From Blind Trial and Error to Precise Prediction: How Mutation Forecasting is Rewriting the Rules of Protein Design

Imagine you have an extremely sophisticated machine with hundreds of buttons and knobs. You want it to run better—faster, more energy-efficient, more durable. But you don't know the specific function of each button, so you can only press or turn them at random and then observe the results. Most of the time, the machine will break down; occasionally, a small adjustment slightly improves performance. In this way, you try thousands of times and only then find a satisfactory combination.

A complex machine

Does this sound absurd? But this is exactly the everyday experience that scientists have gone through over the past few decades when modifying proteins—especially enzymes. And the key to changing all of this is a technology called 'mutation prediction.'

What is mutation prediction?

Proteins are long-chain molecules made up of 20 amino acids connected in a specific sequence. A so-called 'mutation' is a change of one or several amino acids in the gene sequence of a protein. Replacing a single amino acid can completely deactivate an originally stable enzyme, drastically increase its catalytic efficiency by dozens of times, or even give it entirely new functions.

Mutation prediction involves using computational models to 'simulate' on a computer what effect a particular mutation will have on the protein—will it become more stable or more fragile? Will its activity increase or decrease? Has its selectivity for different substrates changed? With such predictions, researchers do not have to blindly create tens of thousands of mutants in the lab and test them one by one. Instead, they can select only a few that are most likely to succeed for experimental verification.

Amino acid mutation

From "Blind Man Touching the Elephant" to "Following the Map"

Before AI intervened, mutation prediction mainly relied on two types of methods. The first type is physicochemical models, which attempt to calculate minute changes in protein structure after mutations at the atomic level. This method is fairly accurate but requires enormous computation, only targeting small proteins with dozens of amino acids, which is almost impossible for common enzymes with hundreds or thousands of amino acids.

The second type is statistical methods, which learn patterns by analyzing large amounts of known mutation data. For example, if a charged amino acid appears at a certain position and is more likely to cause protein inactivation, the model will give such mutations a low score. But this approach heavily relies on "seen" similar data, and its predictive power is very limited for protein families that have never been studied.

It's like trying to predict a stranger's personality: the first method is to try to analyze every brain cell — theoretically feasible but not realistic; The second method is to refer to data from everyone you know—but if this stranger comes from a cultural background you know nothing about, the prediction will be inaccurate.

How does AI make mutation predictions reliable?

The real turning point came after the emergence of deep learning and large-scale protein language models. The basic idea of protein language models is: treat the amino acid sequences of proteins as a kind of "language," with each amino acid equivalent to a "word," and the evolutionary history of the entire protein family is a vast "text." By performing self-supervised learning on massive natural protein sequences, the model can automatically capture deep patterns such as "which amino acids cannot be easily changed," "which amino acids frequently appear together," and "which mutations are tolerated in evolution."

Based on such models, the accuracy of mutation prediction has greatly improved. Between 2024 and 2025, multiple independent studies have shown that state-of-the-art protein language models (such as ESM-2, ProGen2, etc.) can achieve correlation coefficients of 0.6–0.8 with experimental results when predicting the impact of single-point mutations on protein stability or function, whereas traditional methods often only reach 0.3–0.4. More importantly, these models do not require known structural information of the target protein and can make quite reliable judgments directly from the sequence.

From Single-Point Predictions to Combinatorial Mutations: Bridging the Gap of the 'Epistatic Effect'

However, the predictive capability for single-point mutations cannot be directly extended to combinatorial mutations. When two or more mutations are introduced simultaneously, they often produce 'epistatic effects' — meaning that the combined effect is not simply the sum of their individual effects. Sometimes, two harmful mutations combined can actually restore function, while two beneficial mutations can cancel each other out. This makes predicting combinatorial mutations much more difficult than single-point mutations.

In practical protein engineering, it is usually necessary to introduce multiple mutations at once to achieve the desired effect. For example, to make an industrial enzyme remain active at 80°C, it may be necessary to optimize the stability of several regions simultaneously. But the search space introduced by combinatorial mutations grows exponentially: for a protein of 300 amino acids, if each position has 19 alternative amino acids considered (excluding the original), just for double mutations there are more than 300×300×19×19 ≈ 32.5 million possibilities, and triple mutations reach astronomical numbers. Traditional methods cannot exhaustively explore this.

AI-driven mutation prediction can address this challenge through various strategies. A common method is to first score each single-point mutation with a language model in a zero-shot manner, and then estimate the effects of combinatorial mutations through specific combination rules (such as pseudo-energy summation based on predicted structures). More advanced research directly fine-tunes pre-trained models using sequence-function datasets, allowing the model to automatically learn interactions between mutations and thus predict higher-order combinatorial mutation fitness end-to-end. The maturation of these technological approaches is enabling mutation prediction to move from theory to industrial application.

Combinatorial mutation

Industrial Application of Mutation Prediction

Nowadays, mutation prediction is no longer an exclusive tool of academic laboratories, but has begun to be widely applied in fields such as industrial enzymes, antibody drugs, and biocatalysis. For example, in the pharmaceutical industry, predicting 'immunogenic mutations' in antibodies can help reduce adverse reactions of drugs in the human body; in the food industry, predicting thermal stability mutations in glycosidases can lead to the development of high-temperature-resistant processing enzymes; in the field of green chemistry, predicting enantioselective mutations can produce biocatalysts with high stereoselectivity.

Applications of mutation prediction

These cutting-edge mutation prediction models are moving from academic research to industrial applications. For example, Matwings Technology has launched the conversational protein R&D intelligent agent MatwingsVenus™ (Xiaowu™), which combines the aforementioned language model-driven mutation prediction capabilities with natural language interaction and automated wet lab scheduling, providing researchers with a closed-loop platform from 'asking an idea' to 'getting results.' Researchers only need to describe their goals in everyday language—for instance, 'I want to improve the half-life of this enzyme at 60°C'—and the agent can invoke the underlying mutation prediction engine, recommend several high-potential mutation site combinations, and automatically plan subsequent experimental validation paths.

From Prediction to Evolution

The maturation of mutation prediction technology is quietly changing the fundamental model of protein engineering. In the past, we could only slowly optimize an enzyme through repeated random mutations and screening; today, we can quickly evaluate thousands of mutation combinations on a computer and only bring the most reliable predictions to the lab for validation. This means that the time and cost required to evolve a natural enzyme to the performance level needed for industrial applications are drastically decreasing.

When prediction accuracy is high enough and the search of combination space is smart enough, protein 'directed evolution' will no longer rely on luck and extensive repetitive work, but will become a discipline that is predictable and iterative, much like engineering design.