AUTHOR=Cotugno Nicola , Santilli Veronica , Pascucci Giuseppe Rubens , Manno Emma Concetta , De Armas Lesley , Pallikkuth Suresh , Deodati Annalisa , Amodio Donato , Zangari Paola , Zicari Sonia , Ruggiero Alessandra , Fortin Martina , Bromley Christina , Pahwa Rajendra , Rossi Paolo , Pahwa Savita , Palma Paolo TITLE=Artificial Intelligence Applied to in vitro Gene Expression Testing (IVIGET) to Predict Trivalent Inactivated Influenza Vaccine Immunogenicity in HIV Infected Children JOURNAL=Frontiers in Immunology VOLUME=11 YEAR=2020 URL=https://www.frontiersin.org/journals/immunology/articles/10.3389/fimmu.2020.559590 DOI=10.3389/fimmu.2020.559590 ISSN=1664-3224 ABSTRACT=

The number of patients affected by chronic diseases with special vaccination needs is burgeoning. In this scenario, predictive markers of immunogenicity, as well as signatures of immune responses are typically missing even though it would especially improve the identification of personalized immunization practices in these populations. We aimed to develop a predictive score of immunogenicity to Influenza Trivalent Inactivated Vaccination (TIV) by applying deep machine learning algorithms using transcriptional data from sort-purified lymphocyte subsets after in vitro stimulation. Peripheral blood mononuclear cells (PBMCs) collected before TIV from 23 vertically HIV infected children under ART and virally controlled were stimulated in vitro with p09/H1N1 peptides (stim) or left unstimulated (med). A multiplexed-qPCR for 96 genes was made on fixed numbers of 3 B cell subsets, 3 T cell subsets and total PBMCs. The ability to respond to TIV was assessed through hemagglutination Inhibition Assay (HIV) and ELIspot and patients were classified as Responders (R) and Non Responders (NR). A predictive modeling framework was applied to the data set in order to define genes and conditions with the higher predicted probability able to inform the final score. Twelve NR and 11 R were analyzed for gene expression differences in all subsets and 3 conditions [med, stim or Δ (stim-med)]. Differentially expressed genes between R and NR were selected and tested with the Adaptive Boosting Model to build a prediction score. The score obtained from subsets revealed the best prediction score from 46 genes from 5 different subsets and conditions. Calculating a combined score based on these 5 categories, we achieved a model accuracy of 95.6% and only one misclassified patient. These data show how a predictive bioinformatic model applied to transcriptional analysis deriving from in-vitro stimulated lymphocytes subsets may predict poor or protective vaccination immune response in vulnerable populations, such as HIV-infected individuals. Future studies on larger cohorts are needed to validate such strategy in the context of vaccination trials.