Stuck at a plateau with industrial enzyme screening and modification? Try AI-assisted design

Have you ever run into a situation like this—after one round of directed evolution, your enzyme activity goes up by 30%, the second round only gives you 5%, and by the third round it’s barely progressing. You know there’s still room for improvement, but you don’t know which site to tweak next. Keep doing error-prone PCR? The cost-to-benefit ratio keeps dropping. Switch to rational design? But you’re afraid the predictions might be off and you’d just be wasting experiments.

If you’re working on industrial enzyme screening and modification, this scenario probably feels all too familiar. Low efficiency of natural enzyme screening, the plateau in directed evolution, and high reliance on experience for rational design—these three “stubborn” problems are pitfalls almost every enzyme engineering research group hits.

1. The three major “bottleneck” problems in industrial enzyme screening and modification

Three Major Challenges in Industrial Enzyme Screening and Engineering

First, the efficiency of natural enzyme screening isn't very high. Although the natural enzyme library is huge, the enzymes that truly match specific substrates and reaction conditions are likely very rare. Building the library, screening it, and validating it—after a full round, the number of positive clones is limited, and candidates that meet performance standards are even scarcer. The root of the problem is that enzymes in nature aren't made for industrial settings—they can easily lose activity when the temperature rises, their activity changes with pH, and some enzymes even experience substrate or product inhibition at high substrate concentrations. Natural enzymes are often 'favorites in the lab, challengers in industry.'

Second, directed evolution often hits a plateau. Classic directed evolution methods like error-prone PCR and DNA shuffling gradually show less improvement in enzyme activity after multiple rounds. Researchers can usually feel that there's room for optimization but have trouble pinpointing the key sites. Continuing blind screening becomes less and less efficient, leaving projects stuck in a 'you can see the ceiling but can't find the way through' situation.

Third, rational design relies heavily on experience. Rational design based on structural biology and homology modeling usually requires deep expertise, and predicted results don't always match experimental verification—the mutation strategies reported in literature can have very different effects when applied to different enzyme systems or reaction conditions.

So, why do traditional methods often hit dead ends? The essence of these challenges is that the sequence space of enzymes is huge, while experimental throughput and intuition have limits. An enzyme with 300 amino acids can theoretically have 20³⁰⁰ sequence variants, far beyond what can be explored experimentally. Traditional trial-and-error modification strategies are usually inefficient when facing complex performance optimization goals.

3. How AI is Rewriting the Rules of Industrial Enzyme Screening and Engineering

In recent years, with the rapid development of protein structure prediction and AI design technologies, the technical path for industrial enzyme screening and engineering is undergoing profound changes.

Structure prediction tools like AlphaFold have greatly lowered the barrier to obtaining enzyme structures, allowing even enzymes that are difficult to crystallize to get fairly reliable 3D models. On this basis, a new generation of AI tools is providing even further solutions: predicting mutation effects based on large-scale protein language models, identifying functional sites with deep learning, and virtual screening combined with molecular simulations.

Recent research is validating the feasibility of this approach. In 2026, the CABBI center under the U.S. Department of Energy used an AI-driven autonomous enzyme engineering platform to achieve a 16-fold and 26-fold activity increase for two industrial enzymes within just four weeks. Another study published in Nature Communications demonstrated the potential of generative AI—researchers used the GenSLM protein language model to directly generate new variants of tryptophan synthase TrpB, some of which performed even better on non-natural substrates than versions obtained through lab-directed evolution. Another research team used AI to redesign botulinum neurotoxin protease as a starting point for evolution, and the final variants exhibited over 79 times higher substrate selectivity compared to the best variants derived from the wild type.

In short, industrial enzyme screening and engineering are evolving from an "experiment-driven" approach to a new paradigm of "computation-guided, experiment-validated"—first using AI calculations to drastically narrow the candidate pool, then experimentally testing the most promising dozens to hundreds of variants, concentrating experimental resources on the directions with higher success rates. This isn’t about replacing experiments, but about making them more targeted and efficient.

4. From "Finding a Needle in a Haystack" to "Precision-Guided": What Does a Practical Workflow Look Like?

Suppose you have an industrial enzyme and need to improve both thermal stability and catalytic efficiency. Traditional methods might involve: running a round of error-prone PCR, screening hundreds of clones, picking a few high-activity ones, doing a second round... repeating the cycle, with diminishing returns.

With AI platforms like MatwingsVenus™ (Xiaowu™), the workflow could look like this: you input your task in natural language—such as "improve the catalytic efficiency of a certain enzyme at high temperatures"—and the system automatically calls a protein language model to predict multi-site cooperative mutations, combining structural bioinformatics analyses to prioritize mutation sites. You then pick a few dozen candidate variants recommended by AI for experimental validation. The whole process goes from "finding a needle in a haystack" to "precision-guided."

MatwingsVenus™

MatwingsVenus™ (XiaoWu™) is a conversational AI for protein research developed independently by Matwings Technology. It has three core capabilities in industrial enzyme screening and engineering:

First, intelligent sequence mining — based on a protein database with billions of real labels and a sequence similarity network, it helps quickly screen for potential functional enzymes from the natural sequence space, narrowing down the experimental screening range.

Second, multi-site collaborative mutation design — it integrates 200 protein design tools and combines protein language models with structural bioinformatics to intelligently predict and prioritize mutation sites for multiple targets like enzyme activity, stability, and substrate specificity, providing reference to overcome the optimization bottleneck of single-point mutations.

Third, end-to-end evidence chain support — from structure modeling and molecular simulation to assessing the developability of mutants, each recommended site has traceable computational evidence, so you can judge, question, and make your own adjustments based on it.

This product just won the top honor of "Treasure of the Pavilion" at the 2026 World Artificial Intelligence Conference, being the only AI4S (Scientific Intelligence) product selected this year, sharing the stage with achievements from tech giants like Baidu and Alibaba.

The core value of MatwingsVenus™ (XiaoWu™) lies in the fact that it doesn’t just give you a black-box prediction. Instead, it integrates protein data retrieval at the billion-level scale, 200 protein design tools, structure modeling, molecular simulation, and literature evidence into a single workflow — making your design decisions traceable, freeing researchers from large-scale initial screenings, and letting them focus on more valuable validation and deep optimization.

5. The Next Five Years of Industrial Enzyme Screening and Modification

Looking ahead, the field of industrial enzyme screening and modification may show several clear directions:

First, AI-first approach. More and more research groups will integrate AI predictions into the early stages of experimental design, using computational methods to narrow down the search space before validating with experiments. "Calculate first, experiment later" will gradually become a common path in enzyme engineering research.

Second, multi-objective synergistic optimization. In industrial applications, enzymes are often not needed for just "the best single performance" but rather a balanced combination of enzyme activity, stability, expression level, substrate spectrum, and other factors. AI's potential in multi-objective synergistic optimization will be further unleashed.

Third, from individual enzymes to system-level design. Industrial enzyme screening and modification is no longer limited to optimizing a single enzyme; it gradually extends to metabolic pathways, enzyme combinations, and system-level design of host cells. AI agents are expected to play a larger role in this direction.

Final Thoughts

If you are engaged in industrial enzyme screening and modification research, ask yourself: where is the current project stuck? Is it that you can't find the right enzyme, or have you hit a platform stage after multiple rounds of modification?

When traditional methods reach their limits, AI-assisted design may offer a new entry point. Intelligent tools like MatwingsVenus™ (Xiaowu™) are making "calculate first, experiment later" the new normal in enzyme engineering research. On the road of enzyme engineering, choosing the right methods and tools often makes your work much more efficient.

What do you think about AI-assisted enzyme modification? Have you ever faced situations where you "couldn't screen" or "couldn't modify" enzymes? Feel free to share your experiences in the comments.