AI Finds Sustainable Plant Protein Alternatives to Emulsifiers

Emulsifiers are the ingredients that keep oil and water mixed together and stable, essential to everyday products from mayonnaise and ice cream to cosmetics and pharmaceuticals. Many widely used emulsifiers today come from animal proteins, including milk-derived caseins and whey, and researchers at the University of Leeds have developed a new computational approach combining artificial intelligence and statistical physics to identify plant-based alternatives, an approach that has already flagged nearly 800 promising plant proteins for further investigation.

 

Why Finding New Emulsifiers Requires Solving a Massive Search Problem

 

For a protein to function as an effective emulsifier, it needs to physically attach itself at the boundary between oil and water and help stabilise that mixture, preventing the two substances from separating back apart over time. Millions of plant proteins exist across the natural world, and any one of them could theoretically possess this specific interfacial behaviour, but conventional laboratory testing can only evaluate proteins one at a time, an approach that is both expensive and slow when the underlying pool of candidates to search through is this large.

That combination, a genuinely enormous number of possible candidates paired with a slow, costly method for testing each one individually, is precisely the kind of problem computational screening approaches are well suited to addressing: rather than testing every possible protein directly in a laboratory, a computational model can rapidly narrow down which candidates are most likely to actually work, allowing researchers to focus their limited laboratory testing capacity on the smaller subset of proteins the model identifies as most promising.

 

Read more: i2Cool Raises Tens of Millions of Yuan for Electricity-Free Cooling Technology

 

How Combining Statistical Physics and Machine Learning Actually Works

 

The research team, led by Dr Simha Sridharan and Professor Anwesha Sarkar at the University of Leeds, alongside machine learning expert Dr Rik Sarkar at the University of Edinburgh, built their approach around two distinct computational techniques working together. First, they used a simulation model grounded in statistical physics, a branch of physics dealing with how large numbers of particles behave collectively, to model how proteins physically interact with the boundary, or interface, between oil and water. This simulation step essentially recreates, computationally, the physical process that determines whether a given protein would actually behave as an effective emulsifier in the real world.

Second, the team applied machine learning specifically to identify which particular sections and structural characteristics of a protein's overall composition most strongly influence that interfacial behaviour. Rather than treating each candidate protein as an unknown black box, this machine learning step allows the researchers to understand which specific structural features tend to correlate with successful emulsification behaviour, information that can then be used to scan across a much larger pool of proteins and flag those sharing similar promising characteristics, without needing to run the full physics-based simulation individually on every single candidate.

 

Why the Nearly 800 Identified Proteins Represent a Genuinely Novel Contribution

 

The model's headline output, nearly 800 plant proteins flagged as having potential emulsifier properties, is notable specifically because the research states many of these proteins "had not previously been considered for this purpose." That detail matters because it suggests this computational approach isn't simply confirming what food scientists already suspected about a small, already well-studied set of candidate plant proteins, but is instead surfacing genuinely new candidates that conventional research approaches, likely constrained by the same laboratory testing bottleneck this computational method is designed to address, had not previously identified or prioritised for investigation.

To validate these predictions against real-world results, the researchers tested several commercially available plant proteins the model had flagged, finding that pea and potato proteins in particular demonstrated effective emulsification properties consistent with what the computational model had predicted. That validation step matters considerably for assessing how much confidence to place in the model's broader output, since a computational prediction method is only genuinely useful if its predictions actually hold up when tested against real laboratory results, and this initial validation using already-known, commercially available proteins provides at least preliminary evidence the underlying approach produces genuinely accurate predictions rather than purely theoretical results disconnected from real protein behaviour.

 

Explore OneStop ESG Marketplace: Alternative protein

 

Why This Matters for the Broader Plant-Based and Sustainable Food Industry

 

Professor Anwesha Sarkar, NAPIC Co-Director at the University of Leeds, described the findings as demonstrating "how AI could help researchers identify promising new ingredients much faster than traditional experimental approaches." For companies developing plant-based food products specifically, this kind of computational screening approach could meaningfully compress the early-stage research timeline involved in finding new functional plant-derived ingredients, since a company or research team could potentially use similar computational methods to identify promising candidate proteins from a much broader pool before committing laboratory resources to testing them directly.

The research also illustrates a broader pattern relevant to sustainable food ingredient development generally: bringing together genuinely distinct fields of expertise, food science, protein chemistry, statistical physics and machine learning, to address a specific bottleneck (the sheer scale of possible candidate proteins) that no single field could efficiently solve alone. For NAPIC specifically, described in the release as involved in developing "the next generation of alternative protein technologies and ingredients," this work represents a template for how similarly interdisciplinary computational approaches might accelerate other areas of sustainable food ingredient discovery beyond emulsifiers specifically, wherever a similarly large search space of candidate molecules or materials exists.

 

Key Takeaways

 

Computational screening methods combining physics-based simulation with machine learning can dramatically narrow the search space for identifying functional plant proteins, reducing dependence on slow, expensive trial-and-error laboratory testing. The University of Leeds research identified nearly 800 candidate plant proteins for emulsifier applications, with initial laboratory validation of pea and potato proteins supporting the model's predictions. This interdisciplinary approach offers a potential template for accelerating discovery across other areas of sustainable food ingredient development facing similarly vast candidate pools.

 

Source: University of Leeds