MAbs and soluble CD4.

Monoclonal antibody (MAb) b12 was produced by Quality Biological, Inc. (Gaithersburg, MD). MAbs 2G12, 2F5, and 4E10 were produced by Polymun Scientific (Vienna, Austria). MAbs PG9 and PG16 were gifts from D. Burton (Scripps Research Institute, La Jolla, CA); MAbs VRC01, VRC02, and VRC03 were gifts from J. Mascola (National Institutes of Health [NIH], Vaccine Research Center, Bethesda, MD); and MAbs to the C1 region of HIV-1 Env (A32) (37) and the CCR5 bs (17b) (38) were kindly provided by James Robinson (Tulane Medical School, New Orleans, LA). MAbs to the V2 loop of HIV-1, 697D (39) and 2158, were kindly provided by Susan Zolla-Pazner. Soluble CD4 (sCD4) was obtained from the NIH AIDS Research and Reference Reagent Program (Bethesda, MD). T8 is a murine MAb that binds to the gp120 C1 region (a gift from P. Earl, NIH).

Blue native polyacrylamide gel electrophoresis (BN-PAGE) and SDS-PAGE analysis of recombinant HIV-1 Envs.

BN-PAGE analysis of recombinant gp140 Env proteins was carried out as described previously (9, 40–42). Briefly, purified HIV-1 Env glycoproteins were diluted in a buffer containing 50 mM MOPS (morpholinepropanesulfonic acid), 50 mM Tris-HCl (pH 7.7), 20% glycerol, and 0.05% Coomassie blue. Protein samples were loaded onto a 3 to 8% Tris-acetate NuPAGE gel (Invitrogen, Carlsbad, CA), and electrophoresis was carried out for 1.5 h at 150 V with 50 mM MOPS–50 mM Tris-HCl (pH 7.7)–0.03% Coomassie blue as the cathode running buffer and 50 mM MOPS–50 mM Tris-HCl (pH 7.7) as the anode buffer. To determine what forms of oligomers were recognized by select HIV-1 MAbs, after BN-PAGE analysis and transfer onto nitrocellulose filters, HIV-1 gp140 proteins were probed with MAbs 2F5, PG9, CH31, and 17B and detected by with goat anti-human whole IgG specific (heavy- and light-chain)-AP (1:3,000 dilution) (Sigma, St. Louis, MO). The immunoblots were developed with Western-blue substrate (Promega, Madison, WI). HIV-1 gp140 proteins were also analyzed in SDS-PAGE under nonreducing and reducing conditions to ensure no major degradations of the HIV-1 gp140 preparations used in the study.

ELISA.

The reactivities of HIV-1 Envs with select HIV-1 MAbs and serum antibodies of immunized guinea pigs with autologous HIV-1 Env and select HIV-1 V3 peptides were tested in ELISA as described previously (43). To probe the antibody specificity induced by HIV-1 Env proteins, sCD4 protein, MAbs 2F5, PG9, CH31, and 17B were biotinylated and used as targets. Blocking the binding of biotinylated sCD4 protein and MAbs to HIV-1 Envs by immunized guinea pigs serum antibodies was also determined by ELISA. Briefly, purified HIV-1 Env proteins (1 μg/ml) in coating buffer (0.1 M NaHCO₃ [pH 9.6]) were used to coat 384-well high binding ELISA plates at 15 μl/well at 4°C overnight. The plates were washed with superwash (1% Tween 20, phosphate-buffered saline [PBS]) and blocked with 40 μl of superblock (4% wheat protein, 1.5% goat serum, 0.5% Tween 20, and 0.05% sodium azide in PBS)/well at room temperature for 1 h. The blocked plates were washed with superwash and then incubated with 10 μl/well antibody in superblock at room temperature for 1.5 h. All MAbs were tested in serial dilution with a typical starting concentration of 50 μg/ml and no less than 20 μg/ml. After incubation, these plates were washed twice with PBS and then incubated with horseradish peroxidase (HRP)-conjugated goat anti-human IgG (1:5,000 dilution in 5% goat serum-PBS) at room temperature for 1 h. These plates were then washed four times with PBS and developed with 30 μl of TMB substrate (SureBlue Reserve, Gaithersburg, MD)/well. The HRP reaction was stopped with 30 μl of 1 M HCl/well, and the well optical density was determined at 450 nm. To determine antibody titers, pre- and postimmunization serum samples of guinea pigs were tested in 3-fold serial dilutions (from 1:30 to 1:5,314,410). Half-maximal effective concentrations (EC₅₀) of antibody binding titers were determined as described previously (43).

SPR.

Surface plasmon resonance (SPR) assays were performed on a BIAcore 3000 instrument and analyzed with BIAevaluation 3.0 software (BIAcore, Inc., Uppsala, Sweden) as described previously (9, 44). Briefly, anti-human IgG Fc antibody (Sigma Chemicals) was immobilized on a CM5 sensor chip to about 12,000 to 14,000 response units (RU), and each antibody was then captured to about 1,000 to 2,000 RU on individual flow cells of the same sensor chip. Env proteins were injected in the solution phase at 100 μg/ml for 1 min over each of the flow cells. Nonspecific binding of Env proteins to the captured control Synagis (anti-RSV) MAb was subtracted from each binding curve (9, 44).

Animals and immunizations.

Outbred guinea pigs were purchased from Jackson Laboratory and housed in the vivarium at Duke University Animal Research Facility. Animals were studied under a protocol approved by the Animal Use and Care Committee of Duke University. Guinea pigs (four animals per immunogen group) were immunized intramuscularly on day 0 and weeks 3, 6, 9, and 12 with 100 μg/injection of recombinant HIV-1 Env gp140 adjuvanted with type B oCpG in Emulsigen (MVP Technologies, Omaha, NE). Serum samples were collected 10 days after each immunization and stored at −30°C until use.

Neutralization assays.

Serum neutralizing antibody levels were measured against a panel of pseudotyped HIV-1 viruses containing Envs from subtype B (B.QH0692, B.SF162, B.SS1196, B.BaL, B.BG1168, B.3988 B.WITO [from acute HIV-1 infection], and B.6101 [from acute HIV-1 infection]), subtype C (C.TV-1, C.DU123, C.DU172, C.92ZM233M, C.92BR025, C.92ZM197M, C.96ZM651, C.DU151, C.97ZA012, C.DU422, C.DU156, C.02ZM233M, and C.02ZM197M), subtype A (A.92RW010, A.92UG37, A.93TH976.01, A.Q23, A.Q168, A.Q259, A.Q461, A.Q769, and A.Q842), and a single circulating recombinant form CRF01_AE Env, AE_93TH976 (45). Neutralization assays were based on reductions in luciferase (Luc) reporter gene expression after a single round of virus infection with pseudotyped HIV-1 viruses in TZM-bl cells (25, 46–48). For each assay, serum samples were collected prior to vaccination (prebleed) and postvaccination (postbleed) from guinea pigs and assayed against individual pseudotyped HIV-1 viruses and also tested against murine leukemia virus (MuLV) as a negative control.

We performed parallel analyses of neutralization data using two categorizations of neutralization data. The first was a traditional strategy, typical of what we and others have used in the past. Here, we considered a neutralization value positive only if the neutralization titer was 3-fold over both the prebleed level and the level seen with the MuLV control, and the value of the experimental neutralization titer was ≥30. Samples that did not meet these criteria were considered equivalent to below the threshold of detection (<20). We refer to this as the “traditional approach.” When applying this method, we noted that we were frequently disregarding low levels of postbleed activity that were reproducibly above background in all four animals in a group, but not 3-fold above background, and it is not known whether such low-level responses may be useful in a vaccine context. To have greater confidence in measures of low-level activity, for every guinea pig that had a higher than background prevaccination spurious NAb activity in its sera, IgG was purified from both pre- and postbleed, and purified IgG was used instead of sera from those animals to test the NAb response to the immunogen. The IgG levels were brought to a concentration equivalent to the IgG level in serum, so the titers were equivalent. A high prebleed background was essentially removed by this step. Thus, we also analyzed the data considering all 50% infective doses (ID₅₀) and values above the background, including the IgG purification step for all animals with high prebleed values. We refer to this as the “inclusive approach.” Using this approach, low-level responses against the panel of Envs were more reproducible within a group of animals that received the same immunogen than when using the traditional approach (see Results). We included both the traditional approach, because it is the way such data have been analyzed in the past, and the inclusive approach, because the results were more consistent between animals and the increased sensitivity might be relevant to human studies, in parallel analyses presented throughout the present study. In our inclusive analysis, values that were recorded as “20” and so were below the threshold of detection were replaced with “10,” as were any values less than the controls. Left censored data points were replaced with their lower limits (i.e., 540 and 43,740, respectively). Our “traditional analysis” was similar except that any postimmunization data point that was <3 times the preimmunization bleed was also replaced with a 10.

Analysis and statistical methods. (i) Hierarchical clustering.

Neutralization data were organized by hierarchical clustering so that similar patterns of NAb susceptibility among pseudovirions and similar patterns of NAb response among immunogens could be grouped and displayed using a heat map and dendrogram clustered according to distances between vectors of log ID₅₀ NAb scores using the Los Alamos HIV database tool heat map (http://www.hiv.lanl.gov/content/sequence/HEATMAP/heatmap.html). Euclidean distances and the average clustering method were used to create the dendrograms. This tool uses the statistical package R (49).

(ii) Nonparametric statistics used to compare distributions of values among consensus, T/F, and chronic immunogen sets.

The analysis was conducted using a Kruskal-Wallis test for omnibus effects and Wilcoxon tests for pairwise effects. We used these nonparametric methods to obviate any potential violations of normality assumptions. The z approximation was used to determine significance when there were ties in the data. These standard tests were performed using either the R package (49) or SAS v9.2. Benjamini-Hochberg (50) false discovery rates (FDR) to adjust for multiple tests were computed using PROC MULTTEST in SAS v9.2.

(iii) GLM analysis.

Because the data were heavily censored, they did not have a log-normal distribution but were markedly skewed with a bolus of points recorded as <20 or below the threshold of detection. (There were 3,015 tests performed, and 1,209 of 3,015 (40%) were recorded as <20; a few other measurements lay above the threshold of the assay, with 36 of 3,015 recorded as >540 [1.2%] and 5 of 3,015 recorded as >43,740 [0.17%]). We compared a Gaussian model, an inverse Gaussian model, and several models designed to address censored data that are discussed below. All modeling and model comparisons were done using the R package (49). For these data, using both the conserved and inclusive cutoffs, the inverse Gaussian model was significantly better by Akaike's information criterion; thus, it was used for the generalized linear model (GLM) comparison. The two approaches we tested designed to handle censoring were the “lmec” (linear-mixed-effects model for left censored data implemented in the R lmec package) and the MCMCglmm algorithm (for fitting generalized linear mixed effects models for left, right, or interval censored data in the R MCMCglmm package). The results obtained in each censored data model were consistent with those obtained in the normal fit and the inverse Gaussian mode, but the latter provided a better fit.

For the model, let y_i be the log₁₀ titer observed for guinea pig i. Also, let the regression coefficient β_k,0 represent the effect of the k^th immunogen relative to a specified reference immunogen, where β_k,0 = 0 represents no difference between the k^th immunogen and the reference immunogen, β_k,0 > 0 indicates that a stronger immune response was elicited by the k^th immunogen relative to the reference, and β_k,0 < 0 indicates a weaker elicited response. Finally, let G_i be an intercept, which augments the immunogen effect for a particular guinea pig. The model of the mean log₁₀ titer response is: E[y_i] = β₀ + β_1,0 Immunogen₁ +[r] β_2,0 Immunogen₂ + … +[r] β_19,0 Immunogen₁₉ + G_i, where Immunogen_k is an indicator variable, so that Immunogen_k = 1 if guinea pig i received the reference immunogen and is zero otherwise. Further, Immunogen_k = 0 for all k if guinea pig i received the reference immunogen. In the following development, we consider models whose reference immunogens were those that produced the strongest responses. CON-S gave the highest overall levels of response, so all other vaccines are measured relative to CON-S.

Binding reactivity of HIV-1 MAbs to Env gp140 proteins.

The antigenicity of the three groups of HIV-1 gp140 Envs were tested and compared using ELISA (Table 1) and SPR (Table 2) for their ability to bind a panel of HIV-1 Env antibodies including CD4 binding site antibodies, b12, and VRC01, VRC02, and VRC03 from a related lineage; CCR5 coreceptor binding site (CCR5 bs) MAb 17b; glycan antibody 2G12; quaternary antibodies PG9 and PG16, a pair from a related lineage; and MPER MAbs 2F5 and 4E10. The relative binding affinity of each of the individual HIV-1 Envs to specific antibodies measured by ELISA (EC₅₀, nM) and SPR (response unit) are summarized in Tables 1 and ​and2.2. Of note, consensus proteins are artificial constructs, a concatenation of the most common form of each amino acid in each position in an alignment. Thus, HIV-1 population consensus sequences contained combinations of amino acids that are not found in nature, and there is always a concern that they may not fold appropriately. Therefore, the consistency of binding of antibodies with discontinuous epitopes to the consensus sequences was reassuring. CON-S gp140 Env bound with the highest affinity among the consensus sequences, and the only MAb tested that did not bind to CON-S was VRC03 (Tables 1 and ​and2).2). VRC01 and VRC02 are derived from the same clonal lineage as VRC03 (57), and these two MAbs did bind to CON-S with high affinity. The only consensus Env that did not bind to most MAbs tested was A1.con.

Table 1

Binding reactivity of HIV-1 Envs to HIV-1-specific MAbs in ELISA^a

Open in a separate window

^a*, No binding detected at 20 mg/ml.

Table 2

Binding reactivity of HIV-1 Envs to neutralizing MAbs in SPR analysis^a

Open in a separate window

^a*, Below the threshold of detection using SPR.

We next compared the overall binding scores of the consensus, chronic, and T/F Envs to the panel of MAbs in shown in Tables and ​and11 and ​and2.2. We first used the nonparametric Kruskal-Wallis omnibus test to determine whether, for a given antibody, one of the three groups was distinct from the others, followed by Wilcoxon pairwise comparisons. VRC01, VRC02, and VRC03 were all derived from the same clonal lineage, with VRC01 being most potent neutralizing MAb among them (57). PG9 and PG16 are similarly clonally related, with PG9 being the most potent (17). Thus, only VRC01 and PG9 were used in the statistical comparisons, to represent the two lineages and minimize the number of tests. Because multiple tests were performed, we performed an FDR comparison. The T/F binding tended to be higher in number of epitopes expressed, and the binding of MAb 17b was significantly different between the three groups; 17b had significantly higher binding affinity for T/F viruses (Tables 1 and ​and22 and see Table S2 in the supplemental material, false discovery adjusted P value = 0.02 for ELISA, 0.03 for SPR). 17b is a CD4-inducible antibody that binds near the CCR5-coreceptor binding site (38) and, interestingly, increased T/F binding affinity was observed even though sCD4 was not present in the binding assay. Similar reactivity patterns of MAbs 2F5, PG9, CH31, and 17b with HIV-1 Envs detected by ELISA and by SPR were also observed in BN-PAGE Western blot analysis. When these MAbs did bind to HIV-1 Envs, MAb 2F5 had the strongest reactivity and recognized all forms of gp140 oligomers, whereas the PG9, CH31, and 17B primarily recognized dimers, trimers, and oligomers but also recognized the gp140 monomer weakly or not at all (see Fig. S2 in the supplemental material).

Immunogenicity of T/F, chronic, and consensus Envs.

Next, the 20 gp140 Envs in Table S1 in the supplemental material were used to immunize guinea pigs. Most consensus gp140 sequences were based on the database from 2003; two B clade consensus Envs (B.con_01 from 2001 and B.con_03 from 2003) and two group M consensus Envs (CON-S from 2001 and CON-T from 2003) were included in the immunogenicity studies to ensure that we were not missing minor Env changes that might lead to induction of the breadth of responses. The immune responses elicited were indeed somewhat different when using immunogens from different years (Fig. 2). The T/F Envs were from 2005 to 2007 isolates (22). NAb responses to these HIV-1 Envs were evaluated using a panel of 36 Env-pseudotyped viruses from clades A, B, C, and CRF01_AE (see Table S3 in the supplemental material).

Open in a separate window

Fig 2

Neutralization activity of guinea pig groups immunized with HIV-1 Envs. Geometric means of IC₅₀ neutralization titers (y axis) for groups of guinea pig that received the indicated HIV-1 Envs (grouped into consensus, T/F, and chronic families) were determined against a panel of subtype A, B, C, and E pseudoviruses using the traditional approach (A) or the inclusive approach (B). The same color codes were used for each subtype for both Env immunogens, as shown in the lower right portions of panels A and B. Pseudoviruses are ordered the same way in each panel, according to the key at the bottom, to facilitate comparisons between vaccine responses.

Bar plots illustrating the relative potency of ID50 NAb responses elicited by each immunogen against each of the pseudovirus Envs are shown in Fig. 2 and Fig. S1 in the supplemental material (Fig. 2 shows the geometric mean response for each immunized group of four animals to each Env; Fig. S3 in the supplemental material illustrates the same data, but the results for each animal are shown individually). As described in Materials and Methods, we analyzed the data using two strategies: a traditional approach (Fig. 2A), in which a positive neutralization value of immune guinea pig serum had to be at least three times higher than both the prebleed serum and the negative control virus MuLV-SVA to be considered positive, and an inclusive approach where IgG was purified to minimize background, and then any responses that were above both negative controls were considered positive (Fig. 2B). By definition, more reactions were deemed positive when using the inclusive method (Fig. 2A compared to Fig. 2B). Using both data interpretation strategies, a pattern is evident that is associated with each of the classes of immunogens. Consensus sequence immunogens tended to have the greatest responses to tier 1 Envs (45) and elicited a moderately high level of response to a subset tier 2 Envs. T/F immunogens gave low levels of response across most Envs tested, but with greater breadth than consensus immunogens. Chronic Env immunogens showed moderate to low-level responses to some tier 1 Envs and sporadic low responses to tier 2 Envs (45). Within-clade matched responses are indicated by the colors in Fig. 2. Little clade specificity was manifested in these data, although occasionally it is evident. For example, the B.con.03 responses favored B clade pseudoviruses. Of note, the B and C clade T/F viruses did not elicit markedly stronger responses within clades than between clades (Fig. 2).

Heat maps, organized by two-dimensional hierarchical clustering, further illustrate the over distinctive patterns between consensus T/F and chronic immunogens. The traditional threshold data heat map is shown in Fig. 3A, with the inclusive method of analysis in Fig. 3B. In these heat maps, pseudotyped Envs with similar susceptibility patterns are clustered together and, likewise, immunogens that elicited antibodies that had similar neutralizing susceptibility are clustered together, regardless of their classification as T/F, consensus, or chronic. Between three and seven animals per immunogen group were tested. The three patterns associated with chronic, T/F, and consensus immunogens are clearly discernible (Fig. 3). The chance that the animals who received the same type of immunogen would cluster according to reactivity pattern to the extent observed in the heat maps is very small if you consider each animal separately (P < 2 × 10⁻¹⁶, the Fisher exact test for a 3×3 table for the traditional method, and P = 0.0000002 for the inclusive method), and still low when animals given the same immunogen are considered as a group rather than single animals, and the group is assigned to the cluster that most of the animals in the group fall into (P = 0.003 for the conserved approach, P = 0.01 for the inclusive). Although the distinct patterns were statistically significant, there were a few exceptions to this overall pattern. The T/F founder virus, C.1086, behaved more like a consensus immunogen than the other T/F strains, in that C.1086 did not have great breadth, and it generates more intense tier 1 than tier 2 responses (for example, the C.1086-vaccinated animals have a NAb profile very similar to that of the G consensus-vaccinated animals). Also, the two subtype B consensus sequences from different years were subtype specific and elicited responses similar to the chronic B clade isolate B.JRFL.

Open in a separate window

Open in a separate window

Fig 3

Heat maps reflecting HIV-1 neutralization activity of each individual guinea pigs immunized with HIV-1 gp140 Envs. The neutralizing activity of individual guinea pigs immunized with a particular HIV-1 Env is presented as a heat map, with stronger neutralizing activity indicated by color, as shown in the key at the upper left. Serum samples from individual guinea pigs are identified by a string beginning with the name of the HIV-1 Env used to immunize the animal, followed by a unique guinea pig ID number. The panel of 36 HIV-1 pseudoviruses is listed at the bottom of the heat map, arranged in order of decreasing susceptibility to neutralization. The neutralizing activity of individual guinea pig sera was analyzed using both the traditional approach (A) and the inclusive approach (B).

Hierarchical clustering strategy used to create the heat map brought the replicate animals in 12 of the 20 immunogen groups that received the same immunogen most closely together, when the inclusive strategy was used (see brackets in Fig. 3B). This compares with only 2 of 20 immunogen group clusters using the traditional method (see brackets in Fig. 3A) (P = 0.002, calculated using a Fisher exact test comparing inclusive versus traditional strategies). This reflects the greater consistency of the response data between animals that received the same immunogen when the inclusive cutoff was used and is an indication that the inclusion of low-level responses does not just add noise but rather captures consistent reproducible patterns in the Env immunogen-induced neutralizing responses.

Comparison of neutralizing breadth induced by T/F, chronic, and consensus Env immunogens.

For an initial comparison of neutralizing breadth responses induced by the three classes of Env immunogens, we counted the number of times an immunogen group of animals had at least two responding animals that had detectable neutralizing titer for each of the 36 Envs and compared the distribution of counts for groups of T/F, consensus, and chronic immunogens (Fig. 4A and ​andBB illustrate the breadth of Env responses among the 36 Env pseudovirions for each immunogen using the traditional and inclusive strategies, respectively). The omnibus Kruskal-Wallis test indicated that there was a significant difference among the distributions using the traditional approach (P = 0.0027), and T/F Envs ranked significantly higher than consensus Envs, which in turn were significantly higher than chronic Envs (all P values of each pairwise comparison are provided in Fig. 4). Using the less stringent inclusive strategy (Fig. 4B), many more low responses against a spectrum of Envs were observed within all groups, and so the groups were not statistically distinguishable, although there was a trend again suggesting that T/F Envs induced a greater breadth of response than did chronic Envs (P = 0.061).

Open in a separate window

Fig 4

Comparison of breadth of antibody responses induced by HIV-1 T/F, consensus, and chronic Env immunogens. The breadth (the vertical axis shows the percentage of the panel of 36 pseudoviruses neutralized, averaged across all immunized animals) for each of the individual HIV-1 Env immunogens (x axis) using the inclusive approach (A) and the traditional approach (B). The immunogens were classed as T/F, consensus, or chronic Envs, and the P values are for comparisons between these classes.

Comparison of the magnitude of responses to single immunogens.

We initially used a generalized linear model to compare the potency of the neutralizing antibody responses using the different immunogens. The model selection is described in Materials and Methods; the model which fit the data the best was an inverse Gaussian model, and animal-to-animal immune system differences were accounted for by treating the interanimal variation as a random effect. The analysis requires as input both the neutralization titers (one for each measured combination of animal and isolate) and a separate list of responses that should be regarded as definitely positive, based on whether the traditional or the inclusive approach described above is used. The CON-S immunogen showed the highest overall levels of response using both traditional (Table 3) and inclusive (Table 4) threshold approaches and so was used as the reference strain, and all other immunogens were measured relative to CON-S. Using the traditional approach, CON-S Env induced significantly greater antibody responses than all other HIV-1 Env immunogens, and chronic responses, along with T/F B.9201, ranked the lowest (Table 3). Using the inclusive approach, CON-S again ranked the highest but was not statistically distinguishable from other consensus immunogens, including CON-T, A1-con, C-con, and B-con_01; the CON-S response was significantly higher than all T/F and chronic Env immunogens.

Evaluation of neutralizing breadth versus strength of responses induced by consensus, T/F, and chronic Env immunogens.

To reexamine the question of breadth and depth in a probabilistic framework, given our uncertainty about near-threshold responses, we developed a new class of models designed to separate the breadth of the NAb response—i.e., the number and variety of strains to which an immunogen induces a response—from the strength of that response—i.e., the raw neutralizing power of the sera from those animals in which the immunogen induces a positive response. This model also incorporated a self-consistent estimation process for the censored data. The resulting model has two aspects: a binomial regression that produces a probability of positive response for each combination of immunogen and isolate and then, again for each immunogen-isolate combination, a pair of linear regression models—one for the positive responses and another for the negative ones—that predict the logs of the neutralization titer.

The novelty of the approach is that we require internal consistency between the two aspects of the model: if some immunogen-isolate combinations produce only modestly strong positive responses, then it is possible that some near-threshold responses, initially classified as negative, may contribute to a better-fitting model if they are reclassified as weak, but positive responses. After an initial round of parameter estimation, we assign a suitable positive weight to such near-threshold responses and fold them back into the analysis, using the EM-algorithm (58) to repeatedly fit parameters and reweight the data until we obtain a self-consistent model.

As before, we fit models based on the two different positivity criteria, the conserved and inclusive strategies. Using this model, the conserved data fit better (in the sense of having the higher log-likelihood), and so the parameters of conserved model are illustrated in Fig. 5 and ​and6.6. Figure 5 shows the fitted probability of a positive response as a function of immunogen-isolate combination. Using this measure and modeling the probability of a response being positive (in contrast to our previous measure of breadth in Fig. 4), the broadest neutralizing responses were induced by the consensus Env AE.con, followed closely by the immunogens derived from the T/F strains B.63521, B.0040, and C.089. The reason underlying this reordering compared to Fig. 4 can be understood by looking closely at the data in Fig. S3 in the supplemental material; while AE.con has fewer detected responses, many of the responses to the T/F immunogens are borderline, and when AE.con does have a response, it tends to be a higher value and so is more likely to be rescored as positive in this probabilistic model. Three of the four immunogens with the highest probability of response to the most Envs are T/F viruses, and these three consistently yield great breadth by both measures (Fig. 4 and ​and55).

Open in a separate window

Fig 5

Heat map reflecting breadth of antibody responses induced by HIV-1 consensus, chronic, and T/F Env immunogens. Heat map showing the fitted probability that a guinea pig in each of the individual immunization groups (rows) would have shown neutralizing activity against the panel of 36 HIV-1 pseudoviruses (columns). A higher probability of a positive response (traditional criterion) is indicated by a darker color, as detailed in the key at the upper left.

Open in a separate window

Fig 6

Heat map showing the fitted strength (predicted log₁₀ titer) of positive responses induced by HIV-1 immunogens. As in Fig. 5, rows correspond to immunogens, whereas columns are labeled by the pseudovirus strain. The key at the upper left shows both the color scale (with the strongest predicted responses corresponding to titers of 10,000, as seen for some tier 1 Envs) and the distribution of fitted values: many lie near log₁₀(20) ≈ 1.3, indicating that some responses near the threshold of the assay may be weak but positive.

The strengths of positive responses induced by different Envs were also compared using this model and were consistent with the findings reported above based on an inverse Gaussian GLM. The consensus Envs as a class were by far the most effective at eliciting potent responses (Fig. 6). If we again make suitably adjusted multiple comparisons between the two best immunogens from each category, the best consensus Env, CON.S, produced significantly stronger responses than any of the others (P = 0.0233 for C.con, the second best consensus, while P < 0.001 for all T/F and chronic Envs).

This study was supported by grant U19AI067854 from the Center for HIV/AIDS Vaccine Immunology, by grant UM1AI100645 from the Center for HIV/AIDS Vaccine Immunology-Immunogen Discovery of the Division of AIDS, NIAID, NIH, and by Vaccine Discovery Center of the Collaboration for AIDS Vaccine Development Program grant OPP1033098s from the Bill and Melinda Gates Foundation.