Dissecting Polyclonal Vaccine-Induced Humoral Immunity against HIV Using Systems Serology

Abstract

While antibody titers and neutralization are considered the gold standard for the selection of successful vaccines, these parameters are often inadequate predictors of protective immunity. As antibodies mediate an array of extra-neutralizing Fc functions, when neutralization fails to predict protection, investigating Fc-mediated activity may help identify immunological correlates and mechanism(s) of humoral protection. Here, we used an integrative approach termed Systems Serology to analyze relationships among humoral responses elicited in four HIV vaccine trials. Each vaccine regimen induced a unique humoral "Fc fingerprint." Moreover, analysis of case:control data from the first moderately protective HIV vaccine trial, RV144, pointed to mechanistic insights into immune complex composition that may underlie protective immunity to HIV. Thus, multi-dimensional relational comparisons of vaccine humoral fingerprints offer a unique approach for the evaluation and design of novel vaccines against pathogens for which correlates of protection remain elusive.

Figure 1. System Serology Analysis

This Systems Serology platform allows for the broad characterization of the polyclonal extra-neutralizing IgG immune profile induced by vaccination. IgG was purified from subjects enrolled in four different HIV vaccine trials (RV144, VAX003, HVTN204 and IPCAVD001). Six Fc effector functions and 58 biophysical measurements were assayed (complete list described in Table S1). All 64 parameters were collected to create an extra-neutralizing serological signature for the four vaccine trials using an array of unsupervised and supervised machine learning algorithms. See also Tables S1 and S2.

Figure 2. Hierarchical clustering of vaccine trial profiles by biophysical properties and functional responses

Data was compiled for the four different vaccine trials. Each column represents the full Ab profile of an individual subject. Colored bars along the bottom correspond to the vaccine trial for each subject. Ab properties are grouped by generalized features (Function, FcR affinity, Bulk IgG, IgG1, IgG2, IgG3, IgG4) indicated by the colored bars on the right. Specific features are listed in Table S2. See also Table S1.

Figure 3. PLSDA and LASSO identify unique combinations of features that differentiate vaccine trial Ab profiles

(A) The scores plot represents the RV144 (red) and VAX003 (blue) vaccine profile distribution for each vaccinee tested (dots) from the LASSO and PLSDA. Remarkably, as few as, 7 Ab features, listed on the loadings plot (B), separated the vaccine profiles with 100% calibration and 97% cross-validation accuracy. LV1 captured 61% of X variance and 72% of the Y variance. (C) LASSO and PLSDA of all 4 vaccine profiles identified 15 Ab features (D) able to discriminate between the distinct vaccine regimens (red, RV144; blue, VAX003; green, HVTN204; and yellow, IPCAVD001) with 84% cross validation accuracy. Together LV1 and LV2 captured 57% of the X variance and 45% of the Y variance, respectively. See also Tables S1 and S2.

Figure 4. Correlation networks of vaccine trial-elicited humoral immune responses probe immune complex dynamics

Correlation networks were generated for VAX003 (A), RV144 (B), HVTN204 (C), and IPCAVD001 (D). Each node (circle) represents either a biophysical feature or an effector function. Nodes are connected with an edge (line) if they are significantly correlated. The different Ab isotypes are identified by different colors as indicated. Edge thickness and color intensity of the connecting lines are directly proportional to statistical significance and edge weight, respectively (thicker and brighter network interactions represent a stronger correlation). The size of each node is directly proportional to its degree of connected-ness (ie. the number of features to which that node is connected). See also Figure S1 and Tables S1 and S2.

Figure 5. Identification of V1V2^high-associated signatures within RV144 vaccine responses

(A) RV144 vaccinees were classified within the IgG V1V2AE^high (red) (top 30%) or IgG V1V2AE^low (blue) groups. (A) LASSO identified a profile of 16 features that differentiated the two groups with 100% calibration and 80% cross-validation accuracy. The loadings plot (right panel) illustrates the features that separated IgG V1V2AE^high or IgGV1V2AE^low responders. Together LV1 and LV2 captured 33% of the X variance and 94% of the Y variance, respectively. (B) The same analysis was repeated for RV144 vaccinees classified as IgG3 V1V2^high/IgG3V1V2^low (B) with 92% cross-validation and 100% calibration accuracy. (D) LASSO identified a signature of 10 features that best separated these two groups. Together LV1 and LV2 captured 39% of the variance in X and 84% of the variance in Y, respectively. See also Tables S1 and S2.

Figure 6. Defining novel signatures of protection in the RV144 case:control data

(A) The PLSDA shows the distribution of all case:control data including all infected and uninfected placebos as well as infected and uninfected vaccinees using 101 humoral features (described in Table S3). LV1 accounted for 68.1% of all variance, separating most placebos from the vaccines, while LV2 only contributed to 4.5% of the variance. (B) Further insights into the distribution of IgA gp120, IgG V1V2 and IgG3 V1V2 levels were analyzed using histograms demonstrating unique multi-modal differences in feature distribution among the infected and uninfected vaccinees. (C) The scatter plot, in the central panel, represents the bivariate distribution of IgA gp120 and IgG V1V2 in the vaccines, and is framed by the histogram distributions for uni-dimensional reference. The blue and red dash-lined boxes represent quadrants within the data that comprise of the fewest cases:controls (low risk, blue) or the highest ratio of cases:controls (high risk, red). (D and E) LASSO and PLSDA identified 9 features that split low and high risk profile separation with 97.8% accuracy in cross-validation. Together LV1 and LV2 captured 70.4% of the X variance and 30.1% of the Y variance, respectively. Correlation networks were generated for both the low risk (F) and high risk (G) groups. See also Figures S2 and S3, and table S3.