#COVIDisAirborne: AI-enabled multiscale computational microscopy of delta SARS-CoV-2 in a respiratory aerosol

Abstract

We seek to completely revise current models of airborne transmission of respiratory viruses by providing never-before-seen atomic-level views of the SARS-CoV-2 virus within a respiratory aerosol. Our work dramatically extends the capabilities of multiscale computational microscopy to address the significant gaps that exist in current experimental methods, which are limited in their ability to interrogate aerosols at the atomic/molecular level and thus obscure our understanding of airborne transmission. We demonstrate how our integrated data-driven platform provides a new way of exploring the composition, structure, and dynamics of aerosols and aerosolized viruses, while driving simulation method development along several important axes. We present a series of initial scientific discoveries for the SARS-CoV-2 Delta variant, noting that the full scientific impact of this work has yet to be realized.

Keywords: AI; COVID-19; Delta; GPU; HPC; SARS-CoV-2; aerosols; computational virology; deep learning; molecular dynamics; multiscale simulation; weighted ensemble.

Figure 1.

Overall schematic depicting the construction and multiscale simulations of Delta SARS-CoV-2 in a respiratory aerosol. (N.B.: The size of di-valent cations has been increased for visibility.)

Figure 2.

Individual protein components of the SARS-CoV-2 Delta virion. The spike is shown with the surface in cyan and with Delta’s mutated residues and deletion sites highlighted in pink and yellow, respectively. Glycans attached to the spike are shown in blue. The E protein is shown in yellow and the M-protein is shown in silver and white. Visualized with VMD.

Figure 3.

Image of RAV with relative mass ratios of RA molecular components represented in the colorbar. Water content is dependent on the relative humidity of the environment and is thus omitted from the molecular ratios.

Figure 4.

SMA system captured with multiscale modeling from classical MD to AI-enabled quantum mechanics. For all panels: S protein shown in cyan, S glycans in blue, m₁/m₂ shown in red, ALB in orange, Ca²⁺ in yellow spheres, viral membrane in purple. A) Interactions between mucins and S facilitated by glycans and Ca²⁺. B) Snapshot from SMA simulations. C) Example Ca²⁺ binding site from SMA simulations (1800 sites, each 1000+ atoms) used for AI-enabled quantum mechanical estimates from OrbNet Sky. D) Quantification of contacts between S and mucin from SMA simulations. E) OrbNet Sky energies versus CHARMM36m energies for each sub-selected system, colored by total number of atoms. Performance of OrbNet Sky versus DFT in subplot (ωB97x-D3/def-TZVP, R²=0.99, for 17 systems of peptides chelating Ca²⁺ (Hu et al., 2021)). Visualized with VMD.

Figure 5.

Delta variant spike opening from WE simulations, and AI/haMSM analysis. A) The integrated workflow. B) Snapshots of the “down,” “up,” and “open” states for Delta S-opening from a representative pathway generated by WE simulation, which represents ∼ 10⁵ speedup compared to conventional MD. C) Rate constant estimation with haMSM analysis of WE data (purple lines) significantly improves direct WE computation (red), by comparison to experimental measurement (black dashed). Varying haMSM estimates result from different featurizations which will be individually cross-validated. D) The first three dimensions of the ANCA-AE embeddings depict a clear separation between the closed (darker purple) and open (yellow) conformations of the Delta spike. A sub-sampled landscape is shown here where each sphere represents a conformation from the WE simulations and colored with the root-mean squared deviations (Å) with respect to the closed state. Visualized with VMD.

Figure 6.

WE simulations reveal a dramatic opening of the Delta S (cyan), compared to WT S (white). While further investigation is needed, this super open state seen in the Delta S may indicate increased capacity for binding to human host-cell receptors.

Figure 7.

D-NEMD simulations reveal changes in key functional regions of the S protein, including the receptor binding domain, as the result of a pH decrease. Color scale and ribbon thickness indicate the degree of deviation of Cα atoms from their equilibrium position. Red spheres indicate the location of positively charged histidines.

Figure 8.

GROMACS performance across 1–8 A100 GPUs in ns/day (thicker, blue lines) and the fraction of maximum theoretical TFLOPS (thinner, green lines); production setup shown with solid line, and runs with the GPU-accelerated thermostat in dashed.