Learning Molecular Representations from Imperfect Experimental Data: Models, Datasets, and Benchmarking for Biochemical Systems
This dissertation develops machine learning methods for learning molecular and biomolecular representations directly from imperfect experimental data and for evaluating such models in a principled and reproducible manner. The first part focuses on representation learning for the Nuclear Magnetic Resonance (NMR) spectra, including a contrastive learning framework for isomer discrimination and molecular retrieval using 1D NMR technology, and solvent-aware Graph Neural Network (GNN) models for the prediction of 2D NMR chemical shifts for small- to medium-sized molecules. These models map local chemical environment to global spectral signals, enabling scalable alternatives to traditional computational or rule-based pipelines. The second part focuses on data curation and evaluation framework, including the curation of a large-scale 2D NMR dataset, 2DNMRGym, used as a node-level regression task to systematically evaluate the state-of-the-art models; as well as the development of a retrieval-oriented evaluation framework for antibody–antigen binding affinity prediction that incorporates pairwise comparisons and statistical tests reflective of real-world antibody discovery workflows. Together, these components form a coherent methodological contribution that advances representation learning, dataset curation, and systematic benchmarking for biochemical systems.