Antibodies represent a multi-billion dollar pharmaceutical market, but the majority of antibody therapies on the market belong to the IgG1 subclass, with other types and subclasses left majorly unexplored as therapeutics options(6). Each subclass has unique features conferring niches in fighting the breadth of encountered pathogens. IgG3 has the strongest effector functions of all IgG sub classes, and represents a subtype of antibodies which is strongly associated with malaria immunity(3,4). However, the short in vivo half-life of IgG3 due to low binding with the neonatal Fc gamma receptor (FcRn) limits its use as a therapeutic agent. We propose to examine the binding of IgG3 to FcRn, known for its role in transplacental transfer and as an IgG salvage receptor. We aim to improve IgG3’s binding affinity to FcRn in silico by computationally modeling mutations derived from biochemical examination of the binding interface, and comparison to previously discovered natural allotypes and engineered IgG’s. There are 29 allelic variants of IgG3 reported across human populations(2). A few of these variants contain an arginine to histidine mutation at position 435, greatly enhancing FcRn binding interactions, and resulting in an increased half-life from roughly one to three weeks. Alongside increased half-life, R435H IgG3 variants promote increased transplacental transfer of malaria-specific IgG3, leading to a reduced risk of malaria during infancy(4). These half-life extending, transcytosis enhancing allotypes, alongside potent complement activating abilities, make IgG3 an interesting candidate for antibody engineering(2,5). IgG1 will serve as the primary model for IgG engineering, as it has been engineered for tighter binding to FcRn to the effect of greatly increased in vivo half-life.
There are multiple data sets surrounding the engineering of IgG1 which can serve to test our workflow(1,8). Wild type IgG sequences and structures are provided by the PDB's 5JII (IgG1) and 6D58 (IgG3). Published IgG1 Fc engineering work is listed in Table 1. The Fc mutant YTE enhances IgG1-FcRn binding affinity 10-fold at pH 6.0, resulting in a 2 to 4-fold increase in serum half-life in humans, providing an excellent starting point for in silico engineering of IgG3(6). Increasing the binding affinity between IgG3 and FcRn at pH 6.0-6.5, while maintaining primary structure binding affinity at physiological pH is critical to successful IgG3 therapeutic use. FcRn half-life extension works under conditions of tight binding at low pH, such that when an is present in an endosome it binds tightly to FcRn, and is carried back to the cell surface(1,3). The second part of this mechanism is FcRn:IgG release at physiological pH once FcRn is once again on the cell surface(1,3). This mechanism is due in large part to the protonation of histidines at low pH, resulting in tighter binding due to salt bridges. Maintaining structure is essential in order to ensure other Fc-mediated functions such as complement activation remain intact since IgG3 is primarily of interest due to its strong existing Fc-mediated functions. Once assured of the capabilities of our workflow with IgG1, we will confirm its efficacy using IgG3 R435 and R435H allotypes. In silico design and engineering of IgG3 will be achieved by altering altering surface residues at the IgG3:FcRn interface. Small surface residue changes can facilitate tighter binding by the formation of hydrogen bonds or complementarily hydrophobic patches.
Figure 1: IgG3 (Left) with a longer and more flexible hinge region, and IgG1 (right)
There are multiple crystal structures available with different truncations at the N and C terminus, and in the case of IgG3, different allotypes. The first step prior to any engineering or workflow testing was defining the sequence we will use as wild type (WT) for both IgG1 and IgG3. These sequences are shown in Figure 2. The Figure was generated using the NCBI Blast tool. Figure 2 also describes which residues were considered critical or relevant for binding to occur between Fc and FcRn. This was based on the work of Lim et al.(6). In order to stay consistent, we defined the low pH environment by specifying that H310, H433, and H434 were protonated, while at physiological pH they were not. This generated the most consistent results when compared to literature.
Figure 2: Wild type sequence alignment of the Fc portion of IgG1 and IgG3. The residues in black represent the amino acids which do not align between IgG1 and IgG3. Highlighted residues are pulled from Lim et al., residues highlighted in gray are defined as loosely relevant for binding while yellow highlighted residues are critical for binding. The histidines with an asterisk below them were specified as protonated in order to simulate binding at low pH.
In order to establish our workflow, we first modeled the in silico Kd of WT IgG1 against a positive and a negative control. An H310A IgG1 variant with severely decreased binding to FcRn was used as a negative control, and the M252Y S254T T256E (YTE) IgG1 variant(15) was used as a positive control since it was included within Monnet et al's mutagenesis experiment for improved IgG1:FcRn binding(8). These three conditions were tested at "acidic", or "mock endosomal" pH via the manual protonation of histidine residues within the HADDOCK software (pH 6.0), and at "physiological" or "extracellular" pH, which did not include protonated histidines within the predicted binding interface.
Our workflow captured the pH-dependent binding characteristics wild type IgG1 Fc at low versus physiological pH, and, as predicted, the mutant YTE IgG1 Fc possessed greater binding affinity compared to WT at low pH. This aligns with experimental published data. Additionally, the binding was worsened by an order of magnitude when the critical residue H310 was mutated to alanine. The initial proof of concept data generated with IgG1 and FcRn is summarized below in Table 1. The published data by Monnet et al. represents a ratio of the specified variant over the Fc-QL variant, captured by fluorescence-based ELISA. Therefore, the increased ratio indicates higher binding, captured by a lower Kd in our workflow, but the fold-change is not directly translatable to changes in real or predicted Kd.
To further this proof of concept work, we tested several mutations which were also experimentally examined by Monnet et al. These mutations were all noted to improve binding via phage-ELISA assays, but only N315D and T256N improved binding using our in silico approach. The other two single mutations tested were T307A and N434Y, which were both expected to improve binding but actually worsened the in silico Kd. It is possible that residues which are less critical or directly involved in binding are more subtle in their mechanism of action and therefore are not able to be captured by our workflow. One example might be the mutation of a residue which is next to a more critical residue and plays a role in directing the critical residues side chain into a more favorable conformation for binding. It is also likely that these mutations influence what would be considered "active" and "passive" restraints within HADDOCK, and as HADDOCK relies on AIR definitions to improve docking results, inconsistencies between well-characterized IgG1 samples (like WT and the YTE variant) versus in silico, single-point mutations that lack structural information may lead to poor binding predictions. Ultimately, these methods do not capture the flexibility and random motion of proteins and thus lack the ability to fully capture the implications of every mutation.
Table 1: Literature binding of FcRn to Fc variants. Monnet et al described a phage-ELISA assay ratio between mutants and either WT Fc or Fc-QL. A higher ratio represented tighter binding. The Kd binding predictions by Haddock2.4 at either pH 6.0 (where we specified protonated Histidines) or physiological pH are represented in columns E and F.
Our final IgG3 mutations are shown in Table 2 below. We successfully generated 3 mutations which improved the binding of IgG3 Fc and FcRn. The mutation F436Y was pulled out of the sequence alignment as a residue which differed between IgG1 and IgG3, and was in the relevant binding residue pocket. Other mutations were selected based on improved binding in IgG1. Interestingly, only the single point mutation F436Y within IgG3 R435 improved binding, increasing affinity 1.8-fold at pH 6.0. The other mutations, though beneficial in IgG1, were not capable of improving IgG3 binding on their own. However, once these mutations were tested within the IgG3 H435 allotype model, 2 out of the 4 tested mutations improved binding. These included H435/N315D (2.2-fold improvement) and H435/T256N (2.1-fold improvement), which similarly improved IgG1 binding within our single-point mutation validation step. Perhaps these mutations alone are not capable of initiating better IgG3:FcRn binding but instead improve IgG3 binding once the histidine residue at position 435 undergoes protonation at low pH, initiating critical binding between FcRn's glutamate(position 115) and aspartate(position130) residues. The F436Y mutation in our H435 IgG3 model decreased FcRn binding despite improvement at the single-point mutation step, and may suggest some sort of negative interaction between histidine and tyrosine in our in silico model, leading to decreased binding at low pH and increased binding at physiological pH.
There were multiple limitations we identified and tailored our methodology to circumvent. Initially we intended to modify the sequence and crystal structures for our mutation candidates all within PyMol using the Wizard mutagenesis tool. This approach resulted in dissociation constants generated by Haddock2.4 which did not align with the published experimental data. Our theory was that the slight backbone and side chain shifting which would occur due to these mutations was not captured by PyMol. We shifted to modifying the sequence and then using the new amino acid sequences to generate an entirely new fold and PDB file which would be used as the Fc input to Haddock2.4. Both AlphaFold2 and ESMfold were evaluated as options to generate these new structures, but we found that only AlphaFold2 generated structures aligned with the experimental data.
The next limitation we encountered was the incorporation of more than one mutation at a time. When modeling IgG Fc's with single amino acid substitutions, our workflow was able to predict improved binding roughly in alignment with literature data. However, modeling an IgG Fc with multiple mutations resulted in contradictory data in relation to literature (Table 1). This could have been due to one single mutation which was not handled well by our workflow, or because the limitations already discussed regarding the method for a single amino acid substitution increase in severity when stacking multiple mutations. Haddock scores and RMSDs displayed poor results for combined mutations compared to single-point mutations which is likely contributing to this discrepancy. For example, the combined T256N, A378V, N434Y IgG1 mutation resulted in an RMSD for the best cluster of 23.8 +/- 0.4, much higher compared to the low-digit RMSDs of single point mutation models, which likely exacerbated its poor binding activity compared to Monnet et al.'s experiment. Future modeling iterations should work to 1) improve initial PDB models of mutated IgG1 before docking and 2) determine how "active" and "passive" restraints within HADDOCK fare within mutagenesis experiments. Recognizing this restraint, we evaluated only single mutations to IgG3, and suggest that further experimental evaluation would benefit from combination of the successful in silico mutations.
Our HADDOCK-modeling process also relies on a "pH shortcut" of manually protonating specific histidine residues which does not take into account how pH changes impact other amino acid residues within IgG1-Fc and FcRn. Lim et al.'s 2020 paper exploring in silico modeling of IgG1 and FcRn relied on a molecular dynamics simulation known as a constant pH molecular dynamics (CpHMD) simulation to test binding at pH 6.0 and pH 7.5, which would provide more accuracy at a higher computational cost. If we were to further validate mutations within our project we would likely need to incorporate a similar pH-sensitive molecular dynamics simulation to understand the global impact of pH on IgG Fc-FcRn, rather than simply focusing on residues in the binding interface. However, as a computationally cheap workaround, defining our low pH by simply protonating the three histidines of greatest interest generated binding data which consistently improved at low pH, roughly in alignment with experimental data.
Overall, despite the numerous limitations, this method is quite useful and could be used to design and test mutants. Given more time and expertise, it is likely we could have worked through some of these limitations. It seems like we just scratched the surface of all that bioinformatics platforms make possible for in silico protein design. As with all computational approaches, the results must be taken with a grain of salt, and an experienced experimentalist would need to thoroughly examine each AlphaFold structure and Haddock binding profile.