A Machine Learning Approach to Predict Functional Performance From Measurable Protein Structural Characteristics: A...

Area of Science:

• Food Science
• Biochemistry
• Computational Biology

Background:

• The food industry increasingly uses specialized protein ingredients for gelling, thickening, and emulsifying. Protein structure significantly influences these functional properties, but the relationship is complex.
• Understanding the link between protein structure and function is crucial for developing novel food ingredients. Traditional methods can be time-consuming and may not capture the intricate interactions.
• Plant-based proteins are gaining traction as sustainable alternatives, necessitating efficient methods for characterizing their functionality.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for predicting key functional properties of plant proteins.
• To identify the most influential structural predictors for protein solubility, emulsifying activity, emulsifying capacity, and gel strength.
• To assess the performance of ML algorithms in modeling the structure-function relationship of plant proteins.

Main Methods:

• Utilized various ML algorithms, including Gaussian-based Support Vector Regression, to predict protein functionality.
• Employed structural predictors such as surface hydrophobicity, zeta potential, undenatured protein content, water holding capacity, soluble protein polymer content, and β-sheet content.
• Assessed model performance using R², Mean Absolute Error (MAE), and Root Mean Squared Error (RMSE), ensuring physical constraint non-violation.

Main Results:

• The Gaussian-based Support Vector Regression model demonstrated high accuracy in predicting solubility (R²=0.8906), emulsifying activity index (R²=0.7383), emulsifying capacity (R²=0.7978), and gel strength (R²=0.8822).
• Specific structural features were identified as key predictors for different functional properties: surface hydrophobicity, zeta potential, and undenatured protein content for solubility and emulsifying activity; surface hydrophobicity, solubility, and undenatured protein content for emulsifying capacity; and solubility, undenatured protein content, water holding capacity, soluble protein polymer content, and β-sheet content for gel strength.
• The study confirmed the potential of ML for predicting plant protein functionality from a limited set of macromolecular structural characteristics.

Conclusions:

• Machine learning algorithms offer a powerful and efficient approach to predict plant protein functionality.
• Predictive models based on structural characteristics can significantly aid in the selection and application of protein ingredients in the food industry.
• These ML tools can streamline ingredient development and optimize food product formulations by accurately forecasting protein behavior.