A row of bottles sits on display in Matwings’ showroom. Some of the clearest and most brilliant were made from discarded water bottles and old bedsheets.

Behind them is a plastic-degrading enzyme—a protein that catalyzes the breakdown of plastic. It breaks plastic waste down into terephthalic acid and ethylene glycol, which are then repolymerized into resin pellets and made into crystal-clear bottles for skincare products and cosmetics.

Hong Liang, co-founder and Chief Scientist of Matwings, calls the material “bio-crystal.” Its impurity content is approximately 90% lower than that of petroleum-derived feedstocks. The resulting bottles offer exceptional clarity and a heat deflection temperature comparable to that of virgin material, fully meeting brands’ performance requirements. “This may be the first biomaterial made with AI,” Hong said.

As AI for Science, or AI4S, gathers momentum, this tangible example may be more persuasive than any model specification. Proteins underpin industries across life sciences and healthcare: antibodies in biopharmaceuticals, collagen in aesthetic medicine, enzymes in in vitro diagnostic reagents, industrial enzyme preparations, and proteins for animal feed.

Yet protein R&D has long followed much the same approach: expert judgment combined with high-throughput screening, followed by repeated rounds of trial and error. Development can take years, costs are high, and the chances of finding a protein with the desired properties remain low.

Matwings, which focuses on AI-driven, end-to-end protein R&D, aims to change that process. It understands that the ability to design proteins with greater stability and tolerance to demanding conditions can open doors across a wide range of downstream industries.

Breaking Technical Barriers: AI4S Takes Protein R&D from Computational Design to Industrial Application

Traditionally, protein engineering has followed an indirect route: start with the amino acid sequence, predict the three-dimensional structure, and then infer function from that structure. Each step introduces uncertainty, with repeated experimental validation needed along the way.

Founded in 2021, Matwings took a different approach: bypassing structure as an intermediate step to design functional proteins directly around industrial requirements.

Speaking to reporters from STAR Market Daily and other media outlets, Hong explained why structure-based approaches had become so prevalent: functional data were scarce.

“When functional data are limited, and human analysis can only account for so many dimensions, researchers have little choice but to start with structure to understand what is happening,” he said. Once Matwings had accumulated enough sequence–function relationships, however, a large model could learn that mapping directly.

Data are the foundation of this approach, and an independent benchmark offers one indication of its strength.

According to information obtained by STAR Market Daily during the interview, a recent preprint from Harvard University evaluated 95 specialized models and 17 general-purpose large language models on the same set of tasks, assessing their ability to rank the functional performance of protein variants in a zero-shot setting.

General-purpose models trained solely on literature and patents achieved scores of around 0.4. Matwings’ VenusREM scored 0.52, ranking first among the 108 models cited in the comparison and serving as a reference benchmark for the general-purpose models. The latest version has reportedly improved that score to 0.56.

These results also help explain why data can create a lasting competitive advantage. As general-purpose models continue to improve, algorithms alone are becoming less likely to sustain a company’s lead. The more durable distinction lies in training data that competitors cannot readily obtain.

This is one of Matwings’ core strengths. To date, the company has accumulated tens of billions of protein sequences, including 6.7 billion with functional labels. The dataset was selected for China’s first batch of exemplary high-quality datasets, providing a foundation for the company’s ability to design proteins for specific functions.

In April this year, Matwings launched MatwingsVenus™ (晓鹜™), a conversational AI agent for protein R&D. Researchers do not need programming expertise or extensive training in structural biology. They can describe their objectives in natural language, and the agent can handle the workflow—from industry research, searches of functionally annotated databases, and model-based analysis to protein design and automated experimental validation. The aim is a continuous process in which designs move into experiments and experimental results inform the next round of design.

“When we first developed MatwingsVenus™, we wanted it to act as an assistant that translates a user’s language into computational instructions and returns the results they need,” Hong said. “Researchers developing an enzyme or another protein can complete the workflow on the platform and then connect it to experiments, reducing the time and money spent on trial and error.”

The central advance is the connection between computational design and physical experiments, creating a complete loop in which computation guides wet-lab work and experimental results inform subsequent computation. Once AI has designed a protein, users can place an experimental order directly through the platform. The task is then routed to Matwings’ fully automated laboratory, where robots carry out sample preparation, protein purification, functional testing, and other wet-lab procedures.

MatwingsVenus™ represents more than an upgrade to a research tool. It makes protein R&D capabilities once largely confined to leading companies and top research institutions available as infrastructure that individual researchers can readily use. In doing so, it moves the field toward a model in which advanced protein research is accessible at the individual level.

Turning a model into a product is the next step. Hong described the process in straightforward terms:

“MatwingsVenus™ screens or engineers sequences from a pool of hundreds of millions to tens of billions, then provides a set of candidates for users to select for targeted laboratory testing. If the results are good, we move toward production. If not, the results go back into the model for another round of design. Typically, two or three rounds are enough to complete the design of a functional protein.”

The platform’s practical capabilities have already been demonstrated in specific projects.

For example, a fusion protein therapeutic designed by Matwings for macular disease has received authorization from China’s National Medical Products Administration to enter clinical trials for three indications. The protein targets three disease-associated ligands in the VEGF family and simultaneously blocks the signaling pathways they mediate. Comparable molecules already on the market or in Phase III clinical trials block only two of these ligands. Preclinical animal studies showed greater efficacy for the Matwings-designed fusion protein.

The design itself offers an instructive example of AI’s contribution. Of the seven engineered mutation sites, three are located at the target-binding interface, while the other four are distant from it. The effects of these distant mutations are difficult to explain through conventional structural biology reasoning, yet AI identified their association with binding affinity.

Patient Capital and Shanghai’s Innovation Ecosystem Support Foundational Research

Matwings began by choosing a technically demanding path in foundational AI4S research. That choice has also attracted institutional investors.

In March this year, the Shanghai Future Industry Fund, part of Shanghai State-owned Capital Investment, co-led the company’s Series A+ financing of more than RMB 200 million. In July, Matwings completed a Series A++ round worth several hundred million renminbi, co-led by Shanghai Guotou Xiandao and SDIC Innovation. Shanghai Guotou FOF Capital participated in both rounds, alongside several other established investment institutions.

Within six months, Shanghai State-owned Capital Investment had backed Matwings twice through its investment platforms. The relationship began with the Shanghai Future Industry Fund, a government-backed fund of funds with a policy mandate to support “zero-to-one” innovation and invest in frontier technologies.

Since its establishment, the fund has focused on six future industries, including information technology, energy, space, and health. AI4S is, in its words, “a very important investment area.”

For Cai Jinxu, Executive General Manager of the fund’s Investment Department, the value of AI4S begins with its ability to advance the underlying science.

“It can significantly accelerate research, whether that involves discovering new materials, designing devices for new energy applications, or developing enzymes and proteins for biomanufacturing,” Cai said. “All of these depend on scientific R&D. AI4S changes how that research is done, so it can accelerate development across materials, information technology, energy, and other fields.”

Cai also sees investment opportunities emerging as the technology evolves. AI4S is moving from structure prediction toward product design. Research agents are beginning to help scientists review literature, plan studies, and analyze data, while automated laboratories and integrated computational–experimental workflows are gradually becoming a reality.

“Each wave of technological change can create new opportunities for entrepreneurs,” he said.

The growing number of AI4S ventures generally follow one of three business models. Some provide tools and platforms, selling model and computing services to pharmaceutical, materials, and other companies through software or technical service fees. Others develop proprietary pipelines and products, retaining the resulting value in their own assets. A third group combines the two, using a platform to support R&D and generating revenue through products.

As these approaches diverge, identifying which AI4S companies merit investment has become an increasingly important question. Shanghai State-owned Capital Investment has made AI for Science a priority and established a dedicated fund in the field. Its model pairs investment managers with managers responsible for helping portfolio companies access resources and operational support. The aim is to support companies throughout their development and help new technologies secure their first trials and commercial applications in Shanghai.

Cai outlined four main considerations in the Shanghai Future Industry Fund’s assessment of AI4S investments.

The first is data. Gaps in model performance can be closed relatively quickly, but data advantages are harder to replicate. Investors examine whether a company has scarce data assets, such as exclusive sample sources or a laboratory that continuously generates new results. The key question is whether those data can keep accumulating as the business grows, creating a reinforcing cycle, rather than remaining a static, one-off acquisition.

The second is experimental validation. AI-generated designs ultimately have to be tested. The central measures are the turnaround time and cost of the design–experiment–feedback cycle. Faster cycles allow models to improve more quickly and generate more data. Investors therefore also consider whether experimental validation can be completed within a manageable timeframe.

Commercial performance is another essential test. The fund prioritizes actual orders and repeat business over expressions of interest or demonstrations. Customers’ willingness to pay for results is the most direct evidence of a technology’s value.

Finally, there is the team. AI4S is inherently interdisciplinary, requiring expertise in scientific research, AI algorithms, engineering, and the market. A founder’s commercial judgment is particularly important. Strong research credentials do not automatically translate into the ability to identify market needs or manage a company.

“If a project relies mainly on benchmark rankings or a single model metric to tell its story, but has little substance in its data or practical applications, we tend to be cautious,” Cai said.

Matwings stood out against these four criteria.

“We first met Matwings in 2025, when AI4S was just beginning to gain momentum,” Cai recalled. At the time, the investment team believed that algorithms and models were evolving so quickly that companies needed to establish defensible positions in both training data and commercial execution. Matwings, he said, “was already ahead on both fronts.”

On data, Matwings had amassed a training collection of tens of billions of protein sequences. A substantial portion came from extreme environments such as the deep sea and salt lakes, with associated information on pressure, pH, salinity, and other environmental conditions. These data help differentiate its predictive capabilities for demanding industrial applications.

On algorithms, its approach to predicting function directly, without first generating a structure, distinguishes it from many other models.

On commercialization, Matwings had already secured orders for products including proteins used in chromatography resins and industrial enzymes. In Cai’s assessment, that placed the company close to commercialization at scale.

For early-stage deep-tech companies, funding is essential, but it is rarely the only constraint. Many investors in the sector now go beyond financial backing and place greater emphasis on practical support after investment. The Shanghai Future Industry Fund is developing a “fund + community” model, using a coordinated range of services to help startups overcome specific obstacles.

For Matwings, Shanghai State-owned Capital Investment has drawn on its industrial and government networks to connect the company with resources for experimental validation, communications, the joint development of industry platforms, and computing infrastructure.

“Each of these things may seem small on its own,” Cai said. “But for an early-stage company, any one of them can save several months.”

For example, obtaining the testing required to bring a material to market had previously been time-consuming. The Shanghai Future Industry Fund helped Matwings connect with testing providers, clearing a route to market. It also introduced the company to leading cloud providers to help address another common constraint for AI4S startups: access to computing capacity.

Cai identified four features of Shanghai that make this approach possible.

First is the concentration of research institutions. Shanghai Jiao Tong University, Fudan University, Shanghai Artificial Intelligence Laboratory, and Lingang Laboratory are advancing both AI and scientific research, giving companies access to talent in both areas.

Second is the breadth of industrial applications. Shanghai’s strong biopharmaceutical sector, together with its concentration of chemical, advanced materials, semiconductor, and advanced manufacturing businesses, provides markets in which AI4S research can be applied and commercialized.

Third is its AI infrastructure. The city’s continued investment in large models, computing capacity, and data resources supports the integration of AI capabilities with scientific research.

Fourth is its capital and wider support network. Shanghai State-owned Capital Investment, for example, has a range of funds covering early-, growth-, and later-stage companies, allowing it to provide continued backing as businesses develop.

From a founder’s perspective, Hong highlighted three advantages: a high concentration of talent, investors willing to take risks, and abundant opportunities to put technology to work.

“Capital upstream, talent in the middle, and applications downstream,” he said. Together, these give Shanghai a strong foundation for advancing AI in biopharmaceuticals and biomanufacturing.