| dc.description.abstract | Antibodies (Abs) are proteins of the adaptive immune response that bind to and
neutralize body-foreign particles (antigens). Specific binding to a wide variety of antigens is
achieved by the tremendous variability inherent to the variable region. Engineered Abs that
specifically bind to clinically relevant targets are administered as vaccines and therapeutics.
Antibody vaccines and therapeutics need to meet special requirements important for their
development, and must be subsequently approved by the Food and Drug Administration (FDA).
These criteria include their expressability under experimental conditions and low risk for
eliciting adverse effects. Technologies developed for this dissertation make use of the observed
antibody space (OAS), which contains antibody sequences obtained from healthy, human blood
donors. Recent advances in Next Generation Sequencing (NGS) dramatically increased the OAS
up to hundreds of million sequences per healthy human blood donor. Novel computational
approaches were developed to process large NGS datasets and to assess their human-likeness
using the OAS. High human-likeness is generally associated with a low risk of an immunogenic
response. It could be demonstrated that our technology is able to differentiate between human,
non-human, and mixed human-like Ab sequences. In addition, our approach enables us to
generate human-like nucleotide sequences with a sequence recovery of up to 97.2%. To support
the development of Abs under experimental conditions, a Deep Learning approach was
developed to classify between Abs likely, and unlikely, to express in Chinese Hamster Ovary
cells, with an average AUC score of up to 0.71± 0.04. In combination with the Rosetta protein
modeling suite, the Ab expressability and human-likeness could be increased via structural re-
design of the Ab proteins. Our human-likeness modeling approach also increased the human-
likeness of the CDRH3 region in 8 out of 28 cases, which is inherently difficult to model due to
its high diversity. We further introduce a novel Rosetta design approach that is capable of
conserving functionally relevant residues in proteins without their explicit knowledge and
hypothesize, that this approach has the potential to functionally characterize antibodies for future
computational antibody discovery. | |
