Panels of HIV-1 Subtype C Env Reference Strains for Standardized Neutralization Assessments

ABSTRACT In the search for effective immunologic interventions to prevent and treat HIV-1 infection, standardized reference reagents are a cost-effective way to maintain robustness and reproducibility among immunological assays. To support planned and ongoing studies where clade C predominates, here we describe three virus panels, chosen from 200 well-characterized clade C envelope (Env)-pseudotyped viruses from early infection. All 200 Envs were expressed as a single round of replication pseudoviruses and were tested to quantify neutralization titers by 16 broadly neutralizing antibodies (bnAbs) and sera from 30 subjects with chronic clade C infections. We selected large panels of 50 and 100 Envs either to characterize cross-reactive breadth for sera identified as having potent neutralization activity based on initial screening or to evaluate neutralization magnitude-breadth distributions of newly isolated antibodies. We identified these panels by downselection after hierarchical clustering of bnAb neutralization titers. The resulting panels represent the diversity of neutralization profiles throughout the range of virus sensitivities identified in the original panel of 200 viruses. A small 12-Env panel was chosen to screen sera from vaccine trials or natural-infection studies for neutralization responses. We considered panels selected by previously described methods but favored a computationally informed method that enabled selection of viruses representing diverse neutralization sensitivity patterns, given that we do not a priori know what the neutralization-response profile of vaccine sera will be relative to that of sera from infected individuals. The resulting 12-Env panel complements existing panels. Use of standardized panels enables direct comparisons of data from different trials and study sites testing HIV-1 clade C-specific products. IMPORTANCE HIV-1 group M includes nine clades and many recombinants. Clade C is the most common lineage, responsible for roughly half of current HIV-1 infections, and is a focus for vaccine design and testing. Standard reference reagents, particularly virus panels to study neutralization by antibodies, are crucial for developing cost-effective and yet rigorous and reproducible assays against diverse examples of this variable virus. We developed clade C-specific panels for use as standardized reagents to monitor complex polyclonal sera for neutralization activity and to characterize the potency and breadth of cross-reactive neutralization by monoclonal antibodies, whether engineered or isolated from infected individuals. We chose from 200 southern African, clade C envelope-pseudotyped viruses with neutralization titers against 16 broadly neutralizing antibodies and 30 sera from chronic clade C infections. We selected panels to represent the diversity of bnAb neutralization profiles and Env neutralization sensitivities. Use of standard virus panels can facilitate comparison of results across studies and sites.

patterns of neutralization response diversity that were not otherwise included (12). In contrast, here we describe a clade C panel of 12 Envs intended to detect relatively weak or potentially clade-specific tier 2 neutralization responses. Vaccine sera that yielded any detectable response(s) could be identified for further evaluation. Ultimately, both the clade C and M group panels are intended for use in vaccine trials and in other settings.

RESULTS
Antibody neutralization. Neutralization titers are typically determined as point values (e.g., 50% inhibitory concentration [IC 50 ] and IC 80 values) to summarize distributions from a series of reagent concentrations. The antibody concentration ranges tested in neutralization assays often produce censored neutralization IC 50 titers, where the range of concentrations does not yield 50% neutralization. Censored outcomes are represented as "Ͼx," where x is the greatest concentration used, or "Ͻy," where y is the lowest concentration used. These cutoffs can differ across assays, generally due to practical constraints of limited serum or antibody availability. Such censoring is an issue for quantitative analysis, because standard practice would use a constant placeholder value for censored outcomes; e.g., an IC 50 value above 50 ("Ͼ50") is replaced with a value of 100. Censoring thresholds of 10, 20, 25, and 50 g/ml were used for different bnAbs (Fig. 1), and it was sometimes necessary to use different thresholds for even a single bnAb, such as 3BNC117. Most of the IC 50 titers corresponding to 3BNC117 were not censored (n ϭ 158 Envs). However, the 3BNC117 values were reported as Ͼ20 g/ml for 38 Envs, and the values were Ͼ50 g/ml for 4 Envs. To standardize the comparisons and to compare different bnAbs against the 200-virus panel, we used a consistent censoring cutoff of 10 g/ml across all assays, and IC 50 s below 0.01 g/ml were censored at 0.01.
Magnitude-breadth panels. The Envs downselected for magnitude-breadth characterization sampled the spectrum of bnAb reactivity patterns from the full set of 200 Envs (Fig. 2). Heat maps show IC 50 titers for the full neutralization panel (Fig. 2a) and    Fig. 3 compares neutralization magnitude-breadth distributions of the full panel of 200 clade C Envs with the distributions of the downselected panels. In most cases, the magnitude-breadth distributions show a high degree of overlap, which means that the downselected panels are a good representation of the properties of the full set. A slight shift toward greater neutralization sensitivity is apparent for some bnAbs, where distributions of selected Envs are biased toward IC 50 values slightly lower than those seen with the excluded Envs. This small bias resulted from favoring more-sensitive viruses in choosing alternate rows in the heat map, i.e., by starting with the most sensitive virus rather than skipping it for the next most sensitive.
The concordance of the breadth-potency curves was very high and consistent across bnAbs for the sets of 100 and 50 Envs (see Table S1 in the supplemental material). Downsampling further to obtain a 12-Env panel increased the bias in favor of some bnAbs and against others and gave only a rough approximation of the full set of 200 Envs (see Fig. S1 in the supplemental material). Also, downsampling to 12 Envs greatly increased the area between the magnitude-breadth curves versus the area seen with the full set ( Fig. S2), thus accounting for part of our rationale not to use downsampling to select a 12-Env panel. We instead considered other approaches.
Serum screening panel (12 Envs). Figure 4a summarizes Env sensitivity to neutralization by plasma, calculated as geometric mean ID 50 among 30 chronic plasma samples, together with the number of bnAbs that neutralized each Env. This coarse measure of sensitivity across all bnAbs was significantly associated with sensitivity to plasma (Kendall's , ϭ 0.338, P ϭ 3.34 ϫ 10 Ϫ11 ). We used this association to select Envs from principal-coordinate analysis (PCA) of bnAb neutralization data via computational guidance.
Informed by the results from testing each Env against multiple bnAbs, we sought to represent the diversity of different bnAb specificities, to reduce the risk of missing neutralization signal by overrepresenting the most common bnAb specificities. For this reason, we selected 12 Envs to represent a range of neutralization sensitivities to polyclonal plasma and monoclonal antibodies. Figure 4b shows the cumulative distribution of Env sensitivities to plasma. Env colors indicate the number of bnAbs with an IC 50 value below 10 g/ml from Fig. 4a. Where plasma and bnAb sensitivities are closely associated, the progression of Envs appears in an order consistent with the progression of rows in Fig. 4a. An overall trend is apparent for an association between serum and bnAb sensitivity, though small inconsistencies across Envs reflect wide variations in neutralization titers against sera and in the number of bnAbs to which each Env is sensitive. Figure 4c compares plasma ID 50 distributions between the Envs in the candidate panel and the remaining Envs. The candidate panel was intentionally chosen to avoid extremely high or low geometric mean ID 50 titers among chronic plasma samples, both to reduce false negatives and to exclude tier 1 neutralization responses, which are readily obtained in induced form and do not correlate with immune protection (13,14). We found no evidence that the geometric mean ID 50 s of the Envs in the selected panel (n ϭ 12) and the remaining Envs (n ϭ 183) were sampled from different distributions (two-sided, two-sample Kolmogorov Smirnov test, P ϭ 0.53). The candidate panel Envs were neutralized by different numbers (Fig. 4) and subsets (Fig. 5) of bnAbs, rather than the Envs being sensitive to all the bnAbs studied, and we confirmed that multiple Envs that were well targeted by each major monoclonal antibody epitope specificity tested were included.
To simplify the diverse outcomes of Env sensitivity to neutralization by different antibodies and to facilitate the selection of 12 Envs that covered a range of distinctive neutralization profiles with respect to the 16 bnAbs tested, we used PCA, which flattens the neutralization data into orthogonal (minimally correlated) sets of linear combinations of bnAbs (Fig. S3). The first two principal components together explained about half (47.6%) of the variance in the bnAb IC 50 data. Adding the third principal component accounted for 64.6% of the total variance. As detailed in the supplemental material, the first three principal components were strongly associated with combinations of CD4 binding-site (CD4bs), V2 glycan, and V3 glycan bnAb specificities. On the basis of comparisons of the alternative clustering methods, we favored the use of Ward's method (15) with squared Euclidean distances (ward.D2) for clustering. Ward's method was best able to cluster distinctive patches of serum and virus specificities within the broader gradient of plasma neutralization sensitivities. The resulting clustered heat map of serum neutralization ID 50 titers ( Fig. 6) is annotated to identify the candidate panel of 12 Envs. The panels identified automatically (lasso and k-medoids), as described in the supplemental material, are also shown for comparison. All three sets of Envs represent a range of average neutralization sensitivities, as reflected   (12), and (c) candidate clade C panel from the manuscript. As noted in the Fig. 1 legend, all assay results were censored above 10 and below 0.001 g/ml to standardize dilution ranges across different experiments. NA, no data. Data for historical panels a and b were computed as geometric means from the CATNAP database, as detailed in the text. Env names in panels a and b are shortened per CATNAP, and panel c lists short names from Data Set S1. by their dispersal from the top to the bottom of the heat map, which corresponded to more-resistant and more-sensitive Envs, respectively. The candidate panel, chosen with computational guidance, covers a more limited range of sensitivities than the automatically chosen Envs. This was done intentionally to avoid both highly sensitive and very resistant viruses during the iterative procedure described above.

PHRU07
Other clustering methods can yield quite different outcomes, and the correlation coefficient between cophenetic distances (16) summarizes similarity among clusters obtained using alternative algorithms (Fig. S4).
Ordering ID 50 s by geometric mean titer reveals the continuum of neutralization responses (Fig. S5) that are characteristic of the polyclonal mixture of antibody potencies and/or specificities found in plasma samples (17). This continuum further emphasizes the benefit of using bnAb sensitivities, rather than plasma responses, for computationally guided panel selection, given that we do not know whether a range of antibody sensitivities or of various antibody potencies dominates the neutralization response of any plasma sample. Fig. S6 summarizes serum neutralization responses among the 12-Env panels identified by 3 automated methods (downselection, lasso, and k-medoids) versus computationally guided selection. Because computationally guided selection avoided individual Envs that were sensitive to all bnAb specificities, the candidate panel does not merely reflect the continuum of neutralization responses, as do the panels identified by automated methods. Sensitive and informative detection of tier 2 neutralization responses, not modeling the full distribution of Env plasma sensitivities, is the main purpose intended for the candidate 12-Env panel.
Information about the 12 candidate Envs, including the geographic region and year sampled, is summarized in Table 1. Other information is tabulated to summarize genetic attributes of these sequences, including the glycosylation state (presence or absence of a potential N-linked glycosylation motif) at sites relevant to antibody binding susceptibility, hypervariable loop lengths and net charges, and the infection stage from which the virus was sampled. Table 2 summarizes the IC 50 neutralization titers by 16 bnAbs. In the candidate panel, ZM233M and Ce703010010_C4 were resistant only to PGT128. Another Env, Ko243, was sensitive to all bnAbs shown. The selection of Envs sensitive to specific bnAb families is evident in the last three rows (Table 2).
Data Set S1 in the supplemental material lists the properties summarized in Table 1  and Table 2 (19,20) with the GTRϩ⌫4ϩI substitution model and were rooted on HXB2, though the distances to HXB2 were excluded from panel APD calculations. Both panels were designed to represent acute and early infection following heterosexual transmission. Because increasing southern African clade C diversity is associated with reduced cross-reactive neutralization between sera and circulating HIV strains (9), a more divergent, more contemporary clade C panel better reflects the modern state of    the epidemic. Such samples are difficult to obtain, and it takes years to acquire and evaluate them experimentally, so an even more recent sampling to assess vaccine trials that are under way is infeasible.
We also compared bnAb neutralization titers from viruses in each panel and summarized neutralization data for the 2006 clade C panel (Fig. 5a) and the global panel (Fig. 5b) from the CATNAP database (21). For the previously published panels, we extracted data available from CATNAP as of May 2017 (http://hiv.lanl.gov/catnap). Data from Envs with multiple published results are summarized as the geometric mean IC 50 among unique values. That is, if an assay had been published three times with the same value and once with another value, only the two distinct neutralization values were averaged. This was done to avoid biased estimates resulting from the use of data from papers that reproduce results from earlier papers without repeating the experiment. One Env (ZM233M) was included in both clade C panels, identified in the figure by an asterisk. The candidate 2017 clade C panel (Fig. 5c) that we described above is no less sensitive to known bnAbs and is intended as an update to the 2006 clade C panel, for sensitive and informative plasma screening.
By design, several Envs in this new clade C panel shared patterns of reactivity to members of distinct bnAb classes. For example, B005582 is particularly sensitive to V3-glycan (V3g) bnAbs, Ce2103 to V1/V2-glycan (V2g) bnAbs, and 2969249 to CD4bs bnAbs (Fig. 5c). Detecting neutralization in plasma samples that have responses to one or more of these viruses would provide clues about antibody specificities therein and would provide information for follow-up experiments that map specificities or isolate monoclonal antibodies.

DISCUSSION
To enhance scientific rigor, improve reproducibility, and unify efforts against HIV diversity, the use of standardized reference reagents for immunological assays is highly beneficial. Standardized reagents enable comparisons between different studies. We have described selection of standardized virus panels from HIV-1 clade C for several anticipated types of investigation, which include screening large numbers of sera from vaccinees for immune-induced neutralization responses and characterizing the magnitude and breadth of neutralization responses by newly isolated monoclonal antibodies.
Guided by the anticipated uses for these panels, we have described practical selection criteria, which utilize available information to obtain appropriately representative Env panels. We have described the use of hierarchical clustering and a simple but elegant downselection method to identify subsets of 100 and 50 clade C Envs from a panel of 200 well-characterized viruses. The panels performed better than randomly selected panels at characterizing magnitude-breadth distributions in aggregate across 16 bnAbs. For particular bnAbs, rather than the overall aggregate, moderate to almost no deviation appeared between the magnitude-breadth distributions revealed by our downselected panels and the full set of 200 Envs. This suggests that the smaller virus panels can be used in place of the full set to characterize bnAb magnitude-breadth distributions. Consequently, the use of smaller virus panels accelerates the rate at which bnAbs can be characterized. To avoid bias in favor of some bnAbs and against others, use of even smaller, 12-Env panels in magnitude-breadth studies is not recommended.
We used PCA of 16 bnAb IC 50 neutralization titers to project 200 Env-pseudotyped viruses onto simplified coordinate systems for computationally guided Env selection. Using this representation, we identified a panel of 12 viruses that covered diverse bnAb sensitivity profiles on reduced dimensions. During panel selection, iterative refinement ensured that the 12 had a representative range of sensitivity to 30 chronic plasma samples.
We also tried automated methods (downselection, lasso, k-medoids) but favored the panel identified with computational guidance, because it does not merely reiterate the plasma neutralization continuum. The diversified detection strategy embodied by the candidate panel may therefore utilize limited sample materials more effectively than the automatically chosen Env sets, each of which contains closely related, and therefore redundant, neutralization profiles.
Clade-specific panels may be better able to detect relevant neutralization responses than nonspecific panels. In a previous study that tested South African plasma samples from individuals with C clade infections from the CAPRISA cohort, a panel of tier 2 clade C viruses showed greater sensitivity to neutralization than tier 2 virus panels from clades A and B (22). Similar findings have been reported in other studies (18,23). We will not know how the two panels will compare with vaccinee sera until there is a vaccine that generates some measurable activity against tier 2 viruses. The earliest success at generating tier 2 virus neutralization could reflect partially matured bnAbs, and it is not known how these immature bnAbs might be differentially detected with clade-specific versus global virus panels.
On the other hand, we do not necessarily expect the candidate panel to perform "better" than the 2006 panel with HIV-1 sera. In fact, some of our previous data suggest the panels could perform similarly (12). The underlying scientific issue concerns potential differences in panel performance with vaccine-elicited antibodies, which cannot be assessed at the moment, because no vaccine yet tested elicits sufficient tier 2 virus neutralization responses. With this in mind, our goal was to design a panel of clade C viruses that are more contemporary and are selected on the basis of more-robust analysis methods to ensure the best possible representation of the current epidemic in southern Africa. The 2006 panel did not use neutralization phenotype data to guide its selection but instead included what was known and available at the time regarding Env genetic variation and reported neutralization assay results for the selected panel. We incorporated neutralization phenotypes throughout panel selection and selected from a very large, clade-specific neutralization panel. We expect the useful phenotypic characteristics of this new panel to emerge in subsequent work.
Our panel of 12 C clade Envs is intended as an update to the panel reported in 2006 (18). The 2006 panel was selected from a small subset through convenience sampling, whereas the 2017 panel was rationally selected from a much larger collection of viruses. The 2017 clade C panel contains more recently sampled Envs, deliberately includes sensitivity profiles that are characteristic of the currently known bnAb families, and includes greater genetic diversity than the earlier panel. This is important, because within-clade cross-reactive neutralization tends to decrease as genetic distance increases (23). Also, to help identify weak clade-specific responses without detecting the nonspecific antibody neutralization that is typical of a tier 1 response (24), the candidate panel includes a range of plasma sensitivities and favors neutralization-sensitive Envs without inclusion of known tier 1 Envs. Consequently, the 2017 clade C panel should be more informative and may be more sensitive than the 2006 clade C panel.
While we think that the candidate screening panel might provide hints about antibody specificities in plasma samples, it is intended for screening and not for epitope mapping, which would be performed to characterize samples that give positive test results for tier 2 neutralization activity. Further analysis would be needed to differentiate between possible specificities in a serum, and "next-generation" fingerprinting methods (25) could be useful for such purposes.
In contrast to the 12-virus global panel of multiclade viruses described in an earlier publication (12), we planned these panels to be used for screening sera and bnAbs from vaccinees where clade C infections predominate and clade C vaccines are being tested. We did not formulate a single quantitative metric to choose the virus panels proposed here for standardization. Instead, we considered a range of current needs for standardized reagents and selected sets of Envs that together satisfied these needs as we thought best. An extremely large number (6 ϫ 10 18 ) of alternative 12-Env panels is possible. We have described several methods to select useful sets of sequences that are intended to represent diversity in a large neutralization assay panel (6,000 plasma ID 50 and 2,600 antibody IC 50 titers). The Env panels we propose are reasonably representative of the diversity of the population from which they were chosen, by several different criteria. They represent distinctive bnAb sensitivity patterns and generally reflect the diversity of neutralization responses seen among sera from infected individuals.
HIV-1 clade C, which constitutes about half of all infections worldwide at present, represents formidable genetic diversity. As long as virus evolution continues, the ability to induce and detect immune responses against this highly diverse pathogen will be of sustained significance.

MATERIALS AND METHODS
The CAVIMC-CAVD HIV-1 Clade C Virus Neutralization Phenotype Study was reviewed and approved by the research ethics committee of the Faculty of Health Sciences of the University of Cape Town (168/2007; 513/2012). All participants provided written informed consent for study participation (9).
Magnitude-breadth panels (50 and 100 Envs). Large virus panels are useful to characterize the magnitude and breadth of neutralizing antibodies, but panel size limits the rate at which results can be obtained. Using large neutralization panels is very expensive and may consume excessive reagent resources. The trade-off is that excessively small panels may not contain a sufficient amount of the information needed to make fair assessments across different bnAbs. We therefore downselected representative sets of 50 and 100 Envs to facilitate studies of antibody magnitude and breadth.
We used a simple strategy to select subsets of viruses that represent the diversity of responses in the full set. To compare Env profiles, we used the Euclidean distance between vectors of 16 bnAb 50% inhibitory concentration (IC 50 ) neutralization titers and then hierarchically clustered the 200 Envs. We weighted the resulting dendrograms by geometric mean IC 50 to obtain a gradient from most sensitive Env to least sensitive Env (within the constraints of the dendrogram branching structure). We used Ward's method (15) for hierarchical clustering but also considered other methods. A simple downselection procedure alternated through the rows of the dendrogram-ordered neutralization heat maps by inclusion of one Env and exclusion of the next. We repeated this procedure to downselect from the full panel of 200 Envs and obtain smaller panels comprised of 100 or 50 Envs. We kept the same row and column order in neutralization panels during downselection rather than recluster and reorder.
For each of 16 bnAbs, we compared the magnitude-breadth distributions of the full panel of 200 clade C Envs with those of the downselected panels. The area between curves (ABC) quantified the difference between the two cumulative distribution functions. We used resampling to evaluate further the ABC values from downselected panels. Random panel selection characterized the null distribution of ABC values to reveal whether dendrogram-based downselection gave significantly lower values than could be obtained by chance. We randomly sampled 100-Env panels from all 200 Envs (without replacement) 10 4 times. From each of these, we also sampled a random 50-Env panel. We computed resampled ABCs against the distribution from 200 Envs and compared these with values from the downselected panels.
We repeated the downselection procedure to obtain an even smaller panel of 12 Envs. Serum screening panel (12 Envs). For the purpose of screening sera from vaccinees, we tried several approaches to select a small panel of viruses, intended to include Envs sensitive to a variety of neutralizing antibodies and sera. This smaller "candidate" panel includes 12 pseudoviruses chosen to detect neutralization responses in vaccinees and to suggest possible antibody specificities therein. Virus selection was guided by neutralization titers from assays against bnAbs and chronic sera from each of 200 Envs. Tier phenotyping (24) of these Envs demonstrated 1.3% tier 1A (n ϭ 2), 8.5% tier 1B (n ϭ 17), 75% tier 2 (n ϭ 150), and 15.5% tier 3 (n ϭ 31) Envs. We excluded the two tier 1A Envs and three highly sensitive tier 1B Envs (geometric mean 50% infective dose [ID 50 ] titers above 250 reciprocal dilutions) from panel selection because they seemed unlikely to be useful in distinguishing the protective responses from those that are nonprotective (13,14).
Our strategy was to select Envs using bnAb IC 50 s to ensure that all specificities were included and to compare the data to ID 50 s from plasma samples from patients with chronic infection. We used principalcomponent analysis (PCA) to simplify high-dimensional data from neutralization assays by projecting them onto fewer dimensions. The overall effect of dimension reduction is achieved by decomposing correlations among the data into principal components (42). This approach has recently been used for unsupervised learning to characterize high-dimensional immunological data from HIV Env antigens (43).
In a computationally guided procedure, we iteratively selected candidate Env panels and then reviewed their distributions in lower-dimensional projections of bnAb IC 50 s. Where the candidate panel contained clusters of Envs rather than dispersed Envs, different Envs were chosen to increase the separation between them and to increase coverage of known specificity profiles with the least overlap possible. This approach enabled us to select 12 candidate Envs that captured the diversity of known bnAb specificities, while ensuring low redundancy among the specificity profiles. We think it is important to sample the diversity of natural antibody responses to heterologous virus isolates because we do not know a priori the nature of the neutralizing antibodies that may be elicited and that may correlate with HIV-1 Clade C Neutralization Panels Journal of Virology