Predicting changes in neutralizing antibody activity for SARS-CoV-2 XBB.1.5 using in silico protein modeling

The SARS-CoV-2 variant XBB.1.5 is of concern as it has high transmissibility. XBB.1.5 currently accounts for upwards of 30% of new infections in the United States. One year after our group published the predicted structure of the Omicron (B.1.1.529) variant’s receptor binding domain (RBD) and antibody binding affinity, we return to investigate the new mutations seen in XBB.1.5 which is a descendant of Omicron. Using in silico ico modeling approaches against newer neutralizing antibodies that are shown effective against B.1.1.529, we predict the immune consequences of XBB.1.5’s mutations and show that there is no statistically significant difference in overall antibody evasion when comparing to the B.1.1.529 and other related variants (e.g., BJ.1 and BM.1.1.1). However, noticeable changes in antibody binding affinity were seen due to specific amino acid changes of interest in the newer variants.


Introduction
In late November 2022, the United States Centers for Diseases Control stated that they began tracking a new SARS-CoV-2 variant known as XBB.1.5. At that time, XBB.1.5 was responsible for around 3% of all infections. Since then, XBB.1.5 has grown to represent 30% of all infections by January 2023 (1,2). XBB.1.5 is characterized by 40 mutations in the Spike protein, 22 of which are in the receptor binding domain (RBD) (3). The highly prevalent mutations in the RBD are shown in Table 1 below. A health concern is that XBB.1.5 may evade existing antibodies derived from therapeutics, vaccination, and or previous Omicron (B.1.1.529) infection. It has been proposed that XBB.1.5 is a recombinant strain of the virus from BJ.1 and BM.1.1.1 as portions of the mutated Spike protein appear to be from each parent strain (4,5). However, alternative hypotheses such as convergent evolution may also explain the similarity of portions of XBB.1.5's mutated regions to those seen in other variants (6,7). In our previous work on the prediction of the receptor binding domain (RBD) structure of the Omicron variant, our process provided robust predictions, having a root mean square de- viation of atomic positions (RMSD) of 0.574Å between the predicted and empirically derived Omicron RBD structure (PDB: 7t9j) (8). Furthermore, our previous study proved useful as a predictive gauge of antibody efficacy several weeks prior to when empirical validations of the Omicron-antibody binding changes could be performed (9).
In this study, we use the methodology in our previous work to investigate XBB.

Structural Changes in Antibody Binding Affinity.
Of the ten antibodies tested in this study, we focus on the structural bases in which the antibodies LY-CoV555, LY-CoV1404, and AZD8895 work. The neutralization mechanisms of three antibodies have been extensively studied. These three antibodies have been available as therapeutics for treatment against COVID-19 infections (either currently or previously in the United States under Emergency Use Authorization) (10-12).
Bamlanivimab (LY-CoV555). As shown in Figure 3, we see a consistent interaction between Bamlanivimab (LY-CoV555) and the variant RBDs at R/Q493. This differs from Jones et al. (13), which states that F490 and S494 in the RBD are the interfacing residues in this region.

P R E P R I N T
The PyMOL structural visualizations of the potential interaction residues coincides with the overall metrics returned from the HADDOCK analyses shown in Table 2 Figure 3, the latter three complexes show a higher number of interfacing residues overall than in BJ.1, thus supporting the reported affinity metrics. Bebtelovimab (LY-CoV1404). Westendorf et al. (14) demonstrated that Bebtelovimab (LY-CoV1404) antibody binding affinity may not be affected by RBD mutations at E484, F490, Q493. Shown in Figure 4, we see consistent interactions from this antibody across all four variants around most of these positions in spite of mutations. These findings for LY-CoV1404 are congruent with the reported affinity metrics from the HADDOCK analyses shown in Table 3. Overall, HADDOCK scores are stable across the four variant complexes. The antibody LY-CoV1494 is predicted to have a slightly weaker interaction with XBB.1.5 compared to the other three variants. Tixagevimab (AZD8895). For tixagevimab (AZD8895), as reported in Dong et al. (15), there is a critical contact residue at F486 on the RBD. We see this residue being interfaced in B.1.1.529, BJ.1, and BM.1.1.1. However, the F486 residue is mutated to proline at this position in XBB.1.5, though interactions from the antibody to the adjacent RBD residues at G485 and N487 of the RBD still occur. See Figure 5. From the HADDOCK metrics shown in Table 4, this F486P mutation increases the binding affinity with the AZD8895, especially in terms of Van der Waals and electrostatic energies. Interfacing residues are abundant across all four of these AZD8895-RBD complexes (in addition to those shown in Figure 5), thus providing additional agreement to the strong affinity metrics reported by HADDOCK.

Methods
Our in silico modeling approach includes the curation or generation of the RBD structures for four SARS-CoV-2 variants and ten neutralizing antibody structures. Next, each antibody structure was docked against each RBD structure and binding affinity metrics were collected for comparison. were predicted with these sequences using AlphaFold2 (ColabFold-mmseqs2 version) (17,18). Next, the most confident structure of each was used in docking analyses.

Antibody Structures.
Representative antibody structures were collected from various Protein Data Bank entries ranging from antibodies derived from infected patients (or patients with breakthrough infections) or commercially available antibodies used in the treatment of COVID-19. See Table 5. Only a fragment antigen-binding (Fab) region of the antibodies was used in the docking analyses.

Antibody
Other

Docking.
To prepare the Fab structures, we renumbered residues according to HADDOCK's requirements such that there are no overlapping residue IDs between the heavy and light chains of the Fab's .PDB file. Residues in the Fab structures' complementarity-determining regions (CDRs) were selected as "active residues" for the docking analyses. Residues in the S1 position of the RBD were selected as the "active residues" of the RBD structures. Since all of the input RBD .PDB files were renumbered to numbers 339-528, all of the input RBD files share the same "active residue" numbers. Each of the ten antibody structures where docked against each of the four RBD structures using HADDOCK v2.4, a biomolecular modeling software that provides docking predictions for provided structures (27).
The HADDOCK system outputs multiple metrics for the predicted binding affinities and an output set of .PDB files containing the antibody docked against the RBD protein.
PRODIGY, a tool to predict binding affinities using Gibbs energy, reported as ∆G in Kcal/mol units), was also run on each of the complexes (28). This process resulted in forty sets of docked structures. Each set contains many antibody-RBD complex conformations, from which we selected the top-performing structure for each antibody-RBD pair. We used this top-performing complex for subsequent structural investigations into interfacing residues and docking positions. These analyses were performed on the antibody-RBD structure pairs shown in Figure 1. The multiple metrics were used to assess the overall binding affinity changes between SARS-CoV-2 variants across multiple representative antibodies. Further, the docked Protein Data bank Files (PDB) were manually reviewed using PyMOL (29) to search for interfacing residues and polar contacts between the RBD and Fab structures that may indicate neutralizing activity.

Conclusions
Building on our previous work (8) in studying Omicron's structure, we have continued to demonstrate the utility of in silico modeling for predicting whether antibody binding affinity changes with the evolution of new SARS-CoV-2 variants. Given that in vitro assessment of protein structure and antibody binding experiments are costly and take an extended time, in silico computational modeling provides a more economical and faster method near or at empirical resolution. Our previous in silico results were confirmed via an empirical study reported by VanBlargan et al. (9). With computational modeling we rapidly preditct the potential severity of a new variant and provides predictions on antibody binding affinity. These predictions inform public health considerations and provide a method of rational drug design based on expected therapeutic and vaccine (and booster) efficacy. Computational modeling can be used to rapidly infer the public health consequences of a new variant in terms of the loss of efficacy of antibodies, such as breakthrough infections and associated healthcare burden.  (15) reports that aromatic residues from AZD8895 CDR loops form a hydrophobic pocket with the RBD residues G485, F486, and N487. Note that in XBB.1.5 there is a F486P mutation and, interestingly, the adjacent residues (G485 and N487) are interfaced in our predicted complex. It is possible the proline at position 486 provides less steric hindrance than phenylalanine, thus allowing surrounding residue interaction. This result can be tested in future studies.
The increased binding affinity of XBB.1.5 for ACE2 may lead to increased transmissibility at the population level (30). The results here do not indicate that we can expect increased disease severity on an individual level for patients that avail themselves of therapeutics and vaccination. The climb in cases of COVID-19 disease linked to XBB.1.5 indicates that XBB.1.5 could be a very serious subvariant of Omicron. While other studies are needed to assess transmissibility, virulence, pathogenicity, and other facets of viral severity and epidemiology, this study predicts that many current therapeutic and infection-derived antibodies provide antibody binding affinities similar to B.1.1.529 for XBB.1.5. Thus, the results indicate that the health care outcomes should be positive for the patients that avail themselves of vaccines and therapeutics.
Limitations and Future Work. This work estimates potential changes in antibody neutralization effects or antibody neutralizing affinity using in silico protein modeling and computational docking analyses. Given the computational and predictive nature of this study, empirical investigations are necessary to validate these findings. However, these computational approaches provide an economical, scalabl, and rapid methodology to understand the severity of new viral variants while the empirical work is being completed. Also, while HADDOCK is considered state-of-the-art in terms of protein docking, there are other docking tools that could pose different results for the comparisons While we tested 10 representative neutralizing antibody structures against four variants of SARS-CoV-2, there are many more and antibody-variant RBD complexes to be assessed. In future work, we shall improve and automate our docking pipeline to enable to large-scale prediction of antibody binding affinity changes across any future SARS-CoV-2 variants of interest. Also, given the dynamic nature of protein structure conformations, alternative conformations may exist that show other polar contacts and antibody-RBD interfaces than those shown by the best performing HADDOCK complexes. All docking outputs and results, including those not shown in the body of this article, are available in the Supplementary Materials.

Conflict of Interest
Author CTF is the owner of Tuple, LLC. The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Supplementary Note 1: RBD Structural Alignment and Comparison
Using PyMOL's alignment tool (with 50 cycles and a cutoff of 2.0Å) (29), we superimposed the RBD structures as shown in Figure 6 and show RMSD metrics in Table 6.  Note that none of the mutations are predicted to disrupt the overall RBD tertiary structure in the AlphaFold2-generated structures. There are minor secondary structure changes as to where alpha helices or the anti-parallel beta sheet may begin or end, but overall the structures are very similar.
Looking at the main loop structure at the top of the S1 region, shown in Figure 7, the residue side chains are in similar positions, differing only by slight angular changes with the exception of the F486P mutation as previously mentioned. This proline mutation does not change the overall loop's conformation, but provides rigidity at this location.  Table 7. CPORT-predicted active and passive residues of the four RBD structures.

Supplementary Note 2: RBD Active Residue Predictions
Also, if we look at the residues that were predicted to be active visually, we can see that the majority of these residues concentrate around the S1 area of the RBD. Specifically, the loop structure on the top of the RBD that has been discussed Ford et al. | Is XBB.1.5 evading neutralizing antibodies? bioRχiv | 7 P R E P R I N T heavily in this study and previous works is consistently predicted to contain multiple active residues across all four variant structures. See Figure 8. There is considerable agreement between the CPORT predictions and the HADDOCK results listed in the main article. Nearly all of the interfacing residues detected in the complexes shown in Figures 3, 4, and 5 are predicted to be active residues from CPORT. Furthermore, many of these predicted active residues are also mentioned in Jones et al. (13), Westendorf et al. (14) and Dong et al. (15), thus further supporting that these residues on the top of the S1 region stand as the likely epitope between the RBD and various neutralizing antibodies evaluated in this study.

Supplementary Note 3: PRODIGY Results
In Figure 1, we show the top performing antibody-RBD complexes' HADDOCK and PRODIGY scores. That is, the metrics for the top performing complex from the top performing HADDOCK cluster. In Figure 9 below, we show the individual PRODIGY scores of the top four complexes from the top performing HADDOCK cluster for each antibody-RBD experiment. There is agreement that no statistically significant drop in overall antibody binding affinity in terms of ∆G as predicted by PRODIGY.  Fig. 9. PRODIGY binding affinity metrics for the top four clusters in each antibody-RBD complex that resulted from the HADDOCK docking process.