Abstract:
Despite being treatable and preventable, malaria remains a major global health problem, responsible
for over 600,000 deaths in 2022 alone. To date, the RTS,S/AS01 (RTS,S) and R21/Matrix-M (R21)
vaccines are the first two licensed malaria vaccines recommended by theWorld Health Organization
for the prevention of Plasmodium falciparum(P. falciparum)malaria in children living in endemic
areas. Both, RTS,S and R21 are subunit vaccines that provide partial protection for about a year.
Despite the success of these first-generation vaccines, there remains an urgent need to developmore
effective malaria vaccines. These new vaccines should ideally be suitable for broader use, including
for pre-exposed adults in endemic regions, pregnant women, and malaria-naïve travelers. Moreover,
they should confer long-lasting sterile protection and help reduce malaria transmission.
Notably, sterile protection against P. falciparum has been achieved with new vaccine platforms, such
as radiation-attenuated P. falciparum sporozoites (PfSPZ Vaccine) and chemo-attenuated PfSPZ
(PfSPZ-CVac) inmalaria-naïve volunteers. However, these vaccines generally show lower efficacy
in pre-exposed residents of endemic regions. The protection induced by these SPZ-based vaccines
primarily occurs during the pre-erythrocytic stages and is largely mediated by cellular and antibody
responses. Therefore, a deeper understanding of the vaccine-induced immune response in both,
malaria-naïve individuals and those living in endemic regions, as well as the identification of new
potential antigen, is crucial for the development of improved vaccine candidates.
Machine learning has become increasingly popular in medical research over the past few decades,
significantly improving our understanding of complex medical data. However, machine learning
faces significant challenges when applied tomedical data obtained in early stage clinical development,
particularly in the context of vaccine efficacy assessment against P. falciparum. Such data is
often high dimensional, limited in sample size, and heterogeneous. Consequently,machine learning
in malaria research requires methods that can effectively handle small, high dimensional, and heterogeneous
datasets.
Moreover, identifying and interpreting P. falciparum-specific antigens involved in vaccine-mediated
and naturally acquired protection necessitates interpretation techniques capable of managing highdimensional
datasets while accounting for correlations between features. This thesis addresses both aspects by developing machine learning methods tailored to small, high-dimensional datasets and inventing
interpretation techniques to evaluate informative P. falciparum-specific antigens associated
with protection frommalaria. To this end, a multitask support vector machine approach was adapted
to combine time- and PfSPZ-CVac dose-dependent data, enhancing vaccine efficacy prediction
compared to state-of-the-art methods in a study population of malaria-naïve volunteers. Additionally,
a kernel-based interpretation technique was developed to evaluate P. falciparum-specific
antigens in the classification of protected versus non-protected volunteers. Another task-specific
contribution was the development of a benchmark for training machine learning models to explore
P. falciparum protein vaccine and drug candidates. Furthermore, this thesis presents a workflow
for appropriately applying machine learning methods to predict vaccine efficacy across different
vaccine types, comparing pre-exposed and malaria-naïve adult volunteers. While the results of
this thesis show promising advancements in the application of machine learning to the analysis of
malaria vaccine-induced humoral immunity, they also highlight the challenges of applying machine
learning methods to small, high-dimensional datasets derived from early stage clinical trials.
