A major limitation for reconciling thresholds of protection (Table) is lack of a standardized assay for measuring in vitro neutralization titers. Although an international standard has been established (7), reported titers seem affected by the assay used, as would be expected from differences in cells, virus, and outcomes measured (8). For example, even when neutralization titers are measure against the same stocks of pooled convalescent-phase plasma (e.g., the World Health Organization [WHO] 20/130 standard), different assays produced geometric mean neutralization titers (GMT) that varied from 120 to >12,000 (7). Even after standardizing measurements from different assays into international units (Table), standardized neutralization titers across the assays still differed by up to 50-fold (7). This difference in neutralization titers across different assays is also evident when comparing the 3 studies quantifying the threshold of protection (Table) (2–4). For example, Gilbert et al. reported the GMT for mRNA-1273 as ≈247 IU/mL (4), compared with 1,057 IU/mL reported by Khoury et al. (2) (Appendix). A quick survey of the literature reveals 6 reported estimates of the GMT for mRNA-1273 vaccinees, ranging from 247 IU/mL (95% CI 231–264) to 1,404 (95% CI 795–2,484) IU/mL, depending on the study (Appendix Table 1). Similarly, estimates of the GMT for ChAdOx1 vaccinees ranged from 23 IU/mL (3) to 144 IU/mL (2). When the same neutralization assay is run across different laboratories, then international standards are probably a very effective mechanism for adjusting for interlaboratory variability. However, it is clear from those discrepancies that expression of titers in international units is insufficient for normalizing between different assays and comparing the thresholds of protection reported in these studies (Appendix), which most likely results from differences in the assays themselves (8).

An alternative approach for normalizing neutralization titers between studies is to assume that similar groups of vaccinees should have similar titers. For example, rather than relying on conversion to the WHO international units, we can assume that the mean neutralization for the mRNA-1273 vaccinees is similar in the phase 1/2 trials (as analyzed by Khoury et al. [2,10]) and in the phase 3 trial (as analyzed by Gilbert et al. [4,11]) (Appendix). Normalization is limited because it does not account for differences in baseline characteristics of the cohort vaccinated in each study (e.g., age), which may influence neutralization titers. However, because immunobridging studies also rely on comparing vaccine titers among different groups, this approach is pragmatic for overcoming the limitations of comparing different assays.

Figure 2

Figure 2. Comparisons of the estimated curves for protection from SARS-CoV-2 infection from 2 vaccines: A) mRNA-1273 (Moderna, https://www.modernatx.com) (4); B) ChAdOx1 (AstraZeneca, https://www.astrazeneca.com) (3...

Applying this normalization approach enabled us to compare the protection curves across different immune correlate studies (Appendix). We aligned the data by assuming that the mean titer for mRNA-1273- or ChAdOx1-vaccinated persons is the same between the phase 1/2 trials and the phase 3 trials for each vaccine (Figure 2; Appendix). Although this normalization is independent of the x-axis scale used, we plotted both curves onto a fold-of-convalescent level scale (Table) developed by Khoury et al. (2) for illustration. This transformation enabled a more direct comparison of the protection curve across the 3 studies. Considering the mRNA-1273 breakthrough-infection model (4) (Figure 2, panel A), for example, we saw good agreement with the Khoury et al. model (2) at the higher neutralization levels achieved with mRNA-1273 vaccination (albeit a seemingly slightly lower maximum protection level predicted in the breakthrough-infection model) but very poor agreement at low neutralization levels. This finding is easily understandable considering the distribution of individual neutralization titers in the mRNA-1273 breakthrough-infection study, in which only ≈10% of participants had a neutralization titer less than the mean titer of early convalescent-phase participants (Figure 2, panel A). Thus, neutralization data with which to estimate protection at lower neutralization levels are sparse (hence, the wide confidence intervals in this region of the curve). Similarly, the ChAdOx1 protection curve (Figure 2, panel B) shows good agreement with the Khoury et al. analysis (2) in the region in which neutralization data are available in the breakthrough-infection study (Figure 2, panel B).

The broad CIs and divergence of the models for which neutralization data are sparse suggests the need for caution when extrapolating the relationship between neutralization and protection beyond the ranges of data available in each study. The vaccine comparison approach has the advantage of fitting to a large span of neutralization titers (a 20-fold range in GMT between the 7 vaccines) (2), enabling prediction of the vaccine efficacy over a wide range of neutralization titers. Because none of the reported phase 3 studies of ancestral SARS-CoV-2 infection reported efficacy <50% or >95%, the vaccine-comparison analysis also extrapolates efficacy above and below these levels. However, studies of vaccine efficacy and effectiveness against SARS-CoV-2 variants suggests that the curve remains predictive against the Alpha, Beta, Delta, and Omicron variants, for which lower neutralization titers are observed (12; D. Khoury et al., unpub. data, https://www.medrxiv.org/content/10.1101/2021.12.13.21267748v2).

Figure 3

Figure 3. Breakthrough-infection data and protection from SARS-CoV-2 infection showing the association between neutralizing antibody titer and protection from symptomatic SARS-CoV-2 infection for an individual person. A) BNT162b2 (Pfizer-BioNTech, https://www.pfizer.com)...

The analysis above does not allow direct visualization or comparison of the fit of the data from breakthrough infection to the data from the vaccine-comparison study. We developed a method for estimating unadjusted protection at different neutralization levels from the breakthrough-infection data (Appendix), which also enables inclusion of data from a third breakthrough-infection study of BNT162b2 (Pfizer-BioNTech, https://www.pfizer.com) vaccinees (5). We show data from the 3 breakthrough-infection studies compared with the vaccine-comparison approach (normalized for the mean vaccinee titer in each study) (Figure 3). Data from the breakthrough-infection studies show remarkable agreement with the vaccine-comparison model (within the neutralization ranges for which sufficient data were available for each breakthrough-infection study), despite the fundamentally different data, assays, and approaches used to estimate protection curves in each study. Furthermore, after alignment to the GMT of each vaccine group, we can use the underlying distribution in neutralizing antibody titers along with the protection curves from each of these studies to predict the overall vaccine efficacy for existing vaccines (as has been done for the Khoury et al. model [2)]) (Appendix Methods). That approach reveals good agreement between all models and the observed data, at least in the ranges where data were available to parameterize the models (Appendix Figure 1). This approach provides cross-validation of the protection curves but also provides a lesson that all models should be used cautiously outside the ranges of the data over which they were developed.

Predicting vaccine efficacy or a clinically useful threshold of protection against COVID-19 would be a major advance. The in vitro neutralization titer has been demonstrated by multiple studies to be well correlated with vaccine efficacy and with a person’s protection from symptomatic SARS-CoV-2 infection (2–6,12; D.H. Khoury et al., unpub. data). The 4 studies that found significant relationships between neutralization titers and vaccine efficacy used different methods (2–5), and data from different clinical trials with neutralization data assessed across a range of neutralization assays. Those factors may all contribute to apparent discrepancies between the relationships reported in each study. However, we show that after centering the data from each study on the GMT of the vaccine used in each study, the 4 studies converge on a common prediction of the relationship between neutralization and protection against infection (within the bounds of data available within each study). The agreement of these studies strongly supports the use of neutralizing antibody titers to predict the efficacy of new vaccines or vaccine efficacy against new variants (assuming the fold drop in neutralization titer for the variant can be estimated). Although neutralizing antibody levels are a clear correlate of protection, identifying a protective threshold applicable to a serologic test is more challenging, in part because no such threshold exists, but instead, there is a gradient of vaccine efficacy that increases with neutralization. Furthermore, significant challenges to defining a particular threshold at which a person’s neutralization titer might be deemed to provide high protection from COVID-19 include the diversity of assays used to measure neutralization, the difficulty in translating neutralization levels between assays, the constant emergence of new and more escaped variants, and the uncertainties of estimating individual neutralization titers.

An additional major challenge is adapting assays (and protection curves) to deal with neutralization of current and future SARS-CoV-2 variants. The studies discussed in this analysis primarily deal with neutralization of and protection from the ancestral SARS-CoV-2 strain because the breakthrough-infection data and vaccine efficacy data in most studies was from phase 3 clinical trials (2–4), which studied infection within the first few months after vaccination and which were mainly conducted before variants of concern had a major foothold, except for the Bergwerk et al. study (5), which was conducted during in the Alpha-dominant period. It would be ideal to be able to adapt each model of immune correlates to test its ability to predict protection against variants of concern. However, until recently, only the vaccine-comparison model has been extended to analyze protection against SARS-CoV-2 variants (12,18,19; D. Cromer et al., unpub. data, https://www.medrxiv.org/content/10.1101/2022.06.09.22275942v1), although a recent study has begun to explore this question by using a breakthrough-infection approach (20). The work on the vaccine-comparison model approach has so far shown that this model, which was originally calibrated on data for ancestral SARS-CoV-2 infections, can also be used to predict vaccine effectiveness against SARS-CoV-2 variants and after boosting, as long as one adjusts for the drop in neutralization titers to the variants and rise in neutralization after boosting (12,18,19; D. Cromer et al., unpub. data). However, the need to standardize neutralization assays for SARS-CoV-2 variants presents an ongoing challenge.

In vitro neutralizing antibody titers against SARS-CoV-2 present a clear correlate of protection from symptomatic SARS-CoV-2 infection. Studies of passive administration of neutralizing monoclonal antibodies in animals and humans support that neutralizing antibody titers are a mechanistic correlate of protection (21–23). Indeed, a recent study comparing protective titers in prophylactic and therapeutic studies suggests that the protective titers may be very similar (E. Stadler et al., unpub. data, https://www.medrxiv.org/content/10.1101/2022.03.21.22272672v2). Neutralizing antibody levels are also correlated with protection from severe SARS-CoV-2 infection (2).

In conclusion, our findings show that the different COVID-19 correlate of protection studies, which seemingly report different thresholds of protection, have strong agreement. However, other immune responses may also play a substantial role in protection against progression from symptomatic to severe SARS-CoV-2 infection. The agreement across multiple studies of the relationship between neutralizing antibodies and efficacy against COVID-19 can be useful for planning future vaccine use, determining population immunity, and reducing the global effects of the COVID-19 pandemic.