Plant protein behavior in food systems: Data-driven, molecular, and multiscale approaches

Abstract

Plant proteins play important roles in food systems, yet their functional behavior remains difficult to understand and predict because it is governed by interactions across multiple length and time scales. Their properties in food systems are not dictated by a single molecular feature or processing variable, but emerge from coupled effects across scales. These effects include contributions from amino acid sequence, protein structure, solvent environment, intermolecular assembly, and other system-specific factors. Conventional experimental approaches are essential for characterizing food protein functionality, but they often provide indirect or ensemble-averaged measurements of systems that are structurally heterogeneous and source-specific. This challenge has become increasingly significant as food protein research moves from descriptive characterization toward more predictive and mechanism-oriented approaches. Data-driven modeling, bioinformatics, and structure-based computation offer new opportunities to identify patterns that are difficult to extract from experiments alone. The work presented in this dissertation takes this opportunity as its starting point. Chapter 1 first established the methodological basis for a simulation-based approach by examining the principles, capabilities, and limitations of all-atom and coarse-grained molecular dynamics simulations in food protein research. This chapter emphasized that molecular simulation is most useful when its resolution, force-field assumptions, and sampling limitations are matched to the question being asked. Chapter 2 then extended this framework toward a data-driven prediction approach that integrated global sequence- and structure-level descriptors with residue-level graph representations. Although the model was trained on microbial protein datasets, its application to seed storage proteins suggested that solubility-related sequence and structural patterns could be partially transferred to food and agricultural proteins. In particular, the seed storage protein analysis indicated that predicted solubility was strongly associated with surface electrostatic properties. Chapter 3 applied this broader computational perspective to Osborne protein classification, a classical framework in cereal and seed protein chemistry. By combining physicochemical descriptors, interpretable AI, molecular-level analysis, and simulations, this chapter examined whether empirical protein fractions could be understood at the monomeric scale. The results suggested that albumins and prolamins show relatively distinct monomer-level characteristics consistent with their solvent preferences, whereas the distinction between globulins and glutelins requires consideration of higher-order organization. Chapter 4 provided a more focused experimental and computational case study using zein–lutein encapsulation. In this system, molecular dynamics simulations, dilute-solution measurements, suspension-level characterization, and powder thermal stability tests were used to examine how ethanol-dependent protein behavior translated into encapsulation performance. The findings showed that although monomer-level simulations and dilute-solution measurements provided important insight into ethanol-dependent zein–lutein interactions, these lower-level observations did not directly translate to functional outcomes at higher structural scales, such as lutein retention in suspension or thermal stability in the powder. This finding suggests that molecular-level prediction alone may be insufficient for food-scale applications of plant proteins, particularly when functionality depends on higher-order assembly and processing kinetics. Chapter 5 further extended the findings of Chapter 4 by applying the zein–lutein encapsulation system to corn gluten meal-derived zein fractions. The feasibility of using the ethanol-soluble CGM extract directly for powder preparation was evaluated, and the effects of solvent-reduction pathway on powder thermal stability were compared with those observed for commercial zein. The resulting powders were also incorporated into cookies to assess whether their protective functionality was retained during baking. Beyond the specific findings of individual chapters, a recurring perspective of this dissertation is transferability: how information obtained from one scale, protein source, model system, or measurement can be extended to another context in food and agricultural protein research. Across the chapters discussed, data-driven approaches based on sequence- or structure-level information revealed useful patterns, but their transferability depended on protein source and structural organization. Similarly, molecular simulations provided useful mechanistic insight, but their relevance depended on whether the modeled scale corresponded to the scale of the functional context. Together, these chapters clarify that computational approaches can be valuable in food and agricultural protein research when they are applied to questions and contexts for which their predictions remain meaningful. Overall, this dissertation shows that computational approaches can provide an additional layer of analysis in plant protein research, enabling complex protein behavior to be examined beyond what can be obtained from empirical characterization alone. The findings further demonstrate the potential to extend such approaches to agricultural coproducts and practical food systems.