Neutralization titer biomarker for antibody-mediated prevention of HIV-1 acquisition

The Antibody Mediated Prevention trials showed that the broadly neutralizing antibody (bnAb) VRC01 prevented acquisition of human immunodeficiency virus-1 (HIV-1) sensitive to VRC01. Using AMP trial data, here we show that the predicted serum neutralization 80% inhibitory dilution titer (PT80) biomarker—which quantifies the neutralization potency of antibodies in an individual’s serum against an HIV-1 isolate—can be used to predict HIV-1 prevention efficacy. Similar to the results of nonhuman primate studies, an average PT80 of 200 (meaning a bnAb concentration 200-fold higher than that required to reduce infection by 80% in vitro) against a population of probable exposing viruses was estimated to be required for 90% prevention efficacy against acquisition of these viruses. Based on this result, we suggest that the goal of sustained PT80 >200 against 90% of circulating viruses can be achieved by promising bnAb regimens engineered for long half-lives. We propose the PT80 biomarker as a surrogate endpoint for evaluation of bnAb regimens, and as a tool for benchmarking candidate bnAb-inducing vaccines.

Neutralization sieve analysis showed that PE is strongly dependent on the neutralization sensitivity of an HIV-1 isolate to VRC01, measured as in vitro 80 or 50% inhibitory concentration (IC 80 or IC 50 , respectively). In particular, statistical tests showed that PE is significantly greater against viruses with lower IC 80 or lower IC 50 , and the result was replicated across each of the individual AMP trials in two distinct cohorts. In HVTN 704/HPTN 085 (704/085), which enrolled 2,699 transgender individuals and men who have sex with men in Brazil, Peru and the United States, PE was 73.0% (95% confidence interval (CI) 27.6-89.9%) against viruses with IC 80 < 1 µg ml -1 . Moreover, in HVTN 703/HPTN 081 (703/081), which enrolled 1,924 heterosexual women in Botswana, Kenya, Malawi, Mozambique, South Africa, Tanzania and Zimbabwe, PE was 78.6% (95% CI 17.3-94.4%) against viruses with IC 80 < 1 µg ml -1 . In each trial, PE was near zero for viruses with IC 80 > 1 µg ml -1 .
The parameter IC 80 is a neutralization property of a given antibody clinical lot against a given HIV-1 pseudovirus in vitro. In the absence of a validated nAb titer correlate of protection, summary measures (for example, geometric mean) of a bnAb IC 80 against each pseudovirus in a panel likely to reflect circulating HIV-1s have been used to predict bnAb prevention potential. For example, geometric mean IC 80 and the percentage of viruses neutralized at a specific IC 80 cutoff (that is, neutralization breadth in vitro) against a panel of representative HIV-1 pseudoviruses have been used to help identify the most promising bnAbs, multispecific bnAbs and bnAb combinations to advance into clinical trials for their potential to achieve high prevention efficacy 7,8 . However, because IC 80 contains no information about the concentration of bnAb in an individual's sera it is therefore not sufficient to quantify the ability of that individual's blood to neutralize an exposing virus at a given time because antibody concentration changes over time after administration. Logically, both time-varying bnAb concentration and bnAb potency/breadth, as quantified by distributions of IC 80 , might be required to estimate prevention efficacy. Certain mathematical combinations of these quantities (for example, refs. 9,10 ) have been assumed to be correlates of prevention efficacy in previous modeling studies.
In this work, we analyzed a nAb titer biomarker that integrates IC 80 with serum bnAb concentration, PT 80 , of a VRC01 recipient's serum at a given time to a given reference virus as a correlate of VRC01 prevention efficacy in the AMP trials. Figure 1 depicts how the PT 80 biomarker is calculated using serum bnAb concentration and IC 80 , where the PT 80 of a bnAb against an exposing virus population (rather than a single reference virus) can be determined using the geometric mean of the IC 80 values of bnAb against each exposing virus. The results support the premise that the PT 80 biomarker can be used to predict HIV-1 PE of bnAb regimens, and hence advance this biomarker as a correlate of protection for evaluation of bnAb regimens and as a tool for guiding research on candidate bnAb-inducing vaccines. Table 1 provides information on the types of cohorts that were used for each analysis.

Cohorts used for each analysis. Supplementary
Serum neutralization ID 80 titer is well estimated by PT 80 . The failure of IC 80 as a self-sufficient correlate was observed in the AMP trials, in that no PE was observed against sensitive viruses (IC 80 < 1 μg ml -1 ) after antibody levels had decayed to becoming undetectable (Supplementary Table 2). Therefore, we assessed serum neutralization titer as a correlate. We previously reported that, in VRC01 recipients in the HVTN 104 trial (healthy, uninfected participants at low risk of HIV-1 acquisition) and in VRC01 recipients in the AMP trials who remained HIV-1 negative through at least the week 88 visit, the experimentally measured serum ID 80 titer of VRC01 against a given virus was well estimated by PT 80 , defined as an individual's VRC01 serum concentration divided by the IC 80 of the VRC01 drug product against the same virus (refs. 11,12 , respectively) ( Fig. 1). In VRC01-recipient cases in the AMP trials (that is, participants with confirmed acquisition of HIV-1 infection after enrollment and by the week 80 visit 6 ), a similar result was observed in sera obtained at the study visit (and, for a subset, from the last two study visits) immediately before the first positive HIV-1 RNA PCR test, when assayed against autologous isolates (Fig. 2).
PT 80 associated with VRC01 prevention efficacy. To define the level of VRC01 serum neutralization that correlates with reduced HIV-1 acquisition in AMP, we first used the median VRC01 serum concentration (calculated as the median observed VRC01 concentration at all mid-infusion visits across all noncases in the case-control samples) and the IC 80 of the acquired virus to estimate how VRC01 PE varied with PT 80 against the autologous virus. Achieving 50, 75 and 90% PE required PT 80 values of 32, 82 and 194, respectively; these titers were higher than those needed to achieve the same level of protection in the NHP high-dose challenge model 4 (8,32 and 83, respectively) (Fig. 3). These protection-by-PT 80 curves looked similar in two different sensitivity analyses of the NHP challenge model data, where bnAb and/or SHIV challenge virus data with outlying protection curves were excluded (figure 2 of ref. 4 ; Extended Data Fig. 1).
Population-level PT 80 correlate of protection. The above estimates, which use a representative VRC01 serum concentration (median), are imprecise given that individual VRC01 concentrations at HIV-1 acquisition differ from this representative concentration. Supporting an alternative approach, the even spread of estimated infection dates across infusion intervals in placebo recipients (Extended Data Fig. 2) suggested that HIV-1 exposures occurred approximately uniformly over time. When population pharmacokinetics (popPK) modeling-which incorporates data from all individuals in a given cohort to study PK at the population level-was used to estimate daily VRC01 concentrations, PT 80 values were markedly higher against sensitive versus resistant placebo recipient-acquired viruses (taken to be representative of strains circulating in the trial settings): median over days and viruses (interquartile range) PT 80 was 41.3 (16.3-112.0) against IC 80 < 1 µg ml -1 virus, 9.6 (4.1-23.7) against IC 80 1-3 µg ml -1 virus and 1.3 (0.3-4.3) against IC 80 > 3 µg ml -1 virus (Fig. 4). Notably, for virus sensitivity categories 1-3 and >3 µg ml -1 , many PT 80 values fell below the limit of detection of the TZM-bl target cell neutralization assay (titer = 10) and, in the >3 µg ml -1 category, nearly half the PT 80 values were <1.0 (although PT 80 values <1.0 can be predicted, they were not experimentally tested, which would require concentration of serum).

Low PT 80 values against autologous virus at HIV-1 acquisition.
If serum VRC01 neutralization titer is a true correlate of prevention efficacy, then PT 80 values of VRC01-recipient noncases (their individual-specific median PT 80 values over follow-up) against viruses acquired by placebo recipients must therefore be higher than those of VRC01-recipient cases at acquisition against autologous viruses (assuming that exposures to HIV-1 occur approximately uniformly over time). Noncases are defined as participants who completed the week 88 visit without HIV-1 infection diagnosis. We found that, indeed, autologous PT 80 values were higher in VRC01-recipient noncases, with a geometric mean of 4.3 (95% CI 4.0-4.7) compared with a geometric mean of 1.5 (95% CI 0.9-2.4) in VRC01-recipient cases, a ratio of 2.9 (95% CI 1.8-4.8, P < 0.001); similar results were obtained for individual dose arms ( Fig. 5 and Methods). In these calculations, each individual's HIV-1 acquisition date was derived by a Bayesian procedure that combines an estimate obtained by fitting a Poisson distribution to the gag-pol-nef sequences from the first RNA PCR-positive time points 13 with information from the HIV diagnostic assays that were applied to 4-weekly samples 14   ) are used to calculate the PT 80 biomarker. a, Formula for calculation of PT 80 for a bnAb against a target virus. IC 80 for a clinical lot bnAb product against a target virus, as determined by the TZM-bl target cell assay, is the bnAb concentration needed for 80% reduction in RLU compared with target virus control wells after subtraction of background RLU. Based on the PT 80 biomarker, increasing bnAb serum concentration and increasing target virus sensitivity (that is, decreasing IC 80 ) have an equal impact on increasing PT 80 and hence improvement in potential prevention efficacy. b, Example calculations showing how PT 80 against a target virus differs for three different bnAbs sharing the same IC 80 against the target virus yet are present at different serum concentrations. A similar result would be obtained (differing PT 80 values) if the same bnAb was present at three different serum concentrations. c, Adaptation of the formula shown in a to a scenario where average PT 80 is calculated against a population of exposing viruses. d, Example calculations of average PT 80 over a follow-up period against an exposing virus population for three different bnAbs. The yellow bnAb has characteristics of VRC01 observed in the AMP trials (average serum concentration over VRC01 recipients and over 80 weeks of follow-up 20 µg ml -1 , average IC 80 of exposing viruses 4.0 µg ml -1 , which is calculated as the weighted average of the three IC 80 s: for example, (0.5 µg ml -1 ) × 0.30 + (2.0 µg ml -1 ) × 0.15 + (6.5 µg ml -1 ) × 0.55 = 4.0 µg ml -1 ). If IC 80 is used for comparison of the yellow and blue bnAbs, the results indicate that the blue bnAb is twofold better than the yellow in regard to its potential prevention efficacy, whereas if PT 80 is used for comparison the blue bnAb is fivefold better than the yellow. The PT 80 biomarker is superior on account of its enhanced measurement of neutralization potency against anticipated exposing viruses.
on HIV-1 acquisition. Affirmatively, the 30 mg kg -1 arm had 23% improved efficacy compared with the 10 mg kg -1 arm 6 (calculation shown in Methods) and the 30 mg kg -1 arm had 2.4-fold higher median mid-infusion visit serum VRC01 concentration than the 10 mg kg -1 arm, where the correlate of risk result yielded 25% risk reduction per 2.4-fold increase in concentration.

Prediction of prevention efficacy of a triple-bnAb regimen.
Various bnAb combinations and bispecific Abs, enabling consecutive targeting of different epitopes, are in clinical development for HIV-1 prevention 16,17 . Predicted potency-breadth curves have suggested that a triple combination consisting of one CD4bs-targeting bnAb, one V2 glycan-targeting bnAb and one V3 glycan-targeting bnAb would have the best neutralization coverage against a panel of clade C viruses 7 . One such combination, VRC07-523LS (CD4bs), PGT121 (V3 glycan) and PGDM1400 (V2 glycan), has been predicted to have 99% neutralization breadth (IC 80 < 10 μg ml -1 ) and a geometric mean IC 80 of 0.09 μg ml -1 against a diverse panel of viruses, compared with 76% neutralization breadth and a geometric mean IC 80 of 2.64 μg ml -1 for VRC01 alone 17 . (LS refers to a Fc-modified version to extend half-life.) This specific triple-bnAb combination is in development for potential efficacy testing and has been safely administered to adults without HIV in a recent phase 1 study 18 . We therefore decided to apply PT 80 to predict PE of 20 or 40 mg kg -1 each of PGT121.414.LS, PGDM1400LS and VRC07-523LS delivered together IV. Using available serum concentration data 18,19 , steady-state serum concentrations of each bnAb over 16 weeks were simulated based on popPK modeling (Methods). Because data from the LS forms of PGT121 and PGDM1400 are not yet available, the predictions considered a possible increase of 2.5-fold in elimination half-life conferred by the LS mutation, with 2.5-fold being a conservative assumption based on the previously reported fourfold increase 20 . Our results suggest that PT 80 > 200 of a bnAb regimen to a given exposing virus would provide 90% or higher efficacy to block HIV-1 acquisition with that virus. We applied this threshold (PT 80 > 200) in our definition of neutralization coverage. Neutralization coverage by at least one bnAb in the regimen, averaged over time for a 16-weekly regimen, and to the 47 viruses (all subtype C) acquired by 29 placebo recipients from 703/081 was 71%; neutralization coverage averaged over time and to the 70 viruses (90% subtype B) acquired by 35 placebo recipients from 704/085 was 73% (Fig. 6a,b). In contrast, neutralization coverage of VRC01 (at 30 mg kg -1 ) averaged over time and to the 703/081 or 704/085 viruses was 8%.

Discussion
This study advances the PT 80 biomarker as a probable surrogate endpoint for HIV-1 acquisition. By estimation of which serum bnAb titers will be protective for humans in the prevention of HIV-1 acquisition, regardless of clade, sex or route of infection, these findings address the critical need to establish targets for future preclinical and clinical expectations, for both passive instillation of bnAb combinations and active administration of any candidate HIV-1 vaccine. Notably, the neutralization titers suggested by data from the AMP trials are necessary for high PE (approximately 200 for 90% prevention efficacy, similar to that reported in NHP 4 ) should be achievable. Crucial for meeting this goal will be the use of bnAb combinations, some of which are now advancing to clinical testing (for example, PGDM1400LS + PGT121.414LS + VRC07-523LS), which have been engineered to have increased half-lives (LS versions) and improved in vitro neutralization. The PT 80 correlate can be used for systematic comparison and down-selection of candidate bnAb combinations in development 21 . We estimate through modeling that the PGDM1400LS + PGT121.414LS + VRC07-523LS combination will provide levels of HIV prevention efficacy that greatly exceed those of VRC01, from 29 to 90% (that is, 7.3-fold higher efficacy). Subcutaneous administration, and the possibility of more frequent dosing, could also help maintain higher bnAb titers.
Modeling of vaccine prophylaxis versus passive infusion prophylaxis requires different considerations, because a candidate bnAb-inducing vaccine would induce a polyclonal bnAb response. Despite this difference, the approach to prediction of prevention efficacy based on serum neutralization titers against anticipated exposing viruses over time is applicable to both prevention modalities. For example, for candidate bnAb-inducing vaccines, the primary immunogenicity endpoint of a vaccine study could be the magnitude-breadth of directly measured serum ID 80 titers against a panel of viruses representing the antigenic distribution of strains circulating in a given geographic region/population of interest, averaged over a given follow-up period. Combining these magnitude-breadth data with antibody decay longitudinal mixed-effects models (for example, those in ref. 22 ) would allow prediction of vaccine efficacy during the follow-up period. However, for passive antibody prophylaxis PT 80 would be used because the ability to accurately model bnAb concentrations over time pharmacokinetically enables effective use of this biomarker that is much easier to use than ID 80 .
Another key future potential application of the PT 80 correlate is in provisional and traditional approval decisions for bnAb combination regimens (and, later, potentially bnAb-inducing vaccines). The framework is also broadly applicable-for example, rapid adaptation of the pseudovirus-based HIV-1 neutralization assay to COVID-19 vaccine studies helped provide evidence for establishing a neutralizing antibody surrogate endpoint for COVID-19 vaccines 23-26with applications including a noninferiority, neutralization-based endpoint approach to support vaccine emergency use authorization or approval 27,28 . Moreover, our findings 23 are informing policy decisions (for example, on boosters) and will potentially aid design of future trials, including those of pan-coronavirus vaccines. The near-simultaneous validation of a pseudovirus-based neutralization assay for these two different pathogens (HIV-1 and SARS-CoV-2) is an exciting step forward in biomarker science, and we anticipate that our PT 80 correlate described here will have similar utility in the HIV prevention field. It is important to note that immune markers can be useful as correlates of protection (that is, reliably predictive of prevention efficacy) even if they are not a mechanistic correlate of protection (defined in ref. 29 as "the immune response that is responsible for protection"). For example, there are several approved vaccines for which a nonmechanistic correlate of protection is accepted as a surrogate endpoint for regulatory decisions 3,29 . It is advantageous to use a serum marker as the correlate of protection for ease of use (for example, avoidance of invasive mucosal sampling). However, in general, for a biomarker like PT 80 -that is, one that statistically correlates with protection yet is measured in serum and is thus less likely to be a mechanistic correlate of protection for a sexually transmitted infection (HIV) compared with another biomarker such as PT 80 in mucosal tissues-it is important to study its correlation with the more mechanistic biomarker. For IV administered VRC01, strong positive associations have been reported between serum versus rectal mucosal levels 30 whereas further work with larger sample sizes may be needed to better understand the extent of correlation between serum versus vaginal mucosal levels. Further research is also needed to characterize neutralization activity in serum versus tissues of virus entry among mAb recipients.
The correlations of PT 80 (or PT 50 ) with prevention efficacy in the AMP trials versus the NHP model are similar 31 , with the caveat that we estimated/modeled population-level PE as a function of PT 80 in the AMP trials whereas the NHP challenge models estimated/modeled per-challenge-level PE (Fig. 3). There has been concern that the high-dose rectal challenge NHP model, which is designed to infect all animals after one exposure, may overestimate the required titer needed to protect humans from sexual exposure to HIV, who demonstrate low per-exposure risk for infection (around 0.4-1.0%); even low-titer, repeat-dose NHP challenge models infect up to 30% of animals per challenge. However, our results suggest that the NHP model did not yield a biased overestimate. Moreover, our results caution that low-dose NHP challenge models could overestimate prevention efficacy. Nevertheless, studies of vaccinated NHP showed that protection against autologous virus requires high titers of tier 2 polyclonal neutralizing antibodies 5 . This is in line with the AMP trial indicating that high titers of bnAbs are apparently required for protection. Collectively, these findings pose a major challenge to the development of HIV-1 vaccines expected to depend on stimulation of bnAb formation, and highlight a need to increase vaccine-elicited bnAb levels and improve their durability.
Two analyses of the AMP trial data-a neutralization sieve analysis and a correlate of risk analysis-mutually support the PT 80 biomarker as a correlate of protection and as a key biomarker for modeling predicted bnAb prevention efficacy. Neutralization sieve analysis showed that prevention efficacy decreased with increasing virus IC 80 , because estimated PE was 4.5-fold higher against viruses tenfold more sensitive to VRC01 neutralization (81.6% against viruses with IC 80 = 0.2 µg ml -1 versus 16.5% against viruses with IC 80 = 2 µg ml -1 ) 6 . Correlate of risk analysis estimated that HIV-1 acquisition risk of VRC01 recipients was 2.1-fold lower for every tenfold increase in VRC01 serum concentration at exposure. Our model of PE based on the PT 80 biomarker posits that a change in either VRC01 concentration against an exposing virus, or IC 80 of the exposing virus, translates to the same impact on PE-for example, increasing concentration tenfold or decreasing IC 80 tenfold would alter PE by the same amount. The fact that the 'IC 80 effect' of 4.5 exceeded the 'concentration effect' of 2.1 suggests deviation from this model. However, this result is expected if in fact the model holds, because there is uncertainty in the estimation of HIV-1 acquisition dates. The uncertainty in the actual date of infection biased the estimate of the concentration effect toward 1.0, which is so-called biasing-towards-the-null caused by random/unsystematic measurement error (for example, ref. 32 in the statistics literature), indicating that 2.1 is an underestimate. In contrast, neutralization sieve analysis did not use estimated infection dates and thus is not subject to this attenuating bias. Therefore, the results are consistent with our model for prediction of bnAb regimen PE, which posits that PE can be improved equally by increasing either bnAb concentration or neutralization sensitivity. Note that the AMP trials did not use an LS-modified bnAb; estimations of concentrations at the true acquisition dates would be more accurate for an LS bnAb, because these have a prolonged elimination phase with serum concentrations decaying slowly.
Both IC 80 and VRC01 concentration varied markedly in AMP: the IC 80 of acquired viruses ranged from 0.07 to >10 µg ml -1 (ref. 6 ); the lowest observed VRC01 serum concentration was <0.07 µg ml -1 (the assay's lower limit of quantification) and the highest observed concentration in the AMP trials was 278 µg ml -1 (5 days post-infusion 2, 30 mg kg -1 arm). This dynamic range is key for discerning correlates. However, one limitation of our analysis is that the challenges of infection timing estimation constrain the ability of the concentration correlates analysis to fully leverage concentration variability. In contrast, the fact that neutralization sieve analysis does not rely upon infection time estimation (and accounts for the full range of IC 80 values) renders it more sensitive for characterization of correlates than concentration correlates analysis.
A further limitation of this correlates study in regard to making progress toward validation of the PT 80 biomarker is that we  a,c,e, HVTN 703/HPTN 081. b,d,f, HVTN 704/  HPTN 085. a,b, Tables below each plot provide neutralization coverage averaged to viruses and averaged over the given time frame. Virus exposure was considered covered by 1-, 2-or 3-active bnAbs if the coverage threshold (PT 80 > 200) was achieved by at least one, at least two or all three bnAbs, respectively. All predictions were made under the scenario that PGT121LS and PGDM1400LS have 2.5-fold higher half-lives than PGT121 and PGDM1400, based on modeling of the observed serum concentration data of PGT121 and PGDM1400 (refs. 18,19 ). For each bnAb regimen, geometric mean PT 80 at each time point was calculated as the geometric mean of predicted serum concentration across bnAb recipients at each time point during steady state (simulated based on popPK modeling of each bnAb as described in Methods), divided by the geometric mean of bnAb drug product IC 80 across viruses circulating in the designated AMP trial. The PT 80 of the triple-bnAb regimen was calculated using the Bliss-Hill interaction model of individual bnAb PT 80 titers 10   restricted to the VRC01 antibody. Our prediction modeling of prevention efficacy for antibody regimens involving other antibodies besides VRC01 (some of which are in different epitope classes than CD4bs antibodies) assumes transportability of the PT 80 biomarker correlate learned from VRC01 to the PT 80 biomarker for other antibodies and antibody combinations. More specifically, our prevention efficacy modeling exercise assumes that the mathematical relationship of per-exposure prevention efficacy with the PT 80 biomarker level would be the same for other antibodies or antibody combinations. Transportability assumptions may be more credible within an epitope class than across different epitope classes. Another limitation is that AMP studied PE and correlates for the outcome of HIV-1 diagnosis, not for infection of a single cell, which is not fully observable. Virus features (IC 80 , genotypes) were measured from the first available HIV RNA-positive detection in plasma. Open questions include: were viruses present in hidden compartments before detectability along with delayed seroconversion? What were their features? Did the presence of VRC01 result in subclinical acquisition of HIV?
Cumulatively, our results demonstrate the utility of the pseudovirus-based neutralization assay in infectious disease research more generally, and advance the PT 80 biomarker toward use as a correlate of protection for future HIV-1 bnAb combination trials or bnAb-inducing vaccine trials.

Online content
Any methods, additional references, Nature Research reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/ s41591-022-01953-6.

Methods
Ethical compliance. All work described here complied with all relevant ethical regulations. This work was approved by the Duke University Health System Institutional Review Board (Duke University) through protocol no. Pro00093087. For the National Institute for Communicable Diseases (NICD), the work was approved by the University of the Witwatersrand Human Research Ethics Committee through protocol no. M201105. All participants in the AMP trials provided written informed consent 6 . Participants were compensated to cover relevant trial participation costs for each completed study visit.
TZM-bl target cell neutralization assay. The TZM-bl target cell assay 34,35 , which is optimized and validated 36 , was used to assess in vitro sensitivity to VRC01 (in VRC01-recipient serum samples) of HIV-1 Env-pseudotyped viruses (for brevity, we refer to HIV-1 Env-pseudotyped viruses as 'viruses' in the main text). Full details on pseudovirus stock preparation, the TZM-bl assay and calculation of serum neutralization titers are reported in the appendix of ref. 6 . A brief summary also follows below.
Using RNA samples from AMP participants who had acquired HIV-1 infection, the genomic region of the acquired HIV-1 isolate coding for the complete Env glycoprotein was sequenced and an env-rev cassette plasmid was synthesized by Genewiz. For production of HIV-1 Env-pseudotyped virus, 293 T/17 cells (American Type Culture Collection, no. CRL-11268) were cotransfected with env plasmids and an env-defective backbone vector (pSG3delEnv). TZM-bl target cell neutralization assays were performed in two laboratories using a common standard operating procedure. Assays for HVTN 703/HPTN 081 were performed at NICD in Johannesburg, South Africa; assays for HVTN 704/HPTN 085 were performed at Duke University. Assay equivalency has been established between NICD and Duke. Neutralization was measured as a function of reduction in luciferase reporter gene expression (due to the presence of serum VRC01) after a single round of HIV-1 Env-pseudotyped virus infection of TZM-bl target cells (obtained from National Institutes of Health (NIH) AIDS Research and Reference Reagent Program no. ARP-8129). Before assay, serum samples were heat inactivated at 56 °C for 30 min. Autologous sera were assayed against each HIV-1 Env-pseudotyped virus at a starting dilution of 1:10 using eight three-fold serial dilutions. Neutralization titers are expressed as the reciprocal dilution of serum from a VRC01 recipient at which relative luminescence units (RLU) were reduced by either 50% (ID 50 ) or 80% (ID 80 ) relative to virus control wells after subtraction of background RLU in cell control wells. The VRC01 drug product was used as a positive control; heat inactivation did not affect neutralization activity of the VRC01 drug product when spiked into a normal human serum sample. Data collection was performed with the Victor X Light luminometer (PerkinElmer 2030 software, instrument program v.4.00.5) at Duke up to 11 November 2020. After this date, the Glomax Navigator System luminometer was used for data collection using Glomax Navigator software (v.3.2.3, firmware v.4.92.0). At NICD, the PerkinElmer Victor X luminometer was used for data collection with PerkinElmer 2030 software (v.4).
Additional assays were performed on the VRC01 drug product (Leidos Biomedical Research, Inc./VRC-HIVMAB060-00-AB, lot no. . Neutralization titers are expressed as the concentration of the VRC01 drug product at which RLU were reduced by either 50% (IC 50 ) or 80% (IC 80 ) relative to virus control wells after subtraction of background RLU in cell control wells. The VRC01 drug product was assayed against each HIV-1 Env-pseudotyped virus three times, at starting concentrations of 100 and 5 μg ml -1 , using eight three-fold serial dilutions. HIV-1 PVO.4 Env-pseudotyped virus was included in each assay as a positive control. VRC01 drug product stock concentrations were prepared at Duke and sent to NICD, and thus the two laboratories worked with identical material.
Binding antibody multiplex assay. Serum VRC01 IgG levels were measured on a Bio-Plex instrument (Bio-Rad) using a qualified assay that was subsequently validated using the same conditions. The assay was designed to measure infused VRC01 by its ability to bind anti-idiotype antibody captured on fluorescent magnetic beads. This assay was derived from a standardized custom HIV-1 Luminex assay [37][38][39] . Bio-Plex software (Bio-Plex Manager, v.6.1) provides two readouts: a background-subtracted median fluorescent intensity (MFI), where background refers to a plate level control (that is, a blank well containing antigen-conjugated beads run on each plate) and a concentration based on a standard curve run on the same assay plate, using a five-parameter logistic (5PL) curve fit. Each sample was run in duplicate.
Clinical-grade VRC01 was titrated to create a standard curve that was used to determine the concentrations of the diluted samples. The negative controls were CH58 (Duke Protein Production Facility 40,41 ) and blank beads. Samples with VRC01 concentrations <0.01 μg ml -1 at a dilution of 1:100 were truncated at 0.01 μg ml -1 for plotting purposes. Serum samples were tested at multiple dilution factors to ensure that MFI fell within the linear range of the standard curve. Reported VRC01 serum concentration was programmatically selected as the sample dilution factor where the in-well concentration was closest to the EC 50 of the 5PL standard curve run on the same assay plate. All samples that tested above the upper limit of quantitation at the minimum required screening dilution were successfully titrated to fall within the linear range of the assay, to provide a reportable concentration. All samples tested with serum concentrations above the lower limit of quantitation (LLOQ) were successfully titrated to provided a reportable value. All samples with concentration values below LLOQ were repeated for confirmation. The programmatically selected concentration was confirmed across other sample-matched linear range concentration values by meeting preset 70-130% agreement criteria; the same recovery threshold applied to the standard curve and spiked controls with known concentrations. Two-phase case-control sampling design for measurement of VRC01 serum concentrations. All VRC01 recipient primary endpoint cases (HIV-1 diagnosis by the week 80 visit; same definition as in ref. 6 ) were sampled for measurement of VRC01 serum concentrations at all blood storage visits through to HIV-1 diagnosis. Among VRC01 recipients completing the week 88 visit without HIV-1 infection diagnosis (noncases), a stratified sample of participants was selected into a subcohort for measurement of VRC01 concentrations at all blood storage visits (baseline, every 4 weeks through to week 80, 5 days post second infusion, week 88). All sampled noncases were not likely to have used pre-exposure prophylaxis, defined for 704/085 by self-report and testing of Tenofovir drug levels from all available dried blood spot samples that were stored at all visits, and for 703/081 by self-report. The sampling was restricted to noncases that did not permanently discontinue infusions. A total of 82 noncases, approximately 50% from each trial, were sampled for concentration measurements by sampling strata defined by randomized VRC01 dose arm cross-classified by geographic region, as described in Supplementary Table 3.

Statistics and reproducibility.
In the AMP trials, participants were randomly assigned to treatment arm as described in ref. 6 . Sample sizes of the two AMP trials were predetermined using a one-sided 0.025-level Wald test for comparison of log-transformed one minus cumulative incidences of HIV-1 acquisition between pooled VRC01 groups versus the control group, as described in the protocol and in ref. 42 Power calculations for the case-control study are described in ref. 42 and further studied in ref. 15 .
This study utilized two sets of TZM-bl target cell HIV-1 neutralization assay results. One set was generated with a clinical lot of VRC01 and a second with autologous serum samples. Assays with the clinical lot of VRC01 were performed three times, where each time the samples were tested in duplicate wells. The three titer values were averaged on the log-transformed scale. Assays with autologous serum samples were performed once using duplicate wells. VRC01 drug product was used as a positive control in each assay run. The assay has been formally validated for accuracy, sensitivity, specificity, precision, linearity, range and robustness. For in vitro neutralization measurements (IC 50 or IC 80 ), duplicate values for wells that scored at least 40% neutralization must have agreed within 30% to have passed quality control. Laboratory staff conducting the TZM-bl target cell assays were blinded to group allocation during data collection and analysis.
The binding antibody multiplex assay (BAMA) was qualified, and validation experiments using the same assay conditions were complete at the time the AMP study was performed. Additionally, qualified BAMA-derived serum VRC01 concentrations demonstrated excellent concordance with true VRC01 concentration in a blinded, HIV-1 seronegative, serum-spiked quality control reference panel (Supplementary Fig. 1). Several criteria were used to determine whether data from an assay were acceptable and could be statistically analyzed. The standard curve EC 50 and MFI values were tracked against historical data in Levey-Jennings, and points with MFI > 100 must have had a percentage coefficient of variation <20% between replicates. Any sample without at least two observed concentrations in agreement with each other, or with baseline MFI > 1,000, was repeated to obtain an accurate measurement.
Point estimates of HIV-1 infection time (calendar date) were calculated by blinded analysts for each participant, using the median of the Bayesian posterior distribution. For BAMA measurements, nonlinear mixed-effects models were used to analyze individual-level concentrations over time, based on data up to the visit before the last HIV-negative visit for cases and on data from all available visits for noncases, including data from six noncase participants who were purposefully sampled on the last visit before week 88, to keep the laboratory blinded to the case-control status of the samples (because all noncases would have had the full course of time points until week 88).
Statistical analyses. popPK modeling for estimation of VRC01 serum timeconcentration curves. All PK analyses considered VRC01 serum concentrations measured by BAMA at post-infusion time points, with concentration values subtracted from that observed at the baseline pre-infusion visit of the same participant. The assay LLOQ is 0.07 μg ml -1 , and values less than LLOQ were replaced by LLOQ2 before baseline subtraction. Nonlinear mixed-effects models were used to analyze these individual-level concentrations over time, based on data up to the visit before the last HIV-negative visit for cases and on data from all available visits for noncases, including data from six noncase participants who were purposefully sampled at the last visit before week 88, to keep the laboratory blinded to the case-control status of the samples (because all noncases would have had the full course of time points until week 88). VRC01 PK following IV infusion, with a more rapid decline in the distribution phase and slower decline in the elimination phase, was described by a two-compartment disposition model with first-order elimination from the central compartment 43,44 . The PK model was parametrized in terms of CL, V c , Q and V p , denoting clearance from the central compartment (l d -1 ), volume of the central compartment (l), intercompartmental distribution clearance (l d -1 ) and volume of the peripheral compartment (l), respectively. An exponential between-individual random effect for CL, V c , Q and V p , as well as an exponential interinfusion-interval random effect (that is, interoccasion variability) for CL and V p , were considered based on patterns observed in the data. In the final popPK model, trial (HVTN 704/HPTN 085 or HVTN 703/HPTN 081) and body weight were included as a predictor of V p and CL, respectively. Further details are described in ref. 43 and references therein. Each VRC01 recipient's concentrations on a daily grid, including at the estimated time of infection for cases, were estimated from the final popPK model. The point-wise 95% prediction interval of each daily grid concentration was computed as the 2.5th and 97.5th percentiles of the estimated concentrations of >1,000 parametric bootstrap samples, generated via resampling of the random effects and residual errors from the final popPK model of the observed concentrations. Based on each bootstrap sample of the concentration data at the observed time points, the popPK model was refit and daily grid concentrations were estimated for each bootstrap dataset.
Serum neutralization titer associated with VRC01 prevention efficacy. Because knowing the exact day of HIV-1 acquisition (and hence VRC01 concentration on the day of acquisition) is unattainable in a clinical trial setting, for Fig. 3 PT 80 was calculated as the median observed VRC01 concentration at all mid-infusion visits across all noncases in case-control samples (19.6 µg ml -1 measured by BAMA), divided by the IC 80 of the VRC01 drug product against the acquired virus (n = 162 HIV-1 cases) in both VRC01 and placebo arms from the AMP trials. The median concentration was computed based on the n = 82 noncase VRC01 recipients sampled for PK modeling from both dose arms and both trials. Note that the neutralization sieve analysis for Fig. 3 could not use individual-specific concentration estimates because the sieve method is designed to assess only how PE is dependent on a virus feature (for example, IC 80 ), not on a feature that depends on both host and virus. The analysis was performed similarly for Extended Data Fig. 5, except that PT 50 and IC 50 were used in place of PT 80 and IC 80 , respectively.
The NHP estimated protection curves in Fig. 3 were calculated from logistic regression using day-of-challenge PT 80 from meta-analysis of each of the datasets described in the figure (and similarly in Extended Data Fig. 5 using day-of-challenge PT 50 ).
In Fig. 4, PT 80 was calculated as the estimated serum concentration for each of the 82 noncases in the case-control sample at each day of follow-up, divided by the IC 80 of the VRC01 drug product against the viruses acquired by placebo recipients (n = 19 IC 80 < 1 µg ml -1 ; n = 10 IC 80 1-3 µg ml -1 ; n = 35 IC 80 > 3 µg ml -1 ). These serum concentrations in daily grid over the trial follow-up period (from day 1 to week 80 visit) were estimated based on popPK modeling of observed VRC01 concentrations for each of the 82 VRC01 recipient noncases sampled for PK modeling. Analyses were performed similarly for Extended Data Fig. 6, except that PT 50 and IC 50 were used in place of PT 80 and IC 80 , respectively. VRC01 serum concentration correlates analysis. The population for analysis included all AMP participants from either trial assigned to one of the two VRC01 dose arms. A Cox model with enrollment as the time origin was used to assess the association of VRC01 serum concentration (included as a time-dependent covariate on a daily grid) with the instantaneous hazard of HIV-1 acquisition, which uses empirical inverse probability sampling weights to accommodate the case-control sampling design 15,45 . Regression calibration was used to account for measurement error in estimated daily grid VRC01 concentration from the popPK model. Two covariates, trial (HVTN 704/HPTN 085 or HVTN 703/ HPTN 081) and dose group (10 or 30 mg kg -1 ), were adjusted for in the Cox model. A nonparametric bootstrap procedure was used to estimate standard error and compute 95% CIs for the hazard ratio of HIV-1 infection per tenfold increment in the current value of VRC01 concentration. The bootstrap-based percentiles on a grid of coverage levels were inverted to compute a two-sided P value for whether VRC01 concentration correlated with the instantaneous hazard of HIV-1 acquisition.
PT 80 against autologous acquired virus analysis. The 'marker method' of Gilbert et al. 15 was applied to assess whether PT 80 against the autologous acquired virus at acquisition tended to be low compared with PT 80 against placebo viruses, assuming that exposures had occurred uniformly over the 80-week follow-up period. The results are shown in Fig. 5b. For each VRC01 case, the PT 80 against their autologous acquired virus at acquisition was calculated in three steps: (1) estimation of the time-concentration curve for each case from a popPK model 43,44 over a daily grid spanning all possible dates of HIV acquisition; (2) averaging of (log-transformed) daily concentrations weighted by the Bayesian posterior distribution of daily probabilities of HIV acquisition over the entire grid; and (3) dividing this aggregated concentration in step 2 by the IC 80 of the VRC01 drug product against their acquired virus (Huang et al. 11 , approach 2). For each VRC01 recipient noncase, we calculated their median estimated VRC01 serum concentration over follow-up (estimated on a daily grid from day 1 to the week 80 visit from the PK model), which represents the most typical VRC01 concentration that would occur at the time of HIV-1 exposure. We then divided this median VRC01 serum concentration by the geometric mean VRC01 IC 80 of the viruses acquired by placebo recipients in the AMP trial, which is a typical value of PT 80 against a given exposing virus assuming that HIV-1 exposure follows a uniform distribution over the trial follow-up period (Extended Data Fig. 2).
The geometric mean PT 80 values among cases and the geometric mean of those typical PT 80 values among noncases were calculated with the standard errors estimated by the bootstrap procedure described earlier for daily grid serum concentrations. From these estimates, Wald-based 95% CIs were computed; bootstrap-based percentiles on a grid of coverage levels were inverted to compute a two-sided P value. The analysis was performed for the AMP trials pooled, for the two VRC01 dose arms pooled and for each individual VRC01 dose arm.
Analyses were performed similarly for Extended Data Fig. 7, except that PT 50 and IC 50 were used in place of PT 80 and IC 80 , respectively.
Neutralization coverage of combination bnAbs. In Fig. 6a-d, steady state of a repeated bnAb administration regimen is considered to be reached when the rate of bnAb input is equal to the rate of bnAb elimination and consequently serum concentration-time levels remain the same over continuing cycles. For each single bnAb, a virus exposure was considered neutralized (covered) if PT 80 (calculated as PK modeled concentration divided by virus IC 80 ) exceeded 200. In Extended Data Fig. 8, this threshold was defined as when PT 50 (calculated as PK modeled concentration divided by virus IC 50 ) exceeded 600. Coverage was computed assuming that exposures to HIV-1 follow a uniform distribution over the 16-or 24-week period post-infusion that is considered. Predicted prevention efficacy of combination bnAbs. The prediction of PE for the triple-bnAb combination, PGDM1400LS + PGT121LS + VRC07-523LS, was carried out in four steps. First, serum concentration of each single bnAb over time (16 or 24 weeks in steady state) was predicted based on the popPK model of each single bnAb assuming an extended half-life of the LS-version bnAb over the parental form as described earlier. Second, the corresponding PT 50 and PT 80 values of each bnAb against either the panel of viruses acquired by 703/081 placebo recipients or the panel of viruses acquired by 704/085 recipients were calculated using the predicted serum concentration obtained in step 1 divided by the IC 50 and IC 80 , respectively, of each virus. Third, the PT 50 and PT 80 values of the combination bnAb regimen against each virus were calculated assuming a Bliss-Hill interaction model 10 (Supplementary Note 1). Similar results were obtained by assuming an additive interaction model ( Supplementary Fig. 2). Last, the PE of the combination bnAb over time was predicted based on more conservative protection-by-PT 50 curves and protection-by-PT 80 curves than those shown in Extended Data Figs. 5 and 3, respectively, assuming the PT 50 or PT 80 value needed for a fixed level of protection is twofold higher in human than in NHP set C (that is, shifting the NHP set C curves to the right by twofold).
Estimation of HIV-1 acquisition dates. Point estimates of HIV-1 infection time (calendar date) were calculated by blinded analysts for each participant, using the median of the Bayesian posterior distribution. This was computed by independent combination of the estimated posterior distributions of infection date after conditioning on: (1) diagnostic curves, as described in Grebe et al. 14 for the infection dating tool method, by a combination of window period distributions (defined in Pilcher et al. 46 ) of last negative and first positive HIV-1 diagnostic tests; and (2) the distribution of pairwise Hamming distances (as described in Giorgi et al. 13 for the Poisson Fitter method), calculated from nucleotide sequence data obtained from the first available HIV-positive samples, following the steps listed in Supplementary Note 2 (i-vii) for sequence-based timing estimates. We also employed Bayesian posterior estimates of infection time by day to assess the probability that the infection occurred in the first 4 weeks after an infusion. These were also used to describe 95% credible regions for the time since infection in the process described in Supplementary Note 2.
Our approach employed Bayesian procedures to combine independent inputs from three data sources: diagnostics timing estimators, gag-pol-based sequence timing estimators and rev-env-nef-based sequence timing estimators. The sequence-based estimates were first fit using the frequentist methodology implemented in Poisson Fitter by employing the steps described below, with the aim of optimizing the estimate of infection time. These Poisson Fitter estimates were then converted to corresponding Bayesian Poisson process regression models to yield probability assessments for incorporation into downstream analyses, maintaining high concordance in both point estimates and uncertainty intervals across the Bayesian and frequentist implementations. When the two sequenced regions yielded distinct time estimates (that is, the 95% posterior credible intervals did not overlap), the gag-pol region was chosen for the final time estimate since it does not encode proteins targeted by the VRC01 bnAb, and therefore the rates of evolution over this region of the viral genome may be impacted to a lesser extent by targeted selection.
The process of running Poisson Fitter 2.0 included production of separate timing estimates for each curated first time point alignment for the two sequenced regions and then integrating the outputs downstream. Only the first time point sequences were used for the primary analysis, except for cases in which the first time point yielded fewer than five sequences; in such cases, the second available time points were analyzed instead. For each sequenced sample, we created Poisson Fitter estimates and then corresponding Bayesian posterior estimates of time since infection, given (1) gag-pol and (2) rev-env-nef region alignments. The process is described in further detail in Supplementary Note 2.
Reporting summary. Further information on research design is available in the Nature Research Reporting Summary linked to this article.

Code availability
Code implementation methods described in Gilbert et al. 15 are publicly available at http://faculty.washington.edu/peterg/programs.html? All other code implementing the methods used to generate the figures and tables of this work is available at Github (https://github.com/HVTN-SDMC/AMP_NeutTiterBiomarker). Instructions for installation and use are given in the accompanying README file.  Fig. 4 | Agreement between predicted vs experimental serum neutralization ID50 titer. Sera from samples from the last visit (and for a subset from the last two last visits) prior to the first positive HIV-1 RNA PCR test were assayed against autologous isolates from 64 VRC01 recipients who acquired HIV-1 infection (cases) (90 isolates, 164 titers). Predicted serum ID50 titer (PT50) values are plotted against the experimental ID50 for each sample and each isolate [Lin's concordance correlation coefficient 33 = 0.92 (95% CI 0.80 to 0.97)]. PT50 was calculated as popPK model-predicted concentration divided by IC50. Thirty-two of 164 experimental ID50 titers were below the limit of detection when the PT50 was greater than 10 (range 10.2 to 36.3). Zero of 164 experimental ID50 titers was at or above the limit of detection when the PT50 was less than 10. Dashed horizontal and vertical lines at a value of 10 show the experimental ID50 limit of detection; concordance of predicted vs. experimental values above vs. below 10 was 132 of 164 (80.5%). Fig. 7 | Predicted serum ID50 titers (PT50s) to autologous acquired viruses at HIV-1 acquisition among VRC01 arm cases, and to placebo-recipient acquired viruses among VRC01 arm non-cases. (a) Violin plots for VRC01-recipient cases vs. non-cases, where Approach 2 of Huang et al. 11 was used to calculate PT50 at a given time point against a given virus. For each VRC01-recipient case, PT50 at the estimated date of HIV-1 acquisition (red dots) was calculated as the estimated VRC01 concentration at acquisition divided by the VRC01 drug product IC50 against the autologous virus. For each of the 82 sampled VRC01-recipient non-cases, PT50 at each day of follow-up against each placebo-recipient acquired virus was calculated as the estimated VRC01 concentration divided by the VRC01 drug product IC50 against the virus (blue dots). The lower bound, horizontal line, and upper bound of the vertical rectangle boxplots show the 25 th , 50 th , and 75 th percentiles, respectively. On each side of the boxplot is a kernel density estimation of the distribution shape of the PT50. (b) By VRC01 dose arm and across dose arms pooled: geometric mean PT50 at HIV-1 acquisition in VRC01-recipient cases against the autologous acquired virus, geometric mean PT50 in VRC01-recipient non-cases (their individual-specific medians over follow-up) to placebo-recipient acquired viruses, and their ratio. Error bars represent 95% confidence intervals. Fig. 8 | (A, B) Neutralization coverage (defined by PT50 > 600) (averaged to viruses) and (C, D) geometric mean predicted serum ID50 (PT50) titer against viruses circulating in each of the AMP trials for the bnAb regimen PGT121LS + PGDM1400LS + VRC07-523LS 20 + 20 + 20 mg/kg delivered intravenously every 16 weeks and evaluated in study cohorts of the same sizes as the AMP trials. In (A, B), the tables below each plot provide the neutralization coverage averaged to viruses and averaged over the given time frame. A virus exposure was considered covered by 1-active, 2-active, or 3-active bnAbs if the coverage threshold (PT50 > 600) is achieved by at least 1, at least 2, or all 3 bnAbs. All predictions were made under the scenario that PGT121LS and PGDM1400LS have 2.5-times higher half-lives as PGT121 and PGDM1400, based on modeling of observed serum concentration data of PGT121 and PGDM1400 18,19 . For each bnAb regimen, geometric mean PT50 at each time-point was calculated as the geometric mean of predicted serum concentration across bnAb recipients at each time-point during steady state (simulated based on population PK modeling of each bnAb as described in Methods) divided by the geometric mean of bnAb drug product IC50 across viruses circulating in the designated AMP trial, that is (a, c) the m = 47 viruses acquired by n = 29 703/081 (Sub-Saharan Africa) placebo recipients; (b, d) the m = 70 viruses acquired by n = 35 704/085 (Americas + Switzerland) placebo recipients. The PT50 of the triple-bnAb regimen was calculated using the Bliss-Hill interaction model of the individual bnAb PT50 titers 10 . μ μ