Enzyme Directed Evolution: Can AI End the Era of Trial and Error?

Enzymes are core catalytic tools in biotechnology, widely applied in industrial production, biopharmaceuticals, environmental remediation, and many other fields. However, natural enzymes are products of evolution, adapted to the mild physiological conditions within living organisms rather than the extreme conditions of industrial applications—they commonly suffer from poor thermostability, weak tolerance to extreme pH, low catalytic activity toward non-natural substrates, and insufficient specificity. How can we overcome these inherent defects of natural enzymes and endow them with novel catalytic functions to meet industrial demands? Enzyme directed evolution provides an efficient and well-established solution to this core challenge.

I. What Is Enzyme Directed Evolution?

Enzyme directed evolution is a protein engineering strategy that mimics the principles of natural selection in the laboratory. Its core logic involves introducing artificial mutations and recombination into the target enzyme gene, constructing a large-scale mutant library, and then employing high-throughput screening to identify enzyme variants with significantly improved catalytic activity, stability, substrate specificity, and other properties from the vast pool of mutants.

Unlike traditional rational design based on protein structure analysis, the greatest advantage of directed evolution is that it does not require prior knowledge of the enzyme's three-dimensional structure or catalytic mechanism. Researchers need only establish an iterative closed loop of "mutagenesis-screening," applying artificially imposed selective pressure to drive enzyme molecules toward the desired functional goals. This technology can accomplish in the laboratory what would take millions of years in nature.

From the broader perspective of protein engineering, directed evolution and rational design are two complementary strategies: directed evolution relies on high-throughput experimental screening and belongs to the "trial-and-error optimization" category, while rational design relies on protein structure and catalytic mechanisms for precise engineering and belongs to the "precision engineering" category. Artificial intelligence is now breaking down the barriers between these two approaches, enabling the integration of their respective advantages.

II. The Two Major Bottlenecks of Traditional Directed Evolution

The Two Bottlenecks of Directed Evolution

Despite its many successful industrial applications, traditional directed evolution still faces two fundamental bottlenecks:

Limited screening throughput and susceptibility to local optima. The protein sequence space is immense—for a protein of length N, there are 20ᴺ possible sequences, yet only a minute fraction possess the desired function. Traditional methods often employ a "mutagenesis and screening" local search paradigm, which, in fitness landscapes with strong epistatic effects, can easily become trapped in local optima. Starting from a natural enzyme and accumulating beneficial mutations round by round, this approach can only explore a limited local sequence space. Epistatic interactions between mutations can interfere with each other, causing the evolutionary process to stagnate at local optima and fail to reach the global optimum.

Long experimental iteration cycles and high costs. Traditional directed evolution relies on random mutagenesis and manual screening. Each round of the "mutagenesis-screening" cycle takes weeks to months, and complete projects can span years, severely limiting the development speed of novel enzyme preparations.

III. From Trial and Error to Precision

The deep integration of AI and protein engineering is reshaping directed evolution from three key dimensions.

3.1 Active Learning: Streamlining Experimental Rounds

AI-Powered Active Learning Workflow

Active Learning-assisted Directed Evolution (ALDE) is an iterative machine learning workflow: models perform virtual predictions on massive mutant libraries, selecting only high-potential mutants for wet-lab experiments, then feed experimental data back to refine models, forming a "prediction-validation-iteration" closed loop. A study in Nature Communications demonstrated that targeting five key residues with significant epistatic effects in the enzyme active site, just three rounds of wet-lab experiments improved the yield of a non-native cyclopropanation reaction from 12% to 93%, transforming directed evolution from "broad-spectrum screening" to "precision targeting."

3.2 Protein Language Models: Data-Driven Mutation Design

The EVOLVEpro framework published in Science integrates protein language models with active learning layers. In low-N regimes, it requires only minimal initial data to initiate multi-objective optimization, simultaneously optimizing enzyme activity, stability, specificity, and other properties—offering a new pathway for enzyme directed evolution in small-sample scenarios.

3.3 AI Redesigns the Evolutionary Starting Point: Breaking Through Natural Enzyme Limitations

AI Redesigns the Evolutionary Starting Point

A 2026 Nature study proposed a new approach: researchers used the AI model ProteinMPNN to redesign the full sequence of botulinum neurotoxin serotype E (BoNT/E)—with PROSS used for cross-validation—generating variants with enhanced stability while retaining catalytic efficiency as the evolutionary starting point, followed by phage-assisted continuous evolution (PACE). In head-to-head comparisons targeting ataxin-2, proteases evolved from the AI-redesigned starting point achieved more than 79-fold greater selected specificity for the target compared to the best-performing variant evolved from the wild-type BoNT/E starting point (Source: Nature, 2026).

The core conclusion: the starting point of directed evolution determines the ceiling of the final outcome. Natural enzymes are not necessarily the optimal starting points; AI can construct "superior evolutionary starting points" that outperform nature, breaking through performance bottlenecks from the very beginning.

IV. AI-Powered Enzyme Mining: Building the Resource Arsenal for Directed Evolution

AI Mining and Evolution Closed Loop

Enzyme mining serves as the source of "discovering novel enzymes" that provide the feedstock for directed evolution. Traditional mining relies on sequence homology alignment and is highly inefficient.

MatwingsVenus™ (Xiaowu™), developed by Matwings Technology, has integrated two core capabilities: AI-powered enzyme mining and AI-driven directed evolution. According to information from the Matwings Technology official website, the platform supports retrieval from billions of labeled protein sequences, integrates over 200 protein design tools and more than 30 expert-optimized Skills. Users can simply input tasks in natural language, and the system automatically completes the entire workflow of enzyme mining, directed evolution, de novo design, and automated wet-lab coordination. In enzyme mining scenarios, the platform can predict protein function directly from sequence, breaking through the traditional "sequence-structure-function" prediction paradigm to achieve closed-loop design from functional requirements to primary sequences, enabling high-throughput discovery of high-performance enzymes with extreme thermostability, pH tolerance, and other properties.

AI enzyme mining and AI-directed evolution form a complete closed loop: AI mining identifies high-quality candidates from vast sequence databases, solving the problem of limited starting substrates; AI-directed evolution precisely iterates and optimizes candidates, solving the problem of insufficient native enzyme performance—dramatically shortening the R&D pipeline from "novel enzyme discovery to performance engineering to industrial adaptation."

V. Advancing Protein Engineering into a New Intelligent Era

AI protein design, enzyme directed evolution, and traditional protein engineering are converging into a deeply integrated framework. AI redesign can both create entirely novel protein sequences and provide optimal starting points for directed evolution, enabling the integrated "design + iteration" paradigm.

Taking MatwingsVenus™ (Xiaowu™) as an example, the platform has further built a "dry-wet closed loop" R&D model upon its AI enzyme mining and directed evolution capabilities. The platform employs AI models to drive high-throughput design of protein sequences and structures, followed by precise validation via an automated wet-lab platform, forming an efficient "design-validate-feedback-upgrade" cycle that compresses traditional protein R&D cycles from years to months. Users propose functional requirements in natural language; after AI completes sequence design, the system automatically interfaces with the automated laboratory, driving robots to perform critical steps including sample preparation, protein purification, and functional assays. Experimental results are then iteratively refined through multiple rounds of AI optimization, achieving full-process "conversational dry-wet closed loop" intelligent R&D. This model frees researchers from relying on extensive trial-and-error experimentation, allowing the vast majority of mutant screening and performance prediction to be completed in virtual space, with only high-potential candidates requiring laboratory validation—achieving an order-of-magnitude improvement in R&D efficiency.

VI. Outlook

Enzyme directed evolution stands at a pivotal turning point. Traditional approaches rely on random mutagenesis and manual screening, suffering from long cycles, high costs, and limited performance ceilings; AI-powered capabilities make the evolutionary process precise, efficient, and controllable. From active learning streamlining experimental rounds, to protein language models guiding mutation design, to AI redesign creating superior starting points—AI has fundamentally restructured directed evolution from the core.

AI-driven platforms represented by MatwingsVenus™ (Xiaowu™) are deeply integrating intelligent enzyme mining with AI-directed evolution, continuously expanding the resource pool; meanwhile, AI protein design keeps raising the starting point and performance ceiling of enzyme engineering. These three elements complement each other, forming a continuous iteration closed loop of "AI mining discovers candidates → AI directed evolution optimizes performance → AI protein design upgrades starting points."

The essence of this transformation is upgrading protein engineering from labor-intensive trial-and-error R&D to compute-intensive intelligent R&D. As AI prediction accuracy and design capabilities continue to advance, the efficiency, success rate, and performance ceiling of enzyme directed evolution will undergo a qualitative leap. This is not merely an iteration of a single technology, but a fundamental paradigm shift of the entire field from experience-based to intelligence-based approaches—providing stronger technical support for industrial catalysis, biopharmaceuticals, green manufacturing, and other industrial applications.