Global disparities in COVID-19 vaccine coverage associated with trajectories of SARS-CoV-2 adaptation Authors: Lingyue Yang, Ziqin Zhou, Qiushi Lin, Yuyang Chen, Yamin Sun, Zengmiao Wang Published: 2026 ABSTRACT AbstractBackgroundVaccination serves as an effective intervention for health promotion and disease prevention across the socioecological systems and has played an important role during the COVID-19 pandemic. However, global disparities in vaccine coverage have increased uncertainty about the trajectories of viral adaptation, and the potential interplay between SARS-CoV-2 adaptation and vaccine rollout warrants further quantification.MethodsUsing over 13 million SARS-CoV-2 genomes across 86 countries from March 2020 to September 2022, we analyzed nonlinear associations between SARS-CoV-2 adaptation and vaccination coverage, considering public health and social measures, international travel, and infection dynamics, before and after the emergence of Omicron. Additionally, we examined the relationship between SARS-CoV-2 adaptation and COVID-19 mortality.ResultsDuring the pre-Omicron period, we found positive associations between nonsynonymous to synonymous divergence (dN/dS) ratios in the S1 subunit and medium levels of adjusted vaccine coverage (effect size: 0.96 [95% CI 0.47, 1.45]), while the association became insignificant at high levels (effect size: -1.89 [95% CI -4.20, 0.43]). However, no significant associations were found when Omicron dominated, possibly due to the immune escape ability of Omicron variants and the complex immune landscape shaped by mass hybrid immunity. Moreover, we observed evidence of dynamic interdependence and positive correlations between COVID-19 mortality and SARS-CoV-2 adaptation, with COVID-19 mortality interpreted as a proxy for uncontrolled viral spread.ConclusionsOur findings suggest a complex nonlinear relationship between vaccine-induced immunity and SARS-CoV-2 adaptation, with high vaccine coverage potentially linked to lower positive selection. We also observed directional coupling between COVID-19 mortality and SARS-CoV-2 adaptation. This may have implications for fair and fast vaccination in pandemic preparedness and response.Clinical trial numberNot applicable.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s12879-026-13826-1.Keywords: COVID-19, SARS-CoV-2 adaptation, Vaccine disparities, Vaccines, Natural immunity Introduction Vaccines represent one of the most effective interventions for preventing infectious diseases. During the COVID-19 pandemic, vaccination is estimated to have prevented more than 14.4 million deaths due to COVID-19 between December 2020 and December 2021 [1]. However, COVID-19 vaccine coverage showed substantial global disparities by income level [2], with vaccination rates reaching approximately 75% in high-income countries but remaining below 2% in some low-income countries [3]. While the adverse consequences of unequal vaccine coverage for saving lives have been well investigated [1–4], the broader effects of global disparities in vaccine coverage on shaping viral adaptation have yet to be extensively quantified. Since late 2020, a global shift in the selection pressure on SARS-CoV-2 has been observed, marked by a substantial increase in the number of SARS-CoV-2 codon sites undergoing positive selection, primarily in highly susceptible individuals [5]. Subsequently, the nonsynonymous to synonymous divergence (dN/dS) ratio within the S1 subunit exceeded 1 in late 2020 and 2021, indicating that this region was undergoing positive selection [6]. The S1 subunits are major targets for both infection and vaccine-induced immune responses, such as mRNA and adenovirus vector vaccines [7]. In addition, mutations outside S proteins, such as ORF6 and ORF9b proteins, have also been shown to play a pivotal role in the adaptation of SARS-CoV-2 to humans [8–10]. With the rollout of vaccines, the interaction between viral selective forces and the heterogeneous levels of population immunity has become increasingly intricate [11]. Beyond natural immunity, vaccination can also drive immune escape, possibly leading to a more predictable immune pressure [12]. Concerns have been raised regarding the possibility that immunity provided by vaccines might exert selective pressure on the virus [13, 14]. Specifically, moderate levels of immunity could result in increased viral adaptation compared to low and high levels of immunity, from a phylodynamic perspective [15]. Moreover, the unequal distribution of vaccines has heightened concerns about the emergence of new strains [4, 16]. Uneven distribution of vaccines may lead to continued virus transmission in areas with low vaccine coverage, increasing infection numbers, and in turn providing further opportunities for virus evolution [3, 17]. The heterogeneity in global vaccination campaigns has prompted the question of how vaccination efforts might affect viral adaptation and its potential role in the emergence of immune-evasive variants. In this context, the COVID-19 pandemic has enabled unprecedented vaccine distribution and large-scale genomic surveillance, creating a unique opportunity to explore these questions in real-world settings. Here, we conducted a comprehensive investigation to assess the global impact of COVID-19 vaccination on SARS-CoV-2 adaptation, and to evaluate how these effects differed following the emergence of the Omicron variant. Our analysis leveraged over 13 million SARS-CoV-2 genome sequences and vaccination information, while accounting for public health and social measures (PHSM), international travel, and infection numbers. Using COVID-19 as a case study, we aimed to provide empirical information on how population immunity induced by vaccines is associated with SARS-CoV-2 adaptation. Materials and methods Data collection SARS-CoV-2 sequences and associated metadata (n = 13,243,751) were downloaded from the sequence repository of Global Initiative on Sharing All Influenza Data (GISAID) [18](https://www.gisaid.org/) as of September 24, 2022. We initially analyzed all the genome sequences downloaded from GISAID by using Nextclade CLI (Version 2.8.0) (https://clades.nextstrain.org) [19]. Nextclade is a pipeline that includes several steps such as sequence alignment, translation, mutation calling, and phylogenetic placement, as well as assigning sequences to a clade and calculating some quality control metrics. In this study, the reference sequence was the SARS-CoV-2 isolate Wuhan-Hu-1 (GISAID accession number EPI_ISL_402125). We further conducted sequence quality control and subsampling of sequences (see Supplementary Information for details). Data about daily full and booster dose coverage are available from the global COVID-19 vaccinations database [20]. We estimated the monthly adjusted vaccine coverage by summing full and booster dose coverage and then adjusted it for vaccine effectiveness (VE) against symptomatic infection for mRNA (93.39%), viral vector (74.73%), and inactivated (65.89%) vaccines [21], based on the monthly cumulative administered proportion for each vaccine type (Table S1). Then, to consider the potential immunity waning six months after vaccinations [22–24], the VE against infection was assumed to decline to zero at six months after vaccination. Daily epidemiological data were extracted from Our World in Data [25] to calculate infections and deaths on monthly records per country. In addition, taking into consideration the underestimation of the reported cases, we also calculated the monthly infection dynamic for each country based on the estimated infection data from the Institute for Health Metrics and Evaluation (IHME) dataset [26]. We applied a method similar to that used for calculating adjusted vaccine coverage to estimate monthly natural immunity coverage. In some countries, adjusted vaccine coverage and natural immunity coverage were strongly correlated, with absolute correlation coefficients reaching up to 0.93 (Figure S1). Data to calculate international travel was collected from the travel pattern data between countries on Facebook Data for Good [27], and we calculated the monthly international travel based on the mean daily volume of a single country-to-country travel route on the monthly records [27]. We extracted the daily stringency index for non-vaccinated people from the Oxford Coronavirus Government Response Tracker (OxCGRT) [28], and calculated the monthly PHSM strength indicators to represent the implementation and intensity of government policy responses in different countries. We also collected country-level socioeconomic and demographic variables from the World Bank, including gross domestic product (GDP) per capita, Gini index, and the proportion of the population aged 65 years or older [29]. Please refer to the Supplementary Information for more details on the data process. Phylogenetic analysis We used the Nextstrain [30] (Version 6.2.0) pipeline to generate phylogenetic reconstructions of SARS-CoV-2 genome sequences for a specific country or region based on the above subsampled sequence dataset. The Nextstrain pipeline further filters, processes, and performs phylogenetic analysis on the subsampled sequences for specific countries or regions [6]. The phylogeny was constructed for the specific country using subsampled sequences and a small set of reference sequences, totaling 274 sequences maintained by the Nextstrain team as the genetic context (Figure S2). These sequences and metadata can be found at https://data.nextstrain.org/files/ncov/open/reference/metadata.tsv.xz and https://data.nextstrain.org/files/ncov/open/reference/sequences.fasta.xz. Augur [31] (Version 21.1.0), IQ-TREE [32] (Version 2.1.2), and TreeTime [33] (Version 0.9.5) were used to infer time-resolved phylogenies. The joint inference mode in TreeTime was employed to infer ancestral sequences. Given that the strategy of phylogeny reconstruction may impact the robustness of results, we collected a set of publicly available SARS-CoV-2 phylogenetic analyses for specific countries. These SARS-CoV-2 phylogeny are produced and maintained by the Nextstrain groups and include countries from Africa (https://nextstrain.org/groups/africa-cdc, created as part of [34]), Europe and North America (https://nextstrain.org/groups/neherlab). In total, there are 50 countries aligning with the scope of our investigation (Table S2), and we utilized them for sensitivity analyses. Calculation of the ratio of nonsynonymous to synonymous divergence, mutation accumulation and mutational fitness We divided all internal branches in the phylogeny based on their dates and selected internal branches inferred from the specific country. Subsequently, we calculated the temporal dynamics of dN/dS ratios within 2 month time windows with overlap by 1 month, using a previously described approach [6]. Specifically, the nonsynonymous and synonymous Hamming distances were calculated between the root sequence (Wuhan-Hu-1 reference sequence) and each internal branch within each gene. These distances were then normalized by the total number of possible nonsynonymous or synonymous sites within the gene to obtain nonsynonymous and synonymous divergence, and the dN/dS ratios was calculated for each internal branch as the nonsynonymous divergence divided by the synonymous divergence. The dN/dS ratio for each time window was then defined as the mean of the dN/dS ratios calculated for all internal branches within that time window, and was used to assess selection pressure for each gene. The ratio greater than 1 suggests positive selection, and the ratio less than 1 suggests negative (or “purifying”) selection. We calculated the monthly number of nonsynonymous mutations over genome regions per country based on the accumulated mutations between the root of the tree and the internal branches in the phylogeny within 2-month time windows, consistent with dN/dS ratio calculations. Mutations were then categorized for each gene according to the Wuhan-Hu-1 reference sequence to identify whether they were synonymous or nonsynonymous. Deletions were grouped together with nonsynonymous substitutions, and deletions spanning multiple adjacent amino acids were treated as a single mutation. Additionally, mutations leading to a premature stop codon were counted as individual mutation events. The scripts used in this study for calculating dN/dS ratios and quantifying the number of nonsynonymous mutations largely rely on previous methods developed by Kistler and colleagues [6]. The metric of mutational fitness was also used to assess positive selection. It was computed based on internal branches of the phylogeny per country, using the same time windows as the quantification of dN/dS ratio. Mutational fitness was performed using Nextstrain pipelines (Version 6.2.0), estimated based on hierarchical Bayesian multinomial logistic regression that calculates the relative impact of individual mutations on the growth advantage of SARS-CoV-2 lineages [35]. Statistical analysis The distributed lag non-linear model (DLNM) was fitted to investigate the potential non-linear association between vaccine immunity and SARS-CoV-2 adaptation over genome regions (including S1/S2 subunits, as well as S, E, M, N, ORF1a/1b, ORF3a, ORF6, ORF7a/7b, ORF8, and ORF9b proteins). The analysis was conducted for two study periods: March 2020 to October 2021 (pre-Omicron circulating) and January 2022 to September 2022 (during the dominance of the Omicron variant). The modeling framework of DLNM is based on the cross-basis function, which combines lag-response and exposure-response functions to simultaneously describe the non-linear exposure-response associations and lag effects [36]. In this analysis, we selected natural cubic splines (NS) with 2 degrees of freedom and specified the maximum lag as 3 for each covariate in lag-response functions to model the nonlinear lag effects over the cumulative lag of three months. Moreover, we set the linear functions to model the associations between SARS-CoV-2 adaptation and confounding effects of international travel and PHSM for the exposure-response dependencies. Moreover, based on both forward and backward selection procedures on the Minimum Akaike Information Criterion (AIC), cubic functions were selected for adjusted vaccine coverage to model the potential non-linear associations with dN/dS ratios in the S1 subunit. The same functions were applied in the other genome regions of SARS-CoV-2. The first DLNM was employed to examine the nonlinear relationship between adjusted vaccine coverage and SARS-CoV-2 adaptation across 86 countries from March 2020 to October 2021. A time series regression model was built with the R packages “dlnm” [36], “mgcv” [37], and “splines” and is given by: The error term of the model followed the Gaussian distribution. Yc, t represents metrics of SARS-CoV-2 adaptation, including dN/dS ratios and the number of nonsynonymous mutations over different genome regions, and mutational fitness of SARS-CoV-2, in each country c in month t. In addition, α indicates an intercept of the regression model, and β indicates the coefficients for each covariate. CB(AVCc, t), CB(PHSMc, t), and CB(log(ITc, t)) represent the cross-basis functions for adjusted vaccine coverage, PHSM, and international travel for each country c in month t, respectively. Moreover, log(Casesc, t) represents the natural logarithmic transformation of the number of newly reported cases in each country c in month t. Furthermore, represents the random intercept by country. PHSMc, t:AVCc, t is the linear interaction to account for the interaction between adjusted vaccine coverage and PHSM. To investigate SARS-CoV-2 adaptation and vaccines in a detailed manner, we further categorized the 86 countries based on whether they had achieved a 50% vaccination threshold with specific vaccine types: mRNA (n = 49, Table S3), inactivated (n = 14, Table S4), viral vector (n = 13, Table S5) vaccines, or mixed vaccinations where no single proportion of vaccine types meets the threshold (n = 10, Table S6), as of October 2021. We subsequently investigated associations between SARS-CoV-2 adaptation and adjusted vaccine coverage for mRNA and inactivated vaccines from March 2020 to October 2021, using the same model setting for all 86 countries. Moreover, the same analysis was conducted to model associations between adjusted vaccine coverage and SARS-CoV-2 adaptation from January 2022 to September 2022. We fitted another model to examine associations between natural immunity and SARS-CoV-2 adaptation over genome regions from March 2020 to October 2021. We selected 22 countries from low coverage regions to further investigate the relationships where natural immunity may play a leading role due to very low vaccine coverage (Table S7). The regression model was used as follows: where the structure of the dependent variable Yc, t, the intercept α, the random intercept by country, and the cross-basis functions for PHSM and international travel, i.e., CB(PHSMc, t) and CB(log(ITc, t)), were the same as those in the regression model for the association between adjusted vaccine coverage and SARS-CoV-2 adaptation. CB(NIc, t) represents the cross-basis function for natural immunity in country c at month t, with the same functional form as that selected for adjusted vaccine coverage to model potential non-linear relationships. Convergent cross-mapping We used convergent cross-mapping (CCM) analysis [38] from the “rEDM” R package to assess directional coupling between COVID-19 mortality and SARS-CoV-2 adaptation (dN/dS ratios, the number of nonsynonymous mutations in the S1 subunit, and mutational fitness of SARS-CoV-2). The CCM analysis was conducted per country from March 2020 to October 2021, and we first normalized reported COVID-19 deaths per million and the metrics of SARS-CoV-2 adaptation. Subsequently, we applied simplex projection to determine the optimal embedding dimension (E) ranging from 2 to 4 for both COVID-19 mortality and SARS-CoV-2 adaptation. We selected the E based on maximizing cross-mapping skill (measured by correlation coefficient, ρ) while minimizing absolute error. Examining convergence in cross-map skill, i.e., improving cross-map skill with increasing data, is a general method for distinguishing cross-mapping from spurious correlation [39]. However, we employed a more specific approach by developing a null test using surrogate time-series data to assess the significance of directional coupling. Specifically, we used the “random_shuffle” method to generate surrogate time series, and this process was repeated 1,000 times to establish a null model distribution. Cross-mapping skill was considered significant when it exceeded random expectations. Note that negative cross-mapping skill values were always set to 0. We used Fisher’s method to test the meta-significance of the CCM analysis across 86 countries to determine whether there was statistically significant directional coupling between COVID-19 mortality and SARS-CoV-2 adaptation worldwide. Results Global disparities in vaccine rollout and SARS-CoV-2 adaptation In the context of the COVID-19 pandemic, an unprecedented vaccination campaign in history was launched worldwide. Countries implemented diverse vaccination strategies, leading to variations in the administration of different vaccine types and resulting in heterogeneity in the global vaccination landscape (Fig. 1A). We categorized the 86 countries into three regions based on full-dose vaccine coverage as of October 2021: low (< 25%), medium (25%-75%), and high (> 75%) coverage regions. This classification resulted in 22, 58, and 6 countries, respectively. We observed that the disparities in full-dose vaccine coverage were evident globally as of October 2021 (Fig. 1B). Global disparities in full-dose vaccine coverage and SARS-CoV-2 adaptation during the COVID-19 pandemic. (A) The proportion of types of vaccines administered across 86 countries as of October 2021. The vaccine types include mRNA, inactivated, and viral vector vaccines. We considered nine major vaccine manufacturers (including CanSino, Johnson, Moderna, AstraZeneca, Pfizer, Sinopharm, Sinovac, Covaxin, and Sputnik.V) to estimate the cumulative monthly proportions for three types of vaccines (mRNA, viral vector, and inactivated vaccines) per country. The sum of the proportions for the three vaccine types in each country per month was 1 (see Methods). (B) Global disparities in full-dose vaccine coverage from 86 countries as of October 2021. The countries were classified into low, medium, and high coverage regions, with full-dose vaccine coverage below 25% (22 countries), 25%-75% (58 countries), and above 75% (6 countries), respectively. (C) Global temporal patterns of SARS-CoV-2 adaptation. The metrics of SARS-CoV-2 adaptation included dN/dS ratios and number of nonsynonymous mutations in the S1 subunit, as well as mutational fitness of SARS-CoV-2. The line (dots) and shaded regions represent the average and 95% confidence intervals (CIs) of the three metrics across 86 countries. The black vertical solid lines represent months when the World Health Organization (WHO) designated labels for variants of concern (VOCs), including Alpha, Beta, Gamma, Delta, and Omicron (lineage BA.1) variants [56]. Horizontal dashed lines represent the dN/dS ratio equal to 1. (D) The distributions of dN/dS ratios in the S1 subunit across regions with low, medium, and high full-dose vaccine coverage. Violin plots represent the distribution of these dN/dS ratios in different vaccine coverage regions from March 2020 to October 2021, with points and error bars representing the mean and standard deviation. Wilcoxon rank-sum tests were applied to assess differences between regions based on the distribution of dN/dS ratios, rather than the summary statistics displayed. (E) The scatter plot of full-dose vaccine coverage and total reported COVID-19 deaths per million across the 86 countries as of October 2021. Spearman’s tests were applied to assess their correlation across different vaccine coverage regions. In (A, B, D, and E), the definitions for low, medium, and high vaccine coverage regions are the same, and light, medium, and dark blue represent these countries, respectively Selection pressures on SARS-CoV-2 proteins for each country over time were measured by dN/dS ratios. For a specific country, the ratio was calculated by the accumulating nonsynonymous divergence sites relative to synonymous divergence sites between internal branches inferred from the specific country and the tree root in the phylogeny, using a time window of 2 months with a 1-month overlap. This metric captures persisting mutations contributing to the viral population while excluding those that die out [6]. Furthermore, we measured the number of nonsynonymous mutations (by tallying the accumulation between the phylogeny root and internal branches) and the metric of mutational fitness to suggest SARS-CoV-2 adaptation [35]. These two metrics were calculated within the same time windows of the internal branches used for dN/dS ratios in the phylogeny. The phylogeny was built using subsampled genome datasets from the country with the global genetic context (Figure S2). Moreover, to account for disparities in sequencing capacities across countries, we subsampled genome sequences to obtain a more balanced spatial distribution worldwide (Figure S3). Since late 2020, dN/dS ratios on spike S1 in many countries were greater than 1 (Fig. 1C and S4), consistent with previous global phylogeny results [6]. Additionally, the mutational fitness of SARS-CoV-2 and the number of nonsynonymous mutations in the S1 subunit also revealed a robust indication of adaptation (Fig. 1C). Negative selection pressure on the S2 subunit (Figure S4) and other proteins outside the S region was found, with dN/dS ratios consistently below 1 at all time points (Figure S5). Scatter plots between full-dose vaccine coverage and SARS-CoV-2 adaptation showed two clusters based on the emergence month of the Omicron variant in November 2021 (Figure S6). Furthermore, since November 2021, the adaptation of SARS-CoV-2 seems to exhibit different patterns (Fig. 1C). As of October 2021, dN/dS ratios in the S1 subunit were higher in low vaccine coverage regions compared to both medium and high vaccine coverage regions (Fig. 1D, Wilcoxon rank-sum tests, p < 0.05). This suggests that differences in vaccine coverage may be associated with variation in SARS-CoV-2 adaptation. Additionally, it appears that differences in full-dose vaccine coverage may also be related non-monotonically to the total number of COVID-19 deaths per million (Fig. 1E). In areas with medium vaccine coverage, the total number of deaths due to COVID-19 may be the highest, higher than in countries with low vaccine coverage, possibly related to variations in underreporting among different countries [40]. Considering the observed complex relationship between heterogeneity in global vaccine coverage, viral adaptation, and COVID-19 mortality, further investigations are warranted to understand how vaccine rollout shapes trajectories of SARS-CoV-2 adaptation. SARS-CoV-2 adaptation over different genome regions linked to vaccine rollout To account for the global use of various types of COVID-19 vaccines, including mRNA, inactivated, and viral vector vaccines (Fig. 1A), we adjusted monthly full-dose and booster vaccine coverages for each country based on the VE against symptomatic infection of different vaccine types [21]. Additionally, we accounted for the potential impact of vaccine immunity waning (Figure S7). Considering the selection pressure related to vaccines may differ for the Omicron variant compared to other variants (Figure S6, the study periods were divided into the pre-Omicron period (from March 2020 to October 2021) and the Omicron period (from January 2022 to September 2022), during which the Omicron variant became dominant (Figure S8). The DLNM was applied to quantify the association between SARS-CoV-2 adaptation and COVID-19 vaccine rollout across 86 countries during the pre-Omicron and Omicron period with the adjustment of PHSM, international travel, and infection dynamics. During the pre-Omicron period, our findings showed a nonlinear relationship between adjusted vaccine coverage and SARS-CoV-2 adaptation (Fig. 2). Specifically, for the S1 subunit, dN/dS ratios initially increased with adjusted vaccine coverage and maximized at the medium level of adjusted vaccine coverage, approximately 30%, with effect sizes of 1.64 (95% confidence interval [CI]: 1.31, 1.98) and 0.96 (95% CI: 0.47, 1.45) at low (25%) and medium (50%) levels of adjusted vaccine coverage, respectively (Fig. 3A). We identified a tipping point occurring at approximately 55% to 65% adjusted vaccine coverage when the association transitioned from significantly positive to insignificant, with an effect size of -1.89 (95% CI: -4.20, 0.43) at 75% adjusted vaccine coverage (Fig. 2A). Given genetic hitchhiking, the elevated dN/dS ratio may indicate positive selection or relaxed purifying selection. To further identify positive selection, we assessed the number of nonsynonymous mutations in the S1 subunits and mutational fitness of SARS-CoV-2, obtaining similar results to dN/dS ratios. This indicated that disparities in vaccine coverage could result in variations in viral adaptation, where areas with high vaccine coverage might encounter reduced vaccine-related positive selection compared to areas with low to moderate vaccine coverage. To account for potential effects of socioeconomic and demographic factors, we further adjusted for GDP per capita, Gini index, and the proportion of the population aged ≥ 65 years, and a similar nonlinear association between adjusted vaccine coverage and SARS-CoV-2 adaptation was observed (Figure S9). The associations between adjusted vaccine coverage and SARS-CoV-2 adaptation. (A) Associations between adjusted vaccine coverage and dN/dS ratios in the S1 subunit. (B) Associations between adjusted vaccine coverage and the number of nonsynonymous mutations in the S1 subunit. (C) Associations between adjusted vaccine coverage and mutational fitness of SARS-CoV-2. In (A–C), the analyses were based on data from 86 countries worldwide for the pre-Omicron period (March 2020 to October 2021) and the Omicron period (January 2022 to September 2022, during which the Omicron variant became dominant). Effects on SARS-CoV-2 adaptation were estimated using distributed lag nonlinear models relative to adjusted vaccine coverage = 0, which represents the unvaccinated population in the same countries before the vaccination campaigns. The gray shaded regions represent the non-significant ranges of adjusted vaccine coverage for SARS-CoV-2 adaptation during the pre-Omicron period, which were 54.6–76%, 59.8–76%, and 62.7–76%, respectively. Blue (green) lines and blue (green) shaded regions represent the estimated values and 95% confidence intervals. The vertical dashed lines in light blue, medium blue, and dark blue represent adjusted vaccine coverage at 25%, 50%, and 75% Association between adjusted vaccine coverage and natural immunity and the dN/dS ratio on different SARS-CoV-2 genes. (A) Associations between adjusted vaccine coverage and the dN/dS ratio on each gene. The analyses were based on data from 86 countries worldwide. (B) Associations between natural immunity and the dN/dS ratio on each gene. We estimated monthly natural immunity coverage similarly to calculating adjusted vaccine coverage. This calculation was based on the monthly estimated infections per hundred people from IHME, with further adjusted natural immunity using the effectiveness of preventing reinfections (89.3%) with the average protection of Alpha (90.2%), Beta (85.7%), and Delta (92.0%) [57]. We considered the waning of natural immunity at six months after infection (see Methods). All of the analyses for natural immunity and dN/dS ratios were based on data from the 22 countries with full-dose vaccination rates below 25% as of October 2021 (Table S6). In (A) and (B), effects on dN/dS ratios were estimated using distributed lag nonlinear models. The study period was from March 2020 to October 2021, i.e. the pre-Omicron period. These effects were calculated relative to adjusted vaccine coverage = 0, representing the unvaccinated population in the same countries before the vaccination campaigns. Dots and error bars represent the estimated values and 95% confidence intervals. We used a t-test to assess the difference in the association between adjusted vaccine coverage (natural immunity) and SARS-CoV-2 adaptation across the S region and proteins outside the S region As the inference of SARS-CoV-2 phylogeny for a specific country depends on the sampling strategy and sequences from external sequences, we conducted analyses with the same metrics recalculated based on the publicly available phylogeny for 50 countries (see Methods, Table S2). Similar nonlinear associations between adjusted vaccine coverage and SARS-CoV-2 adaptation were consistently observed (Figure S10). To assess the potential impact of human travel on viral introduction, we selected the countries with low international travelers between countries from March 2020 to October 2021 from above 50 countries and performed the same analysis. Similar nonlinear trends were still observed in the countries with lower traveler numbers (Figure S11). Given that infection-induced immunity is an important component of the immune pressure, we conducted sensitivity analyses using three alternative measures of natural infection dynamics: IHME-estimated infections (Figure S12), natural immunity coverage (Figure S13), and the monthly cumulative number of reported COVID-19 cases (Figure S14). The patterns of associations remained consistent across all three sensitivity analyses (Figures S12–S14). During the Omicron period, our analysis found that the relationship between SARS-CoV-2 adaptation and adjusted vaccine coverage differed from that observed during the pre-Omicron period, showing either no significant association or a weak negative association (Fig. 2). To assess the potential impact of random imputation of missing international travel data, we conducted a sensitivity analysis by directly including missing values in the model without imputation (Figure S15). Because international travel data were completely missing in May 2022, we additionally estimated values for that month by averaging each country’s April 2022 and June 2022 values (Figure S16). Both analyses (Figures S15–S16) yielded results consistent with the main analysis. Moreover, considering the potential impact of assumptions used to calculate adjusted vaccine coverage, we conducted additional sensitivity analyses using a lower VE against symptomatic infection for Omicron [41] (Figure S17) and an exponential decay function to model waning immunity (Figure S18). Both analyses showed that, during the Omicron period, the association between adjusted vaccine coverage and SARS-CoV-2 adaptation was generally not statistically significant. We also identified a positive association between adjusted vaccine coverage at low and medium levels and dN/dS ratios in E, M, N, ORF1a/b, ORF3a, ORF7a/b, and ORF9b proteins (Fig. 3B). However, this association was relatively weaker than the S regions (Student’s t-test, p < 0.001). Sensitivity analyses were performed by replacing dN/dS ratios with the number of nonsynonymous mutations, yielding similar results (Figure S19). To alleviate the potential impact of country selection on our study, we applied filters to the genome sequences and selected those with less than 1% ambiguous nucleotide sites and 0.05% unique amino acid mutations, while excluding insertions or deletions unless confirmed by the submitter. This allowed us to narrow down the scope from 86 to 59 countries. The analysis also revealed that dN/dS ratios over different genome regions were associated with adjusted vaccine coverage (Figure S20). Moreover, the result was similar when we then excluded countries with a low number of COVID-19 cases, including countries from Africa and those that had implemented zero-COVID policies for sensitivity analysis (Figure S21). To investigate the association between SARS-CoV-2 adaptation and adjusted vaccine coverage among different vaccine types, we classified the 86 study countries into mRNA, inactivated, or viral vector vaccine groups based on whether a particular vaccine type had reached 50% vaccination as of October 2021 (Table S3-S5). Certain countries were designated as having mixed vaccinations because no single vaccine type accounted for over 50% of the total vaccinations administered (Table S6). We observed positive associations between SARS-CoV-2 adaptation and adjusted vaccine coverage of mRNA and inactivated vaccines. The pattern observed for mRNA vaccines was similar to that in the main analysis (Figure S22). Considering that immunity induced by natural infection is also an important factor for positive selection, we selected 22 countries from low vaccine coverage regions (Fig. 1B) and treated them as representative of countries with almost no vaccination (Table S7). The dN/dS ratios in the S1/S2 subunits, as well as S, E, M, N, ORF1a/1b, ORF3a, ORF7a/7b, ORF8, and ORF9b proteins, were found to be associated with natural immunity. The associations observed in proteins outside the spike were weaker than S regions (Fig. 3). Implications of the association between COVID-19 mortality and viral adaptation for vaccination coverage Despite the potential links between positive selection and immunity from vaccines, many studies have demonstrated that vaccination can prevent the deaths of thousands of people [1, 3]. However, the relationship between COVID-19 mortality and SARS-CoV-2 adaptation remains unclear. We used CCM to investigate the dynamic interdependence between SARS-CoV-2 adaptation and COVID-19 mortality across 86 countries worldwide until October 2021. Globally, bidirectional coupling was observed, with statistically significant associations in both directions (Fig. 4A; Fisher’s method, global meta-significance p < 0.001, Table S8). Notably, more countries showed significant differences from the null distribution in the direction from COVID-19 mortality to SARS-CoV-2 adaptation than in the reverse direction (Figure S23), and cross-map skill was also higher in this direction globally (Wilcoxon rank-sum tests, p < 0.001, Fig. 4A). Furthermore, we found a positive correlation between COVID-19 deaths per million and SARS-CoV-2 adaptation (Fig. 4B), especially in countries with low vaccine coverage (Fig. 4C). Because dead hosts do not contribute to onward viral transmission or directly exert selective pressure, COVID-19 mortality may be better interpreted as an epidemiological indicator of uncontrolled viral spread in a vulnerable population. Association between monthly reported COVID-19 deaths per million and SARS-CoV-2 adaptation during the pre-Omicron period. (A) Detection of directional coupling between monthly COVID-19 mortality and SARS-CoV-2 adaptation. Box plots summarize directional coupling, quantified by cross-mapping skill (ρ), between monthly reported COVID-19 mortality and SARS-CoV-2 adaptation across 86 countries. Wilcoxon rank-sum tests were used to assess differences in the strength of directional coupling across groups. (B) Spearman’s correlation between monthly COVID-19 mortality and SARS-CoV-2 adaptation at the national scale from March 2020 to October 2021. “x” represents statistically significant results. (C) Summary of correlation between monthly COVID-19 mortality and SARS-CoV-2 adaptation. Points represent Spearman’s correlation between COVID-19 mortality and each SARS-CoV-2 adaptation metric for each country. Violin plots and box plots enclosed within violin plots represent the distribution and summary of Spearman’s correlation values across 86 countries. In (A) and (C), for box plots, the lower and upper hinges (bounds) correspond to the first and third quartiles (the 25th and 75th percentiles) calculated across all countries. The black bars represent the median (50th percentile) calculated across all countries. The upper (lower) whisker extends from the hinge to the largest (smallest) value, not exceeding 1.5 times the interquartile range (the distance between the first and third quartiles). In (A-C), the dN/dS ratio and number of nonsynonymous mutations in the S1 subunit, and mutational fitness of SARS-CoV-2 were used to quantify SARS-CoV-2 adaptation. The study period was from March 2020 to October 2021, i.e. the pre-Omicron period Discussion Deciphering the immunization factors influencing viral adaptation is critical for understanding preparedness and response to future pandemics. Our analysis suggests that global disparities in vaccine coverage may be associated with SARS-CoV-2 adaptation, implying complex interactions between population immunity and viral adaptation. Our study found that SARS-CoV-2 adaptation across various genome regions was associated with immunity from vaccines and natural infections, suggesting possible interactions between different genes, although the mechanisms may vary. Another study based on genome sequences pointed out that multiple epistatic interactions between proteins, such as S and N proteins, potentially laid the foundation for diversity, virulence, immune evasion, and transmission of viruses [10]. However, for the S1 subunit, an increase in population immunity was more strongly linked to the potential increase in positive selection than other proteins outside the spike. This aligns with previous studies showing the essential role of the S proteins of SARS-CoV-2, especially the S1 domains, in the adaptive evolution of viruses [6, 42–44]. Our analysis revealed that SARS-CoV-2 adaptation on S1 became insignificant when adjusted vaccine coverage was high during the pre-Omicron period, suggesting that disparities in vaccine coverage might lead to variations in virus adaptation. The nonlinear association between different level of adjusted vaccine coverage and SARS-CoV-2 adaptation offers valuable insights into equitable vaccine distribution. We observed no significant association between adjusted vaccine coverage and SARS-CoV-2 adaptation during the Omicron period, in contrast to the pre-Omicron period. This may partly reflect the shorter analysis window (January–September 2022) during the Omicron period, compared to the longer pre-Omicron period (March 2020–October 2021), potentially limiting statistical power to detect associations. It may also be attributable to both the immune escape ability of Omicron variants [45] and the profound shift in the global population immune landscape by early 2022. Widespread Omicron infection waves, superimposed on heterogeneous vaccination histories, generated substantial hybrid immunity and markedly reduced the pool of susceptible individuals. In addition, protection against SARS-CoV-2 reinfection became weaker and less durable during the Omicron era than in the pre-Omicron period, suggesting that the evolutionary pressure on SARS-CoV-2 shifted from intrinsic transmissibility to immune escape [46]. Population immunity shaped by both vaccination and infection continues to influence SARS-CoV-2 adaptation [47, 48]. Given the complex immune landscape during the Omicron era, it became more difficult to isolate the independent effect of vaccine-induced immunity on viral adaptation. Dose-sparing strategies, which aim to enhance vaccine coverage by providing partial immunity (such as using half-doses or extending dose intervals) under vaccine shortage, have been debated during the COVID-19 pandemic [13, 14, 17, 49–52]. These strategies could decrease disease and COVID-19 mortality compared to concentrating available vaccine doses on a small portion of the population [14, 52]. While it is generally believed that implementing such strategies in the short term may decrease the number of infections, the long-term outcome depends on the robustness of immunity, given that partial immunity could enhance the potential for antigenic evolution [13]. Previous theoretical models suggest that dose-sparing strategies may increase the proportion of partially vaccinated individuals, potentially creating a trade-off between an increase in the spread of vaccine-adapted variants and a reduction in cumulative deaths [53]. We observed a positive correlation between COVID-19 mortality and SARS-CoV-2 adaptation. Higher mortality may reflect uncontrolled viral spread in vulnerable populations, which may in turn be associated with sustained viral replication and more opportunities for mutation and adaptation. Given the intricate interplay among the changing immune landscape, epidemiology, and viral adaptation, our findings suggest the importance of fast vaccination to achieve high population immunity. There were some limitations of the study. First, we used reported COVID-19 case and death data for our study; however, the variation in underreporting among different countries [40] may increase the uncertainty of the study. Second, the study was designed at the population level instead of the individual level, and it did not consider the variation in immunity at the individual level and hybrid immunity. Although we attempt to explore the association between vaccine immunity and natural immunity with SARS-CoV-2 adaptation, differentiating these two types of immunity is challenging. Third, there was heterogeneity in the genomic surveillance capacity and data sharing of SARS-COV-2 worldwide. We addressed this by performing quality control and subsampling of sequences to mitigate sampling biases. Subsequently, we constructed separate phylogenies using subsampled datasets for each country with the genetic context. To address the effect of human travel, we chose a long-time window and calculated three adaptation metrics based on the internal branches from the phylogenetic tree built for a specific country. Moreover, given that the inference of phylogeny for specific countries depended on the sampling strategy and sequences from the global context, we downloaded publicly available phylogenetic analyses for 50 individual countries (Table S2). We further selected countries with lower international travelers to conduct sensitivity analysis, deeming them less influenced by human travel. Although the non-linear relationship between adjusted vaccine coverage and viral adaptation from the analysis still holds, further studies are warranted to investigate the impact of travel on viral adaptation. Fourth, this retrospective study only investigated associations between SARS-CoV-2 adaptation and vaccine rollout until October 2021. It is challenging to accurately distinguish the effects of vaccines on different variants of concern (VOCs), and the present analysis cannot directly predict how adaptation will change in the future. Fifth, vaccination coverage is closely correlated with socioeconomic inequality, making it challenging to disentangle their independent effects on SARS-CoV-2 adaptation. Socioeconomic disparities have been shown to influence transmission dynamics and disease severity even in the pre-vaccination period [54, 55]. Although we adjusted for GDP per capita, Gini coefficient, and population age structure, residual confounding cannot be fully excluded. Sixth, to explore the relationship between adjusted vaccine coverage and SARS-CoV-2 adaptation across different vaccine-type groups, countries were grouped by the dominant vaccine type as of October 2021. This simplified static classification may not fully capture changes in the predominant vaccine type within countries and may introduce temporal bias. This limitation may be particularly pronounced in countries where the dominant vaccine type changed substantially during the study period (e.g., from viral vector vaccines to mRNA boosters). Conclusions Our findings emphasize a nonlinear pattern regarding the relationship between different levels of adjusted vaccine coverage and SARS-CoV-2 adaptation based on global sequence datasets. In future pandemics, rapid and equitable vaccine distribution will be critical for increasing global immune population and for promoting healthy and sustainable socioecological systems. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Acknowledgements We would like to thank GISAID and all of the data contributors, including the originating and submitting laboratories for providing the genetic sequence data, on which this research is based. Abbreviations Nonsynonymous to synonymous divergence Public health and social measures Global Initiative on Sharing All Influenza Data Vaccine effectiveness Oxford Coronavirus Government Response Tracker Distributed lag non-linear model Natural cubic splines Akaike Information Criterion Optimal embedding dimension Convergent cross-mapping Institute for Health Metrics and Evaluation Gross domestic product Variants of concern Author contributions Z.W. designed the study. L.Y., Z.Z., and Q.L. collected, accessed, and verified the statistical data. L.Y. and Z.W. conducted the analyses. L.Y. and Z.W. wrote the first draft of the manuscript. Q.L., Y.C., and Y.S. provided important comments on the draft manuscript and edited the manuscript. All authors read and approved the manuscript. Funding This study was supported by the National Science and Technology Major Project (2025ZD01900203); Key research projects of the Beijing Natural Science Foundation-Haidian District Joint Fund (L232014); Beijing Research Center for Respiratory Infectious Diseases Project (BJRID2026-010); Fundamental Research Funds for the Central Universities (2233300001); National Natural Science Foundation of China (82204160); Science and Technology Projects of XizangAutonomous Region, China (XZ202501JD0012); Research on Key Technologies of Plague Prevention and Control in Inner Mongolia Autonomous Region (2021ZD0006). The funders had no role in study design, data collection and analysis, the decision to publish, or in preparation of the manuscript. Data availability The data used for this study is publicly available. This study is based on 13,243,751 SARS-CoV-2 sequences and associated metadata downloaded from the sequence repository of GISAID. Publicly available phylogenies can be downloaded at https://nextstrain.org/groups/africa-cdc and https://nextstrain.org/groups/neherlab. Data on vaccinations and epidemiology can be downloaded at https://ourworldindata.org/coronavirus, international travel data at https://dataforgood.facebook.com/dfg/tools/travel-patterns, the stringency index at https://www.bsg.ox.ac.uk/research/research-projects/covid-19-government-response-tracker, the estimated infection data at https://www.healthdata.org/covid/data-downloads, and socioeconomic and demographic data at https://data.worldbank.org/indicator. Declarations Ethics approval and consent to participate Not applicable. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement The data used for this study is publicly available. This study is based on 13,243,751 SARS-CoV-2 sequences and associated metadata downloaded from the sequence repository of GISAID. Publicly available phylogenies can be downloaded at https://nextstrain.org/groups/africa-cdc and https://nextstrain.org/groups/neherlab. Data on vaccinations and epidemiology can be downloaded at https://ourworldindata.org/coronavirus, international travel data at https://dataforgood.facebook.com/dfg/tools/travel-patterns, the stringency index at https://www.bsg.ox.ac.uk/research/research-projects/covid-19-government-response-tracker, the estimated infection data at https://www.healthdata.org/covid/data-downloads, and socioeconomic and demographic data at https://data.worldbank.org/indicator.