Identification of Immune Signatures of Novel Adjuvant Formulations Using Machine Learning

Abstract

Adjuvants have long been critical components of vaccines, but the exact mechanisms of their action and precisely how they alter or enhance vaccine-induced immune responses are often unclear. In this study, we used broad immunoprofiling of antibody, cellular, and cytokine responses, combined with data integration and machine learning to gain insight into the impact of different adjuvant formulations on vaccine-induced immune responses. A Self-Assembling Protein Nanoparticles (SAPN) presenting the malarial circumsporozoite protein (CSP) was used as a model vaccine, adjuvanted with three different liposomal formulations: liposome plus Alum (ALFA), liposome plus QS21 (ALFQ), and both (ALFQA). Using a computational approach to integrate the immunoprofiling data, we identified distinct vaccine-induced immune responses and developed a multivariate model that could predict the adjuvant condition from immune response data alone with 92% accuracy (p = 0.003). The data integration also revealed that commonly used readouts (i.e. serology, frequency of T cells producing IFN-γ, IL2, TNFα) missed important differences between adjuvants. In summary, broad immune-profiling in combination with machine learning methods enabled the reliable and clear definition of immune signatures for different adjuvant formulations, providing a means for quantitatively characterizing the complex roles that adjuvants can play in vaccine-induced immunity. The approach described here provides a powerful tool for identifying potential immune correlates of protection, a prerequisite for the rational pairing of vaccines candidates and adjuvants.

Figure 1

Overview of all measurements collected in this study. Samples were collected from blood, liver, lymph node, and spleen. Serology, Fluorospot, cytokine, and flow cytometry assays were carried out for all tissues and different time points for peripheral blood mononuclear cells (PBMCs).

Figure 2

Hierarchical clustering of vaccine-induced immune responses. Hierarchical clustering of immune responses based on their correlation coefficients is shown, colored by immune cluster. Shaded circles below the immune measures indicate statistical significance as a vaccine-induced response or an adjuvant effect. Cluster names are shown, and clusters that predominantly show vaccine-induced measures are highlighted.

Figure 3

Principal component analysis of vaccine-induced immune responses. The first two principal components (PC1, PC2) are plotted comparing subjects with different antigen doses (left) and different adjuvant conditions (right), compared to non-vaccinated controls. Vectors corresponding to the projection of each immune measure along the two components are shown.

Figure 4

Adjuvant-specific differences in the SAPN-based vaccine. The ALFA-specific response in CSP-specific IL5- and IL6-producing cells (top left and right, respectively). ALFQ-biased and ALFQ-specific responses in CSP C-term-specific ELISA and CSP-specific IL12/IL23p40-producing cells (bottom left and right, respectively).

Figure 5

Linear regression model of combination ALFQA adjuvant. Median values for eight representative immune parameters that showed significant differences with respect to adjuvant are displayed in a radar plot using normalized values for ALFA (blue), ALFQ (green), and ALFQA (red). The estimated values based on the linear regression model for ALFQA (pink) is shown along with the 95% confidence interval for the estimated values (shaded pink).