Proteins Are the Executors of Life’s Functions — How AI is Rewriting the Code
Published on November 17, 2026

A feature report on Matwings Technology (MatwingsVenus) and its approach to AI protein engineering
Proteins are the executors of life’s functions. From the enzymes that catalyze metabolism, to the antibodies that defend against pathogens, to the cytoskeleton that supports cells, almost every moment of life’s activity is precisely driven by them.
To “read” and rewrite proteins is nearly equivalent to holding the key to controlling life’s functions — and AI is accelerating this process. Through deep learning, machines can discern the connections among protein sequences, structures, and functions from vast data, transforming R&D into an efficient compute-first-then-verify workflow, and even creating proteins that do not exist in nature according to human needs.
Protein engineering is regarded as a key field in which AI empowers the life sciences, and has become a battleground for leading AI companies. In September of this year, AI giant Anthropic launched a global protein design competition, planning wet-lab validation of 5,000 AI-designed proteins. But this competition was not open to Chinese teams.
After learning the news, a group of young scientists could not sit still. “Everyone was very angry,” said Hong Liang, Distinguished Professor at Shanghai Jiao Tong University and founder and chief scientist of Matwings Technology (MatwingsVenus). So they quickly assembled scientists from nine top universities, including Shanghai Jiao Tong University, Tsinghua University, Peking University, and the University of Hong Kong, as well as three national laboratories, and organized an AI protein design challenge named the “Changjiang River Cup,” open to global teams, offering up to 7,000 proteins of free wet-lab validation.
Although in the relatively short history of AI protein engineering the most prominent teams and products have mostly come from abroad, Chinese teams such as Matwings Technology are also continuously making innovations and breakthroughs across dimensions such as model capability, engineering, and commercialization, becoming a force that cannot be ignored in the process of AI empowering the life sciences.
From Focusing on Structure to Focusing on Function
Proteins are composed of about 20 amino acids. These amino acids link into diverse sequences, then fold into specific three-dimensional structures, ultimately becoming proteins with specific functions.
Both sequence and structure may contain important information about protein function. For a long time, scientists first determined a protein’s amino acid sequence, then resolved its structure using techniques such as X-ray crystallography, and finally mapped these features to function.
Even after accumulating considerable data, predicting structure from sequence — and further, predicting function — remains extremely difficult. In 2020, AlphaFold emerged, bringing the first paradigm revolution: it can predict structure from sequence with unprecedented accuracy, thereby shortening validation time.
“In the past we were always ‘hindsight,’ using computation to explain others’ experimental results; but AlphaFold can be ‘foresight,’ guiding experiments in advance and greatly reducing the cost of trial and error,” Hong Liang once said in an interview. At the time, AI protein research faced three paths: first, structure prediction represented by AlphaFold; second, computational de novo protein design. Hong Liang noted that de novo protein design was innovative but difficult to realize in practice. “After following it for half a year, I found that none of the proteins we made had function.”
Later, Hong Liang founded Matwings Technology, taking the third and more “down-to-earth” direction: protein engineering, which optimizes protein performance by adjusting existing protein sequences. To achieve this, predicting structure alone is not enough — the AI model must also accurately predict function, so that scientists know what adjustments need to be made to the sequence or structure.
“The features humans summarize are often low-dimensional, like judging what animal it is by only seeing two ears behind a bush,” Hong Liang said. In 2022, he and his team began using massive protein sequence data to train protein AI models. Unlike humans, the hundreds of millions of parameters in large models, though hard to explain, form a high-dimensional description of proteins.
If an AI model can predict structural differences based on sequence and functional information, then can it directly predict function from sequence information alone? This might seem to “give up” structural-level information, but Hong Liang believes this is precisely the strength of large models.
“The reason we couldn’t bypass structure in the past is that functional data was too scarce and humans’ analytical dimension was too low,” he explained. “We had to reduce high-dimensional sequence information to relatively fixed three-dimensional structures, and then understand principles such as antibody–antigen interaction. But now, with massive functional label data, we can directly use large models to learn the mapping from sequence to function.”
Hong Liang mentioned that training a language model is like “fill-in-the-blank” (cloze): hide the latter part of a sentence and let the model fill it in. Training a protein model is the same — randomly “delete” part of a sequence composed of hundreds of amino acids and let the model fill it back in. To complete the task, the model must learn the complex associations among amino acids that transcend three-dimensional distance.
“We don’t reify it into a three-dimensional structure; instead, we treat it as a paragraph and do reading comprehension at the level of characters, words, sentences, and paragraphs, letting AI directly read the ‘code of life,’” Hong Liang said.
Hong Liang pointed out that models trained this way are hard to explain through traditional structural biology, yet their prediction performance is surprisingly good. This is not a denial of structure’s value in biological understanding, but a more efficient path discovered in AI engineering practice.
Matwings Technology has accumulated a vast amount of high-quality functional label data, much of it from organisms that survive in extreme environments — high temperature, high pressure, highly acidic, or highly alkaline. These unique training data have greatly enhanced model performance.
Currently, VenusREM — the model developed by Matwings Technology that predicts function based on protein sequence information — has reached the international frontier in prediction accuracy. Hong Liang mentioned that in July this year, VenusREM ranked first among 95 specialized models and 13 general-purpose large models with a score of ρ = 0.523 in the Harvard PG-LLM preprint evaluation, and was set as an authoritative industry reference benchmark for all protein design large models.
Applications Across Multiple Fields
With its distinctive AI large model, Matwings’s computational capability is being transformed into real-world industrial value. Hong Liang noted that proteins designed by Matwings have been applied in multiple fields, including innovative drugs, packed/target proteins, and industrial enzymes.
Tan Yang, a technical scientist in Matwings’s AI Lab, told reporters that AI protein design mainly has two approaches: first, “substitution,” which screens for a product with better performance by changing individual amino acids without altering the protein’s length; second, “generation,” which, like generating text, “spits out” a brand-new protein that does not exist in nature, one amino acid at a time.
Through the large model’s “understanding” of function, researchers can design proteins that meet their needs. Taking plastic degradation as an example, the PET-degrading enzyme developed by Matwings can efficiently decompose discarded mineral-water bottles and polyester clothing into two extremely pure monomer raw materials — terephthalic acid and ethylene glycol — which can then be re-polymerized into brand-new plastic particles, with the entire process being pollution-free. These raw materials have wide uses: they can be used to make automotive airbags, ultra-elastic soles for running shoes, crystal bottles for cosmetics, and even become 3D-printing materials.
“Matwings’s business model is ‘platform technology + underlying products,’” summarized Cai Jinxu, an investor in Matwings Technology and Executive Manager of the Executive Department of the Shanghai Future Industry Fund. They both provide automation solutions externally to validate the platform’s generalization capability, and incubate a series of underlying products to demonstrate the industrial and commercial value of their technology. This “dual-wheel driven” business model has truly brought AI’s value into production scenarios and products.
“Foolproof” Protein R&D
As AI matures in single-point protein design tasks, a new bottleneck has emerged. “Much R&D cannot be solved by a single task,” Tan Yang pointed out. A complete protein design workflow may involve the combination of multiple models, such as backbone design, sequence design, and structure prediction, with scattered intermediate steps and severe information loss. The industry needs a platform to connect these steps.
To this end, Matwings Technology developed MatwingsVenus® , a conversational dry-wet combined protein design agent, which is the gateway to Matwings’s core technical capabilities. “You can think of it as a super researcher,” Hong Liang introduced. Users only need to state their needs in natural language, such as “I want an enzyme with higher activity in an acidic environment,” and MatwingsVenus® can automatically retrieve literature, patents, and market data, determine the R&D direction, and then invoke the underlying protein large model for design.
At the other end of MatwingsVenus®, it connects to an automated laboratory. Sequences designed by users can be directly submitted to the automated laboratory for synthesis and testing.
In the vision of AI for Science (“AI4S”), an agent can autonomously plan and design experiments, then direct robots in an unmanned laboratory for rapid validation, forming an efficient scientific-research closed loop. Tan Yang honestly acknowledged that automated experimentation is still in a developmental stage, and the inconsistent interface standards of equipment from different manufacturers pose a huge challenge for unified AI scheduling.
Despite the many challenges, its industrial significance is equally profound. “When the essential logic and skills of science become standard actions of AI and robotic arms, all anyone needs to do is submit their requirements,” Hong Liang said. An entrepreneur who knows nothing about protein engineering can also refine their own product by conversing with the agent. This will ignite the innovative vitality of countless teams.
AI4S’s Shanghai Moment
Cai Jinxu believes that an exclusive data moat, a dry-wet combined experimental closed loop, a composite team with both AI and scientific backgrounds, and a clear commercialization path have allowed Matwings to stand out from many AI4S projects and win the support of “patient capital.”
It is reported that Shanghai Guotou (state-owned investment) system fund managers have cumulatively invested 42 million yuan in Matwings Technology: in March 2026, Shanghai Guotou Future co-led an A+ round of financing of over 200 million yuan in Matwings Technology, with Shanghai Guotou Futeng following; in July of the same year, Shanghai Guotou Pioneer co-led an A++ round of financing of several hundred million yuan, with Shanghai Guotou Futeng — as an old shareholder — adding again. Within just four months, Shanghai Guotou system funds made two rounds in succession, continuously supporting the company.
Beyond capital support, Shanghai Guotou also continuously tracks the company’s operations and financing progress after investment, participates in shareholder voting at subsequent financing rounds, and helps connect with industrial customers and investor resources.
“We helped connect with government departments to accelerate product approval, connected with mainstream media to expand influence, assisted in hosting the ‘Yangtze River Cup’ competition, and also helped connect with cloud vendors to obtain computing resources,” Cai Jinxu said. “These may look like small things, but each one is a real difficulty in running a business.”
Rich computing power, a strong biomedical industry base, and a complete capital ecosystem together constitute Shanghai’s advantages in AI4S development. And the ultimate decisive factor in this “competition” lies in talent.
The “Hundred Teams, Hundred Projects” is a science-intelligence development special project organized annually by the Shanghai Municipal Commission of Economy and Informatization, starting last year. Teams applying for projects must be composed of three types of talent: AI scientists, domain scientists, and engineers; and the project leader must in principle be a young scientist no older than 40. Projects lacking cross-disciplinary talent cannot enter the support list.
Relevant documents explicitly state that one of the project’s objectives is to “cultivate a group of young scientists who understand science, understand AI, and have industrial thinking.” Under such top-level design, more and more young talent is gathering in Shanghai.
As a technical scientist in Matwings’s AI Lab, Tan Yang was born in 2000 and majored in computer science. Most other team members are also post-1998, with computer science majors accounting for one-quarter, and the rest spanning physics, chemistry, biology, pharmacy, and humanities. Hiring is not restricted by major — “the most important thing is being smart,” Hong Liang said. His definition of smart is “the ability to define and solve problems, and the speed and efficiency of accepting new things and using new technologies.”
“The post-2000 generation has been using AI since college, just like we eat and drink,” Hong Liang said. “In the end, industrial competition relies on their initiative and their efficiency in acquiring resources.”
Tan Yang said that, in the face of the rapidly evolving AI for Science, if young people can join a good platform, they must participate in the most cutting-edge competition and compete with top teams on outcomes — “the returns will be extremely rich.”
Based in Shanghai, rooted in China, but with their sights set on the world and the future, young people are looking ahead. Tan Yang believes that, with AI, most diseases could see good treatment plans within about 10 years. Because data is abundant, AI protein/enzyme design will break through first, and as data expands and investment increases, AI drug design will eventually usher in a moment of “intelligent emergence.”
Hong Liang has a vision for the company: like a chip company, he hopes that one day a “Matwings Inside” mark will appear across the global biotech field, making Matwings Technology’s AI the core engine driving biological innovation.
