Conformational ensemble of the NSP1 CTD in SARS-CoV-2: Perspectives from the free energy landscape

Abstract

The nonstructural protein-1 (NSP1) of the severe acute respiratory syndrome-associated coronavirus 2 plays a crucial role in the translational shutdown and immune evasion inside host cells. Despite its known intrinsic disorder, the C-terminal domain (CTD) of NSP1 has been reported to form a double $\alpha$-helical structure and block the 40S-ribosomal channel for mRNA translation. Experimental studies indicate that NSP1 CTD functions independently from the globular N-terminal region separated with a long linker domain, underscoring the necessity of exploring the standalone conformational ensemble. In this contribution, we utilize exascale computing resources to yield unbiased molecular dynamics simulation of NSP1 CTD in all-atom resolution starting from multiple initial seed structures. A data-driven approach elicits collective variables (CVs) that are significantly superior to conventional descriptors in capturing the conformational heterogeneity. The free energy landscape as a function of the CV space is estimated using the modified expectation maximized molecular dynamics. Originally developed by us for small peptides, here, we establish the efficacy of expectation maximized molecular dynamics in conjunction with data-driven CV space for a more complex and relevant biomolecular system. The results reveal the existence of two disordered metastable populations in the free energy landscape that are separated from the conformation resembling ribosomal subunit bound state by high kinetic barriers. Chemical shift correlation and secondary structure analysis capture significant differences among key structures of the ensemble. Altogether, these insights can underpin drug development studies and mutational experiments that help induce population shifts to alter translational blocking and understand its molecular basis in further detail.

Significance

SARS-CoV-2, the causative agent in the COVID-19 pandemic, encodes 16 mature nonstructural proteins (NSPs) besides the structural proteins that attribute shape to the viral particles. Although an intrinsically disordered region, the NSP1 C-terminal domain (CTD) is reported to block the host mRNA entry channel in the 40S ribosomal subunit by forming a two-helix conformation. Here, we investigate the conformational ensemble of NSP1 CTD to reveal the thermokinetics in its order-disorder transitions. Our analyses determine the existence of two distinct unfolded populations separated by high kinetic barriers from the predominant one that is compatible with complex formation with the ribosome. Structural insights from this study may help identify suitable drugs to impair the ribosomal blocking mechanism.

Introduction

The severe acute respiratory syndrome-associated coronavirus 2 (SARS-CoV2) virus, causative agent of the novel coronavirus (CoV) pandemic, belongs to the genus Betacoronavirus under the subfamily Coronavirinae and family Coronaviridae (1). This pathogen shares many similarities with SARS-CoV and middle east respiratory syndrome-associated coronavirus (MERS-CoV) (2). Unlike other RNA viruses, CoVs contain an ∼30 kb long genome in the form of a positive-sense single-stranded RNA. Starting from the 5′ end, this includes two overlapping open reading frames (ORF1a, ORF1b) followed by the sequence that codes for the membrane, envelope, nucleocapsid, and spike proteins of CoV (3). Sixteen mature nonstructural proteins (NSPs) are generated after proteolytic cleavage of the precursor polyproteins translated from the ORFs. The NSPs’ primary function is related to the replication and transcription of the viral genome and invasion of the host immune system (4,5,6,7). Among them, NSP1 is one of the first sets of proteins that are expressed at the earliest stage of host infection (8). This protein is reported to block the mRNA entry into the ribosomal machinery and help in host mRNA degradation by as yet unknown mechanisms (9). NSP1 is also known to inhibit interferon expression, and other antiviral responses triggered by the host immune system (9,10). Further research and mutational studies suggest that the mRNA translation inhibition mechanism may be specific to SARS-CoV-2, and yet closely related to other members of the Betacoronavirus family owing to their unique NSP1 C-terminal domain (CTD) sequence (2,11).

The critical role of NSP1 in the viral life cycle makes it an important drug target and the subject of studies at molecular resolution aimed at better understanding of the disease (12,13,14). Structural investigations show that it is characterized by a globular N-terminal domain and intrinsically disordered C-terminal region (9,15,16). The presence of a long and flexible linker between these two domains allows the N-terminus to move around 60 Å apart from the other termini (11). Interestingly, despite the disorder in the CTD, it is reported to form a double $\alpha$-helix structure while blocking the translational machinery (9,11). Studies have captured the NSP1 C-terminal interaction with the head and body of the 40S ribosomal subunit and with rRNA to effectively block the mRNA entrance channel (11,17). Moreover, mutation and deletion experiments reveal that the CTD is necessary and sufficient in the translation inhibition and, when docked, the N-terminal domain does not interfere with its function (11). Such observations can propel further investigations of the NSP1 CTD in isolation and its structural response to various environmental factors (16).

Although experiments indicate a stable double helix conformation of NSP1 CTD in complex with ribosomal subunit, the isolated NSP1 is not yet structurally characterized for its conformational disorder (16). It is plausible that the CTD may sample multiple secondary conformations in addition to the one compatible with ribosomal complex formation. A thermokinetic picture of the structural ensembles is necessary for unraveling the binding mechanisms of NSP1 CTD to the ribosomal subunit and identifying potential conformations for the drug target. Recent studies have harnessed molecular simulations in revealing potential interactions between NSP1 and SARS-CoV-2 RNA that helps it evade translational shutdown (11). Other efforts in similar directions have characterized mutations critical for NSP1-ribosomal interactions, and its structural disorder (18). Furthermore, computational studies can provide insights into small-molecule inhibitors and potential drug design against NSP1 and their thermodynamics of binding (19,20,21,22). However, studies on the thermokinetic properties of the intrinsically disordered region (IDR) are sparse; such studies are essential for drug development efforts.

Conformational exploration of disorder-prone proteins such as NSP1 pose manifold challenges. In addition to appropriate sampling, the challenges primarily lie in the identification of key structural elements, in the design of appropriate system descriptors, and in the quantitative estimation of the kinetic barriers that sequester states with varying biological functions. The advent of exascale computation complemented with developments of smart dimensionality reduction techniques has ameliorated many of these hurdles significantly (23,24,25,26). Yet, the paucity of generalized platforms for capturing the key thermokinetic features of intrinsically disordered protein and IDR landscapes may offset the benefits of enhanced computational power (27,28). In the current contribution, we present a molecular dynamics (MD)-based workflow designed to effectively capture the structural diversity and the free energy landscape (FEL) of the NSP1 CTD as a function of low-dimensional geometrical descriptors. We assess the performance of data-driven descriptors, or collective variables (CVs), in differentiating conformational signatures of the NSP1 C-terminal states against traditional descriptors. We further estimate the conformational FEL of the system by incorporating a modified version of the recently developed expectation maximized molecular dynamics (EMMD) (29). Introduced by us for efficient estimation of free energy barriers in rare transitions within biomolecular systems, EMMD determines Bayesian optimized estimates for poorly sampled regions within the FEL. This method was recently leveraged to determine the kinetic barriers demarcating catalytic substates of a tyrosine kinase (30).

Our results indicate that the FEL of the isolated NSP1 CTD accommodates a large number of thermally accessible conformational substates. In addition, we report the emergence of a metastable state isolated from the other basins by large kinetic barriers. We describe the remarkable similarity between this state and the experimentally reported structure of NSP1 in complex with the ribosomal RNA that is correlated with the onset of virulence (9,11). Our results can be expected to help design molecular inhibitors that bias the system against the virulence-prone state of NSP1. They may further help streamline generalized workflows for IDR conformational search. The advent of exascale computing has accelerated the exploration of complex biomolecular substates (31,32). Statistical learning methods such as EMMD can be leveraged to quantitatively examine plausible thermokinetic interconvertibilities and thereby enhance overall information content.

Materials and methods

Modeling an initial three-dimensional atomistic structure of a disordered protein sequence is highly challenging due to the known experimental limitations. In such scenarios, a linear structure can be generated from the sequence using classical molecular modeling tools. However, the generated structure may not be an optimal sample to initiate exploring the native FEL of the system. Here, we have used the PepFold3 web server (33,34), which has been used in earlier studies to generate stable starting conformations of IDRs (16,35,36). PepFold3 uses a coarse-grained force field and performs 200 simulations from which the minimum energy structures are selected (Fig. 1). The subsequent structure was placed in a dodecahedron box with water molecules. The TIP3P water model (37) and Amber99sb-ildn force field were used (38). The system was solvated at 0.15 M NaCl salt concentration and then neutralized by adding counterions. All the simulations in the current work are performed using the Gromacs 2021.1 MD engine (39). The peptide was minimized for 50,000 steps using the steepest descent algorithm. The minimum distance between the atoms of the peptide and the box edge was 12 Å. Particle mesh Ewald (40) was used to calculate long-range electrostatic interaction with a cutoff of 9 Å. Bonds with H atoms were constrained using LINCS (41), and periodic boundary conditions were employed in all the simulations. Sixty-four replicas of the system were selected in the temperature range 300.0–470.3 K such that they offer a replica exchange rate of approximately 0.2 (42). Each replica was first equilibrated at its respective temperature in an NVT ensemble for 300 ps with restraints of 1 kcal mol⁻¹ Å ⁻² and then for another 200 ps with no restraints. A velocity rescaled (modified Berendsen) thermostat (43) was used for temperature coupling. The system was then further equilibrated for 500 ps in the NPT ensemble without any restraints using the Parrinello-Rahman barostat (44) at 1 atm. Thereafter, 200 ns of replica exchange molecular dynamics (REMD) simulations were performed with a time step of 2 fs. An exchange between replicas was attempted every thousand time steps. The energies and coordinates were saved every 10 ps for analysis.

Figure 1.

Flowchart of the methods systematically used in this study. To see this figure in color, go online.

In this work, we utilized the sampling enhancement offered by REMD (45) for the initial exploration of the structural diversity featured in the FEL of NSP1 CTD. The resultant trajectories were used to generate seed structures for further unbiased MD simulations (Fig. 1, step 2). A total of 250 independent simulations (each of 50 ns) were performed starting from different initial conformations selected from REMD trajectories. For these simulations, the energy and coordinates were stored every 2 ps for further analysis. In aggregate, 12.5 $\mu$s of unbiased MD trajectories (Fig. 1, step 3) were analyzed to explore the thermokinetics underlying the conformational heterogeneity of the IDR.

For the purpose of projecting the system’s conformational space over low-dimensional degrees of freedom, we have used the principal-component analysis (PCA) and time-structured independent component analysis (TICA). While generating seed structures from the REMD trajectory, we first performed PCA over the backbone dihedrals (dPCA) extracted from the 300 K ensemble. The dPCA was performed on the dihedrals transformed over the sine and cosine functions, and the trajectory was projected over the first two principal components. In PCA, the following function is maximized from the input data:

$$\mathit{Var}\left(\mathit{x}\right)=\frac{1}{N}\sum _{i=1}^N{\left(x_i-\overline{x}\right)}^2$$ (1)

where Var(x) is the variance in the data along the x dimension, and $x_i$ is the ith timestep in the simulation trajectory. i varies from 1 to N where the simulation has N timesteps (46,47,48). To generate seed structures from the projected conformational space, the ensemble was clustered using the k-means algorithm with 250 clusters. One randomly chosen conformation from each cluster contributed as a seed for further unbiased simulation described earlier.

TICA (49,50) was performed while analyzing the unbiased trajectories, and an FEL was constructed by transforming the resultant data along the independent component (IC) axes. TICA, similar to PCA, is also used as a dimensionality reduction technique. However, it maximizes a time-lagged autocorrelation function given time series input data and finds the slow modes of the dynamical system. The time-lagged autocorrelation function maximized by TICA is as follows:

$$Corr\left(x\right)=\frac{1}{{\sigma }_x^2\left(N-\tau \right)}\sum _{i=1}^{N-\tau }\left(x_i-\overline{x}\right)\left(x_{i+\tau }-\overline{x}\right)\text{.}$$ (2)

Here, Corr(x) is the time-lagged autocorrelation function along x dimension, $x_i$ is the value of $x$ at $i$th time point and $x_{i+\tau }$ is the value of $x$ at some (i + $\tau$) time. Here, t is the lagged interval which is chosen in such a way that it is responsible for capturing the slow degrees of freedom in the data. i here varies from i to (N − $\tau$). ${\sigma }_x$ is the SD along x. In general, the global transitions in complex systems are associated with their slow dynamical modes. Therefore, it is important to assess and compare the extent of slowness associated with each CV. In this regard, the variational approach to the Markov processes (VAMP)-based scoring system is utilized to rank the CV spaces according to their inherent slowness, where a higher score represents a slower CV space. We report the results related to VAMP calculations in the next section, while other details are discussed in the supporting material.

Traditionally, FELs are computed by the discrete histogramming method over a predefined CV space. A poorly sampled region in such landscapes may remain noninformative, and the free energy values for such regions may remain obscured. To resolve this problem, herein, we use expectation maximized molecular dynamics (EMMD) (29,30) (Fig. 1, step 5), a histogram-free approach that provides a maximum likely free energy estimate for the unsampled regions in a CV space without enhanced sampling. In this context, initial conformations of a system are independently explored to yield a collection of disjoint state distributions. A continuous probability density as a function of the CV space is then constructed by fitting Gaussian mixture models (GMMs) over the disjointly sampled metastable basins. This model density function provides estimates for the probabilities of poorly sampled transition regions in the conformational space. At this stage, EMMD utilizes the Boltzmann inversion of the continuous density function for estimating the free energy barriers among various metastable states. Therefore, at absolute temperature $T$, the FEL as a function of the CV space is computed from the effective probability as

$$FE{L}_{CV}=-k_{B}T\ln \left(\sum _{k=1}^M{\pi }_k\mathcal{N}\left(CV\mid {\mu }_k,{\mathrm{\Sigma }}_k\right)\right)$$ (3)

where $k_B$ is the Boltzmann constant, $M$ is the total number of element distributions in the mixture, ${\pi }_k$ is the weight factor of the k-th multivariate Gaussian $\left(\mathcal{N}\right)$ having mean ${\mu }_k$, and covariance ${\mathrm{\Sigma }}_k$. In the currently used modified version of EMMD, we fit a variational Bayesian GMM that incorporates inferences from the Dirichlet process to the previous distributions related to the model, helping to produce a more regularized estimate of the state probability distribution, avoiding under/overfitting issues (51,52). The complete theoretical framework and applications are found in earlier work (29,30). The EMMD program (along with examples) are available for public use at https://github.com/Pallab-Dutta/EMMD.

Results and discussions

Conformational sampling

Theoretically, the structural heterogeneity of an IDR can be well explored with the REMD simulations since it allows conformational exchanges among simulated replicas in parallel. The overlapping potential energy distributions (Fig. S1) for each adjacent temperature pair ascertains the extent of exchanges. To evaluate the performance, the replica indices were tracked throughout the simulation. Fig. S2 depicts the frequent exchanges among different replica indices as a function of time. The average exchange probability of $\sim 0.3$ between adjacent replicas was calculated from the simulation. The increasing radius of gyration (Rg) of the IDR as a function of temperature is indicative of enhanced exploration of extended and disordered conformations. However, at each temperature, the equilibration of simulation is confirmed by observing the backbone root mean-square deviation (RMSD) and Rg with time (Figs. S3 and S4). In this work, the RMSD calculation for each replica is done using the corresponding initial structure of that replica after minimization. This ensures that the REMD-generated conformational ensemble can offer NSP1 CTD structures from the equilibrium population that can be used as seed structures to initiate the next set of unbiased sampling.

Generating initial structures from the REMD-generated structural ensemble of the IDR first requires a low-dimensional representation of its conformational space. In the case of such disordered domains, backbone dihedrals of the polypeptide are considered to be important descriptors of the system as they encode the propensity of different secondary structures (53,54). We perform a dPCA where the variance of the state space is maximized over a linear combination of the sines and cosines of these dihedrals. Although considering trigonometric functions for PCA doubles the dimensionality of the input coordinates, it is a crucial step to avoid unnecessary artifacts otherwise raised due to the periodicity of the dihedrals. The REMD trajectory is then projected over the first two principal components (dPC1, dPC2). Fig. 2 A shows the scatterplot of the state space sampled with REMD and projected onto the dPC1 and dPC2. The structural distribution in Fig. 2 A corresponds to a set of conformational clusters.

Figure 2.

(A) Scatterplot of REMD trajectories obtained at 300 K; projected over the first two dihedral principal components (dPC1 and dPC2). (B) Two hundred and fifty clusters were generated with the k-means algorithm; one structure was chosen at random from each cluster for further unbiased simulations. To see this figure in color, go online.

Before spawning unbiased simulations, we categorized the elements of the state space by k-means clustering, a method used extensively in earlier studies (54). This resulted in 250 distinct cluster centers. The fitted clusters are presented in Fig. 2 B with differently colored patches. Here, the same color is often used in multiple disjoint clusters because of the limited option in visually separable colors. One random member from each of the 250 clusters was selected and recomposed into the corresponding structure of NSP1 CTD. These confirmations were used to initiate the next set of unbiased sampling that yielded a 12.5 $\mu$s of trajectory in aggregation.

CVs and FELs

Compared with the predominantly funnel-shaped FEL of globular proteins, an IDR such as NSP1 is expected to be characterized by multiple minima corresponding to various conformational populations resulting in a relatively shallower FEL. In general, in an IDR, the probability of sampling individual features reduces with the increase in the number and diversity of subpopulations. Whereas a standard discrete FEL can effectively capture the complex pattern of features distributed under a metastable basin, often, the free energies corresponding to poorly sampled barriers remain uncaptured. At this stage, we use the modified EMMD to get the maximum likelihood estimates of the poorly sampled features in the FEL. The modified EMMD fits variational Bayesian GMM over the disjointly sampled unbiased trajectories projected over a CV space and infers a regularized prediction for the free energy values. However, due to regularization that avoids overfitting, the Gaussian mixtures may underestimate the highly complex pattern formed under each metastable basin. Therefore, we present the discrete FEL along with its EMMD estimated form to elicit detailed features of the landscape.

We first considered conventional yet physically interpretable system descriptors generally used to study IDRs. The Rg and the fractional $\alpha$-helical content ($\alpha$H) were considered, given that together they describe the overall shape and secondary structure of the system. In the current contribution, the values of $\alpha$H are determined with VMD (56) via the colvar tools (57). Previous studies have substantially used these variables to construct a CV space while studying small peptides and disordered protein sequences (58,59,60). We report the FEL as a function of the (Rg, $\alpha$H) space in Fig. 3. We notice that the FEL is composed of a deep minimum marked as “a” accompanied by “b” and two shallower minima with a relative free energy difference of $\sim 1k_{B}T$ among them ($k_B$ is Boltzmann constant, $T$ is 300 K). A free energy barrier of $\sim 1-2k_{B}T$ is observed while accessing the minima “b,” “c,” and “d” from the deeper basin. Such a CV selection presents a rather funnel-shaped FEL when projected on the individual variables (Fig. S8, A and B) that does not correctly explain the disorder in NSP1 CTD observed via both experimental and computational studies (16,18). Moreover, by design $\alpha$H cannot differentiate states with similar helical content at different regions of the sequence. This indicates the probable existence of degenerate states that may be distinguished upon the projection of the FEL on more efficient CVs. The results, therefore, highlight the necessity of considering other independent degrees of freedom that can reconcile more specific structural aspects.

Figure 3.

FEL of NSP1 CTD as a function of fractional $\alpha$-helical content ($\alpha$H) and radius of gyration (Rg). The local minima are labeled from “a” to “d.” A representative structure from each local minimum is also shown. The $\alpha$-helix, turn, and random coil are determined via the STRIDE algorithm and colored with purple, teal, and white, respectively (55). To see this figure in color, go online.

As discussed earlier, backbone dihedrals of a polypeptide are considered crucial descriptors for an IDR. Although the CVs generated with dPCA maximize the variance, the reduced dimension may not necessarily capture the global structural transitions that may occur over the slow modes of the system. Moreover, we observed that the projection of the system FEL over the first two dPCs (supporting material) obtained by analyzing the unbiased trajectories could only explain $\sim$16% of the total variance present in the system. The introduction of TICA was one of the earliest attempts toward establishing a method for the determination of slow CVs. Similar to dPCA calculations, here, we consider the sines and cosines of the backbone dihedrals while performing the TICA calculations with a lag time of 10 ps. The first two ICs, dIC-1 and dIC-2, offer the orthogonal degrees of freedom that maximize the autocorrelation of the system. To assess the inherent slowness associated with the CV spaces, we performed VAMP calculations using lag times of 0.1, 0.5, and 1 ns for each of the 250 trajectories (see supporting material). The resultant mean VAMP score for each lag time is comparable for the CV spaces used in this study, indicating that their dynamical slowness is comparable. Furthermore, we compute the correlations among these variables to analyze whether they carry orthogonal signatures of the system. The results show that other than the ICs that are orthogonal by design, Rg and $\alpha$H share a minimal correlation with each other and the two ICs (Fig. S6). This analysis highlights the importance of using the TICA-derived CV space that can contribute independently to the system FEL and may disclose new structural insights.

In Fig. 4 we present the FEL as a function of dIC-1 and dIC-2. This analysis has decomposed the state space into three disjoint basins that are featureful in their own ways. The discrete FEL (Fig 4 A) represents the larger basin as composed of the global minimum “a” and a set of shallow minima (marked “b”–”f”). This population of states are distinctly separated from the other two disjoint basins (marked “g” and “h”) by free energy barriers. Although the discrete FEL is unable to capture the free energy barriers separating the states, EMMD is capable of providing maximum likelihood estimates for them. For statistical benchmarks, EMMD was performed over independent datasets that constituted the cumulative 12.5 $\mu$s trajectory. The estimates for mean FEL and SDs were determined similarly in previous applications (29,30). The average FEL thus generated over the CV space is provided in Fig 4 B, and the corresponding SDs are provided in the supporting material (Fig. S9). Due to the regularization in the fitting procedure, the model limits characterization of the features under each minimum as detailed as the discrete FEL. However, it outperforms the discrete landscape in barrier estimation between distant basins. The EMMD estimated FEL depicts that the major basin is separated from the two disjoint minima (marked “g” and “h”) by high free energy barriers exceeding 14 and 40 $k_{B}T$, respectively. This analysis is indicative of the slow dynamics and rate-limiting transitions that exist in the FEL of NSP1 CTD. In isolation, the system can be kinetically trapped in one of these basins; however, the interactions with the 40S ribosome possibly stabilize the two-helix structure selectively.

Figure 4.

(A) Discrete FEL of NSP1 CTD projected as a function of the first two dihedral time-structured independent components (dIC-1 and dIC-2). The minima are labeled from “a” to “h.” A representative structure from each minimum is shown in Fig 5 in orange. (B) EMMD estimated continuous form of the FEL where free energy barriers among the disjoint basins are captured. To see this figure in color, go online.

Structural analysis

We attempt to elicit the population that is most compatible with the NSP1 CTD conformation (reported residues 18–49) found within the 40S ribosomal subunit. We initially performed backbone RMSD calculations with the cryo-EM structure of NSP1 CTD as a reference. Here, we point out that the two helices (residues 23–48) were considered as the reference for alignment as the other segments encode significant disorder. The most aligned representative structure extracted thereby is represented in Fig. 5 along with the reference conformation and the average backbone RMSD values for each state. The projection of the closest structure over the TICA-derived CV space depicts its location under the minimum “a” as marked in the FEL. Interestingly, this minimum is significantly broader and comparatively more populated than the other prominent minima. This suggests that the structures compatible with the ribosome binding are predominant in the (standalone) equilibrium ensemble under physiological conditions. High structural deviations are further observed for the other basins with respect to the experimentally reported structure, manifesting in high RMSD values relative to state “a.”

Figure 5.

Structural alignment of each conformation labeled as “a” to “h” from TICA-derived CV space (in orange) with the experimentally reported structure compatible with ribosomal subunit binding (in black). Backbone RMSDs with the experimental structure are provided with SDs (in braces). See text for details. To see this figure in color, go online.

At this stage, it is important to understand how different residues of the NSP1 CTD interact to stabilize the structurally variable conformations. An inter-residue, scaled contact difference map with respect to state “a” is thereby constructed for each state, wherein the negative values correspond to loss, and the positive ones highlight the formation of contacts (see Figs. 6 and S12). In agreement with the RMSD results, states “a,” “b,” and “c” show similar contacts, whereas states “f,” “g,” and “h” show distinct differences. The change in residue level fluctuations for such states at specific regions is commensurate with the altered interactions (Figs. 6 and S12). The reconfiguration of internal interactions and fluctuations underscores the intrinsically disordered nature of the NSP1 CTD. This analysis is suggestive of altered enthalpic and entropic interplay underlying the emergence of the other metastable states. To interpret the structural differences among the states corresponding to the separate minima in a physically meaningful way, we performed secondary structural analysis and chemical shift calculations for individual basins in the FEL. Here, we used the Sparta+ program (61,62) to predict the chemical shift values for the atoms N, C, and C^α for each residue for individual structures. For this calculation, we consider 1000 random samples from each marked minimum of the TICA-derived CV space. The analysis results in a multivariate space of shift values for each local minimum. We computed the pairwise Kullback-Leibler (K-L) divergence among the multivariate shift distributions obtained from different states (supporting material). These K-L divergence values are depicted via a matrix in Fig. 7. It essentially shows that the local minima under the same metastable basin in this FEL have relatively close chemical shift values, while different metastable basins have large differences. These findings further assure the physical distinction among structures separated by the TICA analysis while underscoring the inability of the conventional CVs (Rg, $\alpha$H) to capture the structural diversity of this IDR. Moreover, the large differences in shift values among distal states in the CV space indicate a possible existence of high dissimilarities in their inherent secondary structure content.

Figure 6.

$\Delta d_{con}$ is the scaled difference between contacts of a given state and that of the reference (state “a”). Details in supporting material. Gain and loss of contacts reflected in positive and negative values, respectively. Corresponding residue-wise backbone root mean-squared fluctuations (RMSFs) are compared. To see this figure in color, go online.

Figure 7.

K-L divergence among the multivariate chemical shift distributions of the marked states in TICA-derived CV space. States closer in the dIC space have relatively close chemical shift values than those that are further apart. To see this figure in color, go online.

Secondary structure propensity of the polypeptide sequence at each state is reported by analyzing 1000 random structures from each minimum selected during the shift measurements. Secondary structural content of NSP1 CTD was calculated using the dictionary of secondary structure of proteins (63) implemented in the pytraj python package (64). The results from the dictionary of secondary structure of protein calculations depict the scaled frequency distribution of different secondary structures as a function of the residues present in the sequence (Fig. 8). The states “a,” “b,” and “c” share overall similarities in their secondary structure, which agrees with their small chemical shift differences and closer position in the TICA-derived CV space. However, a significant reduction in the $3_{10}$ helix propensity for residues 2–5 can be noticed in “b” and “c” as compared with “a.” The more distant structures from the ensemble “a” in this CV space encode major differences in the secondary structures as well. For instance, states “d” and “e” show a major drop in the C-terminal α-helicity commensurate with an increase in the $3_{10}$ helicity in the same region. Structures “f,” “g,” and “h” are mostly disordered, with a significant decrease in the defined secondary structures compared with the other states. However, the proportion and residue level position of α-helical propensity with respect to turn and $3_{10}$ regions can differentiate these states from each other.

Figure 8.

Residue-wise secondary structure propensity for each marked basin (“a” to “h”) in TICA-derived CV space characterized via DSSP scores. To see this figure in color, go online.

The current analysis not only reveals the conformational heterogeneity of the polypeptide as an IDR, but also highlights the existence of metastable states that are structurally distinct and kinetically demarcated from the state resembling ribosomal bound NSP1 CTD conformer. Moreover, the realignment of internal contacts and greater residue level fluctuations in these conformations are also supported by the reduction of ordered $\alpha$-helicity in their distinct secondary structures. It should be noted that an arrangement of specific residues of NSP1 CTD inside the mRNA entry channel of the 40S ribosomal subunit is necessary for its optimal binding with different ribosomal proteins and rRNA. Therefore, a major structural deviation in this viral protein is expected to impair these intermolecular interactions leading to binding incompatibility.

Conclusions

NSPs perform crucial roles related to the existence and replication of the virus in the host system. SARS-CoV-2 has two open reading frames in its genome dedicated to the translation of NSPs. The NSP sequences are conserved within the coronavirus family, with the exception of NSP1 in SARS-CoV-2. The C-terminal region of this protein is known to impair the host mRNA translational machinery by blocking the mRNA entry channel in the 40S ribosomal subunit. While recent studies have confirmed the disordered nature of NSP1 CTD, it forms two stable helices during the ribosomal block. Therefore, it is likely that the system contains both the ordered and disordered structures in its equilibrium population, or such structural diversity could manifest in folding upon binding. Targeted drug design against SARS-CoV-2 may be aided by a detailed understanding of the structural diversity of this protein, as well as the thermokinetics of the inherent FEL. Accordingly, the current contribution describes an MD simulation-based study designed to explore these aspects.

Initial REMD simulations were utilized to generate seed structures for unbiased trajectories, further utilized for free energy calculations and structural analyses. Owing to the inadequacies of conventional geometrical descriptors in describing the system heterogeneity of the IDR, TICA derived slow and orthogonal modes were incorporated to project the effective FEL composed of multiple shallow minima. Structures resembling the ribosomal subunit bound NSP1 CTD belong to a deeper minimum, indicating to their predominance in the equilibrium population. Importantly, we observe the emergence of two distinct metastable minima separated by large free energy barriers predicted by the EMMD protocol. Correlative structural analyses indicate that denaturation of the C-terminal $\alpha$-helix in the NSP1 CTD can be considered as a common feature in the structures that significantly differs from the two-helix conformation. Such conformations that substantially deviate from the ribosome-bound experimental structure are expected to be of lower compatibility for the binding event.

Insights from the current work may influence mutation or drug design studies targeted to stabilize or destabilize the states that play roles in the translation-blocking mechanism. The combined knowledge from such studies can support the platform for a better understanding of the novel coronavirus disease and developing medications for its prevention. Moreover, the study posits complementing exascale computing methods with data-driven approaches in CV design and FEL estimation. It is noteworthy that, in recent years, exascale methods have yielded unprecedented microscopic and mesoscopic insights into biomolecular systems of high complexity (31,32,65,66). Conformational information of metastable states thus obtained may potentially be combined with statistical algorithms such as EMMD to explore plausible thermokinetic connectivities. The advent of such combined strategies appears to hold promise (67). In turn, such efforts may fuel the developments in molecular therapeutics against other challenging diseases that are caused by specific protein folds.

Author contributions

P.D., A.K., and P.B. designed the research and did the simulations. N.S. streamlined the workflow. P.D. contributed in free energy calculations. P.B. and A.K. performed the structural analyses. All authors have equally contributed in the manuscript preparation.

Editor: Abhishek Singharoy.