Datensatz

Zusammenfassung

The human leukocyte antigen class II (HLA-II) genes encode for a group of surface
proteins expressed predominantly on antigen-presenting cells (APCs) where they
present antigenic peptides to naïve CD4+ T cells. Various alleles and haplotypes at
the HLA-II gene loci are genetically associated with several diseases, such as
chronic inflammatory bowel disease (IBD). Experimental determination of the
interactions between HLA-II proteins and antigenic peptides is an extremely
challenging, laborious and expensive task. Thus, computational tools that can
efficiently predict potentially binding peptides may help to reduce the peptide search
space. Recently, deep learning (DL) has been proven to be a reliable solution to
many challenging problems where big datasets are available. In my thesis, the utility
of a novel high throughput microarray-based technology along with deep learning in
modelling peptide HLA-II interactions have been evaluated. Data for six IBD-relevant
HLA-DRB1 proteins have been included in the study. Coupling deep learning with
microarray technology enabled the production of state-of-the-art models that can outcompete
the current gold-standard tools for peptide-HLA-II interaction prediction.
Furthermore, to increase the efficiency of deep learning model construction and
development, the DeepImmuneTool (DIT) library which abstracts the process of
building, training, and optimizing deep learning models has been developed.
Concurrently to that, the Deep Antigen Screening (DAS) pipeline for large-scale
peptide-HLA-II interaction prediction have been constructed and tested. Future work
should aim at carrying more technical replicates to check for the reproducibility of the
microarray technology, collecting more data about other HLA-II alleles and finally,
fine-tune the developed software infrastructure to increase its efficiency and enhance
its accessibility by other researchers.