Machine learning platform for designing acid-tolerant, thermostable sweet proteins

Overview

This solution is a machine learning based screening and design platform that generates acid-tolerant and thermostable variants of sweet proteins such as brazzein and monellin. The platform addresses a key limitation of natural sweet proteins: their tendency to lose sweetness and structural integrity under acidic and high-temperature processing conditions commonly encountered in food and beverage manufacturing. By leveraging protein language models and generative design, the technology enables the rational discovery of protein variants that retain sweetness while withstanding harsh formulation environments.

The offering is relevant to food and beverage companies, natural sweetener developers, flavor and ingredient suppliers, and nutraceutical manufacturers seeking stable, plant-based, low-calorie sweetening alternatives to sugar and artificial sweeteners.

Technical specifications

Core capabilities:

• Protein language model integration using ESM2 embeddings to represent protein sequences and predict functional properties
• Generative modeling pipeline combining Markov Chain Monte Carlo and protein diffusion based methods to produce novel sequence variants
• Comprehensive training database integrating FireProtDB, ProThermDB, and ProDataTherm, providing millions of data points covering thermostability and pH stability
• Adaptation of validated algorithms (Enz-Pred and CPI-Pred) previously benchmarked for predicting enzyme solubility, activity, binding constants, and fluorescence
• Validation workflow using REVOLVER for experimental confirmation of top candidate sequences
• Target proteins include brazzein and monellin, with extension potential to other sweet proteins

Key features:

• Predictive screening of protein variants before costly experimental testing
• Data-augmented approach combining public databases with literature-derived low-pH stability data
• Generative design that produces candidates optimized for both acid tolerance and heat stability while preserving sweetness function

Technology readiness level

The underlying machine learning methodology has been validated through prior benchmarking studies demonstrating superior performance over state-of-the-art methods for predicting enzyme properties. The foundational database (KinMod) and algorithms (Enz-Pred, CPI-Pred) have been published and validated on independent datasets. The current stage involves extending these validated approaches to sweet protein design, with planned steps including database assembly, model training, generative variant generation, and experimental validation of the top 5-10 candidate sequences. The technology is at an early-to-mid research and development stage, with computational pipeline development underway and experimental validation as the next milestone.