Machine Learning for Predicting Vaccine Immunogenicity

Abstract

The ability to predict how different individuals will respond to vaccination and to understand what best protects individuals from infection greatly facilitates developing next-generation vaccines. It facilitates both the rapid design and evaluation of new and emerging vaccines and identifies individuals unlikely to be protected by vaccine. We describe a general-purpose machine-learning framework, DAMIP, for discovering gene signatures that can predict vaccine immunity and efficacy. DAMIP is a multiple-group, concurrent classifier that offers unique features not present in other models: a nonlinear data transformation to manage the curse of dimensionality and noise; a reserved-judgment region that handles fuzzy entities; and constraints on the allowed percentage of misclassifications. Using DAMIP, implemented results for yellow fever demonstrated that, for the first time, a vaccine’s ability to immunize a patient could be successfully predicted (with accuracy of greater than 90 percent) within one week after vaccination. A gene identified by DAMIP, EIF2AK4, decrypted a seven-decade-old mystery of vaccination. Results for flu vaccine demonstrated DAMIP’s applicability to both live-attenuated and inactivated vaccines. Results in a malaria study enabled targeted delivery to individual patients. Our project’s methods and findings permit highlighting and probabilistically prioritizing hypothesis design to enhance biological discovery. Moreover, they guide the rapid development of better vaccines to fight emerging infections, and improve monitoring for poor responses in the elderly, infants, or others with weakened immune systems. In addition, the project’s work should help with universal flu-vaccine design.