How is food enzyme engineering done? A complete breakdown from AI mining enzymes and directed evolution to de novo design

Introduction

Enzymes are the "invisible workers" of the food industry—from starch hydrolysis and protein modification to juice clarification and lactose breakdown, enzyme preparations have long penetrated every aspect of food processing. However, traditional food enzyme engineering has long been plagued by bottlenecks such as low screening efficiency, high randomness in modifications, and difficulty in optimizing multiple targets simultaneously. In recent years, the involvement of AI and large protein models is driving this field from a "labor-intensive trial-and-error" approach to "design-driven precise optimization." This article systematically reviews the core pain points of food enzyme engineering, three technical paths for AI-driven modification, and practical implementation methods, providing a direct reference guide for researchers in enzyme engineering, synthetic biology, and industrial biotechnology.

1. Three Core Pain Points of Food Enzyme Engineering

The demand for enzymes in the food industry is often multidimensional: they need to be highly active and stable, specific to substrates, and low-cost with high expression. This "want it all" reality reflects the three fundamental challenges of traditional food enzyme engineering:

- Low efficiency in natural screening. Traditional microbial screening and genome mining can only cover a tiny fraction of annotated protein sequences, leaving a vast amount of unannotated potential high-quality enzyme resources untapped.

- High randomness in modification. Traditional rational site-directed mutagenesis often focuses on hotspot regions of the active center and tends to overlook the synergistic effects of distant allosteric mutations. Classic directed evolution relies on random mutation across the entire gene to build large libraries, requiring high screening throughput and cost. It is also difficult to predict the non-additive effects of multiple mutations, which often leads to the dilemma of 'single-point improvement but combined inactivation.'

- Difficulty in balancing multiple objectives. Activity, stability, specificity, expression levels… Food enzymes often need simultaneous optimization across multiple metrics. Manual experiments struggle to balance performance accurately, with development cycles often taking several years and experimental reproducibility being poor with highly fragmented data.

2. Three Technical Paths for AI-Driven Food Enzyme Modification

 Three Technical Approaches to the AI-Driven Modification of Food Enzymes

Addressing the three major pain points mentioned above—low efficiency in natural screening, high randomness in modification, and difficulty balancing multiple targets—AI food enzyme modification has developed three relatively mature technical pathways, each aimed at solving a core issue:

2.1 AI Enzyme Mining: Precisely Exploring the Treasure Trove of Natural Sequences

There are massive amounts of undiscovered enzyme resources hidden in nature's microorganisms, but traditional isolation and cultivation methods only cover less than 1% of them. AI enzyme mining uses protein language models to predict enzymes with specific catalytic functions directly from metagenomic sequences, greatly skipping the long process of cultivation screening.

This approach is especially suitable for the food enzyme field—food processing requires a wide variety of enzymes, often needing to function under specific conditions (acid-resistant, heat-resistant, high-sugar tolerant), and these properties may naturally exist in the genomes of microorganisms from extreme environments. AI enzyme mining can compress the "enzyme hunting" time from years to months.

2.2 AI-Directed Evolution: Precisely Guiding Mutation Design

For food enzymes with existing scaffolds but insufficient performance, AI-directed evolution is currently the most mature application direction. Its core is using AI models to predict the impact of mutations on enzyme function, only recommending the mutation combinations most likely to improve performance and reducing trial-and-error in experiments.

Compared to the traditional "wide net" approach of directed evolution, AI-directed evolution has three clear advantages: first, it’s more efficient—under scenarios supported by high-quality experimental data, usually only a few dozen to a few hundred mutant screenings are needed to find high-quality combinations; second, it can optimize multiple targets simultaneously, improving activity while also considering thermal stability and expression levels; third, it has a shorter cycle, with 2–3 rounds of iteration often achieving what dozens of rounds using traditional methods would.

2.3 De novo Design: Creating Completely New Catalysts from Scratch

The cutting-edge direction is AI de novo enzyme design—not relying on the overall scaffolds of existing natural enzymes, but generating entirely new protein scaffolds capable of carrying the target catalytic function based on a clear understanding of the catalytic mechanism and transition state models. Currently, the overall activity and stability of de novo designed enzymes still lag noticeably behind natural enzymes, and their application in the food industry is still in early laboratory exploration, showing potential only in a few high-value specialized enzymes.

3. How to Choose: Different Technical Approaches for Different Stages

Depending on the stage of R&D and the baseline conditions, different technical approaches apply:

- Early stage of the project, unclear target function → Prioritize using AI to mine enzymes and expand the candidate pool, finding scaffolds from natural diversity.

- Already have a scaffold, need performance improvement → AI-guided evolution is the most cost-effective choice, much more efficient than random mutation.

- Multi-objective optimization (activity, stability, expression level) → Hard to balance with traditional methods, AI models have clear advantages in multi-objective optimization.

- Can’t find the corresponding function in natural enzymes → Can try de novo design, but be prepared for the timeline and cost.

Research and Development Selection and Closed-Loop Wet and Dry Testing

4. MatwingsVenus™ (XiaoWu™): Full-Chain AI-Driven Food Enzyme R&D

Shanghai Tianwu Technology’s self-developed MatwingsVenus™ (XiaoWu™) is a smart protein R&D agent. Based on proprietary protein large models and automated experimental platforms, it offers a full-chain AI solution for food enzyme engineering, covering everything from enzyme mining and directed evolution to experimental verification.

AI Enzyme Mining: Quickly expand the candidate enzyme pool. The platform integrates a massive protein sequence database, covering both conventional and various extreme environments. Using a functional enzyme mining model, researchers can rapidly screen high-potential candidate enzymes from vast natural sequences based on target catalytic functions and property requirements (like optimal temperature, optimal pH, and substrate preference), skipping traditional large-scale separation and screening.

AI Directed Evolution: Multi-objective collaborative optimization. For food enzymes with an existing scaffold but needing performance improvement, the platform’s directed evolution module supports multi-objective optimization for activity, stability, substrate specificity, expression levels, etc., outputting candidate mutation combinations ranked by priority. Compared to traditional random mutations, AI-directed evolution can significantly reduce experimental rounds and screening workload, shortening the R&D cycle.

Dry-Wet Closed Loop: Iterative acceleration from design to verification. The platform connects dry experimental design with wet-lab verification. AI-designed mutants can directly go into the automated experimental platform for cloning, expression, purification, and enzymatic characterization. Once experimental data is fed back, the AI model uses it for the next round of optimized design, forming a "design–verify–iterate" closed-loop R&D model.

This approach has already been validated in various food-related enzyme systems, such as polysaccharide-degrading enzymes and glycosyltransferases. For example, with cyclodextrin enzymes, the platform performed multi-objective optimization via AI-directed evolution (enhancing transglycosylation activity, reducing hydrolysis activity, increasing regioselectivity). In just a single round of prediction, the optimal triple-point mutant combination was obtained, increasing the product ratio from 63% to 98% and improving the transglycosylation/hydrolysis ratio by 12 times. Production feasibility has been verified at a 100 L pilot scale. The results were published in *Bioresource Technology*.

5. Practical Recommendations for Research: Four Actionable Experiences

For teams currently conducting food enzyme engineering research, here are four actionable practical suggestions:

Step 1: Clearly outline goals and constraints first.

Food enzyme optimization is always multi-objective. The first step is to clearly define the optimization goals and constraints (how much activity increase, how high stability is required, what is the minimum expression level). The more specific the goals, the higher the hit rate for AI-driven designs.

Step 2: Use natural scaffolds whenever possible instead of designing from scratch.

If the target function already exists in nature, prioritize mining natural scaffolds and then perform directed evolution, which is usually more efficient than designing from scratch. AI for enzyme mining can greatly accelerate this process.

Step 3: Prefer AI methods for multi-objective optimization.

Traditional methods can handle single-objective optimization, but when it comes to coordinating three or more performance metrics, AI-driven directed evolution shows a significant efficiency advantage.

Step 4: Establish a rapid iteration feedback loop.

The probability of success in one round is low. The key is to establish a fast "design–validate–feedback" iteration mechanism so that each round of experimental data can inform the next round of design, creating a positive cycle.

6. Outlook

Food enzyme engineering is undergoing a paradigm shift from "trial-and-error-driven" to "design-driven." With AI, food enzyme R&D no longer fully relies on expert experience and large-scale screening; experiments can now be guided accurately through predictive models. According to industry reports, the global smart enzyme market is expected to reach $68.74 billion by 2026, with food and beverage processing being the largest application area. For enzyme engineering and synthetic biology researchers, this represents both a technical challenge and a historic opportunity—those who master AI tools will have the ability to develop higher-performance, lower-cost food enzymes in a shorter time, pushing the food industry toward more efficient and greener practices.