Nasopharyngeal swab samples were routinely tested using the Cepheid Xpert Xpress SARS-CoV-2 assay or the Roche cobas SARS-CoV-2 test. The routine diagnostic tests were performed in a CAP-accredited laboratory in a tertiary hospital, in biosafety level 2 plus containment. In brief, laboratory personnel in full PPE (18) performed all sample processing and chemical inactivation within a biosafety cabinet. All nasopharyngeal swabs were chemically inactivated (19, 20) for 30 min prior to transfer to the GeneXpert Infinity (Cepheid) in biosafety level 2 containment, or cobas 6800 System (Roche) in biosafety level 2 plus containment, for the SARS-CoV-2 tests. Remnant clinical samples found to be positive for SARS-CoV-2 RNA, with a Ct value of <31 were subjected to an additional round of chemical inactivation (19, 20) within the biosafety cabinet within biosafety level 2 plus containment, in the same laboratory. Total nucleic acid extraction of the chemically inactivated remnant samples were performed using the NucliSENS easyMag (bioMerieux) platform. The resultant total nucleic acid extract was used for downstream RT-PCR, and sequencing on ONT MinION in accordance with the ARTIC protocol v3 (21), within the same laboratory, in biosafety level 2 containment. A maximum of 25 clinical samples was sequenced on an ONT MinION sequencer in a single run. The laboratory has the capability to obtain SARS-CoV-2 sequence data from 50 clinical samples with two available ONT MinION setups. RAMPART protocol (22) was used to monitor the depth of coverage for each sample and to construct a draft genome. A previously sequenced clinical sample and a negative control (no template control) were included as positive and negative controls in every run.

SARS-CoV-2 genomes collected in Singapore were obtained from GISAID (3,406 genomes as of 30 June 2021; Supplementary File 1). Genomes data collected from non-human hosts and environment were discarded. Missing collection dates were imputed with the first date of the respective month. Clade and lineage information were also downloaded from GISAID, where the clade the trends were calculated based on rolling average (7 days). Specific information of 99 de-identified positive cases used in the case studies was collected in Singapore General Hospital between January to August 2020 (Supplementary File 2).

Each SARS-CoV-2 genome obtained from on-site sequencing or GISAID was aligned to the standard SARS-CoV-2 reference genome (NC_045512.2) (23), using blastn. The alignment output in XML format was used for identifying single nucleotide polymorphisms (SNPs), insertion and deletion (Indels) with respect to the nucleotide position of the reference genome. For the genomes with multiple alignment blocks, the priority for identifying SNPs and Indels in the overlapping region was given to the longer alignment block.

To visualize the overall trends within a country, we employ two different types; a 2D scatter plot via tSNE (24) and an interactive dendrogram. We begin by calculating nucleotide diversity at each position using

where x_i is a frequency of each nucleotide and ∑i∈{A,T,C,G}xi= 1.

We then choose top 50 positions with highest diversity and positions that are predefined in GISAID (25). The high diversity positions were then used for calculating hamming distance among all 3,406 genomes for tSNE and dendrogram plots. For tSNE plot, we also applied k-means clustering, where k was set to 8, which is equal to the number lineages predefined by WHO.

For outbreak tracking, users can input multiple SARS-CoV-2 genomes (the maximum number of input genomes ≤200) obtained from the on-site sequencing platform. The input file should be in fasta format, where each input file contains one genome. The file name must follow SampleID_dd-mm-yyyy_GroupName.fasta, which consists of sampling identifier, collection date and predefined group name such as a location.

Each input genome is aligned to the SARS-CoV-2 reference genome using blastn, and GISAID clade is reported (25) (Supplementary File 3). A standard lineage identification model (version 3.1.15) was installed for predicting PANGO lineage of each input genome (11), in line with the nomenclature used worldwide. CalmBelt also alerts users when the breadth of coverage is <95%, allowing users to carefully analyse low-quality genome data.

To construct a phylogenetic tree of the input genomes, we followed the steps used for determining PANGO lineage by using these specific options “mafft –retree 2 concat.fasta > concat.mafft” and “iqtree2 -s concat.mafft -m GTR+G” (11). Similar to Nextstrain, we applied TreeTime (26) to generate time stamped phylogenies (Nextstrain's transmission tree) of the input genomes. To generate a similarity tree of the input genomes, whole-genome alignment information from blastn was used for calculating normalized hamming distance, and UPGMA clustering was applied on the distance matrix. In addition, as the hierarchical clustering process could introduce ambiguity for a large number of closely related genomes, a heatmap of distance matrix was also displayed to provide a quantitative measure of relatedness (i.e., number of different nucleotide positions).

An interactive similarity tree is drawn based on the whole-genome distance matrix alongside the collection date. Each node is colored according to the group name. Within the interactive tree, CalmBelt reports a subset of mutations and amino acid changes (with respect to blastn result). In the case that there are <3 nucleotides (i.e., frameshift mutations) to translate for an amino acid, the amino acid is marked as “X.” A full list of mutations is reported in a downloadable table. Monthly statistics of each PANGO lineage or sublineage in each country is also provided.

We implemented a portal where users can inspect specific genomic regions and mutations of interest of the genomes (3,406 genomes from GISAID) that have been processed in the previous step. Once users input the region of interest, SNPs and Indels are reported in multiple interactive charts. The charts display both nucleotide and amino acid changes that occurred in each month. To inspect concurrent mutations or variants of concerns, the portal reports the number of occurrences based on a list of mutations from users. These features are useful for detecting diagnostic-escape by inspecting primer binding sites and inspecting concurrent mutations.

For implementing CalmBelt and analyses, we used multiple Python packages including pandas 1.1.3, matplotlib 3.3.2, numpy 1.19.2, scipy 1.5.2, seaborn 0.11.0, plotly 4.14.3, ray 1.1.0, sklearn 0.23.2, biopython 1.78, dash 1.19.0. Additional software packages include Blastn 2.6 (15), Pango 3.1.7 (11), Mafft (–retree 2) (16), IQTREE-2 (-m GTR+G) (17), and TreeTime (26). CalmBelt's source code and an installation guideline are available at https://github.com/BioML-CM/CalmBelt. The demo version and example data is available at https://calmbelt-demo.mtms.dev.

Nasopharyngeal swab samples were routinely tested using the Cepheid Xpert Xpress SARS-CoV-2 assay or the Roche cobas SARS-CoV-2 test. Samples found to be positive for SARS-CoV-2 RNA, with a Ct value of <31 were subjected to total nucleic acid extraction using the NucliSENS easyMag (bioMerieux) platform. The total nucleic acid extract was used for downstream RT-PCR, and sequencing on ONT MinION in accordance with the ARTIC protocol v3 (21). RAMPART protocol (22) was used to monitor the depth of coverage for each sample and to construct a draft genome (Methods; Figure 1A).

Once the draft genome became available, Users dragged and dropped the SARS-CoV-2 genome files into the analysis interface (Figure 1B). Multiple genomes from on-site sequencing were aligned to the reference genome. ClamBelt then identified mutations, calculated whole genome similarity, and constructed multiple trees for transmission tracking (Methods). As the number of mutations per genome continued to increase (~2–3 mutations per month; Supplementary Figure 1), CalmBelt only visualized a subset of mutations while providing the downloadable full list of mutations.

The similarity tree captures hierarchical clustering results based on the whole genome similarity. The x-axis represents collection date, and colors represent case-specific information such as locations, as defined by the users. CalmBelt integrates existing nomenclatures including GISAID clade (9, 25) and PANGO lineages (12, 27). For each case, comprehensive information including mutations at nucleotide- and amino acid-level can be displayed interactively.

As highlighted in the Nextstrain web portal (28), “copying mistakes” that accumulate in the genome could facilitate tracking of transmission routes and dynamics, while a phylogenetic tree could infer locations of each case. CalmBelt therefore visualizes multiple types of trees, as well as a distance matrix capturing pairwise distance between SARS-CoV-2 genomes, to provide multiple perspectives for epidemiology units (Figure 1B; Supplementary Figure 2).

Besides, CalmBelt analyzed overall statistics for a large set of 3,406 genomes from Singapore (Figure 2, Methods). Users can inspect the statistics based on genomic information in terms of GISAID clade, which captures the trends at the beginning of the pandemic, and the newly defined PANGO lineage name according to WHO (Figure 2A). For example, we observed a clear dominant clade from April 2020 to September 2020, corresponding to the foreign worker dormitory outbreak in Singapore (29). Apart from the period of dormitory outbreak, the trends we observed in the Singapore data are more similar to the trend observed across countries in Asia compared to the trends observed in other continents, in line with geographical information (Supplementary Figures 3, 4).

Similar to Nextstrain (8), nucleotide diversity was calculated for each genome position. We observed various mutation rates across different genes (Figure 2B). While mutations in the S gene, which encodes the Spike protein (30) is of particular interest due to its role in infectivity and vaccine efficacy, other genes (e.g., E gene, N gene) are monitored as they are commonly used as diagnostic targets (31). We note that CalmBelt shows information from all genomes without subsampling, a pre-processing step used in Nextstrain.

Genomic loci with high diversity were used for visualization (Figures 2C–F). Overlaying PANGO lineage information onto the 2D visualization, we observed multiple clusters that have the same lineage name (Figures 2C,E). We found that clustering based on high diversity genomic loci could capture the foreign worker dormitory outbreak (C2; Figures 2D,F) and distinguished Delta (B.1.617.2) into three different clusters (C0, C3, C5) that spreaded at distinct time periods, suggesting different sources of the outbreak even though all cases were annotated as Delta based on the newly assigned PANGO lineage.

Four healthcare workers working in the same institution were found to be positive for SARS-CoV-2 within a short span of time. There were concerns of healthcare-associated/occupationally acquired SARS-CoV-2 infections among the healthcare workers. Therefore, on-site sequencing was performed for 4 healthcare workers' samples, along with 6 samples from the community and 2 samples from dormitory residents. A similar pattern was observed across different types of trees (Figures 3A,B, Methods). The similarity and transmission trees had the added advantage of capturing temporal relationships based on collection date information, which are useful adjuncts in outbreak investigations. Based on CalmBeIt's outputs, it was concluded that sequences from the four de-identified healthcare workers belonged to different clusters, suggesting that SARS-CoV-2 transmission did not occur in the healthcare or occupational setting in this instance.

To enhance the ability to identify potential sources of epidemiologically unlinked cases, CalmBelt allows users to query for statistics of a given PANGO lineage in different countries for a specific month. For example, in March 2020, B.6.6 was widespread in Cambodia (2/19, 10.53%), Singapore (36/381, 9.45%) and Malaysia (5/56, 8.93%) (Supplementary Figure 5). This information will allow users to identify potential links for further epidemiological investigations.

We inspected 89 randomly selected, de-identified positive cases from the foreign worker dormitory outbreak in Singapore between April to August 2020. Using CalmBelt's interactive tree, we observed two main clusters (Figure 4). In the first cluster (orange color; left), the majority of the cases were from a single dormitory (D8), two cases were from D10 and one case was from D9. However, there are 5 cases from D8 in the second cluster, suggesting that there were at least two sources of the infection in D8. In the second cluster (mixed colors; right), cases from many dormitories were observed. Cases from the same dormitory tend to form a cluster, highlighting the spread of the SARS-CoV-2 virus within each dormitory. In addition, we observed multiple small clusters in some dormitories (e.g., within D2, S53 + S74 + S128, S136 + S137, S129 + S130 + S77, S157 + S132), which may suggest multiple sources and routes of transmission within D2.

While PANGO lineages are meant to delineate outbreaks, in the setting of a rapidly disseminating local outbreaks, the resolution offered by PANGO lineage assignment alone is likely to be insufficient. For example, in the case of this large-scale dormitory outbreak, the majority of the cases (73%) were annotated as B.6.6. CalmBelt was able to capture whole genome similarity and increased the resolution, thereby delineating possible clusters within each dormitory.

CalmBelt can be used to monitor genomic mutations and allow users to investigate genomic regions of interest. As a use case, we selected three genomic regions in the N gene, based on PCR primers used for the detection of N gene [US-CDC-N1 to N3 (32); Figure 5A]. We found that mutations occurred in <5% of Singapore sequences (18 out of 598 and 6 out of 401) in the regions primed by US-CDC-N1 and US-CDC-N3. However, for the region primed by US-CDC-N2, we observed a higher percentage (19%) of Singapore sequences that harbor a mutation at position 29,179. This tool therefore allows users to monitor for emerging diagnostic escapes, based on the local laboratories' primer sequences and relevant regional SARS-CoV-2 sequence data.

As the number of mutations per genome continued to accumulate, ~2–3 new mutations emerged per month (Supplementary Figure 1). CalmBelt allows users to input a custom list of mutations and reports the number of cases where each mutation is present. For example, we added a list of mutations occurring in B.1.617.2 lineage or Delta (13) including variants of biological significance S:P681R and S:L452R which are associate with viral infectivity and reduction in antibody neutralization in the presence of other mutations in S gene (33, 34) (Figure 5B). Alternatively, users can input a protein name and a range of amino acid loci (Figure 5C). CalmBelt also allows users to input a specific set of mutations to observe the co-occurring frequency in each month (Figure 5D). In our example, we found that the co-occurrence of two variants (S:P681R and S:L452R) is well-preserved, which suggests biological significance that warrants further investigations (35).