Advancing Antibody Discovery through Deep Learning

Deep learning has significantly transformed numerous scientific fields, including computer vision, natural language processing, and, increasingly, the biological sciences. Its applications now extend to cellular image analysis, genome-wide association studies, and the acceleration of drug discovery workflows. While traditional therapeutic development has primarily focused on small molecules, recent advances have enabled the application of deep learning to biologics–particularly monoclonal antibodies (mAbs). These Y-shaped proteins, composed of two heavy (H) and two light (L) chains, achieve antigen recognition through their variable regions (VH and VL), which contain six hypervariable loops known as complementarity-determining regions (CDRs). Monoclonal antibodies have become essential tools in treating infectious diseases, cancer, and autoimmune conditions, especially in cases where small molecules are ineffective. To streamline antibody discovery, the reverse vaccinology 2.0 method has emerged as a powerful approach for this purpose. It enables the isolation of potent antibodies directly from immune repertoires of vaccinated or infected individuals, accelerating the development of new therapies and vaccines. However, reverse vaccinology 2.0 pipelines still face key limitations: they often lack information on the molecular target of isolated antibodies, fail to preserve native VH/VL pairing during sequencing, and rely heavily on rigid structural assumptions that may overlook critical aspects of antibody–antigen (Ab–Ag) dynamics. These limitations present major bottlenecks to rational antibody design and epitope-focused immunogen engineering. This thesis leverages deep learning to enhance in-vitro antibody discovery, using reverse vaccinology 2.0 as a guiding framework. By incorporating both sequence and structural data, it offers three complementary contributions to improve computational antibody modeling. First, we develop a curated dataset and a classification framework to predict VH/VL pairing compatibility, enabling the reconstruction of functional antibody sequences in cases where native chain pairing is not preserved. Second, we present Abtarget, a deep learning model that classifies whether an antibody targets a protein or non-protein antigen based solely on its variable region–offering early functional insights into repertoire datasets. Third, at the structural level, we explore antibody-antigen interactions through geometric fingerprinting and propose a flexibility-aware approach by incorporating predicted Local Distance Difference Test (pLDDT) score as proxies for conformational adaptability. This strategy improves paratope prediction and interface modeling, particularly for antibodies targeting highly variable epitopes. These contributions introduce new tools, datasets, and methodologies that connect sequence-driven prediction with structure-informed design. By integrating deep learning techniques to tackle challenges in in-vitro antibody development, this work advances scalable, flexible, and interpretable strategies for antibody discovery, enabling faster, more targeted, and cost-effective therapeutic development.