Data sources

PROVENT (NCT04625725) was a randomised, double-blind, placebo-controlled, multicentre, phase 3 clinical trial that assessed the efficacy of a single intramuscular 300-mg dose of tixagevimab–cilgavimab compared with placebo (2:1 randomisation; N = 5197) for the prevention of symptomatic COVID-19⁷. The primary endpoint was a reduction in incidence of reverse transcription-polymerase chain reaction (RT-PCR)–confirmed symptomatic COVID-19 with tixagevimab–cilgavimab compared with placebo in SARS-CoV-2–naive participants who had not received a COVID-19 vaccine. PROVENT recruited adults ( ≥18 years of age) with an increased risk of inadequate COVID-19 vaccination response or SARS-CoV-2 exposure. Exclusion criteria included prior SARS-CoV-2 infection, receipt of a vaccine or biologic for prevention of COVID-19, or allergy to any component of tixagevimab–cilgavimab or placebo.

The SUPERNOVA main cohort study (NCT05648110) was a randomised, double-blind, multicentre, phase 3 clinical trial to assess the efficacy of sipavibart relative to comparator (1:1 randomisation; N = 3335) for the prevention of symptomatic COVID-19¹⁰. The study was ongoing at the time of this analysis. SUPERNOVA enrolled participants ≥12 years old who had a qualifying immunocompromising condition. Participants received a second dose of their original randomised study intervention (sipavibart or comparator) after 6 months. Originally the comparator arm comprised of only tixagevimab–cilgavimab; at the request of the regulatory authorities on June 14, 2023, the comparator was changed to placebo, leading to redosing in participants originally receiving tixagevimab–cilgavimab. The primary efficacy outcomes were symptomatic COVID-19 caused by any variant or by non-Phe456Leu–containing variants within 181 days of dosing. The primary analysis data cut was event driven, leading to longer follow-up for participants first receiving tixagevimab–cilgavimab (192.6 days) compared with those receiving placebo (146.2 days).

Both trials were conducted in accordance with guidelines from the Declaration of Helsinki, Council for International Organisations of Medical Sciences, and Good Clinical Practice Council for Harmonisation. The protocols were approved by the appropriate institutional review board or ethics committee at the study sites. The list of institutional review boards and ethics committees that approved the studies are available (Supplementary Files 1 and 2). All participants provided written informed consent. The institutional review boards/ethics committees were not required to approve the present modelling study. Since the purpose of the analysis is to better characterise the efficacy of the mAb in the disease of interest, this is regarded as primary use of the data, and the participating patients directly consented to such use of the data in the informed consent form.

Predicted daily serum mAb concentrations

Daily serum mAb concentrations were generated as input for this analysis based on the previously developed popPK models for tixagevimab–cilgavimab and sipavibart. Both models were two-compartment with linear elimination and first-order absorption following intramuscular administration. The tixagevimab–cilgavimab popPK model was fit to data from eight phase 1, 2, and 3 studies²⁶. The sipavibart popPK model included all available data from ongoing sipavibart studies^10,27,28.

For PROVENT, daily serum mAb concentrations for each participant receiving tixagevimab–cilgavimab with at least one quantifiable post-baseline pharmacokinetic (PK) sample were generated using the tixagevimab–cilgavimab popPK model, with the tixagevimab–cilgavimab concentration calculated as the sum of concentrations of the individual mAbs²⁶. Individual predictions (empirical Bayesian estimates) were generated from time of dosing through day 366 or receipt of a second dose of tixagevimab–cilgavimab, and were available for 3272 of the 3442 participants in the tixagevimab–cilgavimab arm, including all 63 participants with primary events.

In the main cohort of SUPERNOVA, PK samples were collected from approximately 600 participants in the sipavibart arm, with additional samples taken at all illness visits. Daily individual serum mAb concentration predictions were generated using the sipavibart popPK model. Of the 1649 participants who received sipavibart, individual popPK predictions (empirical Bayes estimates, sampled daily) were generated for 768 participants with at least one available post-dose PK sample (at scheduled visits and/or illness visits). For the remaining 881 participants, popPK simulations were generated based on the individual participant dosing records and baseline characteristics¹⁰. The output was the mean of 10 replicates per individual, conducted by sampling from between-subject variability. Of 1649 participants in the sipavibart arm, 122 had a reported SARS-CoV-2 case per the primary endpoint definition for overall efficacy, 115 of whom had individual popPK predictions generated, and seven of whom had no available post-dose PK samples and therefore had PK profiles estimated based on simulation.

In vitro potency and serum nAb ID₅₀ titre assessments

In vitro IC₅₀ of tixagevimab–cilgavimab and sipavibart were assessed, along with observed serum nAb ID₅₀ titres from sera collected from participants in PROVENT in the same pseudovirus neutralisation assay (PhenoSense SARS-CoV-2 nAb assay; Monogram Biosciences, San Francisco, CA). The PhenoSense SARS-CoV-2 nAb assay is based on previously described methodologies^29,30 using a lentivirus system, where human immunodeficiency virus (HIV) pseudovirus virions expressed the SARS-CoV-2 spike protein from the variant of interest. HEK293 cells were transfected with a HIV genomic vector that contained a luciferase report gene plus an envelope expression vector carrying the SARS-CoV-2 spike protein open reading frame. Neutralising titres or IC₅₀ values were measured by assessing inhibition of luciferase activity, following preincubation of pseudovirions with serial dilutions of sera from trial participants or from in vitro dilutions of mAbs.

Following from previous work²¹, the predictive accuracy of nAb ID₅₀ titres derived from serum concentration and in vitro IC₅₀ values for PROVENT was explored using a small subset of the observed serum nAb ID₅₀ titres measured in the same assay as the in vitro IC₅₀ values (Supplementary Methods).

Variant prevalence data

Full-length SARS-CoV-2 genomic sequencing data from the Global Initiative on Sharing All Influenza Data (GISAID) database were used to infer country-specific variant prevalence^31,32. For PROVENT, variants, except for Omicron, were grouped by their World Health Organisation (WHO) nomenclature and regional daily prevalence was calculated per WHO group (Table 2). Omicron variants were grouped by major subvariants BA.1, BA.1.1, BA.2, BA.3, and BA.4/5 (with BA.4 and BA.5 sharing identical spike sequences). Variants with ≥1% daily prevalence were reported and those without a WHO label were omitted (4.7% of total sequenced data). During SUPERNOVA, SARS-CoV-2 surveillance sequencing was significantly reduced, leading to higher variability in prevalence percentages, especially for less prevalent lineages; variants were therefore limited to those with daily prevalence >5%. Additionally, due to fewer lineage IC₅₀ values available, Pango lineage assignment based on spike sequence only (Hedgehog) was utilised³³. This approach allowed IC₅₀ values to be assigned to all lineages sharing an identical spike sequence. Measured IC₅₀ values against the virus with a matched spike sequence of a given lineage were used where possible. Lineages without available measured IC₅₀ values were categorised into three groups for IC₅₀ value assignment: with mutation at 455 (L455X); with mutation at 456 (F456X); without mutation at 455 or 456. Lineages with L455X were assigned an IC₅₀ value of 83.1 ng/mL, corresponding to JN.1-containing L455S. Lineages with F456X were assigned an IC₅₀ of 1000 ng/mL, corresponding to the upper limit of the assay, consistent with loss of potency to variants containing F456L based on measured IC₅₀ values against multiple variants containing this mutation. Lineages without either of these mutations were assigned a surrogate IC₅₀ value of 7.1 calculated based on the average IC₅₀ values of variants without either L455X or F456X. On-study prevalence is presented in the Supplementary Methods.

Prevalence-adjusted in vitro potency

Variant prevalence data were combined with individual-level data by region (country for PROVENT/continent for SUPERNOVA) and calendar date^31,32. Once mapped, the prevalence-adjusted IC₅₀ was derived as prevalence-weighted geometric mean of IC₅₀ values. For each participant and day, we calculated a prevalence-adjusted IC₅₀ as a weighted geometric mean of variant-specific IC₅₀ values (Table 2), based on regional variant prevalence, reflecting each participant’s exposure to variants circulating at that time. Additional details are provided in the Supplementary Methods.

Prevalence-adjusted predicted daily nAb ID₅₀ titres

Daily prevalence-adjusted predicted nAb ID₅₀ titres were generated from the predicted daily mAb (tixagevimab–cilgavimab or sipavibart) serum concentrations and prevalence-adjusted in vitro IC₅₀ values: serum mAb concentration (ng/mL)/prevalence-adjusted IC₅₀ (ng/mL). The predicted nAb ID₅₀ titre formula has been shown to align well with observed nAb ID₅₀ titres following tixagevimab–cilgavimab administration across different SARS-CoV-2 variants²⁶. This method of deriving predicted nAb ID₅₀ titres provides an estimate that is specific to the administered mAb, independent of baseline nAb titres related to previous exposure or vaccination status or changes in nAb titres related to vaccination or SARS-CoV-2 infection during the clinical trial (consistent with the while-on-treatment primary estimands). Further, through prevalence-adjustment it is hypothesised that the associated nAb titre-efficacy relationship is independent of the SARS-CoV-2 variant.

A ToP for clinically relevant efficacy derived from the relationship with prevalence-adjusted nAb ID₅₀ titres

This model is referred to as the PROVENT ToP model. A Cox model with time-varying covariates was fit to RT-PCR–confirmed symptomatic COVID-19 through day 366, with covariates for treatment and its interaction with an intercept term and log₁₀(prevalence-adjusted nAb ID₅₀ titres + 1). This model is similar to that used to establish the associated efficacy–nAb titre relationship for HIV³⁴. Time was measured from randomisation and participants were censored for unblinding for consideration of, or for receipt of, a COVID-19 vaccine⁷.

Four time-varying Cox models were considered. The first was a single-parameter time-varying Cox model that is supported by the biologically plausible assumption that the nAb-efficacy curve passes through the origin (no intercept). The second was a two-parameter model that relaxes this assumption to allow greater model flexibility (efficacy intercept). The third also relaxes this assumption whilst adjusting for baseline covariates (efficacy intercept + baseline covariates). The fourth increased from zero efficacy once a threshold level of nAb ID₅₀ titres was achieved (nAb ID₅₀ titre intercept). The second model was chosen. Further details of the selected ToP model are presented in the Supplementary Methods. Once the appropriate parameterisation was selected further evaluations of the appropriate transformation of prevalence-adjusted nAb ID₅₀ titres were evaluated and the transformation log₁₀(prevalence-adjusted nAb ID₅₀ titres + 1) was applied (Supplementary Methods).

Validation of the PROVENT ToP model

Data from SUPERNOVA was used to externally validate predictions from the PROVENT ToP model using the three approaches described.

As it may be of interest to examine how mAb efficacy changes over time following administration (i.e., to assess relationship with falling serum mAb concentration), an assessment of whether the PROVENT ToP model could predict overall (i.e., attributable to any SARS-CoV-2 variant) instantaneous efficacy from SUPERNOVA was conducted. The PROVENT ToP model was evaluated at the daily geometric mean of the prevalence-adjusted nAb ID₅₀ titres to generate daily instantaneous efficacy predictions and two-sided 95% CIs for the SUPERNOVA study. To compare this with observed data, daily Epanechnikov kernel-smoothed hazard functions were derived with an optimised window, and the observed efficacy was estimated as 100(1 – hazard ratio) (%). Further details are provided in the Supplementary Methods.

In clinical trials, efficacy is commonly aggregated through to data cutoff. Therefore, an assessment of whether the PROVENT ToP model could predict the overall average efficacy from SUPERNOVA was conducted cumulatively for each day since first dose. The geometric mean of prevalence-adjusted nAb ID₅₀ titres up to each day since first dose from SUPERNOVA were evaluated by the PROVENT ToP model. These were visually compared with efficacy that would have been reported at each day since first dose under the pre-defined efficacy analysis from SUPERNOVA¹⁰, referred to here as observed efficacy. Further details are provided in the Supplementary Methods.

Given the prevalence-standardised approach used in the PROVENT ToP model, it is of interest to assess the ability of the model to predict efficacy against a specific variant or a mixture of variants. Planned analyses from SUPERNOVA included the average efficacy of overall, matched non-F456X mutation variants, subvariant-specific estimates, and F456X mutation variants at 3-month and 6-month timepoints relative to any dose¹⁰. The geometric mean of the variant-specific or prevalence-adjusted nAb ID₅₀ titres over the two timepoints were evaluated by the PROVENT ToP model to provide predictions corresponding to these efficacy estimates. External predictive accuracy was assessed using Lin’s CCC separately for the two timepoints. Further details are provided in the Supplementary Methods.

Model robustness to SUPERNOVA data

The model was updated with the inclusion of both PROVENT and SUPERNOVA data to assess the PROVENT ToP model robustness. This model is henceforth referred to as the PROVENT + SUPERNOVA ToP model. The time-varying Cox model was fit to the pooled data with a study-specific stratification factor to allow the baseline hazard to vary. Further details are provided in the Supplementary Methods.

Conversion to international units (IU/mL)

The same pseudovirus neutralisation assay (PhenoSense SARS-CoV-2 nAb assay; Monogram Biosciences, San Francisco, CA) was used for measuring mAb IC₅₀ values and serum nAb ID₅₀ titres. Mean nAb ID₅₀ titres generated by the replicate testing of the WHO first international standard, National Institute for Biological Standards and Control 20/136 using the PhenoSense SARS-CoV-2 nAb Assay (Wuhan D614 variant, 24 measurements across multiple vials, operators, and days) were used to calculate a conversion factor of 0.1428 that enabled the conversion of nAb median infectious dose titre to IU/mL. Although this standard was established for the Wuhan variant at the start of the pandemic, neutralisation titres across variants are frequently presented on the same titre scale and the titre shift (fold-change from one variant to another) is considered reflective of change in neutralising ability of sera or mAbs. Similarly in this work, we combined titres across multiple variants based on variant prevalence. The format of pseudovirus neutralisation assay used here was the same across multiple variants, except for the change in SARS-CoV-2 spike sequence, so any change in titre was the result of change in neutralising ability of a mAb or sera. The same factor could therefore be applied to convert the resulting titre threshold to IU/mL.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.