Detecting Rapid Spread of SARS-CoV-2 Variants, France, January 26–February 16, 2021

Variants of severe acute respiratory syndrome coronavirus 2 raise concerns regarding the control of coronavirus disease epidemics. We analyzed 40,000 specific reverse transcription PCR tests performed on positive samples during January 26–February 16, 2021, in France. We found high transmission advantage of variants and more advanced spread than anticipated.

We used a generalized linear model (GLM) to analyze the binary strain variable (with values wildtype or variant). The covariates were the patient's age, the RT-PCR kit used for variant detection, the sampling date, and the geographic region from which the sample originated (Appendix 2). By using a type-II analysis of variance, we found that all covariates except the type of RT-PCR kit to be significant (Table 1). In particular, the proportion of variants increased with date and decreased with age (Appendix 2 Figure 2) and hospital origin.
To investigate the temporal trends, we fitted a logistic growth model to the fitted values of an analogous GLM only on the data from general population samples (Appendix 2). Assuming that variations in frequencies are driven by transmission advantages, we found that variants have a 50% (95% CI 37%-64%) transmission advantage over wild-type strains ( Figure 1).
The analysis of variance already showed that variant frequency varied across regions (Table 1). We performed the logistic growth fit at the local level for regions for which adequate data was available ( Figure 2). The growth advantage of the variant was more pronounced in some regions. In Ile-de-France, more than half of infections already appeared to be caused by the variants by February 16, whereas in other regions, such as Burgundy, this proportion would not be reached until March 2021. However, some regions were less well represented in this analysis, which could affect local estimates.
Finally, we investigated the correlation between the increase in variant frequency among positive tests in a region and the temporal reproduction number, denoted R t , in that same region. R t was estimated from coronavirus disease intensive care unit admission data by using the EpiEstim method (6) with a serial interval from Nishiura et al. (7), as described in Reyné et al. (unpub. data, https://www. medrxiv.org/content/10.1101/2020.12.05.2024437 6v1) (Appendix 2). We used the Spearman rank correlation test and found a positive but nonsignificant trend (ρ = 0.50; p = 0.09) (Appendix 2 Figure 3).

Conclusions
We used 2 variant-specific RT-PCR tests to detect the fraction of infections caused by SARS-CoV-2 lineages B.1.1.7, B.351, and P.1 in regions in France during January 25-February 16, 2021. We did not find any significant difference between the 2 specific RT-PCR kits used, suggesting that similar data collected in France could be pooled. Our results have several practical implications.
In general, we found that many infections screened were caused by variants, especially B.1.1.7, and the trend increased over time. On the basis of our estimates, by February 16, 2021, more than half of SARS-CoV-2 infections in France could have been caused by variants, although with pronounced spatial heterogeneity. In a conservative scenario, where all uninterpretable tests were assumed to be caused by the wild type, most infections would have been caused by variants by the end of week 7 of 2021, and the estimated variants transmission advantage was 36% (95% CI 26%-48%) (Appendix 2 Figure 4).
Variant-positive samples originated from significantly younger patients, which is consistent with an earlier report (E. Volz et al., unpub. data) but contrasts with Davies et al. (4). Our analysis did not enable us to discriminate between epidemiologic effects (e.g., if variants' transmission chains were seeded in different populations than the wild types), sampling biases, or biologic effects. Additional data from RT-PCR amplification cycles could provide useful insights. Finally, earlier reports have found variant proportion to be associated with higher basic reproduction number (4; E. Volz et al., unpub. data). We found such a trend among regions in France, but it is not statistically significant.
A limitation of this study is that, in spite of its intensity, the sampling was performed retrospectively, which could generate biases if, for instance, transmission chains associated with variants were increasingly sampled. However, we found that samples that originated in hospitals were associated with a lower variant detection. Because testing in the general population is usually performed a week after infection and hospital admissions occur ≈2 weeks after infection (M.T. Sofonea et al., unpub. data, https:// www.medrxiv.org/content/10.1101/2020.05.22.20 110593v1), we expect hospital data would reflect an older state of the epidemic than screening data. RT-PCR does not have the resolution of full-genome sequencing, and other variants of concern could be underestimated or missed with this approach. However, the time scale considered and the relatively slow evolutionary rate of SARS-CoV-2 make this approach appropriate to monitor variant spread. Furthermore, next-generation sequencing performed on 48 samples showed a strong consistency with the specific RT-PCR tests (Cohen κ of 1 for the TIB Molbiol test and 0.87 or 0.88 for the ID Solutions test depending on the variant; data not shown).
These results illustrate that variant-specific RT-PCRs are an option for SARS-CoV-2 epidemic monitoring because of their affordability and rapid deployment. They also demonstrate that SARS-CoV-2 variants spread in France was faster than anticipated (L.D. Domenico et al., unpub. data, https://www. medrxiv.org/content/10.1101/2021.02.14.2125170 8v1), which stresses the importance of swift public health responses.

Zoonotic Infections
To revisit the December 2020 issue, go to: https://wwwnc.cdc.gov/eid/articles/issue/26/12/ The main model was performed by using a GLM assuming a binomial distribution, where the variable of interest was the binary variable strain (i.e., wild type or variant) and the explanatory variables were the sampling date (integer), the individual age (integer), the test kit used (boolean), the location of the sampling (boolean), and the region (factor). We further added an interaction between the region and the date. Odds ratios were computed by estimating a likelihood profile. We use a type II error for the analysis of variance given the uneven sampling between regions (the analysis of variance function in the car package of R).

Logistic Growth Fitting
We used the fitted values of the GLM model applied to the data after removing samples that came from hospitals (the sampling location effect was also obviously removed from the model) to perform the inference of a 2-parameter logistic growth kinetic curve:  (2) and a constant equal to 1. The CI relies on those of the estimated relative growth rate.

Reproduction Number and Variant Increase
The reproduction number was estimated by using the method described in Reyné et al.
To estimate the increase in the proportion of variants among positive tests, we performed a GLM with a binomial distribution to explain the type of infection (wild type or variant) as a function of 2 factors (sampling date and individual age). This was done only on data collected outside hospital settings. The regression coefficients were used to perform a Spearman's rank correlation with the most recent time point reproduction number estimate (February 16, 2021) (Appendix 2 Figure 4).