of NHR-targeting antibodies by ∼1000–5000 fold. , Our group also demonstrated that inhibitors targeting the PHI (both D5_AR IgG and 5-Helix), while only moderately potent compared to best-in-class bnAbs targeting other epitopes, exhibit neutralization profiles that are remarkably consistent between strains and are independent of neutralization tier. Despite these developments, enthusiasm in the field for the viability of the NHR as a vaccine target has been limited by the relatively moderate potency of anti-NHR antibodies compared to other HIV-1 bnAbs, motivating efforts to identify more potent NHR-directed antibodies. This work is an advancement of that effort, with the engineering and characterization of D5_FI, an NHR-targeting antibody that is 5-fold more potent than D5_AR and the most potent NHR-targeting antibody yet.

This work has several limitations, including that the neutralization screening experiments to identify improved variants were performed in HXB2, a lab-adapted strain without direct clinical relevance. In addition, some of the smaller differences in neutralization potency of single amino acid variants used to guide recombination and upon which the protein language model was trained (e.g., 1.5-fold improvement) may fall within expected noise for the TZM-bl neutralization assay, although we attempted to control for this by performing multiple replicates across different days and with different batches of virus.

In summary, while D5_FI is still undeniably less potent than most bnAbs to other targets, this improvement places this antibody at the low end of the spectrum of potencies of bnAbs that have been shown capable of protecting nonhuman primates during in vivo challenge studies. − Taken together with the demonstration that D5_FI is capable of neutralizing replicating R5 and X4 challenge strains, this work sets the stage for experiments investigating the in vivo efficacy of anti-NHR antibodies in challenge experiments, including examination of the in vivo effects of the observed in vitro potentiation due to FcγRI.

Materials and Methods

Antibody Production (Lower Throughput)

VH and VL segments were cloned into linearized pCMVR backbones with 5X In-Fusion HD Enzyme Premix (Takara Bio). Plasmids were transformed into Stellar Competent Cells (Takara Bio) and transformed cells were grown at 37 °C. Colonies were confirmed by sequencing and then maxi-prepped (NucleoBond Xtra Maxi, Macherey-Nagel). Plasmids were sterile filtered using a 0.22 μm syringe filter and stored at −20 °C. Antibody variants used for neutralization assays were expressed in Expi293F cells (Thermo Fisher Scientific) using FectoPRO (Polyplus). VH and VL plasmids were cotransfected at a 1:1 ratio; cells were transfected at 3 × 10⁶ cells/mL. Cell cultures were incubated at 37 °C and 8% CO₂ with shaking at 120 rpm. Cells were harvested 3 days post-transfection by spinning at >4200g for 15 min and then filtered through a 0.45 μm filter. Supernatant was combined with 1/10th volume of 10× PBS and purified using affinity chromatography with ÄKTA pure fast protein liquid chromatography (FPLC, Cytiva) instrument with a 5 mL MabSelect PrismA column using wash steps with 1× PBS and elution with 100 mM glycine (pH 2.8) into one-tenth volume of 1 M Tris (pH 8.0). The eluted proteins were then concentrated using 50-kDa or 100-kDa cutoff centrifugal concentrators and further purified by size exclusion on the same ÄKTA FPLC with Superdex 200 Increase 10/300 GL column (Cytiva). Proteins were then further concentrated using 50-kDa or 100-kDa cutoff centrifugal concentrators and filtered through a 0.22 μm filter and stored at 4 °C before use.

The complete sequence of the variable heavy and light regions (IGHV1–69 germline, IGKV1–5 germline respectively) for D5_AR IgG and Fab are

D5_AR_VH:

QVQLVQSGAEVRKPGASVKVSCKASGDTFSSYAISWV RQAPGQGLEWMGSIIPLFGTAAYAQKFQGRVTITADESTSTAYMELSSLRSEDTAIYYCARDNPTFGAADSWGKGTLVTVSS

D5_AR_VL:

DIQMTQSPSTLSASIGDRVTITCRASEGIYHWLAWYQQ KPGKAPKLLIYKASSLASGAPSRFSGSGSGTDFTLTISSLQPDDFATYYCQQYSNYPLTFGGGTKLEIK

Synthesis of IQN17

IQN17 (sequence RMKQIEDKIEEIESKQKKIENEIARIKKLL QLTVWGIKQLQARIL) was synthesized using standard Fmoc-based solid-phase peptide synthesis on a CSBio instrument. The resin was 250 μmol NovaSyn TGR R resin (Novabiochem) and coupling was performed for 15 min at 60 °C with 4-fold molar excess of amino acids. Dry peptide resin was cleaved with 94% trifluoroacetic acid, 2.5% water, 2.5% 1,2-ethanediol, 1% triisopropylsilane at RT for 3.5 h followed by precipitation in cold diethyl ether. The crude peptide was purified by reversed-phase HPLC on a C18 semiprep column over an acetonitrile (ACN) gradient in the presence of 0.1% TFA and fractions were collected based on liquid chromatography–mass spectrometry (LC/MS) analysis. The pure monomeric protein was dissolved to 1 mg mL^–1 in 100 mM Tris-HCl pH 8.0 and oxidized by air at 37 °C with gentle shaking for 48 h. The peptide mixture was then lyophilized, dissolved into 20% ACN/80% water, and repurified via HPLC. Peptide trimer product whose mass corresponded to that of three IQN17 peptide chains was collected.

Protein Crystallization

D5_AR Fab was expressed and purified as above on ÄKTA FPLC with Protein G HiTrap (Cytiva) and bound to equimolar IQN17 at RT for 2 h. The D5_AR:IQN17 complex was then purified on ÄKTA FPLC with Superdex 200 Increase 10/300 GL column (Cytiva) in tris-buffered saline (TBS), concentrated with 100-kDa cutoff centrifugal concentrators and filtered with 0.22 μm filters and frozen at −20 °C at 7.44 mg mL^–1. 96-well crystallization screens were set up using the Greiner Intelliplate with 0.3 μL sitting drops using ProPlex and MCSG-2 screens. Based on initial hits (0.1 M HEPES, 15% PEG 4000, ProPlex; 0.2 M triammonium citrate 20% PEG 3350, MCSG-2; 0.2 M trimethylamine N-oxide, 0.1 M Tris, 20% PEG 2000, MCSG-2), 24-well plates with 2 μL hanging drops were set up, from which crystals were obtained in several conditions after 6 weeks. D5_AR:IQN17 was crystallized at RT in a hanging-drop vapor diffusion system by mixing 1 μL of the protein complex at 7.44 mg mL^–1 in TBS with 1uL of well solution (10% PEG 2000 MME, 100 mM TMAO, 100 mM Tris pH 8.5). Single crystals were harvested and flash frozen in liquid nitrogen.

X-ray Crystallography

X-ray diffraction data were collected to 2.2 Å at the SLAC National Accelerator Laboratory Stanford Synchrotron Radiation Lightsource (SSRL) beamline BL9–2. The crystal belonged to space group H 3 with unit cell dimensions a = 159.97 Å, b = 159.97 Å, c = 67.97 Å, α = 90°, β = 90°, γ = 120°. Diffraction data were processed using HKL3000, and the structure was solved using Phaser in Phenix , by molecular replacement using the D5 Fab from PDB ID 2CMR and IQN17 from PDB ID 2Q7C as search models. The structural models were further refined using iterative cycles of automated refinement in Phenix Refine, manual model fitting using Coot, and MolProbity validation. Structural images were generated with PyMOL (Version 2.4.2 Schrödinger, LLC), and backbone RMSD values were calculated using PyMOL’s “align” command with parameters set to include only backbone atoms (e.g., “align D5 & backbone, D5_AR & backbone”). Solvent-accessible surface areas were calculated using PDBePISA.

High-Throughput Antibody Production

Small-scale expression and purification of antibodies using the Agilent Bravo was performed as previously published. Briefly, 2.5 mL cultures of Expi293F cells were transfected as above and after 4 days, harvested by centrifugation. The supernatant was moved to the Bravo Agilent instrument and purified using ProPlus PhyTip column tips (Biotage, PTV-92–20–07), with PBS washes and a low pH elution using 100 mM glycine pH 2.8 into 1/10th volume 1 M Tris pH 8. The eluted antibodies were then spun down at 13,000 xg x 3 min to remove aggregates and filtered through 0.22 μm spin filters before use.

Protein Language Models

D5_AR VH and VL sequences were used to query a combination of six language models utilizing a previously published method for single amino acid substituted variants. Briefly, six large-scale masked language models (ESM-1b and five models ensembled to form ESM-1v, trained on the 2018–03 release of UniRef50 and the 2020–03 release of UniRef90 respectively) were queried with antibody VH and VL amino acid sequences to produce a list of sequences containing single amino substitutions with higher language model likelihood than the original sequence as determined by at least one of the language models.

Supervised Machine Learning Model

A supervised machine learning model was trained using the initial language-model screen of 41 variants as training data to learn pseudovirus neutralization potency fold-change from wild-type D5AR as a surrogate fitness function. Heavy chain and light chain variable region sequences were concatenated and featurized using language model embeddings from ESM-1v. Supervised training was implemented using a Gaussian Process model with a linear kernel. To recommend candidate sequences for subsequent experimental evaluation, the supervised model was used to evaluate an in-silico deep mutational scan (DMS), excluding residues included in the training data set, of both the wild-type D5AR sequence as well as the most potent variant in the training data set. The top 20 sequences were acquired from both DMS libraries using both greedy and upper confidence bound acquisition schemes , for a total of 80 variant sequences (74 unique VHs and VLs due to some redundancy in sequences proposed by each acquisition scheme).

Transfection to Produce HIV-1 Pseudotyped Lentiviruses

HEK293T cells were transiently cotransfected with a backbone plasmid as well as a HIV-1 Env plasmid for production of HIV-1 pseudotyped lentivirus production using a previously described calcium phosphate transfection protocol. HEK293T cells were passaged in T75 flasks and incubated at 37 °C at 5% CO₂. The growth medium used for passaging and transfections was Corning DMEM (Dulbecco’s Modified Eagle Medium with 4.5 g/L glucose, l-glutamine, and sodium pyruvate) with 10% fetal bovine serum, 1% penicillin streptomycin (Corning), and 1% l-glutamine (Corning). The backbone plasmid psg3ΔEnv was obtained through the NIH AIDS Reagent Program, Division of AIDS, NIAID, NIH from Drs. John C. Kappes and Xiaoyun Wu: HIV-1 SG3 ΔEnv Noninfectious Molecular Clone (Cat#11051). , The psg3ΔEnv plasmid was propagated in MAX Efficiency Stbl2 cells grown at 30 °C with shaking and Env plasmids were propagated in Stellar Competent Cells grown at 37 °C with shaking. DNA was isolated using a maxi-prep kit (NucleoBond Xtra Maxi, Macherey-Nagel) and sequence-confirmed. In brief, 6 × 10⁶ HEK293T cells were plated in 10 cm Petri dishes in a total volume of 10 mL of DMEM and incubated overnight at 37 °C and 5% CO₂ without shaking. Once the cells reached 50–80% confluency, they were transfected as follows: In a Falcon tube, 20 μg of psg3ΔEnv was mixed with 10 μg of Env plasmid and water for a final volume of 500 μL. Five hundred microliters of 2X HEPES-buffered saline [pH 7] (Alfa Aesar) were added dropwise to the mixture and 100 μL 2.5 M CaCl₂ were subsequently added. The mixture was incubated at RT for 20 min and then added dropwise onto the cells. Next, 12–18 h after transfection, the medium was aspirated from the dish and replaced with 10 mL of fresh DMEM with additives. Virus-containing medium was harvested 48 h after medium swap and centrifuged at 300g for 5 min; the supernatant was sterile-filtered with a 0.45-μm poly­(vinylidene difluoride) filter and stored in 1 mL aliquots at −80 °C.

Viral Neutralization Assays

The neutralization assay was adapted from the TZM-bl assay for standard assessment of neutralizing antibodies against HIV-1 as described. TZM-bl/FcγRI cells (HeLa luciferase/β-galactosidase reporter cell line stably expressing human CD4, CCR5, CXCR4, and FcγRI) obtained from the NIH AIDS Research and Reference Reagent Program as contributed by Drs. John C. Kappes and Xiaoyun Wu were plated in white-walled, 96-well plates (Greiner Bio-One 655098) at a density of 5000 cells/well and incubated for 12–16h at 37 °C with 5% CO₂. Antibodies were diluted into growth medium in 5-fold dilution series, after which HIV-1 Env-pseudotyped lentivirus in growth medium supplemented with DEAE-dextran (Millipore Sigma) at 5 μg/mL was added to the serially diluted antibody wells. Medium was removed from plated cells and replaced with 100 μL of the antibody-virus mixture. After incubation at 37 °C for 48 h, the media was again completely removed from the plated cells and 100 μL of BriteLite assay readout solution (PerkinElmer) was added to the cells and luminescence values were measured using a microtiter plate luminometer (BioTek) after shaking for 10s. Normalized values were fitted with a three-parameter nonlinear regression inhibitor curve in GraphPad Prism 9.1.0 to obtain IC₅₀ value. Fits for neutralization assays were constrained to have a value of 0% at the bottom except for the assay with replicating HIV-1 (Figure C) as noted in the figure legend. Neutralization assays were performed in biological duplicate with technical duplicates. Replicating challenge viruses NSN-FX and CH040 were generously provided by Valerie Rezek, UCLA (Scott Kitchen Lab).

Safety and Biosafety

No unexpected or unusually high safety hazards were encountered. Viral neutralization assays were performed by trained personnel under BSL-2+ conditions.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acschembio.5c00035.

• Additional experiments including neutralization studies comparing other antibody variants and viral strains, as well as tables with neutralization IC₅₀ values (PDF)

○.

Department of Molecular Biophysics and Biochemistry, Yale University, New Haven, Connecticut 06520, United States

◆.

Department of Chemical Engineering, Stanford University, Stanford, California 94305, United States; Stanford Data Science, Stanford University, Stanford, California 94305, United States; Arc Institute, Palo Alto, California 94304, United States

∇.

S.T. and S.K. contributed equally to this work.

The authors declare no competing financial interest.