Back to list

Future Industries in Shanghai | From Waste Plastic to Crystal Bottles: How a Shanghai Company Is Rewriting Protein R&D with AI

Published on November 18, 2026

Future Industries in Shanghai | From Waste Plastic to Crystal Bottles: How a Shanghai Company Is Rewriting Protein R&D with AI

STAR Market Daily, October 17 (Reporter: Shi Shiyun) — A row of bottles stands in the showroom of Matwings Technology. The brightest of them are made from mineral-water bottles and old bedsheets.


They are “made” by a plastic-degrading enzyme — a protein that catalyzes the breakdown of plastic. Waste plastic is enzymatically hydrolyzed into terephthalic acid and ethylene glycol, then repolymerized into pellets and molded into crystal-clear bottles for skincare and cosmetic products.


Hong Liang, co-founder and chief scientist of Matwings Technology, calls the material “bio-crystal.” It contains roughly 90% fewer impurities than petroleum-based feedstock; the resulting bottles offer excellent clarity and a heat deflection temperature on par with virgin material, fully meeting the performance requirements of brand customers. “This may be the first biomaterial made by AI,” Hong Liang said.


At a time when AI for Science (hereinafter “AI4S”) is booming, this detail may be more persuasive than any model parameter. Proteins are the foundational feedstock of the life-and-health industry: therapeutic antibodies, collagen for medical aesthetics, the enzymes used in in-vitro diagnostic reagents, industrial enzyme preparations, and even animal feed protein all depend on them.


Yet for a long time, the way proteins are developed has barely changed. It relies on expert intuition paired with high-throughput screening, and on round after round of trial-and-error iteration that often takes years, carries high R&D costs, and has only a very low probability of landing on the ideal protein.


What Matwings Technology — which focuses on AI-driven, full-stack protein R&D — wants to rewrite and disrupt is precisely this traditional path. It understands that whoever can design proteins with better stability and greater robustness holds the key to a long chain of downstream industries.


▌ Breaking the Technical Barrier: AI4S Reshapes the Protein R&D Pipeline, from Compute-Driven Inference to Industrial Deployment


By convention, protein engineering has followed a zigzag path: start with the amino acid sequence, predict the three-dimensional structure, then infer protein function from structure. The chain advances layer by layer, is full of discontinuities, and requires repeated experimental validation along the way.


Founded in 2021, Matwings Technology — with the fearlessness of a newborn calf that does not dread the tiger — did something rather different: it bypasses structure and moves directly to the precision design of functional proteins aimed at industrial requirements.


Hong Liang explained the rationale to STAR Market Daily and other media: it is because functional data are too scarce. “There isn’t enough functional data, and the analytical dimensions available to humans are limited, so we had no choice but to begin with structure-based, interpretive exploration.” But once Matwings Technology accumulated enough sequence–function correspondences, a large model could learn that mapping directly.


What underpins this approach is data. To see how solid that data foundation is, let the third-party leaderboards speak.


According to what the STAR Market Daily reporter learned in interviews, a recently released Harvard preprint placed 95 specialized models and 17 general-purpose large language models on the same set of tasks, evaluating their ability to rank the functional performance of protein mutants under zero-shot conditions.

General-purpose large models, trained only on literature and patents, achieve an accuracy of only around 0.4, whereas Matwings Technology’s VenusREM model reaches 0.52 — the best result among 108 models — and was adopted as the authoritative benchmark against which general-purpose large models are compared. It is understood that the latest version of the model has raised its accuracy to 0.56.


These numbers also illustrate how data builds a moat. As general-purpose model capabilities keep iterating, algorithms alone can no longer sustain a long-term competitive advantage; what truly creates differentiated competitiveness is training data that others cannot obtain.


And this is precisely Matwings Technology’s core advantage. To date, it has accumulated tens of billions of protein sequence entries, 6.7 billion of which carry functional annotations. The dataset has been selected among the first batch of national high-quality dataset exemplars — and this is the confidence behind Matwings Technology’s ability to achieve functional protein design.


In April this year, Matwings Technology launched MatwingsVenus™, a conversational protein R&D agent. Researchers need not be proficient in programming, nor must they be steeped in structural biology; they simply describe their R&D goal in natural language, and the agent runs the entire workflow end to end: from industry research, annotation-library retrieval and model inference to protein molecule design and automated experimental validation, realizing an intelligent R&D paradigm in which “design is validation, and validation is iteration.”


“When we first developed MatwingsVenus™ (Xiaolu™), we wanted it to serve as a tool assistant that translates the user’s language into the platform’s computational language and then delivers the result the user wants. If a researcher wants to make an enzyme or a protein, they can complete the whole workflow on the platform and then hand it off to the lab along that same workflow, saving the time and money of trial and error,” Hong Liang said.


The biggest breakthrough of this model is that it bridges the gap between the digital world and physical experiments, building a “complete dry–wet closed loop” in which computation drives wet-lab work and wet-lab results feed back into computation. Once AI completes the protein molecule design, users can place an experiment “order” directly on the platform; the task flows seamlessly to the automated lights-out laboratory built by Matwings Technology, where robots autonomously carry out the full suite of wet-lab work — sample preparation, protein purification and functional assays.


The launch of MatwingsVenus™ is far more than the upgrade and iteration of a single R&D tool. It converts protein R&D capability — once the preserve of only leading companies and top research institutes — into foundational infrastructure that ordinary researchers can readily access, shifting the protein R&D paradigm from a “big-platform-driven model” controlled by a few players to a “personal-use era” that anyone can pick up.


With a model in hand, the next step is to turn it into a product. Hong Liang laid out the process in fairly plain terms: “MatwingsVenus™ screens or engineers across hundreds of millions to billions of sequences, and provides a series of candidates for users to take into the laboratory for precise experiments. If a round of results is good, we move to production; if not, we feed the results back to the model for another round of design. Typically, two to three rounds of iteration are enough to complete the design of a functional protein.”


In more concrete projects, MatwingsVenus™’s deployable R&D capability has already been validated.


For example, a fusion protein drug designed by Matwings Technology for macular disease of the ocular fundus has received clinical trial approval from the National Medical Products Administration, covering three indications. The fusion protein can target three key disease-related ligands in the VEGF family and simultaneously block the signaling pathways they mediate. By comparison, comparable molecules currently on the market or in Phase III clinical trials can block only two of these ligands, and preclinical animal studies show that this fusion protein has superior efficacy.


It is worth noting that the AI design logic behind this fusion protein is quite striking: of the 7 engineering sites, 3 fall on the target-binding interface and the other 4 are far from it. The role of these distal mutations is difficult to explain with conventional structural-biology experience, yet AI uncovered their link to molecular affinity.


▌ Capital Empowerment: Patient Capital Backs Foundational Innovation as Shanghai’s Ecosystem Nurtures New Sci-Tech Growth


Choosing to enter through the “hard but right” AI4S foundational-technology track was Matwings Technology’s starting point — and it also brought institutional investors on board.


In March this year, the Shanghai Future Industry Fund under Shanghai Guotou (Shanghai International Group) co-led its A+ round of more than RMB 200 million. In July, Matwings Technology completed an A++ round of financing in the hundreds of millions of RMB, co-led by Shanghai Guotou Xiandao and Guotou Innovation, with Shanghai Guotou Futeng Capital following on in both consecutive rounds and several other well-known institutions also participating.


Within six months, Shanghai Guotou acted twice, adding capital in relay — and the starting point was precisely the Shanghai Future Industry Fund, a functional government fund-of-funds dedicated to “zero-to-one” investment that takes frontier foundational technologies as its core target.


Since its founding, the Shanghai Future Industry Fund under Shanghai Guotou has consistently focused on six major future industries, including future information, future energy, future space and future health, and AI4S is “a very important track” among them.


In the view of Cai Jinxu, executive general manager of the investment department at the Shanghai Future Industry Fund, the value of AI4S is first reflected in breakthroughs at the foundational level of science: “It has a powerful accelerating effect. Whether it is the discovery of new materials, the design of new-energy devices, or the enzymes and proteins used in biomanufacturing, the foundation is all scientific research. What AI4S changes is precisely how scientific R&D itself is done, so it can accelerate R&D across materials, information, energy and many other fields at the same time.”


Cai also noted the investment space opened up by technological change: AI4S is moving from structure prediction toward product design; research agents are beginning to help scientists search the literature, make plans and analyze data; and automated laboratories and dry–wet loops are being implemented step by step. “Every round of technological change can bring new entrepreneurial opportunities.”


At present, a great many AI4S projects have emerged on the market. In terms of business model, they can be roughly divided into three categories:

First, the tool-platform route: providing models and computing services to pharmaceutical and materials companies and charging software or technical service fees;

Second, focusing on self-developed pipelines and products, internalizing value in owned assets;

Third, adopting a hybrid “platform plus product” model, in which platform capabilities empower R&D and monetization is achieved by bringing products to market.


As the track diversifies, exactly what kind of AI4S company is worth betting on has become an important question for investors. As a frontier paradigm in which artificial intelligence empowers scientific research, AI for Science is a frontier direction of priority for Shanghai Guotou. To this end, Shanghai Guotou has established a dedicated AI for Science fund and built a dual-manager model of “fund manager plus enablement manager,” providing full-lifecycle, hands-on support for sci-tech innovation projects, promoting the first trial and first application in Shanghai of a batch of hard-core innovation outcomes, and injecting new AI momentum into the strategic foundation of Shanghai’s modern industrial system.


Cai introduced that the Shanghai Future Industry Fund’s AI4S investments focus mainly on four aspects:

Data is the first dimension of evaluation. Gaps in models are relatively easy to close, but data moats are much harder to replicate. Investors examine whether a company possesses scarce data assets — for example, exclusive sample sources, or data continuously generated by self-built laboratories. The key lies in whether data can accumulate continuously as the business advances and build a growth flywheel, rather than being static data acquired in a one-off way.


Turning AI-designed schemes into reality, in turn, depends on a validation loop. The designs produced by AI ultimately need to be validated experimentally. The core criterion is the turnaround speed and cost of the “design–experiment–feedback” loop. The faster the loop runs, the more efficient model iteration becomes and the greater the volume of data accumulated. Therefore, whether the verification cycle of the track itself is manageable is also brought into the scope of assessment.


Commercial performance is a hard indicator for testing corporate value. Institutions give priority to a company’s real orders and customer repurchases, rather than stopping at cooperation intentions or demo effects. The willingness of market customers to pay for results is the most intuitive standard for validating technical value.


Finally, the team. AI4S is a typically interdisciplinary field, requiring a team to combine research, AI algorithms, engineering implementation and market awareness. The founder’s business judgment is especially critical. Many researchers have outstanding technical strength, but precisely identifying market demand and managing a team well are another set of key capabilities.


“If a project tells its story mainly through leaderboard scores or a single model metric, with thin data and thin deployment, we will be rather cautious,” Cai said.

Apply those four yardsticks, and Matwings Technology comes into view. “We first engaged with Matwings Technology in 2025, when the AI4S track was just emerging,” Cai recalled. Even then, the team judged that algorithmic models iterate extremely fast, and that for a company to hold its ground it must quickly build moats in training data and in the commercial loop — and Matwings Technology “happened to be ahead on both counts.”


On the data side, Matwings Technology’s protein sequence training data reach the scale of tens of billions of entries, a considerable portion of which comes from extreme environments such as the deep sea and salt lakes and carries environmental information such as pressure, pH and salinity. This gives its predictive capability real differentiation in harsh industrial settings. On the algorithm side, its path of bypassing structure generation and connecting directly to function prediction differs markedly from most models. On the commercial side, Matwings Technology has already won orders for products such as filler proteins and industrial enzymes which, in Cai’s words, puts it very close to commercial scale-up.


For early-stage hard-tech companies, capital is certainly key, but it is often not the only bottleneck constraining development. Today, many investment institutions deeply engaged in the hard-tech track have moved beyond the framework of purely financial investment and placed post-investment enablement in an important position. At present, the Shanghai Future Industry Fund is building a “fund plus community” operating model, aimed at relying on systematic post-investment services to genuinely remove the various bottlenecks facing startup companies.


For Matwings Technology, Shanghai Guotou has drawn on its own industrial and government resource networks to provide precisely matched resource connections and implementation support in areas such as experimental validation, brand communication, co-building of industry platforms and access to computing infrastructure, filling in the gaps in the company’s resources on the path to industrialization.


Cai said: “Taken individually, these things may not seem like much, but for an early-stage company, each one can help it save several months.”


For instance, in product validation, the authoritative testing required before a feedstock can go to market had previously been time-consuming; the Shanghai Future Industry Fund helped Matwings Technology connect with testing resources and cleared the product’s path to market. Computing power is a core necessity for AI4S companies and also a bottleneck for many startups; the fund simultaneously connected Matwings with cloud computing services from a leading cloud provider, effectively easing the pressure on the company’s computing supply.


The confidence behind this approach comes from the endowments of the city of Shanghai itself. Cai summarized four advantages:

Dense sources of research. Institutions such as Shanghai Jiao Tong University, Fudan University, the Shanghai AI Laboratory and Lingang Laboratory advance “in step” on both AI and science, so companies can find both AI talent and scientific talent;

Ample industrial scenarios. The biomedicine industry has a solid foundation, and chemical, new-material, integrated-circuit and advanced-manufacturing companies are clustered here, so AI4S R&D results can find market space in which to land and monetize;

Solid AI infrastructure. Shanghai has consistently invested in large models, computing power and data resources, which precisely matches AI4S’s need to fuse AI capabilities with scientific problems in both directions;

Well-developed capital and ecosystem support. Shanghai Guotou’s system, for example, has a fund matrix covering the early, middle and later stages, able to continuously support a company’s development in relay.


From the company’s perspective, Hong Liang also mentioned that Shanghai’s strengths come down to three things: a high density of talent, courageous capital, and abundant application scenarios. “Money upstream, talent midstream, scenarios downstream” — this too gives Shanghai a strikingly strong advantage in developing AI biomedicine and AI biomanufacturing.

Scan to join community

Scan to join community