A Super-Resolution Spatial Atlas of SARS-CoV-2 Infection in Human Cells

Abstract

The spatial organization of viral and host components dictates the course of infection, yet the nanoscale architecture of the SARS-CoV-2 life cycle remains largely uncharted. Here, we present a comprehensive super-resolution Atlas of SARS-CoV-2 infection, systematically mapping the localization of nearly all viral proteins and RNAs in human cells. This resource reveals that the viral main protease, nsp5, localizes to the interior of double-membrane vesicles (DMVs), challenging existing models and suggesting that polyprotein processing is a terminal step in replication organelle maturation. We identify previously undescribed features of the infection landscape, including thin dsRNA “connectors” that physically link DMVs, and large, membrane-less dsRNA granules decorated with replicase components, reminiscent of viroplasms. Finally, we show that the antiviral drug nirmatrelvir induces the formation of persistent, multi-layered bodies of uncleaved polyproteins. This spatial Atlas provides a foundational resource for understanding coronavirus biology and offers crucial insights into viral replication, assembly, and antiviral mechanisms.

Introduction

The SARS-CoV-2 virus infection in mammalian cells arises from one viral RNA molecule comprising the genome, which codes for 29 proteins (Figure 1A). The expression of these proteins yields an array of protein structures and modified endoplasmic reticulum (ER) membranes in various locations within the infected cell, and these structures are responsible for the viral infections which have occurred. It is critical to learn about the shapes, locations, and logistics of these structures in the human cell context. Some studies focused on localizing SARS-CoV-2 proteins in transfected cells using diffraction-limited (DL) optical microscopy¹. The intracellular localization of certain SARS-CoV-2 proteins has also been investigated using DL microscopy in infected cells^2,3, however the ~200 nm resolution available from DL microscopy does not allow one to distinguish key virus-induced structures, such as double membrane vesicles (DMVs) or virions nor endogenous tiny structures, such as ribosomes.

Figure 1: Overview of the genes of SARS-CoV-2, structures observed previously for replication organelles and schematic of the polyproteins.

(A) SARS-CoV-2 genome with targets labeled in this study indicated in black font and remaining targets in gray font.

(B) Model of an infected A549-ACE2 cell with endogenous targets used in this study and with viral replication organelles (ROs) for which the intracellular localization was previously reported.

(C) Schematic models of non-cleaved viral polyproteins pp1a and pp1ab with different NSP domains and their positions with respect to the ER membrane following translation.

SARS-CoV-2 and other Nidovirales transform host ER membranes into special organelles with double membranous walls, called double-membrane vesicles (DMVs). The interior of these vesicles is isolated from the cytoplasm; however, communication between the DMV interior and the cytoplasm is possible via DMV pore structures⁴. Considering the protective environment inside the DMV, it can be suggested that SARS-CoV-2 uses DMVs for the storage and replication of its genome as these processes produce intermediates, such as double-stranded RNA (dsRNA) and non-capped viral RNA that can induce an immune response and be degraded in the cytoplasm. There is an increasing body of experimental evidence that viral genomic RNA (vgRNA) replication is indeed associated with DMVs^5–7. Moreover, some results suggested that the replication-transcription complexes (RTCs) may occur specifically at DMV pores^8,9, however there is conflicting data on whether RTCs are located in the DMV interior^5,8 or exterior⁹.

Super-resolution (SR) optical microscopy allows much better optical resolution of the localization of molecules forming structures, typically down to 10’s of nm when blinking single fluorophores are used. These techniques are known as Single-Molecule Localization Microscopy (SMLM)¹⁰, which includes Stochastic Optical Reconstruction Microscopy (STORM)¹¹ and related techniques. One focused ion beam – scanning election microscopy (FIB-SEM) study of DMVs utilized another SR method, Stimulated Emission Depletion (STED) microscopy, for studies of intermediate filaments that are remodeled during the infection¹². As is well-known, EM methods do see membranes quite well, but other proteinaceous and nucleic acid structures as not as easily seen. Recently, we employed SR with single molecule immunofluorescence (IF) and RNA fluorescence in situ hybridization (FISH) labeling to explore vgRNA, dsRNA, and several key proteins for SARS-CoV-2 infected cells⁵. This work yielded detailed shapes and locations of the replication organelles (ROs), and the greatly improved resolution allowed clear localization of the vgRNA, RNA-dependent RNA polymerase (RdRp) enzymes, dsRNA, and nsp3 within the ROs, along with modified ER membrane providing a protective outer perimeter. Since the results of our previous publication⁵ provide a key foundation for the present work, we show in Figure 1B a reminder of the key results: replication organelles (ROs) form in the perinuclear region, which are delimited by modified ER membrane. Molecular pores formed by nsp3 and nsp4 are present in these ROs, and dsRNA and vgRNA occupy the central internal region.

As is well-known¹³, part of the translation of the vgRNA into proteins produces polyproteins containing all the non-structural proteins (nsp1 to nsp16), some of which are membrane bound, as illustrated in Figure 1C. However, what is unknown is how these polyproteins are processed to yield the positions of the constituent proteins in an infected cell. For example, it is important to know which proteins are inside and which are outside of the DMV, and how this knowledge relates to the infection process. To obtain a more comprehensive view of the details of the SARS-CoV-2 attack on the host cell, we have now labeled and imaged all of the non-structural and structural proteins produced by the viral genome (except one which has an inferior antibody, shown as gray in Figure 1A) along with dsRNA and vgRNA in A549 lung epithelial cells using 2D and 3D SR microscopy. Where possible, the imaging was performed in 3 colors to provide additional context. This represents a novel atlas of the viral infection players and structures, covering many locations in the host cell. Beyond several known protein and dsRNA positions which were noted by DL microscopy, we observe (a) novel connectors of dsRNA that connect various ROs, (b) the organization in the DMV-RO provided by the polyprotein cutting enzyme, nsp5, and the two pore-forming proteins, nsp3 and nsp4, and (c) diverse morphological structures of the Membrane (M) structural protein in various regions of the infected cell. In addition, in the presence of treatment with the nsp5 inhibitor nirmatrelvir, surprising multilayer shell-like structures of membrane are observed containing many of the polyprotein components. Now that a full Atlas of protein and nucleic acid positions for SARS-CoV-2 infection is in hand, deeper insight into the mechanisms of the viral infection should result. With this more complete picture of the SARS-CoV-2 life cycle, new therapeutic targets may be identified.

Results

Labeling and imaging strategy

We used the human lung carcinoma epithelial cell line A549, designed to be more susceptible for SARS-CoV-2 infection by expression of the human angiotensin-converting enzyme 2 (ACE2) located on the plasma membrane¹⁴. Being derived from human alveolar tissue, A549 has advantage over other commonly used cell lines for SARS-CoV-2 research (e.g. African green monkey cell line Vero E6 or Human kidney cell line 293T-ACE2) as it provides a more relevant model to study human respiratory viruses.

For imaging infected cells, we use super-resolution (d)STORM fluorescence microscopy¹⁵ that, unlike DL light microscopy or EM, provides both nanometer-scale resolution and molecular specificity for cellular imaging¹⁶. For 3-color SR microscopy, we use simultaneous dSTORM imaging of two spectrally close far-red fluorophores Alexa Fluor 647 (AF647) and CF680 separated by ratiometric demixing¹⁷, combined with a third, spectrally distinct fluorophore CF583R¹⁸ imaged following the far-red acquisition. This approach benefits from low cellular autofluorescence and high-performance dyes in the red and far-red spectral regions and reduces the density of simultaneously imaged fluorophores by imaging CF583R separately from AF647 and CF680.

To label SARS-CoV-2 and host proteins (Figure 1), we use IF with commercially available antibodies. First, we identified primary antibodies against viral proteins that provide specific signal using our IF protocol that is compatible with biosafety requirements for virus inactivation and preserves the ultrastructure of DMVs and other organelles⁵. We found usefully performing antibodies for 15 NSPs (nsp1…nsp10 and nsp12…nsp16), for 3 structural proteins (S, M, N) and, additionally, for 5 accessory proteins (ORF3a, ORF6, ORF7a, ORF8, ORF9b). For several targets (nsp1, nsp3, nsp4, nsp5, nsp7, nsp12, N, ORF9b), we discovered 2 high-quality antibodies from 2 different species (rabbit or mouse and sheep) that allow for easy multiplexing using corresponding secondary antibodies labeled with spectrally distinct dyes (Table S1).

We detect SARS-CoV-2 dsRNA using the well-known J2 mouse monoclonal antibody¹⁹ as this provides high sensitivity and specificity in our experimental conditions^5,20. RNA FISH labeling of SARS-CoV-2 vgRNA was previously developed²¹ and provides great detail, especially when combined with SR imaging⁵. In the current work, we have also established fluorescence labeling for the human small ribosomal subunit 40S using RNA FISH against 18S ribosomal RNA (rRNA), which provides context for nsp1 and for virus-induced DMVs. We adjusted the labeling protocol to be compatible with both RNA FISH and IF labeling, allowing both types of labeling in the same sample⁵ (see Methods).

Intracellular localization of non-structural proteins

Non-structural proteins (NSPs) nsp1…nsp16 originate from SARS-CoV-2 polyproteins pp1a/pp1ab that are translated from the ORF1a/ORF1ab genes (Figure 1A). The pp1a/pp1ab polyproteins are cleaved into individual NSPs by protease domains found in nsp3 (papain-like protease, PL^pro) and nsp5 (3C-like protease, 3CL^pro, also known as main protease, M^pro). PL^pro cleaves the polyproteins at specific sites between nsp1…nsp4, and 3CL^pro at the remaining sites between nsp4…nsp16²².

Unlike structural proteins that make up infectious virus particles – the virions, the role of NSPs consists in regulating the intracellular life cycle of the coronavirus. This includes remodeling of the cellular environment, generation of ROs/DMVs, and replication, transcription and processing of the viral genome. For this reason, NSPs are localized in the cytoplasm, and many NSPs are possibly associated with viral ROs. While the localization of nsp3 and nsp4 has been precisely determined from the structure of the DMV pore⁴, the positioning of other NSPs remains unclear. Here we use the known localization of nsp3 and nsp4 at the DMV membrane and the localization of dsRNA and vgRNA in the DMV interior^5–8 to precisely localize all NSPs in SARS-CoV-2 infected cells.

Nsp1

Nsp1 is involved in modulating cellular processes, such as inhibition of host protein translation by binding to ribosomes²³ and to the mRNA export receptor NXF1²⁴. DL studies have demonstrated diffuse cytoplasmic localization of nsp1 in SARS-CoV-2 infected cells^3,25,26, however, to our knowledge, no nanoscale localization for this protein has been reported.

First, we co-imaged nsp1 with dsRNA (marker for DMV interior)⁵ and with nsp4 (label for DMV membrane pores⁴). In the resulting 3-color SR images of SARS-CoV-2 infected A549-ACE2 cells, DMVs can be observed as incomplete ring-like structures of nsp4 surrounding dsRNA clusters in the cytoplasm (Figure 2A). Nsp1, however, did not display any particular pattern with respect to the DMVs and mostly localized diffusely in the cytoplasm outside DMVs (Figure 2A). Because nsp1 interferes with mRNA binding to ribosomes by blocking the mRNA entry channel on the 40S ribosomal subunit²³, we labeled 40S subunits using FISH staining of 18S rRNA. 18S rRNA displayed very sparse signal in the nucleus, somewhat denser punctate signal in the nucleolus and, as expected, very dense widely scattered punctate signal in the cytoplasm, corresponding to ribosomes, polysomes and individual 40S subunits²⁷ (Figure 2B). In infected cells, nsp1 localized in the same cytoplasmic regions as 18S rRNA. To quantify the colocalization between these two targets, we calculated bivariate pair-correlation functions $g_{\mathit{12}}(r)$ between 18S rRNA and nsp1. $g_{\mathit{12}}(r)$ is the distribution of distances between all possible pairs of localizations of two species, and a peak in $g_{\mathit{12}}(r)$ at a certain distance r₀ would signify that the two species are more frequently found spaced by this distance⁵. For 18S rRNA and nsp1, $g_{\mathit{12}}(r)$ peaked at r = 0 nm (Figure 2B), meaning that these two species tend to have a minimal distance between each other and demonstrating that nsp1 indeed associates with 40S ribosomal subunits in the cytoplasm of SARS-CoV-2 infected cells.

Figure 2: Nanoscale cellular localization of nsp1, nsp2, nsp3, nsp4 and nsp5.

(A) Representative 3-color SR image of a SARS-CoV-2 infected A549-ACE2 cell in an early infection stage labeled for nsp1 (yellow), nsp4 (magenta) and dsRNA (cyan). Nsp1 localizes diffusely throughout the cytoplasm, while nsp4 encapsulates round dsRNA clusters (double-membrane vesicles, DMVs).

(B) Representative 2-color SR image of nsp1 (yellow) and 18 ribosomal RNA (rRNA) (magenta) indicates diffused localization of both targets in the same regions of the cytoplasm. In the inset, the pair-correlation functions calculated between the localizations of nsp1 and 18S peak at r = 0 nm, suggesting an association between nsp1 and 18S rRNA. The blue dashed line approximates the position of a nucleolus characterized by an increased density of 18S rRNA.

(C) Representative SR image of an infected cell in an early infection stage labeled for nsp2 (magenta), nsp3 (yellow) and dsRNA (cyan). Nsp2 with nsp3 encapsulate dsRNA clusters (DMVs) and localize in between accumulations of DMVs (convoluted membranes, CMs).

(D) In the late infection, a significant portion of cellular nsp2 (magenta) localizes to fragmented and dispersed Golgi bodies (labeled with anti-Giantin antibody, green). The bottom magnified panels (1–3) display selected Golgi bodies, and yellow arrowheads point to additional examples of Golgi bodies found in the large image.

(E) SR image of a cell in the early infection suggests localization of nsp5 (magenta) in DMVs labeled by nsp3 (yellow, DMV membrane) and dsRNA (cyan, DMV core).

(F) SR image of a cell at the late infection stage indicates that nsp5 (magenta) localizes to vgRNA-labeled (green) DMVs. In the inset, the pair-correlation functions calculated between the localizations of nsp5 and vgRNA in cells at late infection stages peak at r = 0 nm indicating a strong correlation between these two targets.

In the pair-correlation function plots in (B) and (F), the faint blue lines represent individual cells, the faint red lines represent $\mathrm{g}12\left(\mathrm{r}\right)$ calculated applying the complete spatial randomness (CSR) for the localizations in these cells, and the bold lines represent the mean values for the corresponding functions calculated from all analyzed cells. The color-framed boxes represent magnified regions from the corresponding numbered (when available) frames in the large images. The white dashed curves approximately delimit the nuclear edge. The white arrowheads in (A), (C), (E) point to selected examples of individual DMVs, while the green triangles point to selected examples of DMV aggregates associated with nsp3-labeled CMs. Scale bars, 2 μm (large images), 500 nm (magnified images in (C), (D)) and 200 nm (magnified images in (E), (F)).

Nsp2

Nsp2 enhances viral infection via interaction with host proteins^28–30, however its functions are not well established to date. Even though nsp2 is dispensable for viral replication³¹, the vgRNA synthesis in cells infected with nsp2-deleted SARS-CoV-2 advances much more slowly compared to WT SARS-CoV-2³⁰. Our SR imaging reveals distinct patterns of nsp2 localization in SARS-CoV-2 infected A549-ACE2 cells at early and late infection time points. In early infection, when ROs have a diameter of ≲ 200 nm⁵, we observe a strong immunofluorescence signal of nsp2 close to nsp3 (Figure 2C). Nsp2 forms ring-like structures that, like nsp3, encapsulate dsRNA, suggesting the localization of nsp2 close to the DMV membranes. Also, nsp2 localizes together with nsp3 in between clustered DMVs, the regions that were previously attributed to convoluted membranes (CMs)^5,7,32. Nps2 localization at both DMV membranes and CMs was previously reported in a different β-coronavirus, murine hepatitis virus (MHV)³³. However, in late infection characterized by large merged ROs⁵, we observe that a significant portion of nsp2 localizes to Golgi bodies that are fragmented and dispersed in the cytoplasm at this time point (Figure 2D) due to the remodeling of Golgi apparatus by SARS-CoV-2^12,34. The localization of nsp2 at dispersed Golgi bodies is a novel observation and might reflect its interactions with additional host targets during infection.

Nsp5

The main function of nsp5 is cleavage of the polyproteins pp1a/pp1ab at 11 sites between nsp4 and nsp16. Inhibition of nsp5 protease prevents viral replication, therefore this protein is an attractive drug target³⁵, and among numerous nsp5 inhibitors, nirmatrelvir (PF-07321332)³⁶ has been approved for COVID-19 treatment in many countries. In many schematic models of the SARS-CoV-2 life cycle, the polyprotein processing is assumed to happen in the cytoplasm outside DMVs^13,37–39, however little experimental evidence exists about this localization.

In early infection, we observe concentric localization of nsp5 around the dsRNA clusters, at a somewhat closer distance to dsRNA than nsp3 (Figure 2E), which suggests possible localization of nsp5 in the DMV lumen or on the DMV membranes. In late infection, the ROs/DMVs are better visible using vgRNA FISH staining that fills whole ROs (DMVs and vesicle packets) at this time point⁵. In our SR images of nsp5 with vgRNA, nsp5 localizes in the same regions as vgRNA, indicating that nsp5 is associated with ROs in late infection, which is also illustrated by the bivariate pair-correlation functions (Figure 2F). These observations suggest that, contrary to the previous models, the nsp5 protease might act in DMVs or on their surface.

Nanoscale localization of all non-structural proteins within replication organelles

DMV membrane pores contain the nsp3 and nsp4 proteins arranged in a rotationally symmetric manner^40–42. Nsp3 is located at the cytoplasmic side of the DMV and its N-terminal domain extends ~20 nm away from the outer DMV membrane⁴. Nsp4 is a much more compact protein, located close to the inner DMV membrane with its C-terminus facing the DMV interior. In this section, we aimed at precisely localizing all SARS-CoV-2 NSPs in the DMV and with respect to the DMV membranes, and we therefore chose the position of the nsp4 C-terminus as a reference point for NSP localization. Because dsRNA fills the interior part of DMVs^5,8, we used the location of the centroid of dsRNA localizations as the origin point for the radial distance analysis of SARS-CoV-2 targets (Figure 3A). We analyzed only cells in early infection stages with well-defined isolated DMVs with a diameter of ≲200nm (Figure 3B). In later infection stages, DMVs grow, merge into vesicle packets and form less regular structures⁵, which would make this analysis more difficult.

Figure 3: Nanoscale positioning of all SARS-CoV-2 NSPs with respect to DMVs.

(A) Schematic of a SARS-CoV-2 DMV with dsRNA in the center and nsp4 with nsp3 at the membrane pores. The position of all remaining components is to be determined with respect to the known positions of these targets.

(B, left panel) Representative 3-color SR image of the cytoplasm of a typical A549-ACE2 cell in the early SARS-CoV-2 infection stage used for analysis. DMVs (dashed circles) can be observed as round dsRNA clusters (cyan) surrounded by nsp4 (yellow) and nsp3 (red) localizations.

(B, top right boxes) Magnified image of a single DMV that corresponds to the white box in the left panel reveals three-layer organization of dsRNA, nsp4 and nsp3 in a DMV.

(B, top rightmost plot) Radial density distribution functions $\mathrm{g}\left(\mathrm{r}\right)$ of dsRNA, nsp4 and nsp3 obtained by averaging n DMVs from 16 different cells.

(B, bottom right panels) Rotationally averaged images of radial distributions of dsRNA, nsp4 and nsp3 reveal the three-layer structure averaged across all analyzed DMVs.

(C-H) Representative 3-color SR images of individual DMVs, radial density distribution functions $\mathrm{g}\left(\mathrm{r}\right)$ and rotationally averaged images of the radial distributions for vgRNA, 18S rRNA, nsp1, nsp2, nsp5 and nsp9 with nsp4 and dsRNA. Data obtained from 10 (C), 15 (D), 26 (E), 13 (F), 25

(G), 22 (H) individual early-stage infected cells. The average peak positions R of $\mathrm{g}\left(\mathrm{r}\right)$ in (B-H), where applicable, are indicated on top of the plots as mean ± 95% CI obtained from Gaussian fitting of the peak portion of $\mathrm{g}\left(\mathrm{r}\right)$ resampled with replacement (bootstrapping). n stands for the number of analyzed DMVs for each labeling combination.

(I) Mean positions ± 95% CI for all NSPs associated with DMVs after rescaling the positions of nsp4 within individual target groups (B-L) to match the global mean position of nsp4. The global position of nsp4 ± 95% CI (yellow error bar) was calculated using 13 groups (B, F-L) with n = 36662 DMVs. The detailed data for remaining NSPs is shown in Figure S1.

(J) Angular bivariate pair-correlation functions $\mathrm{p}\left(\mathrm{\theta }\right)$ of nsp4 with one of the other targets peak at $\mathrm{\theta }=0°$ for all targets, except 18S rRNA and nsp1.

(K) Model of the mature early-stage DMV based on the SR data.

Scale bars, 1 μm (B, overview image), 100 nm (C-H, K and magnified images of individual DMVs and rotational average images in B).

The positioning of nsp3 and nsp4 is known precisely from the cryo-EM structure of the minimal DMV pore⁴. Therefore, to test our experimental approach, we first localized nsp3 (N-term) with respect to dsRNA and nsp4 (C-term). The 3-color SR images of dsRNA, nsp4 and nsp3 reveal a concentric three-layer structure: 1) typical^5,8 round dsRNA cluster in the center; 2) a nsp4 layer encapsulating dsRNA; and 3) a larger nsp3 layer that encapsulates dsRNA with nsp4 (Figure 3B). Next, for each DMV defined as a dsRNA cluster surrounded by nsp4 localizations, we calculated a radial distribution function $\mathrm{g}\left(\mathrm{r}\right)$ of single-molecule localizations with respect to the centroid of the dsRNA cluster. After averaging over thousands of individual DMVs from many infected cells, the radial distribution $\mathrm{g}\left(\mathrm{r}\right)$ for each NSP demonstrated a clear peak that reflects the average radial position of the protein within the DMV. The non-zero density in the center is explained by 2D projection of localizations originating from a 3D spherical object (the DMV). This analysis shows that nsp3 N-term is found approximately 35 nm farther away from the DMV center than nsp4 C-term (Figure 3B), in good agreement with the DMV pore structure⁴.

Previously, it was reported that vgRNA localizes inside DMVs adjacent to dsRNA^5,7,8. Indeed, we observe that within the DMV, vgRNA localizes on average at approximately the same radial positions as dsRNA (Figure 3C), reflecting the accumulation of vgRNA in the interior of the DMV lumen⁵. Ribosomes could be found only very rarely in DMVs^40,43, suggesting that translation of viral proteins mostly happens outside ROs. Our analysis indicates that the average localization densities for both 40S ribosomal subunits and nsp1 are indeed lower in the ROs than in the surrounding cytosol (Figure 3D, E), further confirming that DMVs are mostly not involved in viral protein synthesis.

The localization of SARS-CoV-2 nsp2 with respect to DMVs is largely unknown, however there was evidence of nsp2 positioning at the cytosolic side of the DMV membrane of MHV³³. The radial distribution $\mathrm{g}\left(\mathrm{r}\right)$ of nsp2 peaks at approximately the same radius as nsp4, indicating that SARS-CoV-2 nsp2 is also located very close to the membranes of early DMVs (Figure 3F). Nsp6 is the third transmembrane protein among the coronaviral NSPs that transforms and zippers host ER^44,45 and therefore can be expected to anchor on DMV membranes. $\mathrm{g}\left(\mathrm{r}\right)$ for nsp6 shows that this protein is found slightly deeper into the DMV than nsp4 (Figure S1A), which suggests a possible nsp6 localization at the inner DMV membrane.

SARS-CoV-1 nsp5 has been found protected by membranous structures^32,46 and was reported to localize at convoluted membranes³². However, the precise localization of nsp5 with respect to DMVs has not been measured. Our radial distribution analysis indicates that SARS-CoV-2 nsp5 localizes predominantly in the DMV lumen close to the DMV membrane, on average 16 nm deeper than nsp4 (Figure 3G). Because nsp5 is the main coronaviral protease indispensable for viral polyprotein processing and for viral replication, its localization inside DMVs suggests a new model of coronaviral polyprotein processing combined with DMV formation. We propose that DMVs are possibly formed with mostly non-cleaved polyproteins pp1a/pp1ab that contain the nsp5 M^pro domain, which becomes active and finishes the polyprotein processing after the DMV closure.

The remaining nsp7…nsp16 proteins are all involved in viral RNA (vRNA) replication and processing³⁸, and therefore can be expected to localize and act in the protective DMV environment. Among them, the RdRp catalytic subunit nsp12, and the RdRp accessory domains nsp7, nsp8 and nsp9, have already been localized in the DMV lumen^5,8, while high-resolution data for the remaining proteins is lacking in the literature. Using specific antibody labeling and super-resolution microscopy, we measured the radial density distributions $\mathrm{g}\left(\mathrm{r}\right)$ for nsp7, nsp8, nsp9, nsp10, nsp12, nsp13, nsp14, nsp15 and nsp16 within DMVs. The peak in $\mathrm{g}\left(\mathrm{r}\right)$ for all these proteins lay approximately 5–25 nm deeper than for nsp4 suggesting that nsp7, nsp8, nsp9, nsp10, nsp12, nsp13, nsp14, nsp15 and nsp16 are predominantly located in the DMV interior (Figure 3H, S1).

The average radial position of nsp4 exhibits some variability (73–88 nm) between different measurement pairs due to the variability between the cells in different samples, that are inevitably found at slightly different infection time points. To be able to precisely compare the positions of all NSPs between each other, we aligned the average nsp4 position in each of the measurement pairs to the global average position of nsp4 calculated from all groups and applied the same scaling factor to the second target in each group. The resulting rescaled positions reveal that nsp5…nsp16 are found significantly deeper in the DMV than nsp4, while nsp2 localizes at a similar radial position as nsp4, while nsp3 N-term lies in the cytosol, significantly farther away from nsp4 (Figure 3I).

Because we found nsp2…nsp16 to locate relatively close to the DMV membrane pores in the radial direction, next we asked if nsp2 and nsp5…nsp16 are found close to the DMV pores also in the angular direction around the boundary of the RO. To investigate this, we calculated bivariate angular pair-correlation functions⁴⁷ $\mathrm{p}\left(\mathrm{\theta }\right)$ between nsp4 and each of the remaining targets (Figure 3J). These functions reflect the average angular distance between two targets and would peak at $\mathrm{\theta }=0°$ if the two targets are systematically found at a similar angular position. $\mathrm{p}\left(\mathrm{\theta }\right)$ was normalized such that $\mathrm{p}\left(\mathrm{\theta }\right)=1$ for a non-correlated pair of species. Among all targets, $\mathrm{p}\left(\mathrm{\theta }\right)$ has the highest peak at $\mathrm{\theta }=0°$ for nsp4 with nsp3, reflecting the known co-localization of both targets in the DMV pore, which is directly noticeable in the SR images (Figure 3B). 40S ribosomal subunits and nsp1 both demonstrate very weak correlation with nsp4, while vgRNA, nsp2 and nsp5…nsp16 strongly correlate with nsp4 in the angular direction, with a clear peak at $\mathrm{\theta }=0°$ (Figure 3J). This indicates that nsp2 and nsp5…nsp16 predominantly localize close to the DMV pores not only in the radial but also in the angular direction around the membrane, suggesting that these proteins are attached to or located close to the pores in mature replication-competent DMVs as indicated by the strong dsRNA signal in the core region.

To additionally confirm the concentric localization of NSPs around the dsRNA core of DMVs/ROs and to discard possible data misinterpretation due to 2D projection (3-color 2D dSTORM) of 3D objects (DMVs), we performed 2-color 3D dSTORM microscopy using a double-helix point spread function (3D-DHPSF)⁴⁸ on nsp5, nsp8 and nsp10 with dsRNA. This method allows for a near-isotropic resolution in 3D within a ~2μm sample depth^16,49. For each target pair, we indeed observe NSPs adjacent to dsRNA clusters and surrounding them in 3D (Video S1), which, when projected in 2D under different angles, resemble the ring-like 2D images of individual NSPs (Figure 4A–C).

Figure 4: 3D and 2D SR images of SARS-CoV-2 infected cells with focus on DMVs and other dsRNA structures.

(A-C) Examples of 3D dSTORM images of DMVs acquired using a double-helix point spread function (DHPSF) with nsp5 (A), nsp8 (B) and nsp10 (C) co-labeled with dsRNA in SARS-CoV-2 infected A549-ACE2 cells. Each blue frame combines views of the same DMVs under different angles with axis shown in a corner of each image: red = X, green = Y, blue = Z axis.

(D-E) 3D-DHPSF images of dsRNA in a SARS-CoV-2 infected A549-ACE2 cell. The magnified panel in (E) corresponds to the colored frame in (D). The z position of fluorophores is encoded in the colormaps shown as colorbars with z-values labeled, where the lowest z position is closest to the coverslip.

(F) 2D dSTORM images of dsRNA co-labeled with nsp4 in a SARS-CoV-2 infected A549-ACE2 cell demonstrate examples of thin dsRNA connectors between dsRNA structures. The images correspond to yellow boxes in the whole-cell image shown in Figure S2A

(G) Cell in an early infection stage contains large, approximately round dsRNA granules in the cytoplasm that lack nsp4 but are decorated with nsp9 puncta. The magnified images on the right correspond to the green box in the large image. Additional examples of dsRNA “connectors” are highlighted by green triangles, and selected examples of dsRNA granules lacking an nsp4 shell are indicated by white arrowheads.

(H-J) Examples of large dsRNA granules that mostly lack nsp4 but are decorated with nsp5 (H), nsp13 (I) and nsp8 (J).

The white dashed curves approximately delimit the nuclear edge. Scale bars, 5 μm (whole-cell images in D, G), 500 nm (H-J, magnified box in G) and 200 nm (A-C, E-F).

dsRNA landscape in SARS-CoV-2 infected cells

We observe that some dsRNA clusters are interconnected by fiber-like dsRNA strands (Figure 4C, S2A) which we term “connectors”. The width of the dsRNA connectors, measured as σ of an equivalent Gaussian function, is only (17±5) nm (Figure S2A), i.e., is limited by our single-molecule localization precision (${\mathrm{\sigma }}_{\mathrm{l}\mathrm{o}\mathrm{c}}\approx 15\mathrm{n}\mathrm{m}$, quantified from ThunderSTORM fits) and by the size of antibodies used for dsRNA detection (up to ~ 20 nm, primary + secondary antibodies⁵⁰). This suggests that the dsRNA “connectors” might potentially be helical dsRNA molecules residing in the lumen of ER tubules as those commonly interconnect DMVs as observed by EM studies⁴². Besides human A549-ACE2 cells, we also observed these dsRNA fibers in the Vero E6 cell line (Figure S2B), indicating that these fibers are a general feature of intracellular SARS-CoV-2 infection. dsRNA connectors are also easily visible in 3D SR microscopy (Figure 4C–E, Video S2).

In the cytoplasm of SARS-CoV-2 infected cells, besides typical dsRNA clusters (d ≲ 200 nm) surrounded by nsp3, nsp4 and other NSPs (Figures 2–3), we also observe larger dsRNA clusters (d ≳ 250 nm) that lack the DMV membrane pore proteins nsp3 and nsp4 (Figure 4G–J). These granules or condensed structures typically stay approximately round, however in cells at later infection stages they may adopt less regular shapes and may occupy larger areas of the cytoplasm (Figure S2C–G). Despite the absence of nsp3 and nsp4, these aggregates are typically decorated with sparse puncta of various NSPs, including nsp9 (Figure 4G, S2C), nsp5 (Figure 4H, S2D), nsp13 (Figure 4I), nsp8 (Figure 4J), nsp10 (Figure S2E) and nsp12 (Figure S2F). However, these dsRNA structures mostly lack nsp2 (Figure S2D) and locate in the cytoplasmic regions devoid of nsp1 (Figure S2G). The existence of dsRNA aggregates with the attached replicase proteins but with absent nsp3 and nsp4 indicates that the replicase proteins might possibly exercise their functions also without interaction with DMV pores. In some cases, larger nsp3–4-lacking dsRNA structures are linked to several normal DMVs by dsRNA “connectors” (Figure 4F, S2A), suggesting a hypothetical mechanism of RO filling and vRNA propagation in the cytoplasm via dsRNA “runners”.

Nanoscale localization of structural proteins in SARS-CoV-2 infected cells

SARS-CoV-2 virus has four structural proteins, namely the spike protein (S), membrane protein (M), envelope protein (E), and nucleocapsid protein (N)^51,52. Using IF staining and SR imaging, we investigated the nanoscale localization of the S, M and N proteins in cells infected by the SARS-CoV-2 virus. First, we validated the antibodies using virion labeling and SR imaging that revealed a concentric localization of the S and M proteins with the N protein located mostly in the center of the virion (Figure S3A). The diameter of the S ring in the labeled virions was 86 nm, while for the M protein this diameter was 18 nm smaller, consistent with the virion dimensions and with the positioning of M in its membrane and S on the exterior⁵³.

The M protein is the most abundant structural protein on SARS-CoV-2 virion membranes and plays critical roles in viral particle assembly and egress⁵². We stained M protein using antibodies in A549-ACE2 cells infected with SARS-CoV-2 virus for 24hpi and performed SR imaging to monitor the nanoscale localization of M protein. Importantly, we found four morphological structures of M protein, which have distinct localization patterns from the perinuclear region to the cytoplasmic membrane, possibly corresponding to different stages of viral assembly and egress (Figures 5A and 5B). In the perinuclear region, M proteins aggregate into large unstructured conglomerates where the distribution of M proteins is relatively sparse (Figure 5A, blue square), while in the region away from the nucleus towards the cytoplasmic membrane, M proteins form small, dense aggregates (Figure 5A, green square). In the cytoplasmic region close to the cell periphery, M proteins seem to assemble into virions which aggregate into large, hollow clusters (Figure 5A, red square). On the cytoplasmic membrane, virions are distributed with M proteins localized on the viral envelope to form hollow particles (Figure 5B).

Figure 5: Nanoscale localization of the SARS-CoV-2 structural proteins M, S and N.

(A) Left panel: representative SR image of M protein in a SARS-CoV-2 infected A549-ACE2 cell at 24 hpi reveals distinct morphological structures of M protein. Right panel: zoomed-in regions of the left SR image enclosed by blue (top), green (middle), and red (bottom) rectangles.

(B) Representative SR image of M protein in a SARS-CoV-2 infected A549-ACE2 cell at 24 hpi showing virions arrayed on the cell membrane. Yellow rectangles: zoomed-in virions.

(C) Top panel: representative two-color SR image of M protein (magenta) and Giantin (green) in a SARS-CoV-2 infected A549-ACE2 cell at 24 hpi showing M protein distributed inside Golgi. Bottom panel: zoomed-in regions of the top SR image enclosed by a yellow rectangle.

(D) Left panel: representative two-color SR image of M protein (magenta) and LAMP1 (green) in a SARS-CoV-2 infected A549-ACE2 cell at 24 hpi showing the distribution of M proteins inside and outside lysosomes. Right panel: zoomed-in regions of the left SR image enclosed by yellow (top), red (middle), and blue (bottom) rectangles.

(E) Representative two-color SR images of M protein (magenta) and S1 protein (green) in SARS-CoV-2 infected A549-ACE2 cells at 24 hpi showing M protein distributes together with S1 protein in the hollow (top) and solid (bottom) clusters.

(F) Histogram of the equivalent radius of the solid (blue) and hollow (red) clusters of M protein in SARS-CoV-2 infected A549-ACE2 cells at 24 hpi. Data for the histogram is generated from 276 solid clusters and 180 hollow clusters obtained from 22 cells and 8 independent experiments.

(G) Left panel: representative two-color SR image of vgRNA (magenta) and N protein (green) in a SARS-CoV-2 infected Vero E6 cell at 24 hpi showing they only colocalize in the cytoplasm outside ROs. Right panel: zoomed-in regions of the left SR image enclosed by yellow (top) and blue (bottom) rectangles.

(H) Bivariate pair-correlation functions ${\mathrm{g}}_{12}\left(\mathrm{r}\right)$ calculated between the localizations of vgRNA and N protein indicate their close association.

(I) Left panel: representative two-color SR image of vgRNA (magenta) and S2 protein (green) in a SARS-CoV-2 infected Vero E6 cell at 24 hpi showing they do not colocalize within or outside ROs. Right panel: zoomed-in regions of the left SR image enclosed by yellow (top) and blue (bottom) rectangles.

(J) Bivariate pair-correlation functions ${\mathrm{g}}_{12}\left(\mathrm{r}\right)$ calculated between the localizations of vgRNA and S2 protein indicate nanoscale anti-correlation between them.

Dashed lines in A, C, D, G, and I indicate the boundary of the cell nucleus (large dark region). Color bars in A-E show the number of single-molecule localizations within each SR pixel (20 × 20 nm², except for B, 16 × 16 nm²). Scale bars, 10 μm (large image in A), 5 μm (large images in B, C, D, G, I), 1 μm (magnified images in A and C), 500 nm (E and magnified images in D, G, I).

To explore these diverse morphological structures of the M protein, we co-imaged it with other host cell compartments. Since the M protein is glycosylated⁵⁴ and most glycosylation reactions happen in Golgi, we first co-stained cells with an anti-M protein antibody and an anti-Giantin antibody which labels the rim of Golgi cisternae⁵⁵. We found that the large, unstructured conglomerate of M protein distributed in the perinuclear region is localized within Golgi (Figure 5C), which might be essential for the glycosylation of the M protein. We then co-stained M protein with the anti-LAMP1 antibody which labels the membrane of lysosomes and late endosomes⁵⁶. It has been reported that β-coronaviruses including SARS-CoV-2 virus use lysosomes for egress⁵⁶. Indeed, we found some lysosomes containing dense aggregates of M proteins, while many lysosomes do not contain M proteins (Figure 5D). Interestingly, in some large lysosomes, M proteins accumulate along the inner membrane of the lysosome, making them appear hollow (Figure 5D, yellow square), while in some small lysosomes, M proteins fill in the whole organelle (Figure 5D, red square). We also noticed that some large clusters of M protein aggregates are not localized within lysosomes (Figure 5D, blue square). To determine whether the M protein aggregates inside lysosomes correspond to virions, we co-imaged the M protein with the S1 domain of the spike protein and found that these M protein aggregates assemble with the S1 protein (Figure 5E), supporting that these are overlapping assembled virions. We further quantified the size of two M protein aggregate clusters and found that the small, dense M protein clusters have a median equivalent radius of 601 nm and the large, hollow M protein clusters have a median equivalent radius of 1140 nm (Figure 5F), which is similar to the lysosome size (0.5–1 μm)⁵⁷.

Next, we explored the nanoscale cellular localization of N proteins and S proteins by co-imaging with vgRNA. In early infection, N proteins tend to accumulate in the cytosol outside ROs and around them (Figure S3B), which was reported earlier^5,58 and might correspond to the sites of N protein translation from subgenomic RNA that has been exported out of ROs. Later, N proteins distribute more evenly in the cytoplasm (Figure 5G), indicating that vgRNAs are not assembled with N proteins in the ROs. N proteins distributed in the cytoplasm do colocalize with sparse vgRNA puncta (Figure 5G, yellow square, and 5H), suggesting that vgRNAs are assembled with N proteins into ribonucleoprotein particles (RNPs) in the cytoplasm.

S proteins stained by the S2 domain antibody do not localize inside ROs either, but their distribution in the cytoplasm is not even (Figure 5I, S3C). Instead, S2 proteins aggregate to form some clusters within the cytoplasm and do not colocalize with vgRNA puncta (Figure 5I–J). Similarly to M, perinuclear S2 localizes to Golgi where it is observed starting from early infection (Figure S3C). We also detect S2 subunits in the ER tubules labeled by Sec61β-GFP overexpression, and decorating the nuclear envelope (Figure S3D). Because the Spike protein is likely cleaved by endogenous furin-like proteases in the host cell⁵⁹, next we co-labeled both S1 and S2 domains simultaneously. Interestingly, while both S1 and S2 domains of the Spike protein localize to Golgi and to vesicles with virions (Figure 5E, S3E), unlike S2, S1 does not localize to the ER tubules or to the nuclear envelope (Figure S3E). This suggests that the SARS-CoV-2 S protein appears to disassemble after the furin cleavage in the host cells, with a fraction of the S2 domains remaining in various cellular membranes and the S1 domains following the usual virion assembly pathway to form the spikes anchored in the virion surface.

Transformation of the infection landscape upon Nirmatrelvir treatment of human cells

After establishing the normal SARS-CoV-2 infection landscape of A549-ACE2 cells, we decided to investigate how it changes when the viral life cycle is perturbed by the nsp5 protease inhibitor nirmatrelvir (PF-07321332)³⁶ that is the active component of the anti-COVID-19 medication nirmatrelvir/ritonavir (Paxlovid), approved in the United States, United Kingdom, European Union, Canada and other countries. By preventing the polyprotein cleavage at all M^pro sites acted upon by nsp5, pp1a/pp1ab should stay uncleaved between nsp4 and nsp11/nsp16, and only nsp1, nsp2 and nsp3 would be released as they are cleaved by PL^pro found in nsp3 (Figure 1C). Because nsp7…nsp16 are required for vRNA replication and transcription, the inhibition of M^pro cleavage should stop viral replication, but it is currently unknown how this treatment would affect the transformation of host ER into DMVs as this process can be executed solely by nsp3 with nsp4⁴⁰. The M^pro cleavage inhibition by nirmatrelvir also allows to investigate the process of DMV maturation as it would possibly halt the process in an intermediate state with non-cleaved polyproteins.

First, we added three different concentrations of nirmatrelvir (1×EC₅₀, 10×EC₅₀ and 0.1×EC₅₀, EC₅₀ = 144 nM) to SARS-CoV-2 infected A549-ACE2 cells (MOI = 2) at 0 hpi, 2 hpi or 6 hpi, and we quantified the percentage of infected cells at 24 hpi using confocal imaging of immunofluorescence staining of dsRNA and nsp4 with nsp5 (Figure S4A). We found no decrease in the number of infected cells when treated with 14.4 nM of nirmatrelvir at any of the time points, but when treated with 1440 nM at 0 hpi and 2 hpi, nearly no infected cells were found in the samples. When treated with 1440 nM of nirmatrelvir at 6 hpi or with 144 nM at any time point, we observed a significant (>2.6×) reduction of the number of infected cells compared to the no treatment condition (Figure S4A).

While cells treated with 14.4 nM of nirmatrelvir had no noticeable phenotype change in DL or SR, approximately 33% (11 out of 33 imaged cells, compared to 0 out of 7 imaged cells in the negative control group with no nirmatrelvir treatment) of SR-imaged infected cells treated with 144 nM of nirmatrelvir at any time point or with 1440 nM at 6 hpi displayed multi-layered bodies (MLBs) in the cytoplasm (Figures 6, S4B–D). When labeled for nsp4 and nsp5 and imaged using 2D dSTORM, these MLBs appear to have ring-like or crescent-like shapes with dimensions in the range of 0.5–5 μm (Figures 6A, S4B). The bodies do not display a significant dsRNA signal indicating that they are not associated with vgRNA replication. Within the MLBs, we were able to resolve several layers containing both nsp5 and nsp4 with a spacing of ~75 nm between the layers (Figure 6A, S4B). We then labeled the C-term protein of pp1a, nsp10, and C-term of pp1ab, nsp16, and found that nsp10 and nsp16 also locate in the MLBs (Figures 6B–C). This indicates that the layers are likely to contain the non-cleaved pp1a/pp1ab polyproteins comprising nsp4…nsp16 due to M^pro inhibition by nirmatrelvir.

Figure 6: Multi-layered bodies in SARS-CoV-2 infected cells treated with Nirmatrelvir.

(A) SR image of an A549-ACE2 cell infected with SARS-CoV-2 at MOI = 2, incubated in a medium with 144nM Nirmatrelvir from 0 hpi to 24 hpi and fixed at 24 hpi. Two regions with large multi-layered bodies (MLBs) are magnified in the central panels. Right panels 1–2 demonstrate the density profiles of nsp4 and nsp5 localizations calculated perpendicular to the corresponding numbered white curves, in the direction indicated by the arrows (the localization density in the orthogonal to the arrows direction was averaged). Both nsp4 and nsp5 localize in layered structures with a period of approximately 75 nm.

(B-G) SR images of A549-ACE2 cells infected with SARS-CoV-2 at MOI = 10, incubated with 288nM Nirmatrelvir from 6 hpi to 24 hpi and fixed at 24 hpi. Nsp10 (B), nsp16 (C) and nsp3 (D) localize to the same multi-layered structures as nsp5. (E) Magnified region of (D) demonstrates localization of nsp3 on both sides of nsp5-labeled layers. The averaged density profiles of nsp3 and nsp5 in the right panels calculated perpendicular to the numbered white curves in the direction of the arrows show double structure of nsp3 layers centered on single nsp5 layers. (F-G) Additional examples of MLBs with double structure of nsp3 layers from different cells in the same sample group.

(H) Confocal images of A549-ACE2 cells infected with SARS-CoV-2 at MOI = 10, incubated with 288nM Nirmatrelvir from 6 hpi to 24 hpi and fixed at 24 hpi (left) or 48 hpi (right). Here, Lysosome-Associated Membrane Protein 1 (LAMP1, green) was used as a general cell label. In the left panel, only 6 cells in the bottom right corner display a weak punctate nsp5 signal (magenta) typical for the early infection phenotype, and the remaining cells have only LAMP1 signal and no detectable nsp5. In the right panel, nearly all cells demonstrate strong perinuclear nsp5 signal consistent with late-stage infection.

(I-J) SR images of A549-ACE2 cells infected with SARS-CoV-2 at MOI = 10, incubated with 288nM Nirmatrelvir from 6 hpi to 24 hpi and fixed at 48 hpi. (I) Typical cell in this condition with a late infection phenotype with very dense perinuclear NSP and dsRNA localizations. (J) Typical cell in this condition with cytoplasmic MLBs. Among 32 infected cells imaged in SR, 16 (50%) had MLBs. For similarly treated cells fixed at 24 hpi with examples shown in (B-G), among 41 infected cells imaged in SR, 29 (~71%) had MLBs.

The white dashed curves indicate the approximate position of the nucleus edge. Scale bars, 5 μm (A, left panel, D, I, J); 1 μm (A, middle panels, B, C); 500 nm (E, F, G); 50 μm (H).

To investigate if the MLBs are formed by the altered process of DMV formation, we labeled nsp3 that is necessary for DMV formation together with nsp4. We found that nsp3 locates on MLBs and demonstrates a double structure of layers, where nsp3 N-term is located around 30 nm away from each nsp4/nsp5 layer on both sides (Figure 6D–G, S4C–D). The offset of nsp3 N-term from nsp4/nsp5 is similar to the distance of the cytoplasmic nsp3 N-term from the DMV membranes (Figure 3B). This double-layered pattern suggests that unlike in DMVs where nsp3 is located only on the cytoplasmic side of the double membrane, nsp3 in nirmatrelvir-induced MLBs is found on both sides of zippered roughly planar membranes.

Besides MLBs, apparently normal DMVs can also be observed in the nirmatrelvir-treated cells (Figure 6A, S4B–E). Containing dsRNA and NSPs, these remaining DMVs are likely to be replication-competent and to provide the vgRNA necessary for translation of pp1a/pp1ab and for further formation of the multi-layered aggregates. This observation provides a pathway to reinfection, consistent with the rebound effect which can occur after Paxlovid treatment.

Next, after incubating cells with nirmatrelvir from 6 hpi to 24 hpi, we removed the drug and incubated cells for 24 additional hours without adding the virus. While at 24 hpi few cells had any signal of viral targets, at 48 hpi nearly all cells were reinfected even though no additional virus was added (Figure 6H, S4A). The cells displayed a predominant late-infection phenotype with very bright perinuclear NSP and dsRNA signals (Figure 6H). In SR, these cells had mostly normal late-infection phenotype with dsRNA clusters and very dense NSPs in the perinuclear region (Figure 6I), suggesting that vgRNA from the initial virions likely survived in the cells during the nirmatrelvir incubation, and the virus quickly took over once the M^pro inhibitor was removed. Many cells still contained MLBs (Figure 6J) meaning that the cell machinery cannot clear all MLBs during the 24 h incubation without nirmatrelvir.

Our novel observation of MLB formation upon nirmatrelvir treatment shows explicitly how the cellular invasion is disrupted. This might also have therapeutic implications and might explain possible long-term side effects of the drug if the MLBs persist in some cells after the treatment has ended.

Discussion

Despite the importance of the coronavirus infection in general, many questions about the cellular infection remain unresolved. Several key processes of the cellular viral infection might explain the difference of severity in clinical symptoms, ranging from not symptomatic to life-threatening conditions. While fully elucidating all mechanisms is out of the scope of this work, here we localize numerous key components of SARS-CoV-2 infection in human cells with nanoscale precision (Figure 7A, Table S2) to provide support or lack of support for various models which have appeared.

Figure 7: Summary of findings from the super-resolution spatial Atlas of SARS-CoV-2 infection.

(A) Schematics summarizing novel viral infection landscapes revealed from the super-resolution spatial Atlas of SARS-CoV-2 infection in human cells. (1) Nanoscale localization of NSP proteins within ROs, which is zoomed in on the right panel. (2) The fiber-like connectors physically linking dsRNA clusters, which is zoomed in on the right panel. (3) Large, membrane-less dsRNA aggregates (granules) decorated with replicase components. (4) Diverse morphological structures of M protein in various regions of the infected cell: (a) M protein localized inside Golgi together with Spike; (b) small solid clusters of M protein (potentially assembled into virions with Spike) inside lysosomes; (c) large hollow clusters of M protein (potentially assembled into virions with Spike) inside lysosomes and vesicles containing M outside lysosomes; (d) M protein assembled into virions arrayed on cell membrane. (5) nsp1 colocalizes with ribosomes in the cytoplasm. (6) vgRNA colocalizes with N protein to form RNPs in the cytoplasm.

(B) Suggestive model of DMV formation and polyprotein processing and its disruption by nirmatrelvir. (1) The process starts with vgRNA translation on the ER membrane into polyproteins pp1a/pp1ab. The polyprotein is attached to the membrane by transmembrane domains of nsp3, nsp4 and nsp6. Nsp1 is cleaved by PL^pro in nsp3 shortly after translation. (2) Interaction of the ectodomains of nsp3 and nsp4 lead to membrane bending and to (3) subsequent pore formation⁴ with nsp2-nsp3 found on one side of the double membrane and nsp5…nsp16 on the opposite side. (4) DMV membrane closes and encapsulates vgRNA. The binding of vgRNA to nsp7…nsp16 proteins ensures that vgRNA is inside the DMV. (5) Increased concentration of nsp5 activates M^pro which cleaves pp1a/pp1ab inside the DMV producing a mature RO. (6) NSPs in the DMV assemble into replication-transcription complexes (RTCs) that replicate, transcribe and modify vRNAs. (2a) Adding nirmatrelvir stops polyprotein cleavage that possibly prevents membrane bending and leads to trans-interactions of nsp3 and nsp4 ectodomains, and to zippering of ER membranes that eventually assemble into multi-layered bodies (MLBs) (3a).

Nanoscale localization of NSPs sheds light on DMV formation and maturation

The absence of ribosomes in the DMV interior (Figure 3D) clearly points to translation of SARS-CoV-2 proteins in the cytoplasm outside DMVs. In the current models of SARS-CoV-2 life cycle, it is typically assumed that the processing of pp1a/pp1ab by PL^pro and 3CL^pro also occurs in the cytoplasm outside DMVs^13,37–39. On the other hand, much evidence exists about vgRNA replication inside the DMVs^5–8, and therefore many of the replicase NSPs should localize in the DMV interior. This is supported by our results that demonstrate that all nsp5…nsp16 proteins localize on average deeper than nsp4, towards the DMV lumen (Figure 3, S1). The DMV lumen is isolated from the cytoplasm by a double membrane layer and has very limited and tight membrane pores^4,42. The width of the central channel of the DMV pore is less than 20 Å⁴, which allows for passage of single-stranded RNA but makes the transit of large, folded proteins virtually impossible. The question therefore arises about how do all nsp5…nsp16 enter the DMVs.

Because our data demonstrates predominant localization of the SARS-CoV-2 3CL^pro nsp5 in the DMV interior (Figure 2E–F), close to the DMV inner membrane and to the pores in the early DMVs (Figure 3G, I), it is likely that the polyprotein cleavage might also occur at these sites. Considering these observations, a novel hypothetical model of DMV formation can be suggested, where DMV pores form from the interaction of nsp3 with nsp4 while nsp4 is still attached to the non-cleaved nsp5…nsp11/16 polyproteins (Figure 7B). After the double membrane closure, a trigger (e.g. the increased concentration of NSPs which could favor the nsp5 dimerization necessary for its activity⁶⁰) would activate 3CL^pro that would finish the DMV maturation. Viruses often employ similar mechanisms of polyprotein cleavage that allow them e.g. to produce, from a limited number of genes, protein products with different functions that can be switched based on certain triggers. A well-known example is the maturation of human immunodeficiency virus particles through proteolytic cleavage of the Gag polyprotein that happens only after the virion budding from the host cell⁶¹.

High-resolution structural information about the DMV pore is available only for the minimal nsp3-nsp4 pore without viral infection, therefore the remaining viral targets could not be observed⁴. Moreover, the averaging necessary for obtaining high-resolution EM structures would prevent the observation of any flexible domains, which the nsp5…nsp11/16 are likely to be due to their distant location from the rigid structure of the pore’s core. However, the C-terminal domain of nsp4 where nsp5 would be attached in the non-cleaved polyprotein, is oriented toward the DMV interior⁴, which is compatible with our proposed model of polyprotein cleavage after the DMV formation. The N-terminal domain of nsp3 is located in the cytoplasm⁴ and therefore the PL^pro cleavage of nsp1 and nsp2 from nsp3 (N-term) would release them directly into the cytosol. Nsp1 has indeed a cytoplasmic localization (Figure 3E), and a significant part of nsp2 also localizes outside DMVs, in the CMs and Golgi (Figure 2C–D). The DMV-associated fraction of nsp2 localizes very close to the membrane (Figure 3F), on average closer than nsp5…nsp16 (Figure 3I). Even though our method does not allow one to distinguish between the outer and inner DMV membranes in this case, perhaps nsp2 localizes in the cytosol adjacent to the outer DMV membrane, by analogy with MHV, where nsp2 was localized to the cytosolic face of DMVs and CMs³³. Additionally, it was suggested that nsp2 has transmembrane domains⁶² and therefore might anchor in the outer DMV membrane. It is hard to place nsp2 in the DMV interior since it is cleaved by PL^pro found in the cytosol and nsp2 is also found in the Golgi bodies later in the infection cycle (Figure 2D).

Our suggestive model therefore explains the localization and concentration of all nsp5…nsp16 in the DMV that becomes an isolated protective chamber for vRNA replication, transcription and storage.

dsRNA landscape

An important finding of our study is the discovery of large, spherical dsRNA aggregates or granules that are compositionally distinct from canonical replication organelles (Figure 4F–J, S2). These structures, sometimes exceeding 1 μm in diameter, are devoid of the DMV membrane pore-forming pair nsp3 and nsp4 and of the cytoplasmic nsp1 and nsp2. In contrast, they are decorated with many of the replicase components, from the nsp5 protease to the nsp8 RdRp subunit and the nsp16 methyltransferase (Figure 4G–J, S2). The distinct nature of these bodies—being membrane-less, rich in nucleic acid and replication enzymes, and having a rounded morphology—is highly reminiscent of viral factories or ‘viroplasms’ common in the life cycle of various viruses^63–65 but not previously reported for coronaviruses. It is possible that these are biomolecular condensates formed for dsRNA storage, replication or degradation, such as stress granules formed by the host cell. While RNA-protein biocondensates (speckles) are common in the nucleus⁶⁶, it is striking that for the SARS-CoV-2 virus possibly related structures are found in the cytoplasm.

Another fascinating discovery is the presence of thin, linear dsRNA ‘connectors’ that physically link neighboring DMVs and sometimes tether DMV clusters to the large, membrane-less dsRNA granules (Figure 4C–G; S1A,E; S2A–C). To our knowledge, these inter-organellar bridges are a previously undescribed feature of the coronavirus replication compartment. We propose that they function as conduits for the intracellular propagation of the viral factory, allowing dsRNA synthesized in established DMVs to be transported through the cytoplasm to ‘seed’ the formation of new replication sites. A primary focus for future studies should therefore be to elucidate the molecular composition of these connectors and the biophysical properties of the viroplasm-like granules they feed.

Virion assembly and egress information from structural protein localization patterns

While this study was not designed to fully elucidate all aspects of virion assembly and egress, the SR localization patterns of the structural proteins (Figure 5, S3) do provide some insight. In particular, the M proteins form a set of localization patterns which shift from large, unstructured groups of puncta in the perinuclear region to more-organized round structures partway toward the cell periphery, to hollow rings of clusters of small virion-like objects closer to the cell membrane. The unstructured perinuclear population clearly corresponds to the Golgi bodies (Figure 5C), and the ring-like structures are correlated with lysosomal membranes in a number of cases. However, the M protein also displays a structured pattern in different, often large vesicles that do not contain the lysosome marker LAMP1 (Figure 5D, blue box). SR imaging of Spike confirms that these structures are on the way to the production of large numbers of virions. Previous studies suggested that virions form in the ER-Golgi intermediate compartment before entering the Golgi apparatus^43,67, however our results suggest that M and S are processed and glycosylated in Golgi and assemble into virions only later, possibly in lysosomes and other, LAMP1-negative, vesicles. This indeed agrees with EM studies that proposed virion formation in single-membrane vesicles and virion egress via “large virus-containing vesicles” bypassing the Golgi complex⁶⁸.

We also observe a diffuse non-structured localization pattern of both vgRNA and the N protein throughout the cytoplasm of host cells, and these targets associate with each other, but not with the Spike protein (Figure 5G–J). This pattern likely corresponds to the formation of SARS-CoV-2 RNPs from vgRNA with the N protein, that appears to happen in the cytosol independently of the virion budding from membranous organelles. Because the N protein interacts with M in an RNA-dependent way⁶⁹, the RNPs could then bind to the M protein of the budding virions, ultimately leading to the closure of the virion membrane and to formation of assembled virus particles.

Nirmatrelvir treatment of SARS-CoV-2 infected human cells

Even though nirmatrelvir is an approved anti-COVID-19 drug, its effect on infected human cells has not been investigated at nanoscale. Here, when nirmatrelvir-assisted inhibition of PL^pro occurs, we often observe fascinating multi-layer bilayer structures (Figure 6, S4). As a result, the model of transformation of ER membranes into DMVs needs further refinement. The double-layered localization pattern of nsp3 (N-term) suggests that MLBs contain zippered membranes, possibly due to the trans-interaction of nsp3 and nsp4 ectodomains⁴. Even though the expression of nsp3 with nsp4 is sufficient to generate DMVs, previously it was noticed that the ER membranes do zipper, but DMVs did not form when nsp4 was replaced by a nsp4-nsp5-nsp6 polyprotein that could not be efficiently cleaved⁷⁰. We can suggest that the non-cleaved nsp4-nsp5-nsp6 polyprotein, being anchored in the ER membrane at both nsp4 and nsp6, is not flexible enough to allow for the DMV-forming cis-interaction and bending, and its cleavage at one of the sites between nsp4 and nsp6 is required for the proper DMV formation. Because the DMV pore contains 12 copies of nsp4 in two different conformations⁴, it is possible that e.g. one form is cleaved from nsp5-nsp6 and the other one is not, however, further research is needed to determine this precisely.

Therefore, we can suggest that the 3CL^pro cleavage of the polyprotein might be a molecular switch between the trans- and cis-interactions of the nsp3 and nsp4 ectodomains that precede the membrane bending needed for DMV formation (Figure 7B).

In this work, we observed more than the simple interference on RO formation (greatly reduced ROs, MLB formation). By treating a group of cells with nirmatrelvir exposure and then release followed by continued cell growth in media, the viral infection resumes strongly. It seems that any inhibition of nsp5 must be strong enough to abolish all remaining ROs, otherwise, infection will commence again. The primary composition of the MLBs is now known, and it is possible that disruption of the MLBs as well could enhance the therapeutic effect.

Perspectives and limitations of the study

Although much has been learned about the SARS-CoV-2 infection of mammalian cells in this work, much remains to be discovered. One remaining set of puzzles concerns the single-stranded vgRNA that is likely released from the ROs. What are the partners, or chaperones, which protect this nucleic acid? Another possibility might be modifications of vgRNA which drive cytoplasmic exchange to virion-forming regions⁷¹. Such concepts suggest many additional questions, but with the proper assay and labeling of the key partners, we expect that super-resolution microscopy will help in answering these questions. As has already been demonstrated⁷², the extreme resolution of cryo-EM imaging can profit from the specificity guidance afforded by correlative fluorescence imaging. Specific intermediates could then be identified to enable sub-tomogram averaging. This also would allow for simultaneous observation of membrane structures, which are challenging to image in high resolution in the dense cellular context.

Of course, it will be important in future work to go beyond time-frozen snapshots as much as possible, and to image more of the key players in virion formation simultaneously, especially the membrane, N and M proteins, and vgRNA. This Atlas may also be helpful for studies of protein-protein interactions that can lead to improved drug design²⁹. For example, given the importance of the M protein pathway (Figure 5A–F), an inhibitor of M and the vgRNA/N complex binding could be useful. In any case, we hope that the methods presented in this Atlas are helpful also to future works studying virion creation for SARS-CoV-2.

Methods

Antibodies

Antibodies and the concentrations used are indicated in Table S1. During the initial testing of antibodies, we co-labeled infected cells with the anti-dsRNA antibody and checked for the fluorescence signal in both infected (dsRNA-positive) and non-infected (dsRNA-negative) cells using DL microscopy. If the antibody signal from the infected cells was too low or if the non-specific signal in the non-infected cells was too high, we rejected that antibody. For example, for working antibodies against nsp1, the average specific signal in the late infection stage was approximately 20× higher than the non-specific signal in the cytoplasm of non-infected cells. To confirm that the fluorophore attached to the secondary antibody does not produce artifacts, in several cases we switched the labels by switching the secondary antibodies and found no difference in the SR structures observed.

Culture of cell lines

The A549-ACE2 cells (human lung carcinoma cells expressing human angiotensin-converting enzyme 2, BEI Resources, NR-53821), Vero E6 cells (African green monkey kidney epithelial cells, ATCC, CRL-1586), Vero E6-Sec61β-GFP stable cells described in Ref.⁵, and Vero E6-TMPRSS2 cells (BPS Bioscience, 78081) were cultured in Dulbecco’s modified Eagle medium (DMEM) with GlutaMAX, 25 mM D-Glucose, and 1 mM sodium pyruvate (Gibco, 10569010) in 10% FBS (Sigma-Aldrich, F0926) at 37 °C and 5% CO₂ in a humidified incubator. Cell lines were not authenticated after purchase prior to use. For A549-ACE2 cells, blasticitin (Thermo Fisher, A1113903) was added at a final concentration of 100 μg/ml. For Vero E6-TMPRSS2, Geneticin (G418) was added to the culture medium at a final concentration of 1 mg/ml.

SARS-CoV-2 viral stocks preparation

The SARS-CoV-2 WA 1, isolate USA-WA1/2020 (NR-52281, BEI Resources) was passaged 3 times in Vero E6-TMPRSS2 cells as previously described^73,74. Briefly, a Vero E6-TMPRSS2 monolayer was infected with virus obtained from BEI; post 72 hours of infection (hpi), P1 virus-containing tissue culture supernatants were collected and stored at −80°C. Following titration, P1 virus stock was used to generate a P2 stock by infecting Vero E6 TMPRSS2 monolayers with multiplicity of infection (MOI) of 0.0001 for 72 hours. P2 virus was passaged again in Vero E6-TMPRSS2 cells to obtain P3 stock. Viral titers were determined by standard plaque assay on Vero E6 cells. All the experiments involving virus infection were performed using a P3 SARS-CoV-2 USA-WA1/2020, containing 100% WT population with no deletion in the spike multi-basic cleavage site.

Infection of cells by SARS-CoV-2

At the start of this project, cell infections were performed under BSL-3 conditions as noted below. In early 2025, the safety level for SARS-CoV-2 was reduced to BSL-2 in accordance with CDC regulations, so both conditions are described here as appropriate.

A549-ACE2 cells were plated into 18-well glass bottom μ-slides (ibidi, 81817) with 10k cells per well and cultured in DMEM culture medium containing 10% FBS and 100 μg/mL blasticidin the day before virus infection. Cells were taken to the BSL-3 facility, where they were washed with PBS once and incubated with SARS-CoV-2 WA 1 (USA212 WA1/2020) at an MOI of 2 in 50 μL DMEM culture medium containing 2% FBS per well for 2 hours. The medium containing virus was then removed, and cells were cultured in 100 μL DMEM culture medium containing 2% FBS per well for 24 hours. After that, cells were washed with PBS and fixed by 4% PFA (Electron Microscopy Sciences, 15710) and 0.1% glutaraldehyde (Electron Microscopy Sciences, 16350) in PBS for 1 hour and removed from the BSL-3 facility for further processing. This part of work was conducted at the high containment BSL-3 facility of Stanford University according to the CDC and institutional guidelines.

Infection of cells by SARS-CoV-2 for nirmatrelvir treatment experiments

A549-ACE2 cells were plated into 18-well glass bottom μ-slides (ibidi, 81817) with 10k cells per well and cultured in DMEM culture medium containing 10% FBS and 100 μg/mL blasticidin the day before virus infection. For virus infection, cells were washed with PBS once and incubated with SARS-CoV-2 WA 1 (USA212 WA1/2020) at an MOI of 10 in 50 μL DMEM culture medium containing 2% FBS per well for 6 hours. The medium containing virus was then removed and cells were cultured in 100 μL DMEM culture medium containing 2% FBS and 288 nM nirmatrelvir (MedChemExpress, HY-138687) per well for 18 hours. After that, some cells were washed with PBS and fixed by 4% PFA and 0.1% glutaraldehyde in PBS for 1 hour for further processing. Some cells were washed with DMEM culture medium containing 2% FBS and cultured in DMEM culture medium containing 2% FBS for another 24 hours. These cells were then washed with PBS and fixed by 4% PFA and 0.1% glutaraldehyde in PBS for 1 hour for further processing. This part of the work was conducted in a BSL-2 tissue culture room following Stanford University’s Administrative Panel on Biosafety (APB) protocols and guidelines from the CDC. EC₅₀ for this experiment was determined following a published protocol⁷⁵ in the BSL-3 facility.

Synthesis of the RNA FISH probes

vgRNA FISH probes targeting the ORF1a region of SARS-CoV-2²¹ were prepared using the same procedures as previously published⁵. RNA FISH probes targeting human 18S rRNA were designed using the Stellaris RNA FISH Probe Designer (Biosearch Technologies, Inc., Petaluma, CA) available online at https://www.biosearchtech.com/stellaris-designer (version 4.2). The sequences of the 38 probes are as follows: gagcgaccaaaggaaccata; accacagttatccaagtagg; tcggcatgtattagctctag; gggttggttttgatctgata; tatctagagtcaccaaagcc; gatagggcagacgttcgaat; tatttttcgtcactacctcc; cctcgaaagagtcctgtatt; tccaatggatcctcgttaaa; tcaaagtaaacgcttcgggc; attattcctagctgcggtat; acaaaatagaaccgcggtcc; ccgactttcgttcttgatta; ggtatctgatcgtcttcgaa; catcgtttatggtcggaact; tcccgtgttgagtcaaatta; cacccacggaatcgagaaag; aactaagaacggccatgcac; taaccagacaaatcgctcca; cagagtctcgttcgttatcg; gtcgcgtaactagttagcat; tgttattgctcaatctcggg; cacttactgggaattcctcg; atcaacgcaagcttatgacc; ctggcaggatcaaccaggta; tctttgagacaagcatatgc; tccggaatcgaaccctgatt; ttggatgtggtagccgtttc; ggtcgggagtgggtaatttg; gctttttaactgcagcaact; ctagcggcgcaatacgaatg; cgccggtccaagaatttcac; ggcgggtcatgggaataacg; tagggtaggcacacgctgag; caatcggtagtagcgacggg; atccgagggcctcactaaac; gatagtcaagttcgaccgtc; gttacgacttttacttcctc.

These probes were ordered with 5AmMC6 modifications from Integrated DNA Technologies, Inc. in plate format of 25 nmol scale with standard desalting. Each probe was dissolved in water to a final concentration of 100 μM. The same set of probes was combined with equal volumes of each probe to get a stock of 100 μM mixed probes. The mixed probes were further desalted using ethanol precipitation. Briefly, 120 μL 100 μM probes were mixed with 12 μL 3 M sodium acetate (pH 5.5), followed by 400 μL ethanol. After precipitation at −80C overnight, probes were pelleted through centrifugation at 12,000× g for 10 min at 4°C, washed with precooled 70% (vol./vol.) ethanol three times, air dried, and dissolved in water to make a 100 μM solution of probes. Then, 18 μL 100 μM probes were mixed with 2 μL 1 M NaHCO₃ (pH 8.5), followed by 100 μg Alexa Fluor^™ 647 succinimidyl ester (NHS) (Invitrogen, A37573) dissolved in 2 μL dry DMSO (Invitrogen, D12345). The mixture was incubated for 3 days at 37C in the dark for conjugation and purified for 3 rounds using Monarch PCR & DNA Cleanup Kit (5 μg) (NEB, T1030S) following the manufacturer’s instructions.

RNA FISH and immunofluorescence (IF)

RNA FISH and IF staining were performed following the procedures published previously⁵.

Spinning disk confocal microscopy

Confocal microscopy was performed at the Stanford University Cell Sciences Imaging Core Facility with a Nikon Ti2 Crest spinning disk confocal (SDC) microscope equipped with a Photometrics Kinetix camera, a Perfect Focus (PFS) focus lock system, and 365 nm, 488 nm, 561 nm, and 640 nm lasers. A 20× Plan Apo air objective (NA = 0.80, 0.333 μm/pixel) and a 60× Plan Apo oil objective (NA = 1.40, 0.108 μm/pixel) were used. The pinhole size was 50 μm. Z-series images were taken using a piezo Z-axis stage (Mad City Labs) and the NIS Elements software with Z stacks at 0.3 μm step size.

Confocal image analysis

Image processing was performed in Fiji (ImageJ). Projections of adjacent Z planes showing maximum loci fluorescence were generated. For conditions 1 and 2 in Figure S4A, the number of infected cells and total cells were both manually counted using the “multi-point” function in Fiji. For condition 3, the number of infected cells was manually counted. Each field of view has several hundred cells in condition 3. To count the total number of cells, a Gaussian blur filter with sigma = 3.5 pixels was first applied to blur the fluorescence image of nsp4 and nsp5. The following functions in Fiji “Image->Adjust->Threshold”, “Process->Binary->Fill Holes”, “Analyze->Analyze particles” were sequentially performed to measure the cell number, with the size threshold set to greater than 30 μm². The infection ratio was calculated using the number of infected cells divided by the total number of cells in each condition. Statistical analysis was performed using one-way ANOVA in GraphPad.

Optical setup for 2D 2- and 3-color SR microscopy

2D 2- and 3-color (d)STORM SR imaging was performed on a custom-built microscope⁵, using a Nikon Diaphot 200 inverted frame with an oil-immersion objective 60x/1.35 NA (Olympus UPLSAPO60XO) and a Si EMCCD camera (Andor iXon Ultra 897, 512×512 pixels). The sample was mounted on two stacked piezo stages (coarse U-780.DOS and fine P-545.3C8S, both Physik Instrumente). A 642 nm 1W continuous-wave (CW) laser (MPB Communications Inc.) was used for excitation of AF647 and CF680, and a 560 nm 1W CW laser (MPB Communications Inc.) for excitation of CF583R and CF568. For reactivation of fluorophores, a 405 nm 50 mW CW diode laser (Coherent OBIS) was used. The excitation power was set by adjusting the laser powers and, when necessary, by adding neutral density filters using a motorized filter wheel (FW102C, Thorlabs). All laser beams were coupled into a square-core multi-mode fiber with shaker for speckle reduction (Newport F-DS-ASQR200-FC/PC). The output tip of the fiber (200×200 μm² core size) was imaged with a 10×/0.25 NA objective and magnified to achieve a square illumination region of 47.6×47.6 μm² with a nearly constant intensity in the sample image plane.

The fluorescence was first split from the excitation light with a multi-band dichroic mirror (ZT405/488/561/640rpcv2, Chroma) and then further split into two spectral channels at $\mathrm{\lambda }\approx 690\mathrm{n}\mathrm{m}$ with a dichroic mirror (T690LPxxr, Chroma). The $\mathrm{\lambda }>690\mathrm{n}\mathrm{m}$ channel was filtered with ET700/75M (Chroma) and was used for simultaneous ratiometric imaging of CF680 and AF647. The $\mathrm{\lambda }<690\mathrm{n}\mathrm{m}$ channel was used either as the second channel for the ratiometric imaging of CF680 and AF647, or for imaging of CF583R. When used for ratiometry, this channel was filtered with a combination of ZET635NF and ET685/70M (both Chroma) and when used for CF583R imaging, it was filtered with 561LP and 607/70BP (both Semrock). For remote filter changing, the filters in the $\mathrm{\lambda }<690\mathrm{n}\mathrm{m}$ channel were mounted in a second motorized filter wheel (FW102C, Thorlabs). The sample image was produced with a tube lens ($f=400\mathrm{m}\mathrm{m}$) and then relayed with a set of lenses ($\mathrm{f}=120\mathrm{m}\mathrm{m}$) in both channels to be focused side-by-side on the camera chip with a pixel size of 117 × 117 nm² in the sample coordinates.

Axial drift was compensated with a custom focus lock system using an infrared beam reflected from the coverslip-sample interface as previously described⁵. The microscope was controlled using a custom code programmed in MATLAB (MathWorks, Inc.), which includes the control of the focus lock system and the control of all lasers, laser shutters, filter wheels and the fine piezo stage. The data from the EMCCD camera was acquired using Andor SOLIS 4.31.30024.0.

SR imaging procedure

For (d)STORM, each well of the 18-well chamber was filled with 200 μl of a photoblinking buffer consisting of 200 U/ml glucose oxidase, 1000 U/ml catalase, 10% w/v glucose, 200 mM Tris-HCl pH 8.0, 15 mM NaCl and 50 mM cysteamine. The buffer was prepared freshly by mixing 3 stock solutions¹⁷: 1) 4 kU/ml glucose oxidase (G2133, Sigma), 20 kU/ml catalase (C1345, Sigma), 25 mM KCl (P217, Fisher), 4 mM TCEP (646547, Sigma), 50% v/v glycerol (BP229, Fisher) and 22 mM Tris-HCl pH 7.0 (BP1756, Fisher), stored at −20 °C; 2) 1 M cysteamine-HCl (30080, Sigma), stored at −20 °C; 3) 37% w/v glucose (49139, Sigma) with 56 mM NaCl (S271, Fisher) and 0.74 M Tris-HCl pH 8.0 (J22638.AE, Fisher), stored at +4 °C. For samples with RNA FISH labeling, the buffer was supplemented with 1 U/μl of an RNase inhibitor (302811, LGC Biosearch Technologies) and the catalase from bovine liver (C1345, Sigma) was replaced with catalase from Aspergillus niger (219261, Sigma) as this was found to be less contaminated with RNases. To avoid sample bleaching in wells that are not being imaged during SR imaging of neighboring wells, those were kept in an oxygen scavenging buffer containing 100 U/ml glucose oxidase, 500 U/ml catalase, 4.6% w/v glucose, 93 mM Tris-HCl pH 8.0 and 7 mM NaCl. After removing the imaging buffer or the oxygen scavenging buffer, to prevent bacterial growth, the samples were stored in PBS with 0.02% sodium azide (cs-296028, ChemCruz) at +4 °C.

Before SR imaging, the ROI was selected using standard DL wide-field fluorescence. For 3-color (d)STORM experiments, we first simultaneously imaged CF680 with AF647 in the two spectral channels, using the 642 nm excitation and the power density of ~20 kW/cm², and then switched to CF583R, at ~13 kW/cm² of 560 nm. For 2-color (d)STORM, we first imaged AF647 using only the $\mathrm{\lambda }<690\mathrm{n}\mathrm{m}$ channel, at 9–20 kW/cm² of 642 nm, and then CF583R at ~13 kW/cm² of 560 nm. When the SM blinking density was low, the sample was additionally illuminated with the 405 nm light at up to 50 W/cm² for faster fluorophore reactivation. We used an exposure time of 10.57 ms per frame and the EM gain of 84. The image recording started after the initial shelving phase, after the beginning of clear SM blinking; the blinking movies were acquired for approximately 8 · 10⁴ frames for each fluorophore imaged separately, or for ~1.6 · 10⁵ frames for simultaneously imaged CF680 with AF647.

SR data analysis

SM movies were first processed with the ThunderStorm plugin^76,77 for Fiji⁷⁸ using the following parameters: image filtering with wavelet filter, b-spline order = 3, scale = 2; approximate localization using a local maximum with a threshold of A·std(Wave.F1) and 8-neighbourhood connectivity; sub-pixel localization using an Integrated Gaussian fitting with a fitting radius of 4 pixels, weighted least squares method with the initial sigma of 1.1 pixels. A = 0.9 for both spectral channels for simultaneously imaged CF680 with AF647, or A = 1.5 – 2 for fluorophore imaged separately. After fitting, the data for separately imaged fluorophores was corrected for drift using cross-correlation in ThunderStorm and then filtered using sigma < 200 nm. If drift correction in ThunderStorm failed, we corrected drift for those datasets using SharpViSu⁷⁹. For simultaneously imaged CF680 and AF647, the drift was not corrected at this stage; localizations from both channels were filtered using sigma > 80 nm and sigma < 200 nm, and duplicate localizations close than 100 nm were removed. Localizations of simultaneously imaged CF680 and AF647 were then demixed in SplitViSu¹⁷ and drift was corrected in SharpViSu⁷⁹ using cross-correlation between temporal subsets of the sum of CF680 and AF647 localizations.

For further processing, we kept only localizations with fitted Gaussian sigma between 160 nm and 80 nm to reject out-of-focus localizations. Localizations found within 50 nm on consecutive frames that could originate from multiple localizations of a single molecule were treated in two ways. For SR images, to improve the resolution, these localizations were refined by selecting them from a normal distribution with a mean that equals to the weighted mean of the initial localizations ($\mathrm{w}\mathrm{e}\mathrm{i}\mathrm{g}\mathrm{h}\mathrm{t}\mathrm{s}{\mathrm{w}}_{\mathrm{i}}=\surd \left({\mathrm{N}}_{\mathrm{p}\mathrm{h},\mathrm{i}}\right)$) and a standard deviation (SD) that equals to ${\mathrm{\sigma }}_0\cdot {\left(\sum N_{\mathrm{p}\mathrm{h},\mathrm{i}}\right)}^{-1/2}$, where ${\mathrm{\sigma }}_0$ is the SD of the localization position in the given consecutive series and $N_{\mathrm{p}\mathrm{h},\mathrm{i}}$ is the number of photons acquired from the i-th localization in the series¹⁷. For data analysis other than SR image reconstruction, to suppress overcounting, the localizations of the consecutive series were reduced to a single localization at the weighted mean position. After this correction, to suppress spurious localization of sparse background from photoblued far-red fluorophores⁸⁰ or non-specific binding, the SR data of CF583R was additionally filtered by removing localizations that had 4 or less neighbors within 30 nm.

For image registration, we imaged 200 nm TetraSpeck beads (T7280, Thermo Fisher Scientific) in both the far-red and the CF583R channels and found the positions of the beads by fitting in ThunderStorm as described above. The transformation between the channels was calculated using an affine transformation with help of MATLAB function ‘fitgeotrans’. The calculated transformation was then applied to CF583R localizations using a MATLAB function ‘transformPointsInverse’. After all corrections, final SR images were reconstructed as 2D histograms with a bin size of 20 × 20 nm² (typically used for the whole-cell images in the figures) or 16 × 16 nm² (typically used for the zoomed-in images in the figures).

3D DHPSF imaging

Two-color 3D microscopy was performed with a previously described, homebuilt, 3D single-molecule imaging microscope⁸¹. Briefly, 647 nm, 561 nm, and 405 nm lasers are coaligned and focused onto the fiber input of a multimode fiber de-speckler. The fiber output is imaged into sample plane by reflecting off a quad-pass dichroic and passing through a 100X, 1.4NA oil-immersion objective lens, producing a 35-μm diameter circular illumination pattern at the sample. Collected fluorescence emission passes through the quad-pass dichroic and is relayed to the camera via a two-channel 4f system with a 660 nm long-pass dichroic mirror splitting the emission. Each arm of the emission pathway has a double-helix phase mask⁸² (Double-Helix Optics) placed in the Fourier plane to produce the double-helix point-spread function (DHPSF) response in each channel on the detector. A knife-edge prism mirror is used to direct each channel onto opposite quadrants of an EMCCD. Samples were first imaged with 647 nm illumination (7.5 kW/cm²) with 17 ms exposure and 272 EM gain followed by 561 nm illumination (6.9 kW/cm²) with 15 ms exposure and 272 EM gain. In both cases, 405 nm illumination was gradually increased from 0.6 W/cm² to 60 W/cm² intensity to photoactivate molecules as emitter density decreased.

After cellular imaging, 100 nm TetraSpeck beads (Invitrogen, T7279) distributed in 1% agarose in water (w/v) were imaged for calibrations. First, the bead sample was laterally scanned to image ~3000 beads evenly distributed through the imaging volume. A z scan (3.2 μm range, 50 nm steps) for a single bead in the center of the field of view was then acquired simultaneously in each channel to compute a calibration to map DHPSF lobe angle to z.

Single-molecule DHPSF data was analyzed with homebuilt DHPSF detection and fitting software in PYME available at https://doi.org/10.5281/zenodo.16877694. For each frame, the background is estimated by the median of the previous 30 frames and then subtracted. Candidate DHPSF ROIs for fitting are then extracted via a steerable-filter-based algorithm. Each candidate ROI was then least-squares fit with a model function of the DHPSF with fitting parameters of lobe sigma, lobe separation, lobe amplitude 1, lobe amplitude 2, background, and lobe angle, which was then mapped to z using the calibration. After localizing cellular data, TetraSpeck bead data was localized and used to compute a 3D locally weighted mean quadratic transformation to map 561 nm channel localizations to the 647 nm channel. Localizations were then registered, drift-corrected via cross correlation, and scaled in z by a factor of 0.7 to account for focal shift due to the refractive index mismatch between the aqueous medium and the cover glass. 647 nm channel localizations are then filtered for 140 nm<lobe sigma<275 nm and 900 nm<lobe separation<1400 nm. 561 nm localizations were then filtered for 125 nm<lobe sigma<245 nm and 700 nm<lobe separation<1200 nm.

Bivariate pair-correlation functions

The bivariate pair-correlation functions ${\mathrm{g}}_{12}\left(r\right)$⁸³ were calculated by counting the number of localizations of the second species within a distance between $r$ and $r+dr$ from each localization of the first species⁵. These counts were then normalized by dividing the number of localizations by the area of the corresponding rings of radii $r$ and $r+dr$ and by the average density of the second species in the region. Finally, the obtained numbers were averaged across the localizations of the first species. $r$ was scanned over the range between 0 and 500 nm and $dr$ was set to 1 nm. For the complete spatial randomness (CSR) case, a test CSR dataset was generated with the same average density as for the experimental case across the same ROI. ${\mathrm{g}}_{12}\left(r\right)$ traces were calculated from these CSR datasets as described above. No edge effect correction was performed leading to a slight decrease of ${\mathrm{g}}_{12}\left(r\right)$ at large $r$. Plots in the figures display experimental and CSR ${\mathrm{g}}_{12}\left(r\right)$ for each analyzed cell as faint lines as well as the mean ${\mathrm{g}}_{12}\left(r\right)$ calculated from all cells in bold lines.

Radial density distribution functions for 3-color SR images of DMVs

Radial density distribution was evaluated for each dsRNA cluster using the centroid position of dsRNA as the origin. First, a ROI was manually drawn in a SR image selecting DMV-like dsRNA clusters surrounded by nsp4 localizations. Irregularly shaped dsRNA clusters or those without nsp4 were excluded. The dsRNA centroid positions were calculated by fitting the SR images of dsRNA (pixel size = 20 × 20 nm²) in ThunderStorm using the following parameters: image filtering with wavelet filter, b-spline order = 4, scale = 4; approximate localization using a local maximum with a threshold of 0.5·std(Wave.F1) and 8-neighbourhood connectivity; sub-pixel localization using an Integrated Gaussian fitting with a fitting radius of 5 pixels, weighted least squares method with the initial sigma of 2 pixels. For datasets containing vgRNA, the fitting was performed on images combing both dsRNA and vgRNA. The SR images of both vgRNA and dsRNA were first filtered with a Gaussian filter with $\mathrm{\sigma }=0.5\mathrm{p}\mathrm{x}$, and the combined image was calculated as a sum of pixel values of the two filtered SR images. After fitting, the localizations of the dsRNA centroids were filtered by removing localizations found closer than 3·sigma and keeping only those with sigma > 15 nm and with the number of SM localizations > 200.

For each dsRNA centroid $i$ in a given labeling condition, the number of localizations of the three co-imaged targets within a distance between $r$ and $r+dr$ (where $dr=1\mathrm{n}\mathrm{m}$) from the centroid was counted. These counts were then normalized by the area of the corresponding concentric rings of radii $r$ and $r+dr$, which produces a distribution proportional to the line profile of the SR image that passes through the dsRNA centroid, averaged across all angles. The radial distributions for each $i$ were smoothed using the MATLAB “smoothdata” function with the “gaussian” method and a window size of [6 6]. The resulting radial distribution curves were normalized such that the peak of their sum would equal unity, and the individual curves were additionally multiplied by $n$ (the number of all analyzed clusters $i$ in the dataset) yielding ${\mathrm{g}}_{\mathrm{i}}\left(r\right)$. For each radial position $r$, the 95% CI was calculated as $1.96\cdot \mathrm{S}\mathrm{D}\left[{\mathrm{g}}_{\mathrm{i}}\left(r\right)\right]/\mathrm{\surd }\mathrm{n}$, where $\mathrm{S}\mathrm{D}\left[{\mathrm{g}}_{\mathrm{i}}\left(r\right)\right]$ is the SD calculated over the normalized and rescaled curves ${\mathrm{g}}_{\mathrm{i}}\left(r\right)$ of all clusters $i$ for the given $r$. The final radial density distribution function $\mathrm{g}\left(r\right)$ was calculated as the mean value of ${\mathrm{g}}_{\mathrm{i}}\left(r\right)$. To find the peak position $R$ of $\mathrm{g}\left(r\right)$, a Gaussian function was fitted to $\mathrm{g}\left(r\right)$ in the vicinity of the peak, typically using values down to $~0.9\cdot \mathrm{g}\left(R\right)$ at the left side of the peak and to $~0.6\cdot \mathrm{g}\left(R\right)$ at the right side of the peak, and the fit was checked visually for correct peak fitting. The precision of peak fitting was evaluated using bootstrapping: we randomly replaced ${\mathrm{g}}_{\mathrm{i}}\left(r\right)$ with repetitions 1000 times and fitted the peak positions of the replaced mean distributions ${\mathrm{g}}_{\mathrm{j}}\left(r\right)$, $j=1\dots 1000$ as described above. The SEM of the peak position was calculated as the SD of these 1000 peaks $R_j$ and the CI was represented as $R\pm 1.96\cdot \mathrm{S}\mathrm{D}\left(R_j\right)$.

Bivariate angular pair-correlation functions

Bivariate angular pair-correlation functions $\mathrm{p}\left(\mathrm{\theta }\right)$ were calculated using the same ROIs as for $\mathrm{g}\left(r\right)$. Around each dsRNA centroid, only localizations at a distance $0.5·{\sigma }_{dsRNA}<r<4·{\sigma }_{dsRNA}$ were considered, where $\mathrm{r}$ is the distance of a localization from the dsRNA centroid position and ${\sigma }_{dsRNA}$ is the width of the dsRNA cluster, obtained from the Gaussian fitting in ThunderStorm as described above. The angles of the localizations were calculated using the ‘atan2’ function and the dsRNA centroid as the origin of polar coordinates. The angular distance between all localizations of nsp4 and of the second target was calculated for every dsRNA cluster. $\mathrm{p}\left(\mathrm{\theta }\right)$ was estimated by calculating the histogram of these angular distances with a bin width of $\mathrm{\pi }/100$ and normalizing such that $\mathrm{p}\left(\mathrm{\theta }\right)\approx 1$ for non-correlated localizations.

Statistics and Reproducibility

The experimental measurements were replicated at least 2–3 times for each labeling combination, starting with cell culture and viral infection (independent biological replicates). After initial optimizations, further biological replicates were successful. Each sample preparation resulted in several chambers with up to 18 wells with different target combinations. For each prepared sample well (each label combination), typically 5–15 cells were imaged using SR microscopy. Because some targets appear in many combinations of labels, each of these targets was imaged multiple times. Table S1 lists the number of cells imaged using each particular primary antibody.