Abstract
This dissertation focuses on structure learning, intending to model the relationship between glycan structures and their spectroscopic data by integrating perspectives from computational chemistry and machine learning. It explores the application of structural learning methods, particularly Graph Neural Networks (GNNs), to address fundamental problems in glycoscience and further advances the theoretical foundations of GNNs. The research is organized around the theme of bidirectional structure–spectrum inference. On the forward side, an efficient method is proposed for reconstructing glycan topologies from tandem mass spectrometry (MS/MS) data. On the reverse side, a dataset and benchmark, along with adapted 3D GNNs, are proposed for predicting atomic-level nuclear magnetic resonance (NMR) chemical shifts from glycan structures. Beyond these applications, the dissertation extends GNN theory by formulating and analyzing influence functions for graph-structured data, enabling efficient post-hoc estimation of node and edge importance without retraining. Together, these contributions advance automated glycan analysis, enrich computational tools for glycoscience, and present generalizable methodologies for graph machine learning.