Machine learning-driven identification of serotype-independent pneumococcal vaccine candidates using samples from human infection challenge studies.

Identifying conserved, immunogenic proteins that confer protection against Streptococcus pneumoniae (pneumococcus) colonization could enable development of serotype-independent vaccines. In our controlled human infection model, no individual IgG or cytokine/chemokine response correlated significantly with protection against colonization with pneumococcus, suggesting that effective immunity reflects a coordinated, multi-antigen response. To capture these complex patterns, we trained independent Random Forest models on humoral and cellular datasets. The humoral model identified IgG responses to PdB, SP1069, and SP0899 as predictive of protection. The cellular model revealed that MCP-1 responses to SP1069 and SP0899, and IL-17A production in response to SP0648-3, were associated with protection. Elevated baseline IFN-γ, RANTES, and anti-protein IgG levels were linked to reduced colonization density. We highlight SP1069 and SP0899 as potential serotype-independent vaccine candidates and demonstrate the utility of machine learning to identify immune correlates of protection.