Study: Optimizing variant-specific therapeutic SARS-CoV-2 decoys using deep-learning-guided molecular dynamics simulations. Image Credit: CROCOTHERY/Shutterstock

In a recent study published in Scientific Reports, researchers developed a computational workflow based on molecular dynamic (MD) simulations and artificial neural network (ANN) to assess the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) spike (S) protein receptor-binding domain (RBD)-human angiotensin-converting enzyme 2 (hACE2) binding affinities of SARS-CoV-2 variants.

Study: Optimizing variant-specific therapeutic SARS-CoV-2 decoys using deep-learning-guided molecular dynamics simulations. Image Credit: CROCOTHERY/Shutterstock


Studies have reported that S-hACE2 binding interactions facilitate SARS-CoV-2 entry and subsequent replication in the host. Thus, coronavirus disease 2019 (COVID-19) may be prevented by S-ACE2 binding inhibition.

Accordingly, human soluble ACE2 (hsACE2) that binds to SARS-CoV-2 virions before SARS-CoV-2 entry may prevent COVID-19; however, the approach requires optimization and adaptation to novel SARS-CoV-2 variants.

About the study

In the present study, researchers devised a workflow by combining regular methods with point-cloud-based technology to optimize SARS-CoV-2 variant-specific therapeutic decoy development.

MD simulations were performed to identify human angiotensin-converting enzyme 2 amino acid substitutions that reinforce S RBD-hACE2 interactions, for which an ESF (empirical scoring function) in close relation with the LIE (linear interaction energy) technique was used. In vitro SARS-CoV-2-neutralization assays were performed to assess the inhibition of the SARS-CoV-2 wild-type strain and the Beta variant transmission by hACE2 variants that were in linkage with the fragment crystallizable (Fc) region of human immunoglobulin G1 (hACE2-Fc).

A few variants of hACE2-Fc were also expressed in the Nicotiana benthamiana plant for investigating mass-scale production feasibility. Molecular dynamics run data were combined with hACE2 halos and S RBD halos for ANN (artificial neural network) training. The model was used to estimate binding affinities of SARS-CoV-2 S with hACE2 variants based on the S RBD and hACE2 halos. If a new variant emerged, hACE2 variants could be screened rapidly by the artificial neural network and verified by MD simulations so that COVID-19 treatment strategies could be tailored based on the human soluble ACE2 variant having the greatest binding affinity to the novel SARS-CoV-2 strain.

The potential of the system to estimate the effects of S RBD variant mutations for the same hACE2 decoys was assessed using the SARS-CoV-2 Omicron variant’s BA.1 and BA.2 subvariants as examples. All probable hACE2 mutations were screened, and the 300 most promising estimations were validated by MD simulations. In addition to wild-type hACE2, promising variants pf hACE2, with a C-terminal human IgG Fc tag, were expressed in Chinese hamster-ovary (CHO) cells.

SARS-CoV-2 RNA was quantified by quantitative reverse transcription-polymerase chain reaction (RT-qPCR) and immunohistochemistry (IHC) analysis. The SARS-CoV-2 neutralization potential of hACE2 variants expressed in Nicotiana benthamiana plant leaves (hACE2-Fc K31W_NB) was tested using enzyme-linked immunosorbent assays (ELISA). In-silico analyses were performed to evaluate the binding affinities of hACE2 variants with Omicron BA.3, BA.4/5, and Omicron BA.2.75 RBD proteins.

The crystal structure of wild-type SARS-CoV-2 S RBD bound to hACE2 was downloaded from the protein databank (PDB) database. The model-estimated ΔG value was computed based on electrostatic and van der Waals forces. Sequences used for ANN training comprised S RBD sequences (n=1,165) and hACE2 sequences (n=95) retrieved from visual examination, literature search, or the global initiative on sharing all influenza data (GISAID) database by 4 January 2022.


The hACE2- Fc K31W, hACE2 T27Y_L79T_N330Y_K31W, and hACE2 T27Y_ L79T_K31W hACE2 variants were identified as high-binding affinity candidates. Candidates produced in N. benthamiana showed 5.0-fold lower and 6.0-fold lower IC50 (half-maximal inhibitory concentration) values in comparison to the same variant produced in CHO cells and wild-type hACE2-Fc, respectively. The findings indicated that hACE2-Fc variants with correct folding could be produced in N. benthamiana and plant-produced soluble ACE2 variants represent a promising, cost-effective therapeutic option against SARS-CoV-2.

The ESF estimations were validated in vitro by virus neutralization assays. Experimental data correlated well with estimated ΔGpred (Gibbs free energies) in the model. In comparison to the wild-type of hACE2, the majority of hACE2 variants showed enhanced binding affinities to the SARS-CoV-2 Beta variant, Delta variant, and Omicron’s BA.1 subvariant and BA.2 subvariant. The hACE2-K31W was the only mutant with very less Gibbs free energy, indicating that the K31W mutation may contribute to S RBD interactions. K31W mutation presence was observed in most high-binding affinity mutants.

Variants with 3.0 to 5.0 mutations showed the greatest S RBD binding. The hACE2 T27Y_L79T_K31W and hACE2 T27Y_L79T_N330Y_ K31W showed remarkably high binding affinities for BA.2 S RBD (ΔGpred value −71.0 kJ/mol) in comparison to that for wild-type of hACE2 (−52.0 kJ/mol). With estimated binding affinities of −62.0 and −67.0 kJ/mol, the hACE2 T27Y_L79T_K31W and hACE2 T27Y_L79T_N330Y_K31W variants were the topmost high-affinity variants for BA.3, and the binding affinities for Omicron BA.4/5 and Omicron BA.2.75 were lower. The highest outliers (MD ΔG values of <- 70 kJ/mol) were mapped by the model, to the highest binding affinity value observed.

The findings indicated that ANN was not only able to better estimate values closer to the bulk of the binding affinity distribution than extrapolating from closely related variants but also reliably mapped the high-affinity variants to the highest affinity bracket of −68.0 kJ/mol. The artificial neural network could learn meaningful physical insights from Halos with performance significantly better than simply learning a regression-to-the-mean or copy-function, and the model could combine learned insights from relatively different inputs (distant SARS-CoV-2 sequences). The model identified single mutants comparable to the best hACE2 mutant found in the initial MD runs.

Overall, the study findings highlighted a bioinformatics approach of combining MD simulations, in vitro competitive inhibition assays, live-virus infection assays, and ANN, for rapid, cost-effective and efficient evaluation of the binding affinity of hACE2 decoys to novel SARS-CoV-2 strains at an initial stage, reducing durations of hACE2- decoy adaptation and sample requirements for in vitro selections.

Leave a Reply

Your email address will not be published.