Performing Department
Biological and Agricultural Engineering, Food Science & Tech
Non Technical Summary
Plant-based proteins have emerged as a promising and sustainable alternative source of dietary protein for diverse and novel food products, and are viewed favorably by many consumers. However, despite significant progress in developing liquid and semi-solid foods with plant-proteins (e.g. nut milk or yogurt alternative products), these foods suffer from limited dispersibility of proteins after pasteurization, resulting in sedimentation and an unpleasant mouthfeel when consumed (such as a gritty or chalky texture). In addition, although many consumers view these plant-based milk and yogurt alternatives as healthier, it is likely that their protein digestibility is reduced, which could instead negatively impact consumer health. As a result of these serious limitations in development of dairy alternatives with plant proteins, this project aims to characterize the functionality of plant proteins, and to develop novel protein microstructures that may help to ultimately increase protein functionality. Protein functionality will be assessed in model liquid (a milk alternative or protein shake) and model semi-solid (a yogurt alternative or dip) foods. Functionality will be considered as a combination of stability (how well the proteins remain dispersed in the product after heating), texture (the flow properties and mouthfeel of the product), and digestibility (how much of the protein may ultimately be absorbed by the consumer to benefit health). To allow for translation of the data collected in this project to the US food manufacturing industry, a machine learning model will be developed that will allow for food producers to input the protein concentration and properties into the model and it will predict the level of stability, texture, or digestibility for a given protein and food system. This tool will streamline product development of new foods with plant proteins and allow the food manufacturing industry to deliver higher-quality products to consumers.
Animal Health Component
65%
Research Effort Categories
Basic
35%
Applied
65%
Developmental
(N/A)
Goals / Objectives
The proposed research brings together an interdisciplinary and comprehensive approach that addresses the broader goals of: (a) extracting, fractionating, and characterizing protein fractions from diverse plant sources (pea, almond, oat, and algae); (b) formulation engineering of sub-micron protein structures (e.g., microgels and nanoparticles) to improve the functionality of the protein fractions, such as dispersibility and stability with thermal processing; (c) translation of these protein fractions to develop model liquid beverages and semi-solid foods and characterization of their stability, texture, and digestibility; and (d) development of a machine learning predictive framework to reduce empirical testing for future food product development using plant-based proteins.Ultimately,the overall goal of this project is to provide this information to the US food manufacturing industry to streamline development of novel foods using plant proteins that will have improved quality. This overall goal will be accomplished through the followingspecific objectives of the research plan: 1. Evaluate the role of physico-chemical properties of plant protein fractions (concentrates, isolates, soluble proteins) in their stability in aqueous solutions and as ingredients to develop sub-micron scale assemblies based on physical structuring mechanisms to enhance stability in aqueous solutions.2. Characterize the functionality (stability, texture, and digestibility) of protein fractions and novel protein structures in liquid and semi-solid model food systems.3. Develop a machine learning predictive framework to aid in the development and manufacturing of novel plant-based foods with optimal functionality.
Project Methods
Aim 1: Four plant protein sources will be utilized in this project: pulse (pea), almond, oat and microalgae (Chlorella vulgaris). Pea, oat, and almond flours and Chlorella vulgaris will be obtained from commercial suppliers. Microalgae samples will be freeze-dried, milled, and sieved. Aqueous extraction of plant and algae flours will be carried out using previously developed and optimized extraction methods to yield three protein fractions from each protein source: a protein concentrate; a protein isolate; and a soluble protein fraction. The isolated plant protein fractions will be analyzed for amino acid composition, proximate analysis, and molecular weight distribution of protein fractions using sodium dodecyl sulphate-polyacrylamide gel electrophoresis (SDS-PAGE).Protein stability of each protein fraction will be assessed in an aqueous dispersion as a function of protein concentration (6 levels between 0.5 - 10%) and pH (5 levels between 3 - 7). The stability will be assessed through a combination of measurements: protein dispersibility index, heat coagulation time, and particle size and surface charge.Based on the protein stability measurements, those protein fractions with the highest stability will be utilized to develop microgels, while those protein fractions with lower stability will be utilized to develop nanoparticle dispersions. After the plant protein microgels and nanoparticle dispersions are developed, the stability of these sub-micron scale assemblies will be determined as a function of protein concentration and pH using the same methods as for the protein fractions.Aim 2: Liquid foods will be developed that will serve as a model for a milk alternative or a protein-enhanced beverage/protein shake, depending on the protein content. Those products with 1-4% protein will target a milk alternative product, and those products with 5-15% protein will target a protein shake, as these ranges are similar to currently available products. All liquid model foods will have 2% (w:v) glucose and will be produced at a constant pH (between 6.5 - 8.5). The protein content (3 levels each for milk alternatives and protein shakes between 1 - 15%), and fat (0, 2, or 4%) will be varied to assess the functionality of systems with similar properties as the currently available plant-based milk alternatives, but with higher protein content. The protein fractions or the concentrated protein microgels or nanoparticles will be mechanically dispersed in water at the appropriate concentration for production of the model food systems. The fat content will be added as soybean oil at either 0, 2, or 4%.Semi-solid foods will be developed that will serve as a model for a yogurt alternative or similar product (e.g. dip, sour cream, etc.) produced with the protein fractions and sub-micron assemblies from the plant proteins in Aim 1. All semi-solid model foods will contain 2% (w:v) glucose and 1% pectin (w:v), suspended in deionized water. There will be a minimum total solid content of 10% with a pH of 3.9 - 4.5. The protein (3 levels between 3 - 15%) and fat content (0, 2, or 4%) will be varied to assess the maximum protein content that is functional in the semi-solid model food, and to assess if the addition of the novel protein structures may act as a fat-replacer. To avoid additional variability in model food systems due to fermentation, and to increase the feasibility of testing a large number of experimental conditions, the yogurt alternative products will not be produced using active cultures, but the pH (3.9 - 4.5) will be adjusted with L-(+)-lactic acid or another organic acid to the relevant range to assess the protein functionality. The fat content will be tested at 0, 2, or 4% and will be added as soybean oil.The model liquid and semi-solid food products and control liquid and semi-solid food products developed with whey protein isolate (as a gold standard for a high functionality protein) will be characterized before and after thermal pasteurization for stability, using the methods described for the protein fractions dispersed in water in Aim 1. Product texture will be assessed through a combination of four approaches to determine textural aspects related to the product quality and mouthfeel: 1) measurement of rheological properties (flow and oscillatory behavior); 2) tribology; 3) particle size distribution; and 4) microstructural analysis using confocal scanning laser microscopy. Product digestibility will be assessed through measurement of protein digestion during and in vitro biochemical gastrointestinal digestion. Digestibility will be measured as the degree of protein hydrolysis (determined from free amino group measurement), soluble amino acid release at the end of in vitro gastrointestinal digestion, as well as through band intensity measurement after SDS-PAGE.Aim 3: Machine learning models will be developed for classification of functionality of both plant protein fractions and sub-micron scale protein structures. The inputs of the machine learning model will be compositional and physico-chemical properties of the plant protein fractions and of the dispersion/model food system that will be characterized in this project. The model output for each protein fraction or sub-micron structure:food system combination will consist of three values. These values will represent the classification of Poor, Low, Medium, or High in each of the three categories of plant protein functionality: stability, texture, and digestibility. Multiple machine learning algorithms will be evaluated to discover the optimal model for predicting protein functionality based on the input data. The selected algorithms will include logistic regression, decision trees, random forest, support vector machine, gradient boosting, and neural networks. In addition, ensemble learning approaches will be explored, in which each prediction is produced by a collection of multiple trained algorithms.Feature selection techniques will be utilized to improve the model learning. Established feature selection methods such as genetic algorithms and greedy forward selection will be evaluated for selecting features from the properties measured in Aims 1 and 2. The role of individual selected features in explaining the model will be assessed based on increase in error after permuting a feature while the model remains untouched.To evaluate the machine learning models, we will randomly select 20% of the data generated in Aims 1 and 2 to be held out for model validation, and the remaining 80% of the data will be used for model training and testing. Model training and testing will be done through 5-fold cross-validation, where the data will be split into 5 folds. For each fold, the model will be trained on training datasets and evaluated on test datasets. The default loss function (multi log loss) will be used to construct the objective function. Predictions will be evaluated based on the receiver operating characteristic curve and confusion matrix and will be aggregated for all 5 folds to give the final values. The model will be validated separately on the 20% of the un-seen data that was held out from the model training and testing.