Investigation of target sequencing of SARS-CoV-2 and immunogenic GWAS profiling in host cells of COVID-19 in Vietnam

Two workflows including a framework to align and annotate SARS-CoV-2 and a predictive model for COVID-19 patients have been developed. The first takes as input virus target sequence data from GISAID, the NCBI, and data collected in Vietnam to identify the virus genome sequence variants and provide summary statistics of these sequences. The second integrates PRSs into machine learning models from two sources: (1) GWAS with hospitalized COVID-19 patients and (2) a combination of immune biomarker variants that are associated with severity status and target data.

The data downloaded from GISAID in Vietnam consisted of 361 SARS-CoV-2 samples, shown in Fig. 1. NCBI data and data collected from Vietnamese sites contained FASTA sequences of SARS-CoV-2, a summary of protein mutations, and patient metadata. Other datasets from different countries in Mekong regions were also downloaded. Nextclade (https://​docs.​nextstrain.​org) is a tool that helps to identifies the differences between target sequences and a reference sequence by Nextstrain to assign clades to these sequences [28]. Nextclade was used for alignment to reference SARS-CoV-2 Wuhan-1 (MN908947.3), and for detecting variants and protein mutations on these datasets.

With data from NCBI, we have gathered 322,101 samples collected in Q2 2021 (Quarter 2 of 2021) and 542,275 samples collected in Q3 2021. All samples had a length greater than 29,000 bases and number of Ns is less than 300. These samples were also analyzed by Nextclade to find out which strains and mutations prevail among others.

COVID-19 data from 57,560 patients across 63 Vietnamese provinces (approximately as of July 20, 2021), and other data from almost 19,924 patients were downloaded from public source [29]. Age, sex, status, and other metadata of each patient were included. The patient’s province of residence was a crucial parameter in the model as it represents the environment of coronavirus disease.

Dataset from Whole-Genome Sequencing from Vietnam in The International Genome Sample Resource (IGSR) [30] will be used in the study. The target dataset is a cohort of 99 (of 124 samples) unrelated Vietnamese people in the project (whole-genome sequencing with 30× coverage) from Kinh ethnic group (100 KHV) on GRCh38 [31]. The \(\chi ^2\) goodness-of-fit test for Hardy–Weinberg equilibrium was used on samples with related individuals, and missing genotypes were filtered. We also use the dataset from 1000 Vietnamese Genomes Project (1KVG), a source of genomic variants for Vietnamese population by sequencing the whole genome of 1008 unrelated healthy Vietnamese to a depth of at least 28× [32].

The most common method for calculating PRS is called clumping and thresholding (or pruning and thresholding), applies two filtering steps as shown in Fig. 2. SNPs that weakly correlated with each other were retained. Clumps around SNPs were formed by using the linkage disequilibrium clumping procedure [33].

In PRS analyses can be characterized by the two input data sets: (i) base (GWAS) data: summary statistics (e.g. betas, p-values) of genotype-phenotype associations at genetic variants and (ii) target data: genotypes and phenotype(s) in individuals of the target sample [34]. We investigated the blood and lung biomarkers incorporated into the model for 100 KHV and derived PRS for all individuals. Based on the result, we will reconstruct PRS for a larger dataset and apply several machine learning techniques to predict severity of COVID-19 patients in Vietnam.

GWAS summary statistics of COVID-19 patient variants of Hospitalized vs. not hospitalized were downloaded. The summary was thresholded at \(p = 5e-8\) as standard in QC of GWAS. Summary statistics of COVID-19 were downloaded from open source COVID-19 HGI GWAS (https://www.covid19hg.org). These summary statistics are the result of a meta-analysis of 61 studies from 24 countries, and include the weights (effect sizes) and p-values of 13,498,845 variants, derived from a genotype-phenotype association study with 14,480 hospitalized patient samples and 73,191 control samples. GWAS QC excluded variants with p-value greater than 0.05. The PRS was derived for the training set using a pruning and thresholding method by Plink v1.9 [35]. The best model was selected based on \(R^2\). The PRS was calculated by the below equation of Plink.The effect size of SNP i is \(S_i\), effect allele j is \(G_{ij}\), the ploidy of the sample is P (with humans, \(P = 2\)), the number of SNPs is N, the number of non-missing SNPs in sample j is \(M_j\). In addition, individual phased genotype data of 100 KHV was from a VCF file. Standard quality controls were applied to the KHV VCF files, with missing genotype \(> 0.1\), Hardy–Weinberg Equilibrium \(P > 1e^{-6}\), minor allele frequency \(MAF < 0.01\).

The lack of direct PRS calculations for patients from Vietnam without a genotyping/sequencing profile posed a major challenge. Instead of directly predicting PRS using existing methods, we used a reconstruction method that applies a multivariate linear model to use the PRS calculations of an existing cohort (reference matrix Ref with PRS) with genotyping/sequencing to other cohorts, and showed the model can improve the prediction of severity. The model utilizes covariates captured age, gender, location. The predicted PRSs for 19,924 COVID-19 patients were then derived by a machine learning model to predict severity, and that correlated well with the measurements from clinical readouts.ReconstructPRS is reconstructed PRS using the summation of all fraction (\(Fraction^i_j\)) measured by a covariate i and a sample j in the reference cohort and PRS of sample j in the cohort \(PRS^j_{Ref}\). N is the number of covariates.

SNPs relevant to COVID-19 were then ranked by probability of severity according to a COVID-19-related study from the GWAS catalog (downloaded in August 2021 from https://​www.​ebi.​ac.​uk/​gwas/​). The data show that provinces are highly correlated across datasets. Since the sequencing data of these 19,924 patients is not public, we used 100 KHV results to calculate and reconstruct PRS for the training and testing datasets. The PRS was based on GWAS on GRCh38 from [36], using the pruning and thresholding method as mentioned previously. The predictive model showed how certain non-genetic factors may impact the risk of hospitalization due to the virus. We used three machine learning models including SVM, Random Forest and Elastic Net to predict COVID-19 severity in Vietnamese patients based on PRS and other covariates such as age, sex, location, exercise, and underlying conditions. The training dataset contained 11,814 patients, with status of deceased, active and recovered. These statuses were assigned numeric values in the model. To simplify the models, this was converted to a binary class of ‘Active’ and ‘Recovered’. Deceased patients were excluded.

Figure 3 shows the percentage of SARS-CoV-2 clades in different countries for Q2 and Q3 2021. Overall, the strains in Q2 were quite diverse with common strains such as 20A, 20I (Alpha), 21A (Delta), 20B, 21F (Iota) but in the third quarter, strain 21A (Delta) predominated in most of the countries. In the case of the Delta variant, in the second quarter, a number of Asian countries such as India, Bahrain, Bangladesh, and Uzbekistan recorded the presence of this variant with a significant majority, while some European and American countries such as the US, Switzerland, Germany, this variant appeared but did not prevail. This indicates that the outbreak of the Delta variant took place first in Asian countries, then in European and American countries. Regarding mutations, our analysis results also show that the most common mutations in the third quarter are the typical mutations of the Delta variant such as S:D614G, S:P681R, S:L452R, S:T478,S:R158G. In summary, the data analyzed on NCBI show the emergence and the spread of the Delta variant and its mutations in recent times.

In this part, 3211 FASTA sequences from Thailand in GISAID have been used. We compared Vietnam and Thailand populations as they have similar genetic characteristics in other infectious diseases [37] and their data is widely available in Southeast Asia. Comparison of the number of sequences by each month shows that Thailand had a prevalence of Lineage B.1.1.7 (Alpha) in the second quarter of 2021, while Vietnam had a prevalence of Lineage B.1.167.2 (Delta) from May of 2021. In addition, lineage A.6, B.1.36.16 and AY.30, which first appeared in South East Asia, were detected mostly in Thailand (Fig. 4). The analysis is consistent with the result from Chookajorn et al. 2021 [38] as the spread of the Alpha and Delta variants dominant over the region raised serious problems of the healthcare system. As an emerging epicenter of COVID-19 pandemic, Southeast Asian countries needs to take immediate collaborative actions to resolve these problems. More details of the analysis can be found in Additional file 1: Figs. S1 and S2.

In GISAID datasets, there are 361 FASTA sequences of SARS-CoV 2 from Vietnam. The virus variants are divided into 10 Clades (G,GH,GK,GR,GRY,GV,L,O,S,V). More details on GISAID’s clades can be found on the website (https://​www.​gisaid.​org/​). The result of comparison between all clades from Vietnam shows two common variants D614G (in Spike region), P323L (in NSP12, known as ORF1a region) in almost all clades with prefix G in both groups of Hospitalized and Recovered patients. These 2 mutations overtake frequency of dominant strain. Furthermore, clade GK and GRY have more protein mutations than other clades that can be promising targets for for analyzing protein structure and designing COVID-19 vaccines or drugs (Fig. 5).

We formed nine gene sets associated with severity of COVID-19 patients, as introduced in "Introduction" section. Table 1 reports the immune gene sets, along with the number of genes in each set and the number of SNPs found in 100 KHV.

Allele frequencies for SNPs of genes in each gene set were calculated for both 100 KHV and 1KVG [32]. SNP allele frequencies for all gene panels were highly correlated between sets (Pearson correlation \(p-value < 2.2e^{-16}, R = 0.99\)) (Additional file 1: Fig. S3). 1KVG was able to detect some variants with much lower allele frequency compared with those frequencies of 100 KHV suggesting that using 1KVG (with much larger sample size) to increase the quality of variants, especially in immunogenic and drug targets used in Vietnamese people. These variants were added to the model as “causal” SNPs in the computation of PRS as illustrated in the second workflow in Fig. 1.

The two datasets from [29] (57,560 patients split by province and 19,924 patients with province information provided) were consistent in the distribution of patients between provinces (Pearson correlation \(p = 1.2e^{-15}, R = 0.83\)). This is an important result as location and other phenotypes were used to reconstruct PRS in the training and testing datasets for the machine learning models. The data has been divided by training 70% and testing 30% number of samples in the datasets.

With an average of 100 runs, Random Forest was the best model with AUC = 0.81, followed by Elastic Net with AUC = 0.7 and SVM with AUC = 0.69 in Fig. 6.