Development of A.I.-enhanced peptide array analysis framework to characterize a cross-reactive and polyspecific antibody in Plasmodium vivax and Plasmodium falciparum

Malaria caused by Plasmodium falciparum is among the deadliest infectious diseases and results in more than half a million deaths per year. 12 million pregnant women are infected by malaria in sub-Saharan Africa annually. Vaccine candidates to protect pregnant women from malaria aim primarily to elicit antibodies to P. falciparum virulence factor VAR2CSA that is uniquely expressed during infection in pregnancy. VAR2CSA is expressed on the surface of infected erythrocytes. It binds chondroitin sulfate A (CSA)-specific proteoglycans on the placenta syncytiotrophoblast, disrupting the exchange of oxygen and nutrients from the maternal blood to the fetus. Pregnancy-associated malaria (PAM) is associated with poor maternal and birth outcomes, including severe maternal anemia, low birth weight infants, pre-term birth, and fetal growth restriction, underscoring the importance of developing vaccines specific to this population. Vaccines based on recombinant fragments of VAR2CSA showed promise in pre-clinical studies but failed to elicit broadly inhibitory antibodies in Phase I trials. A critical bottleneck for these vaccines is extensive genetic diversity between VAR2CSA alleles; thus, efforts are underway to identify conserved epitopes for next-generation vaccines. One approach is to study the epitopes recognized by VAR2CSA antibodies acquired naturally during infection in pregnant women. However, antibodies to VAR2CSA were also reported in specific non-pregnant populations. In Colombia and Brazil, men and children had antibodies to VAR2CSA at similar levels to pregnant women of all gravidities. Based on the local epidemiology of malaria, we proposed that exposure to the Plasmodium vivax species, and specifically to the P. vivax Duffy binding protein (region 2; PvDBPII), can elicit antibodies that cross-react with VAR2CSA. A mouse monoclonal antibody raised against PvDBPII, called 3D10, also cross-reacts with VAR2CSA and moderately inhibits the adhesion to CSA of infected erythrocytes expressing different alleles of var2csa. A conserved epitope may be shared between these evolutionarily distinct proteins, which could be targeted in a vaccine against VAR2CSA. Various studies describe the epitope in PvDBPII that is recognized by 3D10 based on mutational analysis, phage display libraries, mimotopes, peptide arrays, and alanine scanning. A discrete binding region was identified within PvDBPII. Yet a discrete homologous epitope in VAR2CSA was not found, suggesting the epitope in the cross-reactive protein may be more complex. This was also evident from our analysis of VAR2CSA peptide arrays that were screened with 3D10 against PvDBPII, the closely related PvEBP2, and two alleles of VAR2CSA (FCR3 and NF54). The array data revealed one primary binding site and several highly reactive peptide clusters within the VAR2CSA DBL domains. My objectives were to develop of a machine learning framework to analyze peptide array outputs and to apply this framework to identify conserved features of 3D10 epitopes within two diverse alleles of VAR2CSA. My framework was used machine learning to extract basic features associated with antibody reactivity from our peptide array data: positive charge, lysine number, etc. Peptide arrays are unique in quantifying antibody reactivity to thousands of simple biomolecules, which lends itself well to identifying these features. These basic features, however, are insufficient for predicting reactivity. Statistical analysis of reactive peptides in contrast to non-reactive peptides can identify patterns of these basic features that are common among reactive peptides. The mutational analysis focuses on modifying the extracted features relative to these complex patterns, which invariably modifies binding. Different mutational strategies to different physicochemical properties associated with antibody reactivity clarify how these features impact reactivity. Now, with the basic features, feature patterns, and experimental validation (the ingredients), we can identify complex motifs and criteria for antibody reactivity (the recipe). To determine whether these criteria predict antibody reactivity, we mapped these criteria onto the surface of proteins of interest and determined if the antibody was reactive to these proteins. In the validation of our criteria, we accurately predicted and observed that 3D10 will only bind DBL-containing Plasmodium proteins and DBL domains consistently satisfy the 3D10 binding criteria. 3D10 behaves like a polyspecific antibody. Despite the presence of multiple potential epitopes, one potential epitope identified on VAR2CSA mapped to the CSA binding channel. Irrespective of the success of vaccines derived from 3D10 epitope mapping, this framework is promising for future peptide array analysis. Refinement by testing new antibody datasets will expand use cases and antibody analysis by peptide array.