AlphaFold2-Enabled Atomistic Modeling of Structure, Conformational Ensembles, and Binding Energetics of the SARS-CoV-2 Omicron BA.2.86 Spike Protein with ACE2 Host Receptor and Antibodies: Compensatory Functional Effects of Binding Hotspots in Modulating Mechanisms of Receptor Binding and Immune Escape

Abstract

The latest wave of SARS-CoV-2 Omicron variants displayed a growth advantage and increased viral fitness through convergent evolution of functional hotspots that work synchronously to balance fitness requirements for productive receptor binding and efficient immune evasion. In this study, we combined AlphaFold2-based structural modeling approaches with atomistic simulations and mutational profiling of binding energetics and stability for prediction and comprehensive analysis of the structure, dynamics, and binding of the SARS-CoV-2 Omicron BA.2.86 spike variant with ACE2 host receptor and distinct classes of antibodies. We adapted several AlphaFold2 approaches to predict both the structure and conformational ensembles of the Omicron BA.2.86 spike protein in the complex with the host receptor. The results showed that the AlphaFold2-predicted structural ensemble of the BA.2.86 spike protein complex with ACE2 can accurately capture the main conformational states of the Omicron variant. Complementary to AlphaFold2 structural predictions, microsecond molecular dynamics simulations reveal the details of the conformational landscape and produced equilibrium ensembles of the BA.2.86 structures that are used to perform mutational scanning of spike residues and characterize structural stability and binding energy hotspots. The ensemble-based mutational profiling of the receptor binding domain residues in the BA.2 and BA.2.86 spike complexes with ACE2 revealed a group of conserved hydrophobic hotspots and critical variant-specific contributions of the BA.2.86 convergent mutational hotspots R403K, F486P, and R493Q. To examine the immune evasion properties of BA.2.86 in atomistic detail, we performed structure-based mutational profiling of the spike protein binding interfaces with distinct classes of antibodies that displayed significantly reduced neutralization against the BA.2.86 variant. The results revealed the molecular basis of compensatory functional effects of the binding hotspots, showing that BA.2.86 lineage may have evolved to outcompete other Omicron subvariants by improving immune evasion while preserving binding affinity with ACE2 via through a compensatory effect of R493Q and F486P convergent mutational hotspots. This study demonstrated that an integrative approach combining AlphaFold2 predictions with complementary atomistic molecular dynamics simulations and robust ensemble-based mutational profiling of spike residues can enable accurate and comprehensive characterization of structure, dynamics, and binding mechanisms of newly emerging Omicron variants.

Graphical Abstract

INTRODUCTION

The wealth of structural and biochemical studies on the SARS-CoV-2 viral spike (S) glycoprotein has provided critical insights into mechanisms of virus transmission and immune resistance.^1–9 The conformational changes within the SARS-CoV-2 S protein, transitioning between closed and open states, are primarily driven by the global movements of a flexible amino (N)-terminal S1 subunit of the S protein that includes an N-terminal domain (NTD), the receptor-binding domain (RBD), and two structurally conserved subdomains—SD1 and SD2. These structural domains coordinate their dynamic changes with a structurally rigid carboxyl (C)-terminal S2 subunit to facilitate diverse functional responses of the S protein through conformational transitions between the RBD-down closed and RBD-up open states.^10–15 Biophysical studies offered comprehensive insights into the thermodynamics and kinetics governing the behavior of the SARS-CoV-2 S protein trimer revealing how mutations and long-range interactions can orchestrate coordinated structural alterations in the dynamic S1 subunit and the more rigid S2 subunit, thereby modulating open-closed population shifts required for interactions with different binding partners, including the host cell receptor ACE2.^16–18 The increasing availability of cryo-electron microscopy (cryo-EM) and X-ray structures of the SARS-CoV-2 S protein variants of concern (VOCs) in diverse functional states and in association with antibodies unveiled that VOCs can induce structural changes in the dynamic equilibrium of the S protein, affect population of functional states and create the diversity of binding epitopes that contribute to the varying binding affinities of S proteins when interacting with different classes of antibodies.^19–28

The cryo-EM structures and biochemical analysis of the S trimers for BA.1, BA.2, BA.3, and BA.4/BA.5 subvariants demonstrated the decreased binding affinity for the BA.4/BA.5 and confirmed the higher binding affinities for BA.2 as compared to other Omicron variants.^29,30 Structural and biophysical studies of the Omicron BA.2.75 variant reported thermal stabilities of the Omicron variants showing that the BA.2.75 S-trimer was the most stable, followed by BA.1, BA.2.12.1, BA.5 and BA.2 variants, exhibiting a 9-fold enhancement of the binding affinity with ACE2 as compared to its parental BA.2 variant and significant antibody evasion.^31–33 Functional studies of the Omicron BA.1, BA2, BA.3, BA.4/BA.5, and BA.2.75 variants revealed a common trend wherein the acquisition of specific mutations that potentially promote immune evasion, possibly at the cost of reduced ACE2 affinity is often counterbalanced by the emergence of compensatory mutations aiming to restore or enhance ACE2 binding, thus potentially increasing the virus’s ability to bind and enter host cells.^34–41 The BA.2 variant produced a second-generation of Omicron variants that display a significant growth advantage and include BQ.1, BQ.1.1 XBB.1, XBB.1.5, and XBB.1.6 subvariants. The XBB.1 subvariant is derived from the BA.2 lineage and emerged by recombination of two cocirculating BA.2 lineages (BJ.1 and BM.1.1.1). The XBB.1.5 subvariant harbors F486P modification and, while it is equally immune evasive to closely related XBB.1 with F486S mutation, XBB.1.5 exhibits growth advantage due to the higher ACE2 binding affinity from a single S486P mutation that can restore most of the favorable hydrophobic contacts.^42,43 The biochemical studies examined the binding affinity of the XBB.1.5 RBD to ACE2 revealing the dissociation constant K_D = 3.4 nM which was similar to that of BA.2.75 (K_D = 1.8 nM) while significantly stronger than that of XBB.1 and BQ.1.1 variants.⁴⁴

The Omicron subvariant BA.2.86 was identified by global genomic surveillance in late August 2023 and exhibits significant genetic differences compared to its predecessors.^45–47 BA.2.86 is derived from the BA.2 variant but has 34 mutations relative to BA.2 (29 substitutions, 4 deletions, and 1 insertion) including RBD mutations I332V, D339H, K356T, R403K, V445H, G446S, N450D, L452W, N460K, N481K, delV483, A484K, F486P, and R493Q (Figure 1).^45–47

Figure 1.

Structural overview of the SARS-CoV-2 S protein and S-RBD for Omicron BA.2 (A) and BA.2.86 variants (B). The S protein is shown on orange ribbons (single monomer) with the S protein trimer structure shown in surface with the reduced transparency. For BA.2 S protein all BA.2 mutational sites are shown in pink-colored spheres and annotated. The positions of BA.2 mutations D339, A484 and R493 mutations are shown in blue spheres. The BA.2 RBD mutations are also shown projected onto crystallographic RBD conformation (orange ribbons) in the BA.2 RBD-ACE2 complex, pdb id 7XB0 (A). The positions of unique BA.2.86 S mutations relative to its ancestral BA.2 variant are shown in blue-colored spheres and fully annotated. The AF2-generated BA.2.86 RBD model (in orange ribbons) and BA.2.86 RBD mutations in blue spheres are shown and annotated.

BA.2.86 is also quite different from XBB.1.5 with 36 unique mutations including 32 substitutions, 3 deletions, and 1 insertion. Along with the shared mutations with XBB.1.5 (T19I, 24–26del, A27S, G142D, 144del, G339H, G446S, N460K, and F486P), additional mutations I332V, K356T, V445H, N450D, N481K, A484K and 483del emerged on BA.2.86’s RBD.^45–47 Some of these mutations such as K356T, R403K, V445H, N450D, L452W, delV483 and A484K are unique to this specific variant (Figure 1 and Table 1). Structural mapping of BA.2 and BA.2.86 specific mutations onto the S protein showed the distribution of mutational sites (Figure 1C,D) highlighting the unique BA.2.86 positions particularly in the RBD region near the binding interface with the ACE2 host receptor. A significant divergence of BA.2.86 subvariant is exemplified by the fact that the genetic distance between BA.2.86 and its predecessor BA.2 is similar to the distance observed between the BA.1 variant and the Delta variant of the virus (Supporting Information, Figure S1).^48,49

Table 1.

Mutational Landscape of the Omicron BA.1, BA.2, XBB.1.5 and BA.2.86 Variants^a

Omicron variant	mutational landscape
BA.1	A67, T95I, G339D, S371L, S373P, S375F, K417N, N440K, G446S, S477N, T478K, E484A, Q493R, G496S, Q498R, N501Y, Y505H, T547K, D614G, H655Y, N679K, P681H, N764K, D796Y, N856K, Q954H, N969K, L981F
BA.2	T19I, G142D, V213G, G339D, S371F, S373P, S375F, T376A, D405N, R408S, K417N, N440K, S477N, T478K, E484A, Q493R, Q498R, N501Y, Y505H, D614G, H655Y, N679K, P681H, N764K, D796Y, Q954H, N969K
XBB.1.5	T19I, V83A, G142D, Del144, H146Q, Q183E, V213E, G252V, G339H, R346T, L368I, S371F, S373P, S375F, T376A, D405N, R408S, K417N, N440K, V445P, G446S, N460K, S477N, T478K, E484A, F486P, F490S, R493Q reversal, Q498R, N501Y, Y505H, D614G, H655Y, N679K, P681H, N764K, D796Y, Q954H, N969K
BA.2.86	Ins16MPLF, R21T, S50L, del69-70, V127F, delY144, F157S, R158G, delN211, L213I, L226F, H25N, A264D, I332V, D339H, K356T, R403K, V445H, G446, N450D, L452W, N460K, del V483, A484K, F486P, R493Q, E554K, A570V, P612S, I670V, H68R, D939F, P1143L

^aFor the BA.2.86 variant, only mutations that are different from its ancestral BA.2 are listed.

Biophysical studies measured ACE2 binding affinities showing that XBB.1.5 and EG.5.1 spikes exhibited comparable affinities to ACE2, with K_D values of 1.34 and 1.21 nM, respectively, as compared to the K_D value of the BA.2 spike (1.68 nM). In contrast, both constructs of the BA.2.86 S protein showed a >2-fold increase in binding affinity, with similar K_D values of 0.54 and 0.60 nM.⁴⁵ Antigenicity and immune evasion capability of the BA.2.86 subvariant was tested on a panel of XBB.1.5-effective neutralizing antibodies revealing that BA.2.86 is antigenically distinct from XBB.1.5 and previous Omicron variants and can escape XBB-induced neutralizing antibodies due to N450D, K356T, V445H, L452W, A484K, V483del, and A484K mutations.⁴⁶ The neutralizing antibody responses against the BA.2.86 variant were considerably lower compared to the responses observed against the BA.2 variant but comparable to those observed against the XBB.1.5, XBB.1.16, EG.5, EG.5.1, and FL.1.5.1 variants.⁴⁷ Functional studies showed that BA.2.86 may exhibit similar growth advantages and immune evasion compared to antigenically distinct Omicron variants EG.5.1 and the FLip variant, which contains the L455F and F456L mutation in the background of the XBB.1.5 variant.^50,51 Comprehensive investigation of the virological characteristics of the BA.2.86 variant demonstrated that the ACE2 binding affinity of BA.2.86 S RBD was comparable to that of XBB.1.5 S RBD and significantly higher than those of the ancestral B.1.1, XBB.1, XBB.1.16, EG.5.1, and the parental BA.2 variants.⁵²

Although these initial investigations examined antigenicity and receptor binding signatures of the BA.2.86 variant, there is still a conspicuous lack of atomistic-level information regarding the structure, dynamics, and binding mechanisms of the BA.2.86 RBD binding with ACE2 receptor and a wide spectrum of experimentally studied antibodies. To our knowledge, there have been no structural or computational investigations conducted to date focusing on this particular variant.

In this study, we combined the AlphaFold2 (AF2) methodology^53,54 with all-atom MD simulations and in silico mutational scanning of binding energetics and stability for AI-augmented atomistic predictions of both structure, dynamics, and binding of the Omicron BA.2 and BA.2.86 RBD complexes with ACE2 receptor. The AF2 approach utilizes evolutionary information by considering multiple sequence alignments (MSAs) derived from related protein sequences as input to capture conserved contacts among evolutionarily linked sequences. This methodology incorporates the transformer architecture, featuring self-attention mechanisms that allow us to discern long-range dependencies and interactions within protein sequences.^53,54 Despite the remarkable success of AF2-based methods and self-supervised protein language models^55,56 that excelled at predicting static protein structures, there are notable short-comings related to their applicability in accurately characterizing conformational dynamics, functional protein ensembles, conformational changes, and allosteric states.⁵⁷ Several recent studies indicated that the structure prediction capabilities of the AF2 methodologies are not trivially expandable for the prediction of conformational ensembles and allosteric landscapes.^58–60 Efforts to optimize the AF2 methodology for predicting alternative conformational states of proteins have primarily centered around manipulating the MSA information which is motivated by the recognition that MSAs carry coevolutionary signals not only for the most thermodynamically probable protein state but also for other functionally relevant conformational states of a protein.^58,59 In one of these approaches, MSAs are randomly subsampled to reduce depth, resulting in shallower MSAs that aim to enhance the diversity and produce a broader range of the AF2 output models, including those representing functionally important and experimentally validated alternative conformational states of proteins.⁵⁸ The alternative AF2-based approach known as SPEACH_AF (Sampling Protein Ensembles and Conformational Heterogeneity with AlphaFold2) involves the manipulation of the MSA through in silico mutagenesis where replacing specific residues within the MSA can induce changes in the distance matrices, ultimately leading to the prediction of alternate protein conformations.⁵⁹ While AF2 relies on a large training set of solved structures and MSAs containing the evolutionary information used to infer structure, the number of experimentally determined alternative protein conformations is relatively small for sufficient training and robust inference. As a potential consequence of this fundamental learning limitation, AF2 methodology was shown to be biased to predict one conformation of fold-switching proteins as 94% of AF2 predictions captured only one experimentally determined conformation but not the other.⁶⁰

In the current work, we adapted and applied AF2 methodology using shallow MSA to predict the structure and conformational variability of the BA.2.86 RBD and the RBD-ACE2 complexes. Using the AF2-based predicted local distance difference test (pLDDT) metric, we can evaluate and rank structural models leading to a robust prediction of the BA.2.86 RBD-ACE2 structures along with pLDDT confidence scores for each residue position allowing us to evaluate structurally stable and more flexible regions. It is important to note that while pLDDT scores provide a valuable metric for assessing the overall confidence in the predicted structures, they may not explicitly capture the full range of dynamic behavior or conformational heterogeneity. The pLDDT score reflects the model’s assessment of local accuracy, but it does not directly represent thermodynamic or dynamic properties.

In the present work, we also leveraged complementary purposes and synergies between AF2 and molecular dynamics (MD) simulations to perform a comprehensive atomistic prediction and analysis of both structure, dynamics, and binding of the Omicron BA.2 and BA.2.86 RBD complexes with ACE2 receptor. MD simulation studies provided important atomistic insights into understanding the dynamics of the SARS-CoV-2 S protein and the effects of Omicron mutations on the conformational adaptability of the S protein and its complexes.^61–69 We recently combined MD simulations and Markov state models to systematically characterize conformational landscapes of the Omicron BA.1, BA.2, BA.3 and BA.4/BA.5 variants⁷⁰ and highly transmissible XBB.1, XBB.1.5, BQ.1, and BQ.1.1 variant complexes.⁷¹

Here, we combined AF2-based structural predictions of the top models for BA.2.86 RBD complexes with subsequent microsecond atomistic MD simulations performed to study the stability and dynamics of the AF2-predicted conformational states. Complementary to AF2-guided structural predictions, MD simulations reveal the details of the conformational landscape and the dynamic effect of BA.2.86 RBD mutations on binding to ACE2. In addition, MD simulations provide equilibrium ensembles of the BA.2.86 RBD-ACE2 structures that are used to perform an ensemble-based mutational scanning of the RBD residues in the ACE2 complexes and characterize conserved and variant-specific binding energy hotspots. Hence, combining AF2 predictions with MD simulations offers a synergistic approach in which AF2 provides accurate multiple structures of the BA.2.86 RBD and ACE2 complexes, while MD simulations leverage these predicted conformations to characterize dynamics and energetics of binding with ACE2 receptor. To understand the mechanisms of antibody evasion for BA.2.86 and characterize the specific role of BA.2.86 mutations in enabling immune escape, we also conduct structure-based mutational scanning of the S RBD complexes with 20 antibodies for the four epitope classes. The results detail energetic mechanisms by which BA.2.86 mutational sites can elicit resistance to neutralization of the RBD antibodies. The results suggest a mechanism in which convergent Omicron mutations can control the interplay between the RBD stability and conformational adaptability, allowing for optimal fitness trade-offs between binding to the host receptor and robust immune evasion profile.

MATERIALS AND METHODS

AI-Based Structural Modeling and Statistical Assessment of AF2 Models.

Structural prediction of the BA.2.86 RBD and BA.2.86 RBD-ACE2 complex were conducted using the AF2 framework^53,54 within the ColabFold implementation⁷² using a range of MSA depths and other parameters. The default MSAs are subsampled randomly to obtain shallow MSAs containing as few as five sequences. We generated structures using shallow MSA depth by adjusting AF2 parameters using ColabFold.⁷² We used max_msa field to set two AF2 parameters in the following format: max_seqs:extra_seqs. Both of these determine the number of sequences subsampled from the MSA (max_seqs sets the number of sequences passed to the row/column attention track and extra_seqs the number of sequences additionally processed by the main evoformer stack). The lower values encourage more diverse predictions but increase the number of misfolded models. Similar to previous studies showing how MSA depth adaptations may facilitate conformational sampling⁵⁸ we manipulated and modified MSA depth by setting the AF2 config.py parameters with a max_seq:extra_seq ratio of 256:512, 128:256, 64:128, 32:64, and 16:32. We found that max_msa in format max_seqs:extra_seqs 16:32 provided the best combination of accurate structure prediction models and properly captured conformational ensembles of the BA.2.86 RBD structures. We additionally manipulated the num_seeds and the num_recycles parameters to produce more diverse outputs. Hence, we use max_msa: 16:32, num_seeds: 4, and num_recycles: 12. AF2 makes predictions using 5 models pretrained with different parameters, and consequently with different weights. To generate more data, we set the number of recycles to 12, which produces 14 structures for each model starting from recycle 0 to recycle 12 and generating a final refined structure. Recycling is an iterative refinement process, with each recycled structure getting more precise. Each of the AF2 models generates 14 structures, amounting to 70 structures in total. In addition, we also predicted one more structure using AF2 with the default and “auto” parameters serving as a baseline structure for prediction and variability analysis. AF2 models were ranked by pLDDT scores (a per-residue estimate of the prediction confidence on a scale from 0 to 100), quantified by the fraction of predicted Cα distances that lie within their expected intervals. The values correspond to the model’s predicted scores based on the lDDT-Cα metric, a local superposition-free score to assess the atomic displacements of the residues in the model.^53,54 Structural models were compared to the experimental structure of BA.2 RBD-ACE2 (pdb id 7XB0) using structural alignment as implemented in TM-align.⁷³ An optimal superposition of the two structures is then built and the TM-score is reported as the measure of overall accuracy of prediction for the models. Several other structural alignment metrics were used including the global distance test total score GDT_TS of similarity between protein structures and implemented in the Local-Global Alignment (LGA) program⁷⁴ and the root-mean-square deviation (RMSD) superposition of backbone atoms (C, C_α, O, and N) calculated using ProFit (http://www.bioinf.org.uk/software/profit/).

All-Atom Molecular Dynamics Simulations.

The crystal structure of the BA.2 RBD-ACE2 (pdb id 7XB0) is obtained from the Protein Data Bank and structures of the BA.2.86 RBD-ACE2 complex are obtained from AF2 modeling. We selected BA.2.86 RBD-ACE2 models based on the pLDDT scores, particularly of the RBM loop residues as the highest scoring models have the lowest RMSD to the cryo-EM and crystal structures of the RBD-ACE2 complexes for Omicron BA.1, BA.2, BA.3, BA.4/BA.5 and XBB.1 variants. Hydrogen atoms and missing residues were initially added and assigned according to the WHATIF program web interface.⁷⁵ The missing regions are reconstructed and optimized using the template-based loop prediction approach ArchPRED.⁷⁶ The side chain rotamers were refined and optimized by the SCWRL4 tool.⁷⁷ The protonation states for all the titratable residues of the ACE2 and RBD proteins were predicted at pH 7.0 using Propka 3.1 software and web server.^78,79 The protein structures were then optimized using atomic-level energy minimization with composite physics and knowledge-based force fields implemented in the 3Drefine method.^80,81 The structurally resolved 2-acetamido-2-deoxy-beta-d-glucopyranose, 2-acetamido-2-deoxy-beta-d-glucopyranose-(1–4)-2-acetamido-2-deoxy-beta-d-glucopyranose, and chloride ions present in the RBD-ACE2 structures were included and optimized.

NAMD 2.13-multicore-CUDA package⁸² with CHARMM36 force field⁸³ was employed to perform 1 μs all-atom MD simulations for the Omicron RBD-ACE2 complexes. The structures of the SARS-CoV-2 S-RBD complexes were prepared in Visual Molecular Dynamics (VMD 1.9.3)⁸⁴ and with the CHARMM-GUI web server^85,86 using the Solutions Builder tool. Hydrogen atoms were modeled onto the structures prior to solvation with TIP3P water molecules⁸⁷ in a periodic box that extended 10 Å beyond any protein atom in the system. To neutralize the biological system before the simulation, Na⁺ and Cl⁻ ions were added in physiological concentrations to achieve charge neutrality, and a salt concentration of 150 mM of NaCl was used to mimic a physiological concentration. All Na⁺ and Cl⁻ ions were placed at least 8 Å away from any protein atoms and from each other. MD simulations are typically performed in an aqueous environment in which the number of ions remains fixed for the duration of the simulation, with a minimally neutralizing ion environment or salt pairs to match the macroscopic salt concentration.⁸⁸ All protein systems were subjected to a minimization protocol consisting of two stages. First, minimization was performed for 100,000 steps with all the hydrogen-containing bonds constrained and the protein atoms fixed. In the second stage, minimization was performed for 50,000 steps with all the protein backbone atoms fixed and for an additional 10,000 steps with no fixed atoms. After minimization, the protein systems were equilibrated in steps by gradually increasing the system temperature in steps of 20 K, increasing from 10 to 310 K, and at each step, a 1 ns equilibration was performed, maintaining a restraint of 10 kcal mol⁻¹ Å⁻² on the protein C_α atoms. After the restraints on the protein atoms were removed, the system was equilibrated for an additional 10 ns. Long-range, nonbonded van der Waals interactions were computed using an atom-based cutoff of 12 Å, with the switching function beginning at 10 Å and reaching zero at 14 Å. The SHAKE method was used to constrain all the bonds associated with hydrogen atoms. The simulations were run using a leapfrog integrator with a 2 fs integration time step. The ShakeH algorithm in NAMD was applied for the water molecule constraints. The long-range electrostatic interactions were calculated using the particle mesh Ewald method⁸⁹ with a cutoff of 1.0 nm and a fourth-order (cubic) interpolation. The simulations were performed under an NPT ensemble with a Langevin thermostat and a Nosé–Hoover Langevin piston at 310 K and 1 atm. The damping coefficient (gamma) of the Langevin thermostat was 1/ps. In NAMD, the Nosé–Hoover Langevin piston method is a combination of the Nosé–Hoover constant pressure method⁹⁰ and piston fluctuation control implemented using Langevin dynamics.^91,92 An NPT production simulation was run on equilibrated structures for 1 μs keeping the temperature at 310 K and a constant pressure (1 atm).

Distance Fluctuation Stability Analysis.

Our approach involved employing distance fluctuation analysis on simulation trajectories to derive residue-based stability profiles. This analysis computed the variations in the average distance between pseudoatoms representing a specific amino acid and those belonging to the remaining protein residues. These distance fluctuations for each residue, concerning all other residues in the ensemble, were then translated into distance fluctuation stability indexes. These indexes serve as measurements quantifying the energy expenses associated with the deformation of each residue throughout the simulations.^93,94 The adaptation of this approach for the analysis of rigid and flexible residues in the SARS-CoV-2 S proteins was detailed in our previous studies.^66–68 The high values of distance fluctuation stability indexes point to structurally rigid residues as they display small fluctuations in their distances to all other residues, while small values of this index would point to more flexible sites that display larger deviations of their inter-residue distances. The distance fluctuation stability index for each residue is calculated by averaging the distances between the residues over the simulation trajectory as follows:

$$k_i=\frac{3k_{\text{B}}T}{\langle {\left(d_i-\langle d_i\rangle \right)}^2\rangle }$$ (1)

$$d_i={\langle d_{ij}\rangle }_{j\ast }$$ (2)

d_ij is the instantaneous distance between residue i and residue j, k_B is the Boltzmann constant, T = 300 K. ⟨⟩ denotes an average taken over the MD simulation trajectory and $d_i={\langle d_{ij}\rangle }_{j\ast }$ is the average distance from residue i to all other atoms j in the protein (the sum over j_* implies the exclusion of the atoms that belong to the residue i). The interactions between the C_α atom of residue i and the C_α atom of the neighboring residues i − 1 and i + 1 are excluded in the calculation since the corresponding distances are constant. The inverse of these fluctuations yields an effective force constant k_i that describes the ease of moving an atom with respect to the protein structure.

Mutational Scanning and Binding Free Energy Computations.

The binding free energies were initially computed for the Omicron RBD-ACE2 complexes using the Molecular Mechanics/Generalized Born Surface Area (MM-GBSA) approach.^95,96 A common strategy applied to reduce noise and cancel errors in MM-GBSA computations is to run MD simulations on the complex only, with snapshots taken from a single trajectory to calculate each free energy component. We also conducted mutational scanning analysis of the binding epitope residues for the SARS-CoV-2 S RBD-ACE2 complexes. Each binding epitope residue was systematically mutated using all substitutions and corresponding protein stability and binding free energy changes were computed. The BeAtMuSiC approach^97–99 was employed that is based on statistical potentials describing the pairwise inter-residue distances, backbone torsion angles, and solvent accessibilities, and considers the effect of the mutation on the strength of the interactions at the interface and on the overall stability of the complex. The binding free energy of the protein–protein complex can be expressed as the difference in the folding free energy of the complex and folding free energies of the two protein binding partners:

$$\Delta G_{\text{bind}}=G^{\text{com}}-G^{\text{A}}-G^{\text{B}}$$ (3)

The change of the binding energy due to a mutation was calculated then as the following:

$$\Delta \Delta G_{\text{bind}}=\Delta G_{\text{bind}}^{\text{mut}}-\Delta G_{\text{bind}}^{\text{wt}}$$ (4)

We compute the ensemble-averaged binding free energy changes using equilibrium samples from simulation trajectories. The binding free energy changes were computed by averaging the results over 1000 equilibrium samples for each of the studied systems. The BeatMusic method demonstrated promising accuracy in various benchmark studies^100–103 and was successfully used in related studies to compute the change in the ACE2 affinity caused by mutations in the S-protein structures.¹⁰⁴ The advantages of this approach are fast and accurate predictions of the effect of mutations on both the strength of the binding interactions and on the stability of the complex using statistical potentials and neural networks. The benchmark studies and our analysis of different approaches for assessment of mutational effects on the S-RBD binding^62–71 showed that the BeatMusic approach is comparable to other knowledge-based structural methods such as D-complex,¹⁰⁵ physics-based FoldX potentials,^106,107 and BindProfX approach.¹⁰⁸ The recent analysis of computational tools for estimating protein–protein binding showed that physics-based FoldX methods often display a similar predictive power to comparable but much faster knowledge-based methods such as BeatMusic which is indispensable for massive mutational scanning experiments.¹⁰⁹ A large-scale in silico mutagenesis study using the BeAtMuSiC approach profiled all possible point mutations in the RBD residues on its stability and binding with antibodies and ACE2 receptor, showing that predictions agreed well with various experimental, epidemiological, and clinical data.¹¹⁰ While more rigorous physics-based approaches may yield more accurate predictions of binding affinity, the requirements for adequate sampling and comprehensive scanning of protein residues make this strategy less plausible in practice. Instead, our approach by combining atomistic simulations and knowledge-based models of protein binding provided the ensemble-average estimates of the binding free energy changes allowing for a comprehensive scanning and meaningful interpretation of the SARS-CoV-2 S binding.

To facilitate readability and underscore the integration of different methods used in this study, we supplemented the detailed description of methods by an additional graphical summary of tools and interfaces between methodologies (Supporting Information, Figure S2).

Electrostatic Potential Calculations.

In the framework of continuum electrostatics, the electrostatic potential φ for biological macromolecules can be obtained by solving the Poisson–Boltzmann equation (PBE)

$$-\nabla \times \in \nabla \phi -\sum _i^M{c}_i{q}_i{e}^{-\beta \left(q_i\phi +V_i\right)}=\rho$$ (5)

$$\nabla \times \left[\epsilon \left(r\right)\nabla \phi \left(r\right)\right]=-4\pi \rho \left(r\right)+\epsilon \left(r\right){\kappa }^2\left(r\right)\sinh \left(\phi \left(r\right)/k_{B}T\right)$$ (6)

where φ(r) is the electrostatic potential, ε(r) is the dielectric distribution, ρ(r) is the charge density based on the atomic structures, κ is the Debye–Huckel parameter, k_B is the Boltzmann constant, and T is the temperature. The electrostatic interaction potentials are computed for the averaged RBD-ACE2 conformations using the APBS-PDB 2PQR software^111–113 based on the Adaptive Poisson–Boltzmann Solver (APBS). These resources are available from the APBS/PDB 2PQR Web site: http://www.poissonboltzmann.org/. The atomic charges and radii are assigned in this approach based on the CHARMM force field.

RESULTS

AF2-Based Atomistic Modeling and Prediction of the BA.2.86 RBD-ACE2 Structure and Conformational Ensembles.

To develop accurate and robust atomistic models of BA.2.86 RBD structure and dynamics, we performed comparative structural prediction of the BA.2 and BA.2.86 RBD-ACE2 complexes using AF2 default settings and AF2 methodology with shallow MSA depth^81–83 within the ColabFold.⁸⁶ We systematically tested the accuracy of predicting the BA.2 RBD-ACE2 structural ensembles by comparing the AF2-derived conformations with the structural and biophysical studies of the S Omicron protein dynamics and binding.^114–116 First, we generated the top five AF2 models of the BA.2 RBD-ACE2 complex with default settings (Supporting Information, Figure S3) showing an excellent structural alignment with the crystallographic BA.2 RBD conformation (pdb id 7XB0) with high confidence pLDDT values. All the top AF2 models of the BA.2 RBD displayed RMSD < 1.0 Å from the crystal structure demonstrating the ability of AF2 to accurately reproduce atomistic details of the BA.2 RBD (Supporting Information, Figure S3). Using this prediction as a reference, we evaluated several combinations of parameters to employ AF2 with varied MSA depth. The objective of this analysis was to characterize the RBD-ACE2 ensembles in which the crystallographic state is the most frequent prediction within the ensemble. Consistent with previous studies, we found changing max_seq and extra_seq allows us to obtain predictions consistent with the crystallographic conformations with a max_seq:extra_seq ratio of 16:32 leading to the most diversity of the RBD loops and sufficient to unequivocally reproduce the experimental structure using pLDDT metric (Figure 2). Importantly, we found that the ensemble of AF2 models selected by the predicted local difference test (pLDDT) displays the greatest structural similarity to the experimental structure. Using several structural similarity metrics such as AF2-based predicted pLDDT score^53,54 TM-score,⁷³ GDT_TS,⁷⁴ and RMSD, we first examined and validated the prediction accuracy of AF2-MSA depth models for the BA.2 RBD-ACE2 complex (Figure 2). The density distribution of the pLDDT values obtained for the AF2-MSA depth ensemble of BA.2 conformations showed a pronounced peak at pLDDT scores of ~80–85 (Figure 2A). The density distributions of other structural metrics TM-score (Figure 2B) and GDT_TS (Figure 2C) that measured similarity between the predicted conformations and the crystal structure similarly displayed sharp peaks signaling the consistent prediction of the crystallographic conformation, while also pointing to moderate variability reflecting fluctuations of the flexible RBD loops. In particular, the distribution featured a major peak for TM-score ~0.9 confirming excellent predictions of the AF2-MSA depth model (Figure 2B).

Figure 2.

Distributions of structural model assessment and structural similarity metrics for the BA.2 RBD conformational ensemble obtained from AF2-MSA depth predictions. (A) Density distribution of the AF2-derived pLDDT structural model estimate of the prediction confidence on a scale from 0 to 100. (B) Density distribution of TM-score measuring structural similarity of the AF2-MSA depth predicted RBD conformations with respect to the crystal structure of the BA.2 RBD-ACE2 (pdb id 7XB0). (C) Density distribution of GDT_TS structural similarity metric between the AF2-predicted RBD conformations and the crystal structure of the BA.2 RBD-ACE2. (D) Density distribution of RMSD between the AF2-predicted RBD conformations and the crystal structure of the BA.2 RBD-ACE2.

Several minor distribution peaks around TM-score ~0.7–0.8 reaffirmed the quality of predictions that also reproduced the flexible RBD regions with high accuracy. The RMSD distribution showed that most of the predicted conformations are similar to the crystal structure with the main peak corresponding to RMSD ~0.75 Å from the crystallographic state (Figure 2D). Overall, the predicted AF2-MSA depth models consistently produced highly accurate predictions of the BA.2 RBD-ACE2 structures. Using the predictions of the BA.2 RBD-ACE2 structure as a validation baseline, we also generated AF2 models for the BA.2.86 RBD and BA.2.86 RBD-ACE2 complex (Supporting Information, Figure S4). The alignment of the top five BA.2.86 models with the crystallographic conformation of the BA.2 RBD showed only small deviations that are primarily associated with the functionally relevant plasticity of the RBM tip loop (residues 475–487) (Supporting Information, Figure S4). Structural mapping of the BA.2.86 mutational sites showed that their backbone positions remain unchanged and most variability may be expected in N481K and A484K sites. We also presented the statistical analysis and confidence assessment metrics of the BA.2.86 RBD predictions (Supporting Information, Figure S5). The MSA is summarized as a heatmap indicating all sequences mapped to the input sequences. The relative coverage of the sequence with respect to the total number of aligned sequences is shown indicating reduced sequence identity to query for flexible RBD regions (residues 480–530) (Supporting Information, Figure S5A). The pLDDT scores per residue for the top five models showed that the RBD core and functionally important mobile regions featured pLDDT values within the 70–90 range, which is indicative of high confidence in the predictions (Supporting Information, Figure S5B). The regions with pLDDT values ~50–70 have lower confidence and must be treated with caution, while pLDDT <50 may be a strong predictor of disorder (Supporting Information, Figure S5B). The heatmaps of the predicted alignment error (PAE) between each residue in the model are shown for the best AF2 five models (Supporting Information, Figure S5C), revealing differences between the high-confidence regions and the low-confidence regions. It is apparent from this analysis that flexible RBD loops 381–394 and especially 475–487 represent regions of lower confidence resulting from their inherent plasticity. Consistent with the ranking of the models, the lower confidence regions are moderately expanded in the flexible loop regions for models 4 and 5, respectively (Supporting Information, Figure S5C). Noteworthy, although pLDDT scores are valuable for assessing the reliability of AF2 predictions and identifying flexible regions, a complete understanding of conformational heterogeneity may require MD simulation approaches that explicitly consider dynamic and thermodynamic aspects of protein behavior.

We then proceeded to generate AF2MSA depth models for the BA.2.86 RBD-ACE2 complex and analyzed structural similarities between the optimized AF2-default model of the BA.2.86 RBD-ACE2 complex and generated conformations (Figure 3). The distribution of pLDDT scores yielded a peak at pLDDT ~82–85 reflecting the prediction convergence (Figure 3A). The distribution densities of structural similarity metrics obtained for the BA.2.86 RBD conformational ensemble showed similar trends to the ones seen for the BA.2 variant, featuring pronounced peaks for TM-score ~ 0.9–0.95 (Figure 3B), GDT_TS ~ 0.9 (Figure 3C) and for RMSD ~ 0.65 Å from the reference structure (Figure 3D). Hence, the majority of predicted BA.2.86 conformations using AF2 with MSA depth are similar to the reference AF2-generated BA.2.86 RBD state and crystallographic conformation of the BA.2 RBD-ACE2 complex.

Figure 3.

Distributions of structural model assessment and structural similarity metrics for the BA.2.86 RBD conformational ensemble obtained from AF2-MSA predictions. (A) Density distribution of the pLDDT structural model estimate of the prediction confidence. (B) Density distribution of TM-score measuring structural similarity of the AF2-MSA predicted RBD conformations with respect to the optimized AF2-generated reference state. (C) Density distribution of GDT_TS structural similarity metric between the AF2-predicted RBD conformations and the optimized AF2-generated reference state. (D) Density distribution of RMSD between the AF2-predicted RBD conformations and the optimized AF2-generated reference state.

Structural alignment of the BA.2 RBD-ACE2 crystal structure and the refined AF2-default model of the BA.2.86 RBD-ACE2 complex (Figure 4A) illustrated a considerable similarity of the RBD conformations showing minor displacements in the intrinsically flexible RBM tip loop (residues 475–487) and peripheral region (residues 520–527). These results are consistent with the cryo-EM and X-ray studies of Omicron S proteins and RBD-ACE2 complexes showing that VOCs and Omicron lineages typically induce small structural changes but may affect the dynamics and binding energetics with the host receptor. Despite a considerable number of BA.2.86 unique RBD mutations relative to BA.2, the AF2-generated structural model of the BA.2.86 RBD is remarkably similar to the parental BA.2 RBD-ACE2 crystal structure, with only minor fluctuations in the RBM tip region (Figure 4A). Structural mapping of the unique BA.2.86 RBD mutations on the AF2-predicted conformations showed that these substitutions can induce only minor structural perturbations of the RBD backbone, including positions of A484K and F486P residues in the RBM region (Figure 4A).

Figure 4.

Structural alignment of the BA.2 and BA.2.86 RBD conformations from complexes with the ACE2 receptor. (A) Structural alignment of the crystallographic BA.2 RBD conformation (in light pink ribbons) and the refined AF2-predicted conformation of the BA.2.86 RBD (in orange ribbons). The BA.2 mutational sites are shown in red spheres and unique BA.2.86 mutations (D339H, R403K, V445H, G446S, N450D, L452W, N460K, N481K, A484K, F486P, and R493Q). (B) Structural alignment of the RBD binding site residues making contacts with the ACE2 receptor. The structural positions of the BA.2 RBD binding site residues (in green sticks) are from the crystal structure of the BA.2 RBD-ACE2 complex. The structural arrangement of the BA.2.86 RBD binding sites residues are overlaid onto BA.2 RBD and are shown in magenta sticks. The RBD binding site residues for both BA.2 and BA.2.86 variants are annotated.

A more detailed inspection of the structural differences in the side chains of the RBD binding site residues depicted more appreciable variations of the positively charged side chains for N440K, N460K, N481K, and also H505 (Figure 4B). Importantly, more significant and functionally important ACE2 binding displacements were observed for side chains of specific BA.2.86 mutational sits R403K, V445H, N450D, L452W, A484K reversal R493Q, and F486P (Figure 4B). These results showed that AF2-generated predictions of the BA.2 and BA.286 RBD-ACE2 complexes can accurately reproduce the experimental structures and capture conformational details of the RBD fold and variant-specific functional adjustments of the RBD binding site residues.

To illustrate the performance of the AF2-MSA depth approach in characterizing conformational ensembles, we performed structural alignment of the predicted BA.2 RBD conformations with high pLDDT values and the crystal structure of the BA.2 RBD-ACE2 complex (Figure 5A). A high degree of similarity (RMSD < 1.0 Å) illustrated the ability of AF2-MSA models to accurately reproduce the experimental structure. Strikingly, we observed that selection of the BA.2 RBD models based on the high pLDDT scores can accurately predict conformations of the RBD flexible loops (444–452, 455–472) and capture moderate yet functionally relevant plasticity in the RBD loop 444–452, RBM tip (residues 475–487) that harbor important mutational sites and peripheral flexible region (residues 515–530).

Figure 5.

Structural alignment of the AF2-predicted BA.2 and BA.2.86 RBD conformational ensembles. (A) Structural alignment of the AF2-predicted BA.2 conformations with high pLDDT values and the crystal structure of the BA.2 RBD-ACE2 complex. The RBD conformations are shown in ribbons and the positions of BA.2 mutational sites D339, F371, P373, F375, A376, N405, S408, N417, N477, K478, A484, R493, Y501, and H505 are shown in red spheres. (B) Structural alignment of the AF2-predicted BA.2.86 RBD conformations with high pLDDT values. The sites of unique BA.2.86 mutations D339H, K356T, R403K, V445H, G446S, N450D, L452W, N460K, N481K, A484K, F486P, and R493Q are shown in blue spheres.

These regions are inherently flexible as revealed in hydrogen/deuterium-exchange mass spectrometry (HDX-MS) studies that informed of residue-specific changes in conformational dynamics induced by mutations and binding of SARS-CoV-2 S protein.^114–116 Our predictions showed that the position of functionally important for ACE2 binding and immune escape F486 residue in BA.2 RBD (F486P in XBB.1.5 and BA.2.86 variants) is preserved despite functional displacements of the RBM loop (Figure 5A). The predicted conformations of other RBD loops (355–375, 381–394, 455–471) are identical to the ones in the crystallographic conformation (Figure 5A). A similar structural alignment of the AF2-MSA depth ensemble was seen in the BA.2.86 RBD (Figure 5B), revealing an appreciable but more homogeneous displacements of the flexible RBM tip. Interestingly, the deletion of V483 in BA.2.86 caused minor conformational changes in the local mobility of the RBM loop (Figure 5B).

The important feature of the AF2-predicted conformations for BA.2.86 RBD is that the deletion of V483 does not disrupt the C488–C480 disulfide bond, and this critical stabilization signature is shared in both BA.2 and BA.2.86 structures (Supporting Information, Figure S6). Despite the V483 deletion in the center of the functional RBM loop, which represents a fairly notable change, and the presence of several other mutations N481K, A484K, and F486P, the AF2 ensemble unveiled a tolerant structural response of the RBM loop in BA.2.86 RBD (Figure 5B, Supporting Information, Figure S6A–C).

The predicted BA.2 and BA.2.86 conformations have precisely overlapped C480–C488 positions and contacts (Supporting Information, Figure S6C), where N481K and A484K mutational sites in the neighboring positions incur no strain on the position and stability of these disulfide bonds thus preserving binding of the BA.2.86 RBD with ACE2. This is consistent with deep mutational scanning (DMS) investigations^117–119 showing that delV483 only moderately affects RBD stability and ACE2 binding. Importantly, the predicted BA.2.86 RBD ensemble showed small functionally relevant variations in the flexible RBD loops 444–452 and 475–487 (Figure 5B) that harbored unique BA.2.86 mutations V445H, G446S, N450D, L452W, delV483, A484K, and F486P representing attractive hotspots of Omicron convergent mutations. These predictions suggested that BA.2.86 mutations would not exert large perturbing effects in the conformational ensemble (Figure 5B). Moreover, variations in the RBD loops 444–452 and 475–487 of the BA.2.86 RBD become more homogeneous than in the BA.2 RBD and represent small loop displacements around the dominant loop state without inducing elements of disorder and preserving ordered “hook-like” folded RBM tip required to minimize binding liabilities of F486P interactions with the ACE2 receptor (Figure 5). These observations echoed the results of MD simulations showing that the RBM loop has an inherent conformational flexibility where ACE2 and antibody binding to this region may elicit a specific distribution of conformations as compared to the unbound RBD form.¹²⁰

The AF2-MSA predictions of the BA.2.86 ensemble also indicated that the F486 side chain may experience some fluctuations which is consistent with the presence of many conserved mutations (F486V, F486I, F486S, F486P) seen in other variants. F486 is also one of the major hotspots for escaping neutralization by antibodies. According to the DMS experiments, among the most common F486 mutations (F486 V/I/S/L/A/P), F486P imposes the lowest cost in RBD affinity loss and has the largest increase in RBD expression.^120–122 The AF2 ensemble predictions may be relevant for understanding the BA.2.86 RBD dynamic responses to binding in which variant-specific mutations may increase the adaptability of the RBD loop and exploit the induced plasticity to boost immune evasion.

To conclude, the AF2 structural modeling produced robust and highly accurate atomistic models of BA.2.86 RBD and RBD-ACE2 complex. The key finding of this analysis is that AF2 models produced with the shallow MSA depth are accurate and can adequately characterize structural ensembles reproducing structural and biophysical experiments of the Omicron RBD-ACE2 complexes.^114–116 The results also convincingly demonstrated that pLDDT statistical assessment of the AF2 models can be reliably used to identify flexible RBD regions and predict major characteristics of conformational heterogeneity of these adaptable regions with atomistic accuracy. Incorporating MD simulations together with AF2-based structural modeling could provide more comprehensive insights by explicitly considering the dynamics and energetics of protein structures, allowing for a more detailed characterization of flexible regions and conformational heterogeneity. To address this, we combined AF2-based structural predictions of the top models for BA.2.86 RBD complexes with subsequent microsecond atomistic MD simulations performed to detail the stability and dynamics of the AF2-predicted conformational states.

Atomistic MD Simulations Reveal Variant-Specific Signatures of Conformational Dynamics and Binding Contacts in the BA.2 and BA.2.86 RBD-ACE2 ACE2 Complexes.

We performed comparative all-atom MD simulations of the BA.2 RBD-ACE2 and BA.2.86 RBD-ACE2 complexes. For the latter, the initial structures were taken as the AF2 predicted and refined conformation of the BA.86 RBD-ACE2 complex. Conformational dynamics profiles obtained from MD simulations were similar and revealed several important trends (Figure 6). The RMSF profiles showed local minima regions corresponding to the structured five-stranded antiparallel β-sheet core region that functions as a stable core scaffold (residues 350–360, 375–380, 394–403) and the interfacial RBD positions involved in the contacts with the ACE2 receptor (residues 490–505 of the binding interface) (Figure 6A). The conformational dynamics profiles revealed marginally greater stability for the BA.2 RBD as compared to BA.2.86 but the differences in thermal fluctuations are minor (Figure 6A). Consistent with AF2 predictions, the profiles displayed displacements in the flexible RBD regions (residues 355–375, 381–394, 444–452, 455–471, 475–487). Of notice are larger fluctuations in the BA.2.86 RBD for the flexible region 460–490, particularly the RBM loop featuring mutations N481K, delV483, A484K, and F486P (Figure 6A). Despite a moderately increased RBM mobility, the RBM tip is maintained in a stable folded conformation that can be described as a “hook-like” folded RBD tip and is similar to the crystallographic conformation of the BA.2 RBD-ACE2 complex. Our previous studies showed that well-ordered and stable “hook-like” conformation of the RBM tip is maintained in BA.2 and XBB.1.5 variants due to hydrophobic interactions provided by F486 (BA.2) and F486P (XBB.1.5).^70,71 The RMSF analysis of the ACE2 residues showed similar profiles across all the examined variants (Supporting Information, Figure S7). The key ACE2 binding motifs correspond to an α helix (residues 24–31) and a beta-sheet (residue 350–356) that display moderate RMSF values in both BA.2 and BA.2.86 complexes.

Figure 6.

Conformational dynamics profiles obtained MD simulations of the Omicron RBD BA.2 and BA.2.86 RBD complexes with ACE2. (A) RMSF profiles for the RBD residues obtained from MD simulations of the BA.2 RBD-ACE2 complex, pdb id 7XB0 (in green lines), and BA.2.86 RB-ACE2 complex (in red lines). (B) Distance fluctuation stability index profiles of the BA.2 RBD residues (in green lines) and BA.2.86 RBD residues (in red lines). The positions of Omicron mutational sites are highlighted in magenta-colored filled circles. The ensemble-average numbers of distinct ACE2 residues making stable intermolecular contacts (>70% occupancy) with the BA.2 RBD residues (C) and BA.2.86 RBD residues (D). Structural mapping of the conformational dynamics profiles for the BA.2 RBD-ACE2 complex (E) and BA.2.86 RBD-ACE2 complex (F). The profiles are mapped onto the crystal structure of the BA.2 RBD-ACE2 and AF2-predicted BA.2.86 RBD-ACE2 complex with the rigidity-flexibility sliding scale colored from blue (most rigid) to red (most flexible).

The conformational ensembles of the S-RBD complexes with ACE2 were subjected to distance fluctuation stability analysis based on the dynamic residue correlations (Figure 6B). A comparative analysis of the residue-based distance fluctuation stability indexes revealed several dominant and common peaks, reflecting the similarity of the topological and dynamical features of the RBD-ACE2 complexes. Despite a similar shape of the distributions for the BA.2 and BA.2.86 RBD variants, we noticed the reduced stability indexes for residues 445–455 in BA.2.86 which implies that this loop becomes somewhat more flexible in BA.2.86 owing to modifications V445H, G446S, N450D, and L452W (Figure 6B). Instinctively, AF2-predicted conformations similarly revealed the greater variability of this RBD loop, showing that AF2-generated ensemble can capture the dynamics signatures of the RBD-ACE2 complexes.

Using conformational ensembles of the Omicron RBD-ACE2 complexes, we also performed a statistical analysis of the intermolecular contacts that revealed several fundamental commonalities and differences in the interaction profiles for the Omicron RBD subvariants (Figure 6C,D). The RBD-ACE2 contacts with occupancy >70% in the MD trajectories were considered as long-lived stable interactions and were recorded for this analysis. The RBD residues can be classified into several groups based on their conservation and respective roles in binding with the host receptor. One of these groups includes identical residues such as Y453, N487, Y489, T500, and G502 and homologous positions (Y449/F/H, F456/L, Y473/F, F486/L, and Y505H), while other group includes more diverse residues undergoing various modifications such as G446/S/T, L455/S/Y, A475/P/S, G476/D, G496/S, K417/V/N/R/T, E484/K/P/Q/V/A, Q493/N/E/R/Y, Q498/Y/H/R, and N501/Y/T/D/S. The conserved residues from the first group make consistent and similar interactions in all Omicron variants, indicating that these RBD positions can function as molecular determinants of the RBD stability and binding affinity (Figure 6C,D). The distribution of contacts in the BA.2 RBD-ACE2 complex indicated that stable interactions are established across the entire binding interface, particularly revealing the formation of salt bridges by R403 with E37, R493 with E35, and D38 (Figure 6C). A considerable number of ACE2 residues engage in interactions with F486, Y489, R493, R498, T500, and Y501 positions. In particular, R498 interacts with D38, Y41, Q42 and L45 ACE2 residues; T500 forms stable interactions with Y41, L45, G326, Q330, K353, G354, D355 and R357; and Y501 is engaged in the favorable stable contacts with D38, Y41, K353, G354 and D355 (Figure 6C). The analysis showed that the interfacial salt bridge interactions in the BA.2 RBD formed by R493 with E37, E35, and D38 residues in ACE2 are partially lost in the BA.2.86 RBD-ACE2 complex. To compensate for this loss the mutated R493Q forms a number of stable interfacial contacts with D30, K31, N33, H34, and E35 including hydrogen-bonding interaction with the K31 side chain and the carboxyl group of E35 from ACE2 (Figure 6D).

To summarize, all-atom MD simulations of the BA.2 RBD-ACE2 and BA.2.86 RBD-ACE2 complexes initiated from the AF2-predicted structures provided a detailed characterization of subtle dynamic differences between the Omicron variants. The key finding of the MD comparative analysis is that AF2 predictions of structural ensembles closely match the conformational flexibility patterns revealed in atomistic simulations. In particular, consistent with AF2 predictions, the profiles displayed larger fluctuations in the BA.2.86 RBD for the flexible region 460–490, and the RBM loop harboring mutations N481K, delV483, A484K, and F486P (Figure 6B). MD simulations confirmed that residues 445–455 that are more flexible in BA.2.86 owing to modifications V445H, G446S, N450D, and L452W (Figure 6B). Consistent with these findings, the AF2-predicted structural ensemble pointed to the greater plasticity of this RBD loop, suggesting that AF2 ensemble predictions may be robustly exploited for dynamic refinement by MD simulations. MD simulations provided a more quantitative characterization of mutational effects in BA.2.86 showing that these mutations can partly reassemble the nuances of the RBD-ACE2 interaction network while preserving a strikingly convergent binding pattern at the protein interface.

Importantly, we found that the key dynamic signatures and dynamic differences between the RBD-ACE2 complexes for variants can be accurately described based on the AF2-MSA depth structural predictions of the conformational ensemble states. We argue that combining the accurate AF2 predictions of multiple conformational states with the detailed dynamic insights from MD simulations is particularly advantageous for studying intricate molecular systems such as S-RBD complexes. Combining AF2 predictions and MD simulations of multiple states also allows for better validation of predictions and assessment of their robustness in a dynamic context. As our predicted AF2-generated ensembles showed a remarkable consistency with the observed dynamic signatures in MD simulations, our results suggest that AF2 ensemble modeling adaptations employing varying MSA depth can be successfully used to identify key dynamic characteristics of different Omicron variants.

Mutational Scanning Heatmaps of the BA.2 and BA.2.86 RBD Residues Identify Common and Variant-Specific Structural Stability and Binding Affinity Hotspots.

Using conformational ensembles obtained from AF2 predictions and MD simulations we performed a systematic mutational scanning of the BA.2 and BA.2.86 RBD residues in the RBD-ACE2 complexes (Figure 7). In silico mutational scanning was done using the BeAtMuSiC approach by averaging the binding free energy changes over the equilibrium ensembles. The resulting mutational scanning heatmaps are reported for the RBD binding interface residues that make stable contacts with ACE2 in the course of simulations. To provide a systematic comparison, we constructed mutational heatmaps for the RBD interface residues of BA.2 RBD-ACE2 (Figure 7A) and BA.2.86 RBD-ACE2 complexes (Figure 7B). Consistent with DMS experiments of SARS-CoV-2 S VOC’s^117–119 the hydrophobic residues Y453, L455, F456, F486, Y489, and Y501 play a decisive role in stability and binding for both BA.2 (Figure 7A) and BA.2.86 complexes (Figure 7B). The large destabilization changes become more particularly pronounced for Y453, L455, and F456 while also revealing high sensitivity of F486, N487, Y489, R493, T500, Y501, and H505 residues (Figure 7). Mutational heatmaps clearly showed that all substitutions in these positions can incur a consistent and considerable loss in the stability and binding affinity with ACE2. In addition, mutational scanning of the RBD residues F486, N487, and H505 showed consistent destabilization changes, placing these residues as another important group of energetic centers (Figure 7). This analysis is consistent with previous studies, suggesting that conserved hydrophobic RBD residues may be universally important for binding across all Omicron variants and function as stabilizing sites of the RBD stability and binding affinity.^70,71

Figure 7.

Ensemble-based dynamic mutational profiling of the RBD intermolecular interfaces in the Omicron RBD-ACE2 complexes. The mutational scanning heatmaps are shown for the interfacial RBD residues in the BA.2 RBD-ACE2 complex (A) and BA.2.86 RBD-ACE2 complex (B). Structural mapping of the RBD binding epitopes of the BA.2 RBD-ACE2 complex (C) and BA.2.86 RBD-ACE2 complex (D). The RBD binding epitope is in green-colored surface. The Omicron RBD mutational sites are shown in red surface. The ACE2 binding residues are in pink sticks. The standard errors of the mean for binding free energy changes and are within ~0.05–0.12 kcal/mol using averages based on a total of 1000 samples obtained from the three MD trajectories for each system.

The mutational heatmap for the BA.2.86 RBD-ACE2 was quite similar showing the major hotspots for Y453, L455, F456, Y489, Y501, and H505 positions that are shared between BA.2 and BA.2.86 (Figure 7B). The direct binding interface is broader for BA.2.86 revealing small variations and tolerance to mutational changes in positions Y473, A475, G476, N477, K484, G485, and P486 (Figure 7B). Notably, unique BA.2.86 mutational positions N450D and L452W are distant from the immediate intermolecular interface and do not form direct contacts with the ACE2 receptor.

To quantify the stability and binding free energy changes for the sites of BA.2.86 mutations, we reported detailed mutational scanning results for these positions (Figure 8). The results showed that binding free energy changes induced by scanning in the BA.2.86 positions H339 (Figure 8A), T356 (Figure 8B), H445 (Figure 8D), S446 (Figure 8E), D450 (Figure 8F), W452 (Figure 8G), K460 (Figure 8H), and K484 (Figure 8J) are small and are mostly destabilizing with ΔΔG < 1.0 kcal/mol. Noticeably, mutational scanning of the K481 position resulted in minor stabilizing changes reflecting favorable stability effects as K481 is not involved in direct intermolecular contacts with the ACE2 receptor (Figure 8I). The initial analysis of the BA.2.86 antigenicity and binding suggested that the higher receptor binding affinity of the BA.2.86 might be partly attributed to the additional positive charges associated with V445H, N460K, N481K, and A484K sites.⁴⁵ However, our AF2-enabled analysis of the RBD-ACE2 contacts derived from simulations of BA.2.86 RBD-ACE2 complex indicated that mutational sites V445H, G446S, N450D, L452W, A484K are not involved in the formation of long-lived specific interactions. Mutational scanning confirmed that these positions displayed a considerable degree of tolerance to modifications and are unlikely to significantly contribute to the binding affinity. These results are consistent with the biophysical analysis of the BA.2.86 mutations suggesting that D339H, K356T, V445H, G446S, N450D, L452W, N460K, N481K, and A484K may have emerged primarily to strengthen immune evasion and mediate the enhanced resistance which implies that infectivity may be traded for higher immune evasion during long-term host-viral evolution.^45–47

Figure 8.

Ensemble-based mutational scanning of stability and binding for the individual BA.2.86 mutational sites in the BA.2.86 background. The profiles of computed binding free energy changes ΔΔG upon 19 single substitutions for the Omicron BA.2.86 mutational sites D339H (A), K356T (B), R403K (C) V445H (D), G446S (E), N450D (F), L452W (G), N460K (H), N481K (I), A484K (J), F486P (K), and R493Q (L). The respective binding free energy changes are shown in maroon-colored filled bars. The positive binding free energy values ΔΔG correspond to destabilizing changes and negative binding free energy changes are associated with stabilizing changes.

Of particular interest is the mutational scanning profile of R403K position where the reversed change K403R is highly unfavorable (ΔΔG = 1.2 kcal/mol), suggesting that R403K in BA.2.86 can lead to the increased binding affinity (Figure 8C). These results agree with the experimental data which examined mutations in the BA.2.86 RBD by generating a set of reverse mutations and demonstrated that K403R reversal significantly increased the equilibrium dissociation constant signaling the reduced binding affinity upon reversed mutation.⁵² Our mutational scanning results are consistent with these biochemical experiments suggesting that K403 mutation in BA.2.86 may contribute significantly to the binding affinity.

Of special significance is the pattern of energetic changes in BA.2.86 positions P486 (Figure 8K) and Q493 (Figure 8L) that correspond to the critical RBD positions involved in both binding and immune escape. We observed that modifications of P486 with the exception of P486F and P486V are destabilizing, confirming the notion that F486P mutation may have evolved to secure productive binding at a minimum loss while allowing sufficient room for immune escape from antibodies targeting this region.^117–119 F486 from the RBM loop is a critical site of convergent evolution that is exploited by BA.2, BA.2.75, BF.7, XBB.1, XBB.1.5, and BA.2.86 variants as one of the hotspots of compensatory binding and immune evasion. Indeed, F486 in BA.2 and BA.2.75 is substituted with F486V in BF.7 and F486S in XBB.1 and F486P in XBB.1.5 and BA.2.86 variants.^121–123

Our mutational scanning analysis is consistent with these studies, suggesting that mutations in these hotspots allow for delicate manipulation of conflicting fitness requirements to balance ACE2 binding affinity and immune evasion. Another critical site of convergent evolution is the reversed R493Q mutation in the BA.2.86 variant. Mutational scanning results showed that modifications at Q493 positions are destabilizing, particularly the back-reversed Q493R modification in the background of BA.2.86 (Figure 8L).

To summarize, mutational heatmaps showed that conserved hydrophobic RBD residues may be universally important for binding across all Omicron variants and function as hotspots of the RBD stability, while the binding affinity of the BA.2.86 RBD may be modulated by balance between R493Q and F486P mutations where R493Q compensates for partial binding loss incurred by F486P mutation which is beneficial for immune evasion. Our results showed that rapid in silico profiling of mutations for the SARS-CoV-2 S protein complexes can yield a fairly robust assessment of major variants and suggest a structural basis for changes in the protein stability and binding.

Forward and Reversed Mutational Scanning of BA.2.86 RBD Mutations Reveals Epistatic Couplings and Synergy between R403K and R493Q Hotspots in the BA.2.86 Background.

To further compare the impact of mutations on the binding affinity to ACE2 receptor, we computed the binding free energy changes that unique BA.2.86 mutations induce in the BA.2 RBD-ACE2 complex (Figure 9A) as well as the effect of the reversed mutations in the BA.2.86 RBD structure (Figure 9B). The specific objective of this comparison was to estimate the role of the BA.2 and BA.2.86 backgrounds on the mutational effects which may indicate synergistic epistatic effects of the BA.2.86 substitutions. It appeared that D339H, K356T, V445H, G446S, N450D, L452W, N481K, and A484K substitutions cause minor changes (ΔΔG < 0.5 kcal/mol) in the BA.2 structure, indicating that these substitutions play no significant role in modulating binding affinity (Figure 9A). Notably, A484K and R493Q mutations can induce more appreciable but still moderate stabilization changes that improve the binding affinity in the BA.2 RBD-ACE2 complex (Figure 9A). The assessment of binding free energies upon reversal mutations in the BA.2.86 RBD-ACE2 complex (the BA.2.86 background) not only reasserted the important contribution of K403 and Q493 to the ACE2 binding but also yielded significantly larger destabilization changes for reversed K403R (ΔΔG = 1.2 kcal/mol) and Q493R changes (ΔΔG = 1.6 kcal/mol) (Figure 9B). Hence, in the BA.2.86 background, K403 and Q493 positions are far more significant for the ACE2 binding. These results echo similar arguments proposed in the recent functional studies of binding and antigenicity of the BA.2.86 variant.⁵² The presented data on forward and reversed mutational scanning of the BA.2.86 mutations in the BA.2 and BA.2.86 backgrounds suggested a potential epistatic effect of R493Q in the context of the BA.2.86 variant.

Figure 9.

Computed binding free energy changes that unique BA.2.86 mutations induce in the BA.2 background in the BA.2 RBD-ACE2 complex (A) and the effect of the reversed to BA.2 mutations induce in the BA.2.86 background in the BA.2.86 RBD-ACE2 complex (B). The computed binding free energies are shown in magenta-colored filled bars. The positive binding free energy values ΔΔG correspond to destabilizing changes and negative binding free energy changes are associated with stabilizing changes.

According to our results, the main drivers of the binding free energy changes in BA.2.86 would result from balancing “dance” performed by hotspots of convergent evolution where a moderate ACE2 binding loss incurred by F486P mutation is compensated by the favorable binding effect of the R403K and R493Q mutations. Similar balancing effects were observed in the Omicron BA.4/BA.5 variant where the R493Q reversion in the BA.4/5 contributed to evading immunity and improvements in the ACE2 binding affinity.⁴¹ These findings lend support to the evolutionary mechanism according to which emerging RBD mutations in the BA.2.86 variant (particularly N440K, N460K, T478K, N481K, A484K, and F486P) may be primarily driven by the immune pressure-driven changes while the improved ACE2 binding affinity relative to BA.2 is mediated via R403K and R493Q mutations.

To summarize the results of forward and reversed mutational scanning of BA.2.86 RBD mutations in the BA.2 and BA.2.86 backgrounds suggested the presence of epistatic couplings between R403K and R493Q hotspots in the BA.2.86 variant. We argue that the BA.2.86 RBD binding affinity may be determined by a combined effect of R493Q and F486P mutations that represent key attractive hotspots of convergent evolution. These results are consistent with the notion that most of the BA.2.86 mutations may have primarily emerged to improve the evasion of acquired immunity by Omicron variants rather than affecting the ACE2 binding. The results suggest that the BA.2.86 variant may have evolved to accumulate convergent immune escape mutations while exploiting potential synergistic epistatic effects of a selected group of hotspots R403K and R493Q to enable sufficient ACE2-binding capability. Several lines of evidence indicated that the observed coordination of evolution at different sites can be due to epistatic, rather than random selection of mutations.^124,125

MM-GBSA Computations of the Binding Free Energies for the Omicron RBD BA.2 and BA.2.86 RBD-ACE2 Complexes.

Using the conformational equilibrium ensembles obtained from MD simulations, we computed the binding free energies for the Omicron RBD BA.2 and BA.2.86 RBD complexes with ACE2 using the MM-GBSA method. The total binding free energy changes showed a more favorable binding affinity for the BA.2.86 RBD (Supporting Information, Table S1 and Figure S8). The breakdown of the MM-GBSA binding energies showed that the electrostatic contribution is more favorable for the BA.2.86 RBD-ACE2 complex but is offset by the unfavorable solvation (Table S1). This reflected the presence of BA.2.86 mutational sites N460K, N481K, and A484K that increased the electrostatic interactions but the net effect of these mutational sites to binding is minor due to the balance of electrostatic and solvation contributions (Table S1). The experimentally measured ACE2 binding affinities for two versions of the BA.2.86 S proteins with K_D values of 0.54 and 0.60 nM, while compared to the K_D value of the BA.2 S protein of 1.68 nM.⁵¹ As expected, the introduction of R403K, N460K, N481K, and A484K mutations enhanced the contribution of electrostatic interactions (Supporting Information, Figure S8) but for N460K, N481K, and A484K, these favorable interactions are largely offset by unfavorable solvation contributions resulting in the net moderately destabilizing interactions. Of particular interest is the analysis of R403K, R493Q, and F486P mutational effects in the BA.2.86 RBD-ACE2 complex (Supporting Information, Figure S8D,E). We found that the binding energy contribution for R403K is more favorable (ΔG = −3.46 kcal/mol) while in the BA.2 RBD-ACE2 complex this contribution ΔG = −1.03 kcal/mol which is due to a more favorable balance of electrostatics and solvation energies in the BA.2.86 complex (Supporting Information, Figure S8D,E). Importantly, MM-GBSA analysis showed that the R493Q mutation in BA.2.86 yielded ΔG = −3.86 kcal/mol while the respective contribution of R493 in BA.2 accounted for ΔG = −2.03 kcal/mol thus confirming a stronger favorable effect of reversal R493Q on binding affinity in the BA.2.86 variant (Supporting Information, Figure S8D,E). It appeared that R493 in BA.2 can enhance the electrostatic interactions, but this contribution is offset entirely by the unfavorable solvation, while R493Q in BA.2.86 produced a more favorably balanced total effect of favorable van der Waals and electrostatic interactions. Notably, the MM-GBSA decomposition analysis also confirmed a highly favorable contribution of F486 in BA.2 to the binding energy (ΔG = −4.87 kcal/mol with a dominant van der Waals component of −5.24 kcal/mol), while F486P in BA.2.86 yielded a significantly reduced binding contribution (ΔG = −1.66 kcal/mol) primarily due to a considerable loss of the packing van der Waals contacts (Supporting Information, Figure S8D,E).

To sum up the main findings, the MM-GBSA analysis provided a useful insight into the mechanism underlying stronger binding of the BA.2.86 variant relative to BA.2, suggesting that differences in the binding affinity are determined by cumulatively better balance of electrostatic and solvation contributions for several binding affinity hotspots including R403K, F486P, and R493Q, while universally important for binding Y489, R498, and Y501 maintain their favorable interactions in both complexes.

Electrostatic Potentials for the Omicron RBD BA.2 and BA.2.86 RBD-ACE2 Complexes.

We also computed and compared the electrostatic potential for the BA.2 and BA.2.86 RBD-ACE2 complexes (Figure 10). The electrostatic potentials on the RBD surfaces for the Omicron BA.2 (Figure 10A,B) and BA.2.86 (Figure 10C,D) complexes are positive with variable charge distributions, showing relatively moderate changes in the overall surface distribution between subvariants. However, we noticed a stronger positively charged potential in the Omicron RBD BA.2.86 (Figure 10C,D) showing positive densities in the RBM-interacting regions and distal from binding RBD regions. Hence, computations of the electrostatic potential surfaces confirmed that the BA.2.86 RBD featured a stronger positive electrostatic potential due to substitutions to basic residues N440K, N460K, T478K, N481K, A484K, and Q498R.

Figure 10.

Distribution of the electrostatic potentials calculated on the molecular surface of the BA.2 RBD-ACE2 complex (A,B) and BA.2.86 RBD-ACE2 (C,D). The crystal structure of the BA.2 RBD-ACE2 complex and the best AF2-predicted model of the BA.2.86 RBD-ACE2 complex are used in computations of the electrostatic potentials. The color scale of the electrostatic potential surface is in units of kT/e at T = 37 °C. Electropositively and electronegatively charged areas are colored in blue and red, respectively. Neutral residues are shown in white.

To facilitate a comparison, we also compared the electrostatic potential using the experimental structures of BA.1, BA.2, BA.2.75 BA.3, BA.4/BA.5 and XBB.1.5 RBD complexes with ACE2 (Supporting Information, Figure S9) revealing that BA.2.86 displayed the strongest positive potential in the binding interface and RBM regions. These results are consistent with recent studies showing that the evolution of the electrostatic surface in Omicron variants shows a gradual accumulation of positive surface charges in the RBD over time and that related clades have similar electrostatic surfaces.^126–128 It was also noted that many antibodies have positively charged S RBD-recognition surfaces.¹²⁷ In summary, based on the mutational scanning data and electrostatic potential analysis, we suggest that the evolution of the electrostatic RBD surface in the latest Omicron variants, including further strengthening of the positive electrostatic potential on the BA.2.86 RBD, may have been primarily exploited to evade antibodies without compromising binding affinity with ACE2. We argue that the BA.2.86 lineage may have evolved to outcompete other Omicron subvariants by boosting immune suppression while balancing binding affinity with ACE2 via through a compensatory effect of R493Q and F486P mutations.

Mutational Profiling of Protein Binding Interfaces with Distinct Classes of Antibodies: Quantifying Functional Role of BA.2.86 Mutations in Immune Evasion.

In addition to an in-depth analysis of the BA.2.86 RBD binding to ACE2, we also examined immune evasion properties of BA.2.86 by performing structure-based mutational profiling of the S protein binding with distinct classes of RBD-targeted antibodies that are experimentally shown to display the reduced neutralization against the BA.2.86 variant. Using this approach, we quantify specific roles of the BA.2.86 mutations in developing broad resistance against different classes of RBD antibodies and eliciting robust immune escape.

We specifically examined a panel of monoclonal antibodies that retain activity against BA.2. XBB.1.5 and EG.5.1 but displayed a markedly reduced or completely abolished neutralization potential against the BA.2.86 variant.⁴⁵ Mutational profiling analysis allowed for direct comparison with the latest experiments that reported fold changes in IC₅₀ of antibody binding with BA.2.86 relative to the BA.2 variant⁴⁵ and enabled to clarify the specific role of unique BA.2.86 mutations in mediating antibody resistance (Figures 11 and 12).

Figure 11.

Structure-based mutational profiling of the S-RBD complexes with class 1 of RBD antibodies. The mutational screening evaluates binding energy changes induced by BA.2.86 mutations in the RBD-antibody complexes. Mutational profiling of the S-RBD complex with S2K146 (A), S-RBD Omicron complex with Omi-3 (B), S-RBD Omicron complex with Omi-18 (C), and S-RBD in complex with BD-515 (D). The binding free energy changes are shown in magenta-colored filled bars. The positive binding free energy values ΔΔG correspond to destabilizing changes and negative binding free energy changes are associated with stabilizing changes. The 3D structures of the RBD-antibody complexes are shown for RBD-S2K146 (E), RBD-Omi-3 (F), RBD-Omi-18 (G), and RBD- BD-515 complexes (H). The RBD is shown in pink-colored surface representation and the antibodies are shown in ribbons (heavy chain in magenta and light chain in green-colored ribbons).

Figure 12.

Structure-based mutational profiling of the S-RBD complexes with class 2 of RBD antibodies. The mutational screening evaluates binding energy changes induced by BA.2.86 mutations in the corresponding RBD positions of the RBD-antibody complexes. Mutational profiling of the S-RBD complex with COV2-2196 (A), S-RBD Omicron complex with XGV347 (B), S-RBD Omicron complex with XGV051 (C), and S-RBD in complex with ZCB11 (D). The respective binding free energy changes are shown in magenta-colored filled bars. The 3D structures are shown for the S-RBD complex with COV2-2196 (E), S-RBD Omicron complex with XGV347 (F), S-RBD Omicron complex with XGV051 (G), and S-RBD in complex with ZCB11 (H). The RBD is shown in pink-colored surface representation and the antibodies are shown in ribbons (heavy chain in magenta and light chain in green-colored ribbons).

The examined structures for class1 RBD antibodies included S-RBD complex with S2K146 (pdb id 7TAS),¹²⁹ S-RBD Omicron complex with Omi-3 (pdb id 7ZF3),¹³⁰ S-RBD Omicron complex with Omi-18 (pdb id 7ZFB),¹³⁰ S beta trimer complex with Omi-42 (pdb id 7ZR7)¹³⁰ and S-RBD in complex with BD-515 (pdb id 7E88).¹³¹ In the analysis of mutational scanning, we specifically focused on the S-antibody binding energy changes induced by BA.2.86 mutations in the corresponding RBD positions (Figure 11). The binding free energy changes associated with BA.2.86 mutations in the complex with S2K146 (Figure 11A) showed an appreciable loss of binding upon F486P and Q493R mutations which implied that F486P in the BA.2.86 is highly deleterious for S2K146 binding while the reversed R493Q in BA.2.86 is favorable for the antibody binding. These observations agree with the experimental screening showing that F486P caused dramatic fold change (>254) in IC₅₀ of antibody binding with BA.2.86 relative to the BA.2 variant.⁴⁵ At the same time, according to these experiments, R493Q mutation may lead to an ~11-fold favorable fold change in the antibody binding. For the Omi-3 antibody, the two experimental hotspots that cause significant loss of binding with BA.2.86 are N460K and F486P mutations⁴⁵ which are accurately captured in the mutational scanning revealing destabilizing ΔΔG = 0.62 kcal/mol for N460K mutation, and ΔΔG = 0.81 kcal/mol for F486P (Figure 11B). Our results also reproduced the experimentally observed R403K-induced favorable change in binding and moderate loss in the Omi-18 affinity caused by R493Q reversal⁴⁵ (Figure 11B). A similar pattern was seen in mutational scanning of the Omicron complex with Omi-18 antibody, revealing the same two energetic hotspots N460K and F486P that cause a significant reduction in the binding affinity (Figure 11C). In this case, both R403K and R493Q mutations appeared to be favorable for the antibody binding but the net effect may be offset by loss in affinity caused by N460K and F486P mutations. In agreement with the biochemical studies⁴⁵ mutational profiling of the S-RBD with another class I RBD antibody BD-515 unveiled a single binding hotspot N460K that can induce a significant loss in binding, while both R403K and R493Q changes can be marginally favorable for the antibody binding (Figure 11D). These results showed that mutations N460K and F486P, also shared by XBB.1.5 and EG.5.1 variants, can mediate resistance to some RBD class 1 antibodies by disrupting key hydrogen bonding between the RBD and antibodies in the case of N460K and reducing the hydrophobic packing with the ACE2-mimicking antibodies. Interestingly, R403K and R493Q that are important drivers of the BA.2.86 binding with the ACE2 are also favorable for antibody binding (Figure 11).

We also examined in detail the class 2 RBD antibodies including the S-RBD complex with COV2-2196 (tixagevimab) (pdb id 8D8Q),¹³² Omicron S-RBD with XGV347 (pdb id 7WED),¹³³ Omicron S-RBD with XGV051 (pdb id 7WTG),¹³⁴ and Omicron S-RBD complex with ZCB11 (pdb id 7XH8).¹³⁵ Consistently, mutational scanning demonstrated that F486 is the key binding hotspot as F486P modification resulted in significant binding affinity loss for this class of antibodies (Figure 12). These findings are in excellent agreement with the biochemical experiments revealing dramatic losses in antibody binding relative to BA.2 (>81 for COV2-2196, >7505 for XGV347, ~298 for XGV051, and >2293 for ZCB11) that are induced by F486P mutation.⁴⁵ Consistent with structural studies of the COV2-2196 antibody,¹³² mutational profiling determined F486P mutation as the dominant hotspot of antibody resistance (Figure 12A). Structural studies also indicated that antibodies XGV347 and XGV051 form extensive hydrophobic interactions mediated by Y453, L455, A475, A484, F486, and Y489 RBD positions.^133,134 For both XGV347 (Figure 12B) and XGV051 (Figure 12C) mutational profiling revealed that BA.2.86 mutation F486P induced a large loss in the antibody binding affinity (ΔΔG > 3.0 kcal/mol) and confirmed a critical role of this convergent evolutionary hotspot in immune escape.

ZCB11 binding is determined by hydrogen bonds, including those formed between N477, K478, N487, and N460 while the second interface between ZCB11 and RBM is stabilized by the hydrophobic interaction mediated by L455, F456, and F486 (Figure 12D).¹³⁵ Here again, the mode of antibody binding is severely affected by F486P mutation, leading to the large destabilizing change of ΔΔG > 2.5 kcal/mol which may explain the experimentally observed complete abrogation of ZCB11 binding to the BA.2.86 variant.⁴⁵ These results supported the notion that F486P mutation in BA.2.86 has a primary fitness effect on the immune escape rather than ACE2 binding.

Mutational profiling was also performed for class 3 RBD antibodies, including the S-RBD complex with antibody A19-46.1 (pdb id 7TC9),¹³⁶ Omicron RBD with S309 (sotrovimab) (pdb id 7YAD),¹³⁷ S-RBD with COV2-2130 (cilgavimab) (pdb id 8D8Q),¹³¹ S-RBD with LY-CoV1404 (bebtelovimab) (pdb I 7MMO),¹³⁸ and S-RBD beta variant complex with Beta-50 and Beta-54 (pdb id 7Q0H).¹³⁹ The results of mutational scanning for this class of antibodies revealed a different group of energetic hotspots where BA.2.86 mutations result in a substantial loss of binding affinity (Supporting Information, Figure S10). For the A19-46.1 antibody, the large destabilizing free energy changes were observed for N450D, L452W, and A484K mutations (Supporting Information, Figure S10A) which is consistent with the experiments showing substantial detrimental losses in IC50 binding of ~5-fold change for N450D, >220 fold for L452W and ~17 fold for A484K mutations.⁴⁵ Structural mapping of the A19-46.1 binding to the RBD illustrated a different mode of interactions for class 3 RBD antibodies (Supporting Information, Figure S11A).

The analysis of S309 binding with Omicron RBD pointed to an appreciable loss of binding due to the K356T mutation (Supporting Information, Figure S10B). These observations reproduced the experimental effects of BA.2.86 mutations on S309 binding in the BA.2 background which demonstrated that K356T is a single dominant hotspot causing almost 20-fold loss in binding.⁴⁵ Structural analysis of RBD-S309 complex (Supporting Information, Figure S11B) showed that S309 forms hydrogen bonds with the RBD residues N334, E340, N343, T345, R346, and K356T of BA.2.86 can singlehandedly cause antibody resistance.¹³⁷ Similarly, a single BA.2.86 mutational change N450D resulted in considerable ΔΔG = 1.5 kcal/mol binding loss while other BA.2.86 substitutions caused only moderate changes (Supporting Information, Figure S10C). These results corroborate with the neutralization data showing >240 fold change in IC50 of BA.2 RBD binding upon a single N450D mutation.⁴⁵ The mutational profiles for S-RBD with LY-CoV1404 (Supporting Information, Figure S10D) and S-RBD beta variant complex with Beta-54 antibodies (Supporting Information, Figure S10E) were similar disclosing the key role of BA.2.86 mutation V445H. This mutational change resulted in a dramatic binding loss of ΔΔG = 2.5 kcal/mol for LY-CoV1404 (Supporting Information, Figure S10D) and ΔΔG = 2.5 kcal/mol for Beta-54 antibody (Supporting Information, Figure S10E).

Structural analysis of the S-RBD complexes with LY-CoV1404 (Supporting Information, Figure S11D) and Beta-54 (Supporting Information, Figure S11E) highlighted a remarkably similar binding mode where both antibodies recognize an RBD region that only barely overlaps with the ACE2 binding interface, particularly at positions V445H and G446S. Strikingly, our analysis suggested that a single BA.2.86 mutation V445H could elicit resistance to the entire class 3 of RBD antibodies. These findings are in accordance with the experimental observations⁵¹ that highlighted >1000-fold loss in LY-CoV1404 binding upon V445H mutation.

To summarize this analysis, structure-based mutational screening of the RBD-antibody interfaces quantified the specific function of individual BA.2.86 mutations in eliciting resistance against different classes of RBD-targeted antibodies. The results of mutational scanning of the RBD-antibodies complexes targeting class 1, class 2, and class 3 epitopes demonstrated how different BA.2.86 mutations, particularly N450D, L452W, N460K, N481K, and F486P, can incur substantial binding free energy losses and lead to immune evasion. These results are in good agreement with the antigenicity characterization of BA.2.86 which demonstrated that N450D, K356T, L452W, A484K, and V445H are responsible for the enhanced immune evasion of BA.2.86 as compared to another highly evasive XBB.1.5 variant.⁴⁵ Our findings also confirmed that both R403K and R493Q modifications can play an important role in both receptor affinity and modulation of antibody binding but have a stronger effect on the BA.2.86 RBD-ACE2 binding. This analysis revealed the molecular basis of compensatory functional effects of the binding hotspots, showing that most BA.2.86 mutations may have primarily evolved to improve immune escape, while binding affinity with ACE2 is modulated cooperative effect of R403K, F486P, and R493Q mutations.

DISCUSSION

The latest biophysical studies proposed possible scenarios underlying the activity and binding of the BA.2.86 S protein by analyzing antigenic cartography and binding data on BA.2.86 neutralization by distinct classes of antibodies.^45–47 These experimental studies suggested that the increased viral fitness of BA.2.86 through the acquisition of a wide range of unique RBD mutations may be largely due to immune pressure that promotes convergent evolution. There is however lack of molecular details on the structure, dynamics, and binding energetics of the BA.2.86 RBD binding with the ACE2 receptor and antibodies to rationalize the experimental data and provide the atomistic basis for the proposed molecular mechanisms. To develop accurate and robust atomistic models of BA.2.86 RBD structure and conformational ensembles, we performed comparative structural prediction of the BA.2 and BA.2.86 RBD-ACE2 complexes using the AF2 approach with shallow MSA depth. Despite a significant accumulation of unique BA.2.86 mutations in the ACE2 binding interface, we observed a relatively moderate structural response of the BA.2.86 RBD that preserves the RBD fold and topology of the ACE2 binding interface.

Our results showed that AF2 models with the shallow MSA depth are robust in accurately capturing the experimental structure and that pLDDT statistical assessment of the AF2 models can be used to identify flexible RBD regions and estimate the extent of conformational heterogeneity in these adaptable regions. It is important however to recognize that while AF2-based pLDDT metrics provide valuable measures for assessing the accuracy of the predicted protein structures, they cannot fully capture the physics-based accurate characterization of protein dynamic changes revealed by MD simulations. Our results confirmed this notion showing that AF2 predictions need to be considered in combination with MD simulations of the predicted multiple conformations to produce the Boltzmann-weighted ensemble that can characterize the relative energetics and probabilities of different conformational states.

By combining AF2-based predictions and atomistic MD simulations, we characterized the conformational ensembles of the RBD-ACE2 complexes used for systematic mutational scanning of the RBD-ACE2 interfaces and detection of variant-specific binding hotspots. The central finding of this mutational analysis is that in the BA.2.86 background, K403 and Q493 positions are far more significant for the ACE2 binding, revealing a potential epistatic effect of R403K and R493Q mutations in the BA.2.86 RBD-ACE2 complex. The results of our study argue that the emergence of BA.2.86 sublineage may have been particularly driven by expanding the scope of mutations to boost immune resistance while employing a focused group of Omicron mutational hotspots (including R403K, R493Q, and F486P) to modulate the ACE2 binding affinity. Recent studies indicated that ACE2 binding can often be synergistically modulated and amplified via epistatic interactions of physically proximal binding hotspots, including Y501, R498, L455, and F456 residues.^140,141

Our study explored complementarities and potential synergies between AF2-enabled structural predictions of protein conformational states and MD simulations showing that integrating information from both methods and combining the strengths of AF2 deep learning predictions and atomistic simulations may produce a comprehensive and accurate characterization of the effects of Omicron variants on structure, dynamics, and binding with ACE2 and antibodies. In this context, it is noteworthy to stress that AF2 predictions are based on deep learning of the experimentally determined structures, and even though this approach can provide highly accurate static structures and top-ranked conformations, the AF2 predictions are not sampled according to the Boltzmann thermodynamic probability distribution. Therefore, the ensemble of structures generated obtained from AF2 predictions does not inherently represent the thermodynamic equilibrium ensemble of conformations weighted by their thermodynamic likelihood. In summary, our results support the notion that combining AF2 predictions with MD simulations can enable robust characterization of structure and dynamic mechanisms by marrying the accuracy of structural predictions with the details of equilibrium ensembles resulting in a more realistic and comprehensive representation of molecular events.

CONCLUSIONS

In this study, we have provided the first comprehensive computational analysis of the BA.2.86 S-RBD protein structure, dynamics, and energetics from the perspective of immune evasion and ACE2 binding affinity properties. We combined AF2 structural modeling, atomistic MD simulations, and systematic binding energetics analysis of the BA.2.86 S protein with ACE2 receptor and a panel of distinct classes of RBD antibodies to quantify molecular mechanisms underlying BA.2.86 binding and immune evasion. Several AF2-based structural modeling pipelines were adapted to predict the structure and functionally relevant conformational ensemble of the BA.2.86 RBD-ACE2 complexes. AF2-based structural modeling with varied MSA depth produced robust conformational predictions of the BA.2.86 RBD-ACE2 complex revealing important functional variations are localized in two RBD loops (resides 444–452 and 475–481) that harbor important mutational sites. Atomistic MD simulations of the BA.2 and BA.2.86 RBD-ACE2 complexes revealed important common and variant-specific dynamic signatures exemplified by modulation of flexibility of functional RBD loops. Importantly, we found that the AF2-predicted ensemble of functional conformations using the pLDDT metric can accurately capture the main dynamics signatures of the RBD-ACE2 complexes obtained from microsecond MD simulations. Using the conformational ensembles obtained from AF2-based predictions and MD simulations we performed a systematic mutational scanning of the BA.2 and BA.2.86 RBD residues in the RBD-ACE2 complexes revealing a group of conserved hydrophobic hotspots and critical variant-specific contributions of R403K, F486P and R493Q mutations. Our data suggested that the BA.2.86 lineage may have evolved to outcompete other Omicron subvariants by improving immune suppression while balancing binding affinity with ACE2 via through a compensatory effect of R493Q and F486P mutations.

The results of mutational scanning of the RBD-antibodies complexes targeting three classes of RBD epitopes demonstrated how different BA.2.86 mutations, particularly N450D, L452W, N460K, N481K, and F486P, can incur substantial binding free energy losses and lead to immune evasion. The study supports a hypothesis that the impact of the increased ACE2 binding affinity on viral fitness is more universal and is mediated through cross-talk between convergent mutational hotspots, whereas the effect of immune evasion could be more variant-dependent and modulated through the recruitment of mutational sites in various adaptable RBD regions. This mechanism is supported by the evidence that high-frequency mutations in the Omicron lineages are more likely to confer a stronger resistance to the potent antibodies against the variant than to materially improve ACE2 binding affinity.¹²⁴ Our study also provided support to a general mechanism in which the acquisition of functionally balanced substitutions to optimize multiple fitness trade-offs between immune evasion, ACE2 affinity, and sufficient conformational adaptability may potentially be a common strategy of SARS-CoV-2 evolution executed by the Omicron variants.

Data Availability Statement

Data are fully contained within the article and Supporting Information. Crystal structures were obtained and downloaded from the Protein Data Bank (http://www.rcsb.org). All simulations were performed using the all-atom additive CHARMM36 M protein force field that can be obtained from http://mackerell.umaryland.edu/charmm_ff.shtml. The rendering of protein structures was done with the UCSF ChimeraX package (https://www.rbvi.ucsf.edu/chimerax/) and Pymol (https://pymol.org/2). The software tools used in this study are freely available at GitHub sites: https://github.com/deepmind/alphafold; https://github.com/sokrypton/ColabFold/; https://github.com/nextstrain; https://github.com/Amber-MD/cpptraj; https://github.com/smu-tao-group/protein-VAE. All the data obtained in this work (including simulation trajectories, topology and parameter files, the software tools, and the in-house scripts are freely available at the ZENODO Web site https://zenodo.org/records/10140418.