Regional and temporal coordinated mutation patterns in SARS-CoV-2 spike protein revealed by a clustering and network analysis

SARS-CoV-2 has steadily mutated during its spread to > 300 million people throughout the world. The WHO has designated strains with certain mutations, “variants of concern” (VOC), as they may have higher infectivity and/or resist neutralization by antibodies in sera of vaccinated individuals and convalescent patients. Methods to detect regionally emerging VOC are needed to guide treatment and vaccine design. Cluster and network analysis was applied to over 1.2 million sequences of the SARS-CoV-2 spike protein from 36 countries in the GISAID database. While some mutations rapidly spread throughout the world, regionally specific groups of variants were identified. Strains circulating in each country contained different sets of high frequency mutations, many of which were known VOCs. Mutations within clusters increased in frequency simultaneously. Low frequency, but highly correlated mutations detected by the method could signal emerging VOCs, especially if they occur at higher frequency in other regions. An automated version of our method to find high frequency mutations in a set of SARS-COV-2 spike sequences is available online at http://curie.utmb.edu/SAR.html.

www.nature.com/scientificreports/ revealed over 350 amino acid variants, about 100 of which occur in the receptor binding domain (RBD) 25,26 . To systematically identify groups of mutations that occur together, we developed a new bioinformatics approach to characterize spatial and temporally correlated mutations within the spike protein based on their similar frequency of occurrence. The highest frequency clusters detected by our automated screening method included all previously known VOCs. In addition, regional and temporal analysis revealed other clusters that were increasing in frequency and could be emerging VOCs. The data also suggest how future vaccine design could be tailored to the local strain ensemble.

Results
Major mutations in the spike protein and relations to VOCs. We collected two groups of full-length spike protein sequences from the GISAID database; a group 1 with sequences from all countries, and a separate group (group 2) from 36 countries where COVID-19 was spreading rapidly. The top ten mutations found in the group 1 and for twelve individual countries are listed in Table 1. Additional data for the top 16 mutations in twelve individual countries are shown in supplementary material (Table Sup.T1), and a complete list of mutations for all 36 countries is shown in an MS Excel file in the supplementary material (Table Sup T2). The top eight of these mutations in group 1, i.e. D614G, P681H, N501Y, T716I, D1118H, A570D, S982A, and A222V are also found as high frequency mutations in most individual countries. Especially, the D614G mutation, found to be the dominating mutation in the first viral variant B.1.1.7 10,27 is also the top ranked mutation in all individual countries. The D614G mutation is responsible for increased viral transmission 10,28 . The second major mutation found in most countries, P681H, which increases the viral fitness in the fusion process, is located near the furin cleavage site between S1 and S2 in the spike protein 29 . In some countries e.g. India, this residue was mutated to P681R. Other major mutations such as N501Y, result in increased infectivity as the mutated residue increases the affinity for the receptor binding domain (RBD) to its receptor ACE2 30 , or the mutation E484K that reduces the immune response 20,31-33 . Other highly ranked mutations in each country are more country specific and may evolve with time and geographic locations under selective pressure. For example, the mutations, S13I and W152C found in the US, could interfere with neutralizing antibodies binding to the N-terminal domain (NTD), as antibody interacting residues W152 and R246 have been identified for the neutralizing antibody 4A8 for SARS-CoV-2 34 . Our list of top mutations correlates with the signature sequences of the major VOC as defined by an expert group of the WHO. For example, the major mutations we found for the UK, D614G, P681H, N501Y, T716I, A570D, S982A and D111H are the signature of the Alpha variant (B.1.1.7). The top mutations (D614G, A701V, E484K, D80A, K417N, N501Y and D215G) in South Africa are the signature of the Beta variant (B.1.351). Of course, in most of the countries a mixture of different variants is circulating, as seen for the US, where the top mutation clusters are from the Alpha variant, and the next most frequent mutations, S13I, W152C, and L452R from the Epsilon variant (B.1.429).
Deletion pattern of the spike protein sequences. The deletions in the N-terminal domain of the spike protein at H69, V70 and Y144 11,23,35 were first found in sequences from UK and may allow the virus to evade the immune response 36 . However, from December on-wards we found this deletion also more frequently in sequences obtained from US, Italy, Germany, Spain, France and other European countries but less in sequences obtained from Brazil, India and South Africa. Table 1. The top 10 mutations (after the D614G mutation, the first divergence from the "Wuhan 2019" sequence which arose rapidly and is present throughout the world at high frequency) in the spike protein listed in order of their frequency in sequences from all countries, and twelve different countries. The most common alterations are P681H (near the furin cleavage site that is a major distinction between SARS-CoV-2 and SARS) and N501Y (in the ACE2 binding region of the RBD). *All SARS-CoV-2 sequences in the GISAID database. protein is synthesized as a single chain protein which is then cleaved by a furin protease into two subunits, known as S1 and S2 (Fig. 1). The N-terminal domain (NTD) and the receptor binding domain (RBD) of the S1 protein are important areas for immune recognition and receptor binding 37 . The RBD is both a target for vaccine 38 and diagnostic 39 design. The helical S2 subunit undergoes large conformational changes during the fusion process. The 3D structure of the peptidase domain of ACE2 in a complex with the spike protein (PDB file: 6M17) 40 shows that ACE2 interacts with at least 22 amino acids in the RBD, including several frequently mutated residues detected by our method: K417, S494, N501, L454, S477 and E484. Compounds that inhibit the interaction between the RBD and ACE2 may be useful treatments for SARS-CoV-2 infection 41 .
Our sequence analysis showed that RBD positions K417, S494 and N501, which form contacts with ACE2, and other three mutations at L452, S477 and E484 are located close to the ACE2 interface, and may play an important role in binding to receptor and antibodies 8,42,43 . In addition, this site is targeted by several antibodies as reported recently [44][45][46] . Most of the mutations on the spike protein found by our analysis are in the RBD and N-terminal domain of the S1 protein, or close to the fusion peptide (e.g. T716, T1027, and Q1071) and CD domain (D1118) of S2 protein. Most of the mutated residues in the RDB, except N501, are surface exposed in the PDB id 6M17 and 6BSB structures according to GetArea 47,48 .
Notably, there were distinct mutation signatures for individual countries. For example, in India we found a major mutation at the P681 position but with an R (9.71%) instead of the H mutation. The R mutation is also found in the Delta variant (B.167.2). This variant is of considerable public concern, as it is increasing with a high rate in many countries, including the US, due to its increased transmissibility and immune evasion 49,50 .

Correlation analysis of frequency mutation indices across the 36 countries.
We performed a correlation analysis, visualized with Cytoscape, of the top 100 mutations in the spike protein by calculating the percentage mutation frequencies in 36 different countries 51 . Using a correlation coefficient cutoff of 0.7, we obtained nine major clusters that contain three or more amino acids (Fig. 2). We empirically used a relatively low threshold of 0.7 to find mutations that arise as groups in different countries, as the number of sequences from each country varies. The clusters found by this analysis are characteristic for clusters that emerge in different geographic locations. For examples the mutations found in cluster 4 (N501Y, A570D, P681H, T716I, S982A, D1118H) occur mainly in European countries, whereas mutations in cluster 6 (D80A, D215G, K417N, E484K, A701V) are mostly from South American countries and South Africa, as seen in Fig. 3.
The mutations observed in cluster 4 coincide with the signature mutations in the Alpha variant as defined by the WHO. Cluster 4 also contains six major mutations with a frequency index above 10% in most countries. These mutations first appeared during the first COVID-19 wave in 2020 and were also present in sequences from Our cluster analysis thus automatically detected five major variants of SAR-CoV-2 and their signature mutations. In addition, we also found a few small clusters that are either country specific or may be due to undetected COVID-19 positive persons traveling to these countries. For example, the mutation S13I in cluster 2 was only present in sequences obtained from US, Mexico, Sri Lanka and Taiwan, while mutations D253G and T859N in cluster 1 were present in sequences from US, Chile and Peru.
Evolution of the mutations in the spike protein over time. In order to understand the evolution of the spike proteins from the first Wuhan sequence, we calculated frequency indices of the mutations for each month, using all sequences in group 2 combined. Correlations of the frequency indices of the mutations over time and cluster analysis (cutoff value 0.98) with Cytoscape revealed three unique clusters of mutations with a high correlation coefficient (Fig. 4a). The mutations from cluster 1 increase in frequency after Jan 2021 (Fig. 4b,  c), and are still circulating in the Beta and Gamma coronavirus variants. The frequencies of mutations in cluster 2 increased suddenly after Nov 2020 and became fixed in the virus profile through April 2021 (Fig. 4d). The mutation D614G is an outlier and displays an earlier unique adaption in the virus population. In contrast, frequencies of mutations in cluster 3 increase after July 2020 and then decrease after Nov 2020 (Fig. 4e). United Kingdom For the UK, a cutoff value of 0.9 identified three clusters, each having a different mutation profile with time (Fig. 5a). The first cluster of the mutations N501Y, A570D, P681H, T716I, S982A, and D1118H, the signature of the Alpha variant B.1.1.7, sharply increased in frequency in Oct-Nov 2020 through April 2021, the last month of our analysis. A second cluster, L18F, A222V, A262S, P272L and E583D peaked in frequency in Nov. 2020 and gradually dropped to its Jun 2020 level in Feb 2021. A third cluster (T95I, G142D, E484K, A701V, Q957R and K417N), has some mutations in common with the Kappa variant (B.1.167.1) first found in India (E484Q mutation), represented only 1-2% of the UK sequences in April, but is gradually increasing in the UK. Thus, our combination of correlation and network analysis can detect mutation clusters with low frequencies and an emergence potential.
United States The cutoff value was increased to 0.98 to obtain well separated clusters of the US data. A high frequency cluster with N501Y, A570D, P681H, T716I, S982A, D1118H and a moderate (10-20%) one, with L5F, T95I, E484K and A701V mutations dominated the data (Fig. 5b). Recent New York and California variants B.1.526 and B.1.427/B.1.429 all contain E484K, with the CA variant also characterized by S13I, W152C and L452R (third cluster). The CA mutations first appeared in Oct, peaked in Jan-Feb 2021, and decreased in Mar-April21.
India Using a 0.98 cutoff value we found one large cluster with amino acids having a low mutation frequency index (< 2%, hence not considered as dominant mutations) and three small clusters with high frequency mutations (Fig. 5c). The first cluster contains mutations that suddenly increase after Jan. 21 (Fig. 5c, left panel). Some of these mutations, such as G142D, E154K, L452R, E484Q and Q1071H are major mutations and may

Discussion
Since the emergence of the SARS coronavirus in Wuhan (China) in December 2019, an explosion of sequence data for SARS-CoV-2 has been collected by many research groups. The GISAID 52 database, first established for sequence data for influenza outbreaks, expanded to become the major database for accessing genomic data of circulating SARS-CoV-2 sequences. We show here that our method can use this data to follow the evolution of the virus globally or within individual countries. Our automated method identified clusters of mutations as a function of time and locations directly from the sequence data base. We also showed that our method is potentially useful to predict future emerging variants of SARS-CoV-2, which can be directly characterized by specific mutation patterns, independent from the use of one of the different nomenclature systems which can be confusing. The US Center for Disease Control and Prevention (CDC) distinguishes the different variants by Greek letters while the Nextstrain team (https:// nexts train. org/ sars-cov-2) suggests using 18 major clades as from 19 A to 21F 19B, and 20A-20I. The PANGOLIN system 53 uses a dynamic lineage system with two clades A and B that are further divided in a hierarchical system with further subgroups with a numerical nomenclature, such as B.1.1. For example, the VOC first detected in the UK is referred to by the CDC as the Alpha variant and in the PANGOLIN system as lineage B.1.1.7 11 . Other  The automated bioinformatics tool we introduce here can help in identifying the major mutation patterns, not just according to their individual occurrence but also to their correlations with other mutations. Our approach, combining correlation and network analysis revealed patterns of distant mutations that occur together, and identified several known VOCs of high importance identified by the CDC. Particularly of interests are the VOCs that might threaten the future use of vaccines based on the original SARS-CoV-2 genome 35,55,56 . We show results of the method here for the spike protein, the major antigen of SARS coronaviruses, but the method can also be useful for other proteins of SARS-CoV-2 and for correlations across different genes.

Methods
Sequence analysis. We downloaded all SARS-CoV-2 spike sequences from the GISAID database (April 2021 release) 52 that contains 1,247,171 SARS-CoV-2 spike sequences archived during the last 16 months until April 2021. All sequences < 1000 residues in length or those with > 100 unknown or missing residues (marked as X) were removed from the data set, to yield 1,180,646 full-length spike protein sequences over all countries (group 1). A separate group (group 2) was selected with spike protein sequences from 36 countries where COVID-19 was spreading rapidly. The spike protein sequences in group 1 and group 2 were analyzed in each month to determine the mutation frequency for each residue. The CD-HIT program 57 (version 4.8.1) was used to cluster the sequences from all groups using a 100% sequence identity. The 100% sequence identity cutoff was used to find all sequences that have at least one mutation. After clustering, one representative sequence from each of the top 1000 clusters (if available, otherwise the maximal number of observed clusters was used) was selected. In each data set, the representative set of sequences were aligned with the spike protein sequence isolated from Wuhan (GenBank id QHR63260.2). A multiple sequence alignment was generated using the MUSCLE program 58 and the data were analyzed to obtain the mutation frequency (P i ) at each position in the sequence alignment. The P i is defined as where N i is the total number of sequences having that mutation in the cluster, and the Total is the number of all sequences in the clusters used in the alignment.
(1) P i = 100 N i Total Figure 5. Mutation index profile from (a) UK, (b) US and (c) from India for SARS-CoV-2 between Jan 2020 to Apr 2021. The data for the D614G is not shown here as it is a dominant mutation since the onset of the COVID-19 pandemic, as shown in Table 1 www.nature.com/scientificreports/ Correlation analysis. In order to characterize the correlations between different amino acid mutations, we calculated the correlation coefficients (r xy ) between the mutations as where x i and y i are the mutation frequency (Pi) of the two mutations. To determine the geographical correlations, the analysis was performed for the indices of the mutations across the 36 countries (i = 1,…,36). The time correlations (i = 1,…,16) were calculated across the 16 months from Jan 2020 to April 2021 for the sequence data for all countries, and separately for each individual country.
Network analysis. We used Cytoscape 51 to cluster and visualize the data obtained from the correlation analysis. For this purpose, we calculated the correlation between all rows where each column in a row represents the mutation frequency index in a country or in a month. For visualization of the correlation, the data was formatted into a three-column format where the first two columns represent the two mutations, and the third column represents their correlation coefficients. The data was clustered using Cytoscape and the clusters were visualized using a circular layout.

Conclusion
In the current work we analyzed over a million sequences of the SAR-CoV2 spike protein to understand their evolution and relationship between various variants circulating worldwide. Using clustering and network analysis, we showed that some of these mutations evolved together and have been shown to have an important impact on transmissibility, receptor binding or immune escape. Mutations with high frequency indices may reflect virus adaptation and can impact a virus phenotype. Using our new computational analysis, we reported several instances where the results of the correlated mutation patterns over several countries or over time coincide with signature mutations of emerging VOCs, in countries such as UK, US, India and Brazil. We also detected mutations that may be future VOCs. Our tool could be also useful for other researchers wishing to detect specific mutation patterns in other areas of the SARS-CoV-2 proteome.

Data availability
Data used for the analysis are available from the GISAID database (https:// www. gisaid. org/). The results from the complete frequency analysis are available as supplementary material. Additional material is available online at the, http:// curie. utmb. edu/ COVID 19/ and the webserver to find VOC's at http:// curie. utmb. edu/ SAR. html. www.nature.com/scientificreports/