Tuning antibody stability and function by rational designs of framework mutations

ABSTRACT

Artificial intelligence and machine learning models have been developed to engineer antibodies for specific recognition of antigens. These approaches, however, often focus on the antibody complementarity-determining region (CDR) whilst ignoring the immunoglobulin framework (FW), which provides structural rigidity and support for the flexible CDR loops. Here we present an integrated computational-experimental workflow, combining static structure analyses, molecular dynamics simulations and in vitro physicochemical and functional assays to generate rational designs of FW mutations for modulating antibody stability and activity. We first showed that recent antibody-specific language models lacked insights in FW mutagenesis, in comparison to approaches that use antibody structure information. Using the widely used breast cancer therapeutic trastuzumab as a use case, we designed stabilizing mutants which were distal to the CDR and preserved the antibody’s functionality to engage its cognate antigen (HER2) and induce antibody-dependent cellular cytotoxicity. Interestingly, guided by local backbone motions predicted using molecular dynamics simulations, we designed a FW mutation on the trastuzumab light chain that retained antigen-binding effects, but lost Fab-mediated and Fc-mediated effector functions. This highlighted the effects of FW on immunological functions engendered in distal areas of the antibody, and the importance of considering attributes other than binding affinity when assessing antibody function. Our approach incorporates interdomain dynamics and distal effects between FW and the Fc domains, expands the scope of antibody engineering beyond the CDR, and underscores the importance of a holistic perspective that considers the entire antibody structure in optimizing antibody stability, developability and function.

Introduction

Antibodies are immunoglobulin molecules that recognize and bind to specific antigens, as a hallmark of the adaptive immune response to clear invading pathogens, or aberrantly expressed molecules in contexts such as cancers.^1–4 Recent advances in machine learning (ML) and artificial intelligence (AI) approaches can be applied to study how antibody responses are mounted in vivo,^5–7 and to design monoclonal antibodies (mAbs) as therapeutics that engage specifically to different clinically relevant targets.^8–10 In particular, protein language models have been applied to evolve antibodies in silico, toward better, more stable antigen binders.^11–13 The advancement of AI approaches also motivated high-throughput profiling of random antibody libraries, to generate experimental data for the training of next-generation antibody ML models.^12,14 The focus of such efforts has typically been in the antigen-binding fragment (Fab) of the antibody, specifically the complementarity-determining regions (CDRs), which are the major contributor to antigen binding (Figure 1a): the CDR loops form a “paratope” of specific three-dimensional shape, which determines its antigen specificity.^15,16 The third CDR loop on the antibody heavy-chain, hereafter CDRH3, is a known major contributor to such specificity.^12,16–18 For example, Chinery et al.¹⁴ generated over 500,000 CDRH3 variants of the breast cancer therapeutic trastuzumab, to benchmark the performance of different computational methods in predicting high-affinity antibodies from these randomly generated sequence libraries and designing new antigen binders de novo. Furthermore, deep learning approaches identified CDR mutations with the potential to enhance the affinity and stability of therapeutic antibodies.^12,13

Figure 1.

Antibody language models (AbLms) offer limited insights for antibody framework (FW) mutagenesis.

(a) Schematic of the role of FW region in antibody structure. (left) Antibody complementarity determining regions (CDR) undergo conformational changes upon antigen binding. These changes are thought to propagate through the FW to the rest of the molecule and can affect effector functions. (right) Illustration of CDR and FW on the antibody variable heavy (VH) and variable light (VL) region structure. Conserved positions in FW (ref. 54) are highlighted with stick representation. (b) Illustration of in silico saturated mutagenesis using AbLMs, a generic protein language model (ESM-2) and a structure-based approach (Rosetta). (c) Heatmap illustrating position-specific probabilities for the wild-type amino acid (columns) along the trastuzumab VH sequence using different language models (rows). FW and CDR regions are delimited below the heatmap. (d-e) Rank of the wild-type amino acid (vertical axis, 1 = best, 20 = worst) along the VH sequences (horizontal axis) of (d) n = 55 human therapeutic antibodies and (e) n = 1,988 paired H-L chains from class-switched memory B cells taken from Jaffe et al. (ref. 53), according to the approaches evaluated in (b). For (e) Rosetta was not attempted due to unavailability of experimentally determined structures for these sequences. Lines depict median and shaded areas depict the inter-quartile range.

However, the function of an antibody is determined not only by its binding affinity to the antigen but also by several other factors. The term “developability” has been used to encompass a range of issues pertaining to the potential success of antibody designs, including stability, solubility, and immunogenicity.^19–22 Computational methods to predict these properties have been developed to streamline in silico antibody design.^22–26 In addition, the therapeutic effects of full-length mAbs do not solely arise from antigen binding, but also from effector functions mediated by the crystallizable fragment (Fc) region, such as antibody-dependent cellular cytotoxicity (ADCC).^27–29 These factors require consideration beyond the CDRs and the antibody–antigen interface.^27,30,31 In antibodies, CDRs can be found in both the heavy and the light chain variable domains, which are composed chiefly of an immunoglobulin fold “framework” (FW) (Figure 1a). The FW region provides structural integrity to support the flexible CDR loops.^15,30–34 Positioned between the antigen-binding CDRs and the antibody constant (C) region, the FW immunoglobulin folds in the heavy and light chains interact with distinct packing geometries, which are thought to mediate the propagation of conformational changes from the CDRs to the rest of the antibody upon antigen binding.^17,35–40 If we consider antibody structures under this holistic perspective,^17,41,42 exclusive engineering of the antigen–antibody interface may yield unintended consequences if these perturbations affect the FW and consequently the propagation of conformational changes throughout the antibody molecule,^{17,35,38,43,44} which could result in adverse effects or a loss of functional efficacy. There is a paucity of detailed molecular understanding on how the FW confers such effects, and how FW mutations could disrupt both antigen binding and Fc-mediated effector functions. Conversely, FW mutations can also be harnessed to optimize these properties without directly altering the antigen-binding site, representing a promising avenue to subtly modulate antibody stability and function.^18,31,33

Here, using the widely used breast cancer therapeutic trastuzumab as a model system, we conducted iterative computational-experimental analyses to investigate mutations on both the trastuzumab heavy and light chains, exploring the potential of FW mutagenesis to tune antibody functionality. Trastuzumab is a well-established therapeutic specific to human epidermal growth factor receptor 2 (HER2), overexpressed in ~15% of breast cancers, and is the first-line treatment to these cancers.⁴⁵ Considering the limitations of existing protein language models trained solely on antibody repertoire data (hereafter antibody language models, or AbLMs) for FW mutagenesis, we opted instead for a structure-based approach to predict stabilizing FW mutations, and verified the stability, antigen binding and downstream functional effects of these mutants using in vitro assays. We used in silico molecular dynamics (MD) simulations both to elucidate insights on how the designed mutants perturbed antibody structure and to suggest new mutants based on antibody local structural dynamics. We dissected the contribution of antibody FWs to antigen-binding and effector functions, fine-tuning antibody stability and function using carefully designed FW mutants.

Results

Antibody language models offer limited insights in FW mutagenesis

We first assessed different computational approaches in their usefulness to suggest mutations in antibody FW regions. In our comparisons, we considered three types of approaches that can generate numerical scores of mutational impact: (1) a biological language model (ESM-2, 150 m parameters) trained on generic protein sequences⁴⁶; (2) antibody language models^5,47,48 (AbLMs) trained on large datasets of antibody sequence repertoires⁴⁹ to learn the sequence contexts specific to antibodies; and (3) a structure-based approach for evaluating mutational effects (Rosetta), using a biophysical scoring function that considers different components of intramolecular forces.^50,51 In total, we considered five computational methods (Figure 1b), inputting either the protein structure or sequence of trastuzumab variable heavy (VH) and variable light (VL) domains and extracting the predicted effects of mutating every amino acid position to every other 19 amino acids, i.e., performing an in silico saturated mutagenesis scan. Figure 1c displayed the probability of the wild-type (WT) amino acid at each position along the trastuzumab VH sequence, according to the four language models evaluated in this analysis. We noticed that all tested AbLMs (AntiBERTy, AbLang2, AntiBERTa2) were uncertain about amino acid identities in the CDR loops, but more confident about the FW region. Specifically, AbLMs consistently gave a probability close to 1 for WT amino acids at the FW positions (Figure 1c) while returning near-zero probabilities for the non-germline residues (Supplementary Figure S1). This trend appeared to be especially strong in AntiBERTy; AbLang2 and AntiBERTa2 displayed a similar pattern, albeit with less bias against the non-germline residues in specific regions, e.g., the CDR1, CDR2, and some positions in the FW3 (Supplementary Figure S1). In contrast, ESM-2, which was not trained specifically on antibody sequences, displayed a more varied pattern. For the VL sequence, similar patterns were observed, with the exception that AbLMs were more confident in the CDR loops than in the case of the VH domain (Supplementary Figures S2 and S3).

We repeated this evaluation over n = 55 human therapeutic antibodies with structural information, which reinforced the patterns seen for trastuzumab (Figure 1d; Supplementary Figure S4); here, similar to ESM-2, structure-based predictions given by the Rosetta point mutant scan application were also more varied in the FW. This could either reflect the fact that the input sequences were well established, manufacturable antibody therapeutics with little room for further optimization in the FW, or that the AbLMs were over-specializing to antibodies whilst failing to generalize to mutations that were rarely seen, such as those in the FW region. We explored this by utilizing a dataset of paired VH and VL sequences from sorted switched memory B cells from three healthy donors,⁵² and observed a similar pattern (Figure 1e; Supplementary Figure S5). Considering the demonstrated abilities of AbLMs to separate naïve and memory B cell receptor sequences,^5,47 we observed accordingly that the probabilities for the WT residues increased with sequence identity to the germline (Supplementary Figures S6 and S7). In one of the tested AbLMs, AbLang2, technical innovations to address the bias toward germline residues were the main uniqueness of the model.⁴⁷ Here, we observed that AbLang2 returned lower probabilities for the WT residues compared to other models, but the correlation between these probability values with sequence identity persisted (Supplementary Figures S6 and S7). Taken together, our results suggested that, despite recent interests in the field of antibody discovery due to their ability for efficient generative designs,¹² AbLMs focus mainly on CDRs in detriment of their capabilities to suggest new FW mutations that would enhance antibody stability. A generic model, such as ESM-2, or indeed a structure-based approach not specific to antibodies, may potentially be more conducive toward FW mutagenesis.

In silico structure-based screening identified VH S85N+R87T as a stabilizing and functional trastuzumab variant

Based on the previous analysis, we opted to pursue in silico FW mutagenesis using a structure-based approach to evaluate the stabilizing effects of mutations, by estimating free energy change (∆∆G_folding) of the antibody, as a measure of folding stability, upon mutation. Using the Rosetta molecular modeling software, we used the point mutant (pmut) scan application to predict ∆∆G_folding for every possible amino acid substitution along the trastuzumab VH and VL sequences (Supplementary Table S1). Most (94.19% in the VH and 94.20% in the VL) Rosetta-predicted mutations were destabilizing (Figure 2a), in line with the fact that trastuzumab is a well-used therapeutic antibody with VH and VL domains close to their optimal properties.

Figure 2.

Structure-based designs of trastuzumab FW stabilizing mutations.

(a) (left) Bar-plot depicting Rosetta-predicted ∆∆G for every possible amino acid substitution in the trastuzumab VH structure, using PDB 7mn8 as input. (right) Filtering approach to select mutations for investigation. (b) Heatmap depicting Rosetta-predicted ∆∆G for the three double mutants selected for further characterization. (c) Illustration of the positions of double mutants listed in (b) in the trastuzumab structure. (d) Melting curves of R50S+R59N and S85N+R87T with respect to the wild-type (WT), determined using differential scanning fluorimetry. PBS-only condition served as a negative control. The melting temperature (T_m) of the constructs is indicated. (e) Binding of WT and mutants against a HER-2-expressing cancer cell line (SKBR3) determined using fluorescence-activated cell sorting (FACS). Statistical significance was evaluated using analysis of variance (ANOVA) models and P-values were obtained by post-hoc Dunnett’s test, comparing mutants to the WT. (f) Viability of SKBR3 cells in vitro in the presence of the antibody constructs. Cell viability was measured as a percentage of cell counts at day 0 of the experiment. Statistical significance was evaluated using ANOVA followed by post-hoc Dunnett’s tests comparing mutants to WT. (g) Antibody-dependent cellular cytotoxicity (ADCC) effect of the antibody constructs, measured on SKBR3 cells (quantified using lactase dehydrogenase [LDH] release as a proxy) upon co-culturing with natural killer (NK) cells isolated from n = 4 healthy donors. The percentage of cytotoxicity was calculated by taking spontaneous release (no antibody control) into account. P-values were quantified using paired t-tests and noted on the plot after false discovery rate correction.

To further filter for candidate FW mutations that optimize antibody stability, we focused on the VH domain and applied the criteria illustrated in Figure 2a. First, we analyzed the occurrence of these predicted mutations in antibody repertoires from breast tumor samples.² We reasoned that naturally occurring residues along the antibody sequence represent tolerable amino acids for stability and function. We observed that 50% of VH stabilizing mutations were observed in >1% of the repertoire sequences (Supplementary Figure S8). Second, we performed an additional round of Rosetta mutational scanning for double VH mutants, to augment the chance to obtain variants with measurable effects on trastuzumab stability and function. We prioritized mutants with consistent Rosetta results between the single and the double mutant scans and designed three stabilizing VH double-mutants located at distinct structural locations (Figure 2b): A40R+K43D, R50S+R59N and S85N+R87T. We obtained probabilities of these substitutions from the language models we had investigated and observed that these models displayed a similar trend as we described: all tested models reported substantially higher probabilities for the WT residues compared to the mutants (Supplementary Table S2), in contrast to the stabilizing predictions from Rosetta. Inspecting these mutations in the context of the antibody structure, R50 and R59 are situated at the apex of the β-fold (Figure 2c), anchoring the CDRH2 loop, with their side chains pointing toward the antigen-binding interface. A40R and K43D have each been predicted as destabilizing individually (Supplementary Table S1), but the combined predicted effect of the double mutant is stabilizing (Figure 2b), potentially due to compensatory effects of the two mutations on the surface charge of the antibody. S85 and R87 are both solvent-exposed positions at the opposite side to the CDR loops (Figure 2c). We note that S85N+R87T introduces a novel N-linked glycosylation site. Our Rosetta pipeline did not consider the presence and structures of ectopic glycans and suggested that the double mutant on its own conferred stabilizing effects.

We next verified the effects of these trastuzumab mutants using in vitro assays. We focused on S85N+R87T and R50S+R59N, since the A40R+K43D double mutant had low production yield insufficient for further characterization. We explored possible reasons for this subsequently in analyzing our MD simulation data (see below). For the remaining two variants, we determined their thermostability using differential scanning fluorimetry. We found that both mutants were stabilizing compared to WT, as indicated by a higher melting temperature (T_m) (Figure 2d), validating our in silico predictions.

We also assessed the capability of each antibody to bind its cognate antigen via fluorescence-activated cell sorting (FACS) against a HER2-expressing breast cancer cell line (SKBR3). We observed a significant reduction of binding for the R50S+R59N mutant; in contrast, the S85N+R87T mutant displayed a similar binding curve as the WT (Figure 2e). When we assessed Fab-mediated effects on cancer cell viability and Fc-mediated immune cell effects against cancer cells in ADCC functional assays, we found that the R50S+R59N mutant also lost capability to inhibit SKBR3 cell proliferation (Figure 2f) and to induce ADCC (Figure 2g). The functional effects of S85N+R87T in terms of ADCC and the inhibition cell proliferation were retained and comparable to WT as expected. The loss of functionality for R50S+R59N is consistent with their structural localization (Figure 2c), being in vicinity to the CDR region as predicted in our structural analysis.

To further elucidate our in vitro evidence from a detailed structural perspective, MD simulations were carried out on the designed trastuzumab variants as well as the WT antibody, in complex with HER2 (Figure 3a). We generated and analyzed MD trajectories of a total of 3 μs for each system (Supplementary Table S3). The overall conformational stability of the Fab was retained for WT as well as the designed variants (Supplementary Figure S9), in agreement with the high T_m measured experimentally. In the WT trajectories, we observed that the loop containing A40 and K43 exhibited substantial flexibility (Supplementary Figure S10); the A40R+K43D mutant could potentially form a salt bridge to rigidify this loop and induce structural instabilities, which could be a possible reason for its low yield in our in vitro antibody production. As a proxy of the binding strengths between our trastuzumab variants and HER2, the distance between the center of mass (COM) of the CDR loops and COM of the HER2 juxtamembrane domain was monitored along the trajectories (Figure 3b). We observed an unbinding event in one of the R50S+R59N trajectories (Figure 3c); in other replicas, we also noticed deviations in the packing geometries of the trastuzumab VH and VL domains (Figure 3d,e). Together with previous reports on the relationship between VH-VL packing and paratope shapes,^17,53,54 R50S+R59N could perturb the trastuzumab paratope and compromise its interaction with HER2.

Figure 3.

Investigating stabilizing trastuzumab VH mutants using molecular dynamics (MD) simulations.

(a) Schematic to illustrate the generation of MD simulation trajectories for WT, R50S+R59N and S85N+R87T constructs. (b) Illustration of the calculation of CDR-HER2 distance from MD simulation frames. Black circles depict the centers of mass (COM) of the HER2 juxtamembrane region and the trastuzumab CDR region. (c) Time evolution of CDR-HER2 distance in the MD trajectories. (d-e) Distributions of (d) VH-VL angle (planar angle defined using COMs of the VH, VL and light chain CDR) and (e) VH-VL distance (straight-line distance between COMs of VH and VL domains) over MD simulation frames for the antibody constructs. The inset schematic illustrated the definitions of these measurements. The replica with the HER2 unbinding event (see (b)) for R50S+R59N was excluded from this analysis. P-values were obtained by comparing mutants against WT using linear-mixed effect models. (f) Illustration of analysis of antibody backbone correlated motion using AlloHubPy [ref. 57]. Each MD simulation frame was encoded using the structural alphabet M32K25 [ref. ⁸⁸], to detect correlated conformational changes between pairwise position (“fragments”) along the antibody structure. This procedure was carried out separately for the WT antibody bound to HER2 (WT_Bound) and that simulated in the absence of HER2 (WT_Unbound) to identify backbone motions which characterized either state. (g) Volcano plot of AlloHubPy fragment pairs whose correlated conformation is found either in the bound (positive log2 fold difference) or the unbound (negative) states. “Hub” positions (absolute log2 fold difference > 2, FDR-adjusted p-values < 0.01) were highlighted. (h) Frequency of AlloHubPy-detected hubs within the trastuzumab heavy chain. Regions corresponding to the sequence position of R50, R59, S85 and R87 were highlighted.

Given the importance of conformational flexibility communicated upon antigen binding from the CDR to more distal sites in the antibody structure,^35,36,55 we asked whether these trastuzumab mutations were associated with local backbone conformational changes important to antigen binding. Our recent computational framework (AlloHubPy56) could be used to analyze MD trajectories to extract conformational flexibility of local protein backbones, and identify correlated motions that are preferred in specific protein states. We used AlloHubPy to compare MD trajectories of WT trastuzumab in the presence and absence of HER2, to identify positions that displayed significant correlated motions associated with HER2 binding (Figure 3f,g, also see Methods). We observed that R50 was located within an AlloHubPy “hub” that displayed significant backbone motions in the unbound state (Figure 3h). This further supported the experimental results that the R50S+R59N mutation could compromise antigen binding.

All together, we established a pipeline to computationally design novel FW mutations, validate their effects in enhancing antibody stability and function, and integrate MD to interpret mutational impact. This approach identified trastuzumab VH S85N+R87T as a stabilizing variant that preserved the capability to bind HER2 and trigger downstream functional effect upon antigen binding.

Molecular dynamics analysis predicted VL Q89 as a mutable site to tune HER2 binding

We next asked whether we can invert our pipeline, to first use MD to design new mutations guided by conformational dynamics and then validate these designs experimentally. The AlloHubPy analysis of the VL domain revealed one communication “hub” that was present only in the antigen-bound state (Figure 4a). This fragment, Q⁸⁹QHT⁹², extends into the CDRL3 loop. Inspection of trastuzumab structures suggested that the first two glutamine (Q) residues were not exposed at the HER2-CDR interface (Figure 4b). Analyzing the preservation of native contacts in the MD simulations of the WT setup, we observed that Q89, but not Q90, retained the majority of its contacts in both the bound and unbound conformational ensembles (Supplementary Figure S11). We therefore explored mutations at the Q89 site to investigate whether this position, despite its lack of direct involvement in HER2 binding, could be tuned to modulate trastuzumab stability and function. We observed in static structures that Q89 formed interactions with surrounding residues in the light chain, including F98 (thus forming the basis of the CDRL3 loop) and Y36 buried in the immunoglobulin fold (Figure 4b). We therefore designed two mutants, Q89H (where the histidine should maintain the local contact network and thus be stabilizing) and Q89A (which would destabilize contacts between this position with the surrounding residues). We generated additional MD trajectories of VL Q89H and Q89A (Supplementary Table S3), and found that Q89H gained stable π–π interaction with F98 (Figure 4c,d). Q89A, on the other hand, lost the majority of local side-chain interactions (Figure 4c), including the contact with F98 (Figure 4e). Further inspecting the MD trajectories, we observed that both Q89A and Q89H did not alter the overall packing geometries between the VH and VL domains (Supplementary Figure S12), suggesting that these mutants should not negatively impact the thermostability of the antibody.

Figure 4.

Local backbone dynamics guided the design of trastuzumab VL Q89A to decouple antigen-binding from antibody effector functions.

(a) Frequency of AlloHubPy-detected hubs within the trastuzumab light chain. The fragment Q⁸⁹QHY⁹² was frequently observed and highlighted on the plot. (b) Mapping of the Q⁸⁹QHY⁹² fragment on the trastuzumab VL structure. Neighboring amino acid contacts (F98, Y36) were highlighted. (c) Number of side-chain interactions mediated by WT (Q89), Q89A and Q89H mutants in the MD simulations. Interactions were quantified separately for hydrogen bonds (H-bonds), π – π and CH-atom/NH-atom interactions (see Materials and Methods). P-values were obtained by comparing the distributions of the mutants against the WT using linear mixed-effects models. (d) Time evolution of the (top) distance and (bottom) angle between Q89H and F98 in the MD simulation trajectory. Dotted lines indicate reference values, below which a π – π interaction occurs. (e) Distributions of the distance between F98 and Q89/A89 in the WT and the Q89A MD simulations. P-value was obtained using a Wilcoxon rank-sum test. (f) Melting curves of WT, Q89A and Q89H determined using differential scanning fluorimetry. Melting temperatures (T_m) of these constructs were noted. Data corresponding to the negative control (PBS buffer) is shown alongside the WT and mutant antibodies. (g) Binding of antibody constructs against HER-2 expressing SKBR3 cell lines quantified using FACS. Statistical significance was evaluated using ANOVA followed by post-hoc Dunnett’s tests comparing the mutants against WT. (h) Effect of antibody constructs on SKBR3 cell viability in vitro. Statistical significance was assessed with ANOVA followed by post-hoc Dunnett’s tests comparing the mutants against WT. (i) ADCC effects (measured using LDH release as a proxy) on SKBR3 cells co-cultured with NK cells from n = 4 donors. The percentage of cytotoxicity was calculated by taking spontaneous release (no antibody control) into account. P-values were quantified using paired t-tests and displayed after false discovery rate correction. (j) Frequency of side-chain interactions (highlighted with stick representation) observed in the Q89A and Q89H MD simulation frames, compared against the WT trajectories.

We expressed Q89A and Q89H in vitro and observed indeed that both mutants maintained a similar thermostability as WT (Figure 4f). We characterized these antibodies in terms of their capability to bind HER2 (Figure 4g), inhibit HER2+ cancer cell proliferation (Figure 4h), and induce ADCC (Figure 4i). We found that Q89H maintained similar functionality as WT. Interestingly for Q89A, at the antibody concentration (10⁻² ng/µl) at which reduced ADCC was observed, this mutant exhibited similar HER2 binding (p = 0.7, Wilcoxon rank-sum test), but inhibited HER2+ cancer cell proliferation (p = 0.1, Wilcoxon rank-sum test) compared to the WT (Figure 4g-i). In other words, the Q89A mutation decoupled antigen binding from downstream effector function of trastuzumab.

We asked whether the MD trajectories provide further molecular details to explain the effects of Q89A and observed stable hydrogen bonds and salt bridges formed at the antibody–antigen interface in the Q89A trajectories (Supplementary Figure S12), including contact mediated by the CDRL3 loop in close proximity to the mutated position, in agreement with the retained affinity of Q89A to HER2 we observed in the experiments. Further, we compared the contact network between the WT and the Q89 mutants, and found that in comparison to Q89H, Q89A displayed wider perturbation of positions in this network, extending from the CDR loops to the VH-VL interface, and propagating to the trastuzumab constant domains (Figure 4j). These analyses highlighted the distal effects of a single mutant and potentially explained the loss of functionality in Q89A that we observed experimentally.

Discussion

Here, we established an integrative pipeline combining computational analyses of antibody sequences, static antibody structures, MD simulations and in vitro physicochemical and functional assays to design novel FW mutations that modulate antibody stability and function. Antibody FW does not directly engage the antigen, lacks conformational flexibility, and has low sequence diversity.¹⁷ The role of FW in antibody function has thus been underappreciated and less extensively studied in comparison to those of the CDRs. Previously, phage display of exhaustive libraries of mutant antibodies has highlighted FW mutations that improve antigen binding and thermostability properties.³⁹ Here, we showed this in silico using a structure-based approach (Rosetta) to predict stabilizing amino acid substitutions. Forming the bulk of the antibody variable domain, solvent-exposed positions in the FW are inextricably linked to issues with antibody solubility, aggregation, and immunogenicity.^{20,21,23,31,57} Given the attention in the field of antibody discovery toward these developability issues, FW mutagenesis is a promising avenue toward optimizing antibody designs to become safe and efficacious therapeutics that can be readily manufactured at a large scale. Our work demonstrates that FW mutagenesis can subtly fine-tune antibody properties, not limited to its stability, but also including its antigen affinity and Fab-/Fc-mediated functions. This approach can substantially expand the scope and possibilities of antibody design, instead of restricting design efforts to the CDRs.

We identified VH S85N+R87T as a stabilizing mutant that retained the downstream effector function of trastuzumab and highlighted the maintenance of packing geometries of the antibody upon mutating S85N+R87T. However, we also noted that these substitutions introduced a novel N-linked glycosylation site. Given the importance of glycosylation in modulating antibody function, we could not exclude the possibility that some of the observed effects of this double mutant may be due to ectopic glycans attached to this site. To further improve and incorporate these considerations, recent approaches in de novo protein designs that combine signals from protein structures and sequences^58,59 may hold promise to optimize design of mutations involving these sites in the future.

A major novelty in this work is the design of a VL mutation (Q89A) that decoupled antigen binding from function. From the perspective of enhancing antibody function, this does not represent a viable mutant for further pursuit. Instead, trastuzumab VL Q89A challenges the theory where high antigen-binding affinity is usually considered as a sign of a functional and efficacious antibody.^42,60,61 Here, this mutation retained antigen-binding, but abrogated the antibody’s ability to inhibit HER2+ cancer cell proliferation and induce ADCC. To the best of our knowledge, this is the first example of such mutation, which allows us to dissect the contribution of different structural elements of the antibody to its functional attributes. Our MD analysis highlighted a substantial loss of local contacts in Q89A, extending beyond the immediate vicinity of the mutated residue throughout the variable region FW and beyond. Given previous reports on how CDR-distal mutations affect antibody yield and antigen-binding,^18,34,40,61 our mutagenesis work provided further molecular insights into how the conformational changes upon antigen-binding may be propagated across the entire antibody molecule. This result highlights the importance of our computational framework in further studies of other antibodies, for mapping disallowed positions for mutagenesis to avoid unintended consequences in antibody effector functions. Elucidating such mechanisms requires a holistic, integrative modeling approach to further probe the importance of different antibody fragments in an antibody’s dynamics,^33,41 specifically on the Fc region, the binding of which to its cognate Fc receptors would induce antibody effector functions.^27,36,44

An interesting finding is the lack of insights provided by antibody language models (AbLMs) in FW mutagenesis. AbLMs have received widespread attention in the field due to their ability to generate promising antigen binders using only sequence data.^5,12,48 Further improvements of these models were proposed, including the minimization of bias toward frequently used germline immunoglobulin gene fragments,⁴⁸ especially given that AbLMs are typically trained from the antibody repertoire sampled from peripheral B cells, most of which are of the naïve immunophenotype.⁶² FW positions encoded in different antibody germline genes are highly optimized in evolution,⁶³ and the sequence conservation at these positions could imply that language models have captured inter-residue contexts to predict the WT residues with high probability. We showed here that amino acid likelihoods returned by AbLMs highlighted substantial certainty in antibody FWs, with the WT amino acids typically returning a likelihood of 1. In contrast, the generic protein language model ESM-2 returned lower probabilities for the WT residues. This was perhaps unsurprising given the lack of sequence diversity in the FW region in antibody repertoires, leaving antibody-specific models to focus on the diverse CDR loops in order to learn abstractions from sequence data. It is important to note that the ability of AbLMs to return accurately the WT amino acids underpinned the promise of using these models for tasks such as sequence clustering and property prediction.⁶⁴ For our purpose, a viable in silico mutagenesis approach should strike a balance between accurate prediction of the germline residues and room for diversity beyond the germline. Here, our analysis suggested that the tested AbLMs focused more on the former and thus offered little insights on how FW positions could be optimized further on their own. This result provides further support to previous research utilizing ESM models (but not AbLMs) to successfully facilitate the design of novel antibody mutations.¹¹ We instead adopted a structure-based approach for FW mutagenesis. In the future, dedicated training strategies, for example, by generating in vitro mutational scanning data specifically on FW substitutions for fine-tuning AbLMs, may help to address this deficiency. Additionally, recent inverse folding models explicitly consider the mapping between sequences and their corresponding structural folds, and hold promise in de novo protein designs.⁶⁵ Further development of AI tools and approaches can potentially streamline the design of FW mutations and shift the paradigm of antibody design to explore different FW positions in optimizing various developability attributes.

In summary, we presented a side-by-side computational-experimental approach to explore the stability and function effects of antibody FW mutations. This approach can potentially be applied beyond antibody FW, for instance, to engineer antibody Fc region to enhance specific effector functions, especially given the focus on Fc engineering in recent years for the design of novel antibody therapeutics.²⁷ Moving forward, considerations beyond the antibody variable domains, including the interrogation of interdomain dynamics,⁴¹ as well as distal effects between the CDR and the Fc, will enable investigations through adopting a holistic view of antibody structural dynamics throughout the entire molecule, and exploring how this can be fine-tuned to achieve desired antibody functions.

Materials and methods

In silico analysis

Antibody sequences and structures

Structures

Wherever applicable, for the trastuzumab Fab structure, we used atomic coordinates from a cryo-EM resolved structure of trastuzumab Fab bound to HER2⁶⁶ (Protein Data Bank [PDB] 7mn8). We noted that in this structure, the coordinates for the HER2 juxtamembrane domain, the only portion of the antigen that directly interacts with trastuzumab, were incomplete. For this reason, the input coordinates of the HER2 juxtamembrane domain were extracted from PDB 8q6j⁶⁷ (residues 507 to 623 according to author-assigned numbering in the structure), where all the residue coordinates were resolved. We noted that these two structures were highly similar (root mean squared deviation [RMSD] <1.5 Å, calculated on the entire trastuzumab-HER2 complex). Additionally, despite the presence of several glycosylation sites within HER2, none of them is within the juxtamembrane domain. For illustration of structures presented here of trastuzumab in the absence of HER2, a better resolution (PDB 4hkz, resolution = 2.08 Å vs 3.45 Å for PDB 7mn8) structure of trastuzumab⁶⁸ was used.

Sequences

For data presented in Figure 1, variable domain sequences of trastuzumab, as well as those of other therapeutic antibodies, were obtained from TheraSAbDab⁶⁹ (accessed May 2024). The list of therapeutic antibodies was further filtered for entries in the format ‘Whole mAb’. This list was overlapped with antibodies with available protein structural data curated in the VCAb database⁷⁰ to obtain the structure input for Rosetta. A total of 55 human antibodies remained for the analysis. Class-switched memory B cell BCR repertoire sequences were obtained from Jaffe et al.⁵² BCR repertoire from intratumoral B cells in breast tumor samples were obtained from Harris et al.² Delineation of CDR and FW followed the IMGT numbering,⁷¹ taken directly from the VCAb⁷⁰ web-tool for sequences extracted from antibody structures and using ANARCI⁷² for sequence data.

Protein language model analysis

We considered the following protein language models: (1) AntiBERTy⁵; (2) AbLang2⁴⁸; (3) AntiBERTa2⁴⁷; and (4) the ‘esm2_t30_150M_UR50D’ version of ESM-2.⁴⁶ For each protein language model, we input the heavy and light chain variable region sequences, and tasked the model to predict the amino acid identity of one specified position by masking this position whilst keeping the remaining sequence unmasked. We obtained from each model the output logits for each possible amino acid at the masked position and converted the logits using the softmax function (in the pytorch package⁷³) to probability scores that are stored. This process was repeated for every position in the sequence, resulting in an n x L matrix, where n represents the total number of amino acids (i.e., 20), and L is the length of the input sequence.

In silico mutational impact prediction

We predicted the stability effects of mutating every VH and VL positions to every other 19 amino acids using tools implemented in the Rosetta software suite⁷⁴ (source code version 2018.33.60351 bundle). The trastuzumab Fab coordinates from PDB 7mn8⁶⁶ were subject to backbone conformational sampling using the relax protocol in Rosetta, to generate n = 200 structures. Each relaxed structure was subject to the point mutant scan (pmut_scan) protocol for predicting every possible amino acid substitution in the VH and VL regions. The pmut_scan application predicted change in free energy (∆∆G), which was taken as an indication of whether the mutation was predicted to be stabilizing (negative ∆∆G) or destabilizing (positive ∆∆G). Mutational impact prediction was predicted for double mutants also using the pmut_scan application, with the additional option ‘-double_mutant_scan’ applied when invoking pmut_scan. The ref15 scoring function⁵⁰ was used throughout these protocols.

Molecular dynamics simulation

MD analysis of trastuzumab Fab WT and the mutants studied in this work (VH R50S+R59N, VH S85N+R87T, VL Q89A and VL Q89H) were performed using Gromacs⁷⁵ and the Amber14sb forcefield.⁷⁶ The input coordinates for the trastuzumab Fab-HER2 complex were extracted from PDB 7mn8,⁶⁶ and subjected to superposition with PDB 8q6j⁶⁷ to address missing residues in HER2 in PDB 7mn8. Histidine residues’ protonation state was assigned through PropKa.⁷⁷ All the systems were prepared using an identical procedure detailed as follows: the structure was placed at the center of a dodecahedron box and filled with TIP3P water. System charge was neutralized by adding counter ions to reach a concentration of NaCl equal to 0.15 M. Three minimization steps were then performed, gradually decreasing the heavy atoms’ position restraints. Equilibration runs were started from the minimized structure in the NVT ensemble, gradually increasing the temperature from 50 K to 300 K. This was followed by NPT equilibration (1 ns) using the Berendsen barostat. Following equilibration, production runs were generated, applying the Parrinello-Rahman barostat to restrain the pressure at 1 bar. The temperature was kept constant at 300 K using the velocity-rescale algorithm.⁷⁸ Equations of motion were integrated using the Leap-Frog integrator and a timestep of 2 fs. Electrostatic interactions were computed using the Particle Mesh Ewald scheme,⁷⁹ with a cutoff of 1.2 nm. The same cutoff value was used for van der Waals interactions.

A table detailing the simulation setups, number of replicates and simulation time can be found in Supplementary Table S3. In brief, a total time of 3 μs was simulated for each system with at least 3 replicates. In R50S+R59N, an unbinding event was sampled in the first 0.2 μs of one replicate and not recovered. The simulation was therefore interrupted at 0.5 μs. An additional simulation of the same length was performed for this system. Mutants were produced by changing the selected residue(s) in the input structure using PyMOL (Schrödinger, LLC).

Trajectory analyses were performed using GROMACS tools and the MDTraj⁸⁰ python library (version 1.10.0). π−π stacking was defined on the distance between the center of mass (COM) of the aromatic rings (hereafter COM_ring) and the angle between their norms.^81,82 XH-ring (with X = C, N) interaction was determined by examining the distance between the atom X and COM_ring, as well as the angle X-H-COM_ring.^83–85 A detailed list of numeric cutoffs for angles and distances used here to identify molecular interactions can be found in Supplementary Table S4. Hydrogen bonds were assigned according to the Baker–Hubbard criterion.⁸⁶ Interaction networks were analyzed using the Networkx library (version 3.2.1).

Identification of local backbone correlated motions upon antigen binding

We used the AlloHubPy⁵⁶ package to analyze our MD simulations for estimating correlated backbone motions within trastuzumab Fab that were associated with antigen binding. We considered the MD trajectories of WT trastuzumab Fab in the absence or presence of HER2, using the last 0.5 μs of each replica. Briefly, each frame in the trajectories was coarse-grained at the level of consecutive four-residue backbone fragments using the structural alphabet M32K25.⁸⁷ AlloHubPy estimated the normalized mutual information (MI) between each pair of backbone fragments while correcting for systematic errors due to finite conformational sampling.⁸⁸ Aiming to identify correlated antibody backbone motions upon HER2 binding, a difference MI matrix was built by subtracting the average MI within MD trajectories of the unbound state from the analogous matrix obtained from trajectories of the bound state. We carried out a differential coupling analysis to identify pairs of fragments (“hubs”) with a log2-fold change in a coupling strength of ≥2 units and an associated p-value < 0.01.

In vitro analysis

Cells

Human leukocyte cones were obtained from the United Kingdom National Health Service Blood and Transplant system using anonymous donor leukocyte cones following provision of written, informed consent. SKBR3 (ATCC HTB-30) cells were cultured in complete DMEM (10% fetal bovine serum (FBS), penicillin (5,000 U/ml), streptomycin (100 μg/ml)).

Cell culture conditions

The HER2+ SKBR3 breast cancer model (RRID:CVCL_0033) was obtained from King’s College London Breast Cancer Now Unit and maintained in high-glucose Dulbecco’s modified Eagle’s Medium (DMEM)-GlutaMAX (ThermoFisher Scientific) with 10% heat-inactivated fetal calf serum. Cells were authenticated by short tandem repeat profiling and used after being tested negative for mycoplasma, up to 30 passages. The cells were maintained in a 5% CO2 humidified incubator at 37 °C.

Flow cytometric analyses

To detect HER2 protein expression levels in vitro, cells were detached with trypsin for 3 min and direct immunofluorescence staining was performed for 20 min on ice using the anti-HER2 IgG1 antibody trastuzumab, followed by a goat anti-human IgG FITC secondary antibody (Vector Laboratories, #FI-3000–1.5). Samples were acquired using the FACSCanto II flow cytometer equipped with BD FACSDiva Software (BD Biosciences) (RRID:SCR_001456) and data were analyzed with FlowJo_V10 software (RRID:SCR_008520) to measure MFI.

In vitro cell viability assay

To measure cell viability, 500 SKBR3 cells per well were plated in 384-well-plates and incubated with treatments for 96 hours at 37 °C. Cell viabilities were detected by CellTiter-Glo Assay (Promega) according to the manufacturer’s instructions. Optical absorbance was read on FLUOstar Omega spectrophotometer (BMG Labtech).

ADCC assays

Natural killer (NK) cells were isolated using the RosetteSep^TM Human NK Cell Enrichment Cocktail (STEMCELL Technologies) following the manufacturer’s instructions. After isolation, NK cells were maintained overnight in RPMI1640 supplemented with 10% FBS. The next day, NK cells were added to SKBR3 cells at an effector-to-target cell ratio (E:T) of 10:1 in the presence of antibody (0.01 µg/ml). Samples were incubated for 4 hours at 37°C, 5% CO₂. A no antibody control served as spontaneous release control. SKBR3 cells only were lysed 45 min prior to the end of the incubation and served as maximum release control. After centrifugation, 50 µL of supernatant were transferred to a clear flat-bottom 96-well plate. Quantification of lactase dehydrogenase (LDH) was performed using CyQUANT LDH Cytotoxicity Assay Kit (Invitrogen) following the manufacturer’s instructions. Absorbance at 490 nm and 680 nm was measured using a FLUOstar® Omega Spectrophotometer (BMG Labtech).

Differential scanning fluorimetry

To determine the melting temperatures of the mAbs, 5 µg of each respective mAb in phosphate-buffered saline were run in triplicate with SYPRO® Orange (final concentration of 5X) on a QuantStudio7TM Flex Real-Time PCR System. Fluorescence was recorded from 25 to 95 °C with a ramp of 0.05°C/s. Melting temperatures (T_m) were obtained by determining the local minima of the first derivative of the fluorescence curves.

Statistics and data visualisation

Statistical analysis was carried out in the R statistical computing environment (v4.1.0). Data visualization was generated with the R ggplot2 package (v3.5.1). Structural visualization was generated with PyMOL (v3.0.0). For dose-dependent measurements, we fitted an analysis of variance (ANOVA) model where the measured values were modeled as a function of dose and the antibody construct. Unless otherwise stated, for other comparisons we used a linear mixed effect model (using the R lme4 package⁸⁹ incorporating the antibody construct as a fixed effect and replicates as a random effect. Statistical significance was examined via post-hoc Dunnett’s test (R multcomp package), comparing coefficients corresponding to a specified mutant and that of the wild-type. For ADCC assays, statistical significance was evaluated using a paired t-test, with data points paired by biological replicates delineated by the donor origin of the NK cells used in the experiment. For the other data presented herein, statistical significance was assessed using a Wilcoxon rank-sum test.

Funding Statement

The authors acknowledge support by the Biotechnology and Biological Sciences Research Council [BB/T002212/1 to F.F. as principal investigator]; a Career Enhancing Grant from the British Society for Immunology to J.C.F.N.; Worldwide Cancer Research [24-0087]; Breast Cancer Now [147; KCL-BCN-Q3] and the CRUK City of London Centre Award [C7893/A29290]. D.G. was supported by a studentship from the China Scholarship Council [number 202008440414]. A.K. was supported by a Summer Studentship award from King’s College London Faculty of Life Sciences and Medicine. This research was supported by the King’s Health Partners Centre for Translational Medicine. The views expressed are those of the authors and not necessarily those of King’s Health Partners.

Disclosure statement

S.N.K. is a founder and shareholder of Epsilogen Ltd. M.G. and J.C. have been employed through a fund provided by Epsilogen Ltd. S.N.K. declares patents on antibody technologies. All other authors declare no competing interests..

Author contributions

J.C.F.N. conceived the project and planned the investigation under the supervision of F.F. and S.N.K. J.C.F.N. coordinated the bioinformatics analyses, contributed by T.C., D.G., A.K. and D.I. M.L.D.S. performed and analyzed the molecular dynamics simulations under the supervision of F.F. and J.C.F.N. A.C. contributed and coordinated the in vitro analyses, with contributions from M.G., A.C., A.M., J.C., Y.L. and S.P. J.C.F.N. drafted the manuscript, with input from F.F. and S.N.K. All authors read, commented and approved the final manuscript.

Supplementary material

Supplemental data for this article can be accessed online at https://doi.org/10.1080/19420862.2025.2532117