Exploring Enzyme–Ligand Interactions in Cofactor Metabolism Using Advanced AI Structure Prediction Methods

Proteins are fundamental biological macromolecules whose three-dimensional structures determine their molecular functions, interactions, and roles in cellular processes. Experimental methods such as X-ray crystallography, NMR spectroscopy, and cryogenic electron microscopy have yielded detailed structures for thousands of proteins, yet remain time-consuming, technically demanding, and resource-intensive. AI-based structure prediction has transformed the field by enabling accurate computational models directly from sequence data. Advances in deep learning, particularly AlphaFold2/3 and related systems, have greatly expanded computational structural biology. AlphaFold3 introduced a unified framework for modelling proteins, nucleic acids, small molecules, and their interactions. Open-source reproductions like Boltz and Protenix have further democratized access to state-of-the-art tools. This thesis benchmarked and applied three modern AI-based structure prediction methods to characterise selected enzymes in Nicotinamide adenine dinucleotide (NAD) cofactor metabolism and their interactions with biologically relevant ligands. NAD is an essential metabolic cofactor and signaling molecule across all life forms. Prior to the main workflow, multiple platforms were evaluated for functionality, computational demands, and suitability for protein–ligand complexes. Protenix was selected as the primary tool for its strong support of multi-chain assemblies, non-standard ligands, diffusion-based sampling, and comprehensive confidence metrics. Boltz-2 was used for targeted drug-discovery applications due to higher structural reliability and fewer ligand geometry issues, while AlphaFold3 served as a supplementary server-based option. Structural models were generated and assessed using confidence scores including pLDDT, pTM, and ipTM. The resulting structures were analysed for protein architecture, ligand-binding interactions, and overall plausibility of the predicted complexes. This enzyme-ligand prediction approach proved instrumental in two key biological contexts: (1) therapeutic coenzyme-boosting strategies (e.g., in aging and neurodegeneration) and (2) the evolutionary arms race between bacteria and phages. These findings highlight the utility of contemporary AI-driven structure prediction tools as accessible and powerful resources for structural characterisation and hypothesis generation in biomedical research, effectively complementing traditional experimental methods in studying protein function and molecular interactions.