AI for Protein: When Artificial Intelligence Learns to 'Create Things,' Protein Design Welcomes Its iPhone Moment

Introduction: A Revolution in the Making

In every cell of our bodies, in the resilient threads spun by spiders, and in the heat-tolerant enzymes of deep-sea vent microbes, nature has long displayed its exquisite craftsmanship as the ultimate molecular engineer. Proteins, these precise molecular machines woven from 20 amino acids, drive almost all life activities—from muscle contraction to immune defense, from photosynthesis to gene editing.

However, over the long course of history, human utilization of proteins has been arguably 'clumsy.' We have only been able to 'pan for gold' in nature, screening through thousands of natural proteins to find barely usable candidate molecules, and then hoping for luck through long processes of random mutation. Changing the position of one amino acid creates about 6,000 possible variations; changing two leaps to 23 million; changing three jumps directly to 50 billion—a scale where an average-sized protein contains 361 amino acids, and the theoretically possible sequence combinations are 20 to the power of 361, a number beyond the total number of atoms in the universe.

In this 'protein universe,' traditional methods are like blind men feeling an elephant. A trained scientist needs over ten years of practice to independently design a functional protein; successfully engineering a protein molecule to reach industrial-level performance often requires 2 to 5 years of trial and error. This is not research and development—it is luck accumulated through manpower and time.

But all of this is being rewritten. The force rewriting it comes from artificial intelligence.

AI-protein design

1. From the 'AlphaFold Moment' to 'Design is Validation'

If there is any event that can be regarded as the starting point of 'AI for Protein,' it is undoubtedly the awarding of the 2024 Nobel Prize in Chemistry. Half of that year’s Nobel Prize in Chemistry was awarded to David Baker in recognition of his pioneering contributions to computational protein design; the other half was awarded to Demis Hassabis and John Jumper for their revolutionary breakthroughs in protein structure prediction.

This award sends a clear signal: AI is no longer just an auxiliary tool in biology but has become the core engine driving the advancement of protein science.

In just a few years, from the AlphaFold series to RoseTTAFold, from RFdiffusion to ProteinMPNN, deep learning models have achieved atomic-level accuracy in protein structure prediction and design. The introduction of generative AI has propelled protein design from the traditional model of 'explicit energy function optimization' into a new era of 'probabilistic generative modeling'—AI no longer merely predicts the appearance of existing proteins but also begins to create entirely new protein molecules that do not exist in nature.

Structural biology is entering a 'second wave' following the breakthrough of AlphaFold2: panoramic prediction of protein conformations and routine de novo design of protein complexes are becoming two entirely new frontiers. AI-driven protein design platforms are transforming processes that were highly reliant on experience and screening into scalable engineering systems, achieving extremely high success rates in experiments.

The underlying logic of this transformation is actually quite simple: when AI can not only 'understand' proteins but also 'design' them, the rules of the game are completely changed.

2. Dry-Wet Closed Loop: From 'Paper Simulation' to 'Building It'

However, an easily overlooked but crucial question is: are proteins designed by AI really "usable"?

This is a soul-searching question. In the field of protein engineering, there has long been a "last mile" gap—the beautiful molecular designs on the computer and the real proteins that can be stably expressed and functional in the lab are separated by a huge distance. Designing the backbone structure of a protein is one thing; actually expressing it in cells, folding it correctly, and ensuring it is active is another.

This is precisely the direction the entire industry is striving to break through: to connect AI design with experimental validation in a cycle.

The complete picture of AI for Protein is not just a collection of algorithmic models, but an upward-spiraling closed loop of "design → validate → feedback → redesign." An important review published in *Nature Reviews Bioengineering* in 2025 provides a practical roadmap, dividing AI tools into seven toolkits that cover the full process from initial design to experimental validation, and points out that AI is fundamentally reshaping protein design — what used to be a process of trial and error is now becoming a predictable and reproducible engineering discipline.

This closed loop is essential because the predictive power of AI models needs high-quality wet lab data to "feed" and calibrate them. Whoever possesses efficient experimental validation capabilities can form a "data flywheel" — each experimental result trains stronger AI, and stronger AI generates more precise designs. Repeating this cycle can make research and development efficiency and success rates far surpass traditional methods.

This means that the competitive dimension in the field of AI for Protein has shifted from a single-point comparison of algorithm performance to a systemic competition in the efficiency of the design–validation loop.

3. China's Power: When protein research and development becomes something that can be resolved through a "dialogue."

Chinese companies have not been absent from the global race in the AI protein race. According to market data, there are currently about 30 Chinese companies investing in the AI protein field, mainly focusing on industrialization—emphasizing the practicality and delivery efficiency of technology, and striving to bring AI protein design from the laboratory to factories and clinical settings.

Within this group, a company from Shanghai—Matwings Technology—is reconstructing protein research and development in a unique way.

Matwings Technology was founded in 2021 by Hong Liang, a distinguished professor at Shanghai Jiao Tong University, with a core team deeply engaged in the intersection of life sciences and artificial intelligence. CTO Liu Hao summed up the pain points of traditional protein R&D in one phrase: "Needle in a haystack." The space for protein sequence combinations is vast and far exceeds the limits humans can exhaustively experiment with. With traditional methods, the positive rate for experiments was only about 0.1%, but with the help of AI large models, Matwings' team increased this figure to around 30%, reducing the R&D cycle from 2 to 5 years to 2 to 6 months.

Supporting this efficiency leap is Matwings Technology's massive dataset containing nearly 9 billion protein sequences—not only covering conventional species but also integrating special sequences from extreme environments such as volcanoes and the deep sea. About 500 million of these include functional labels such as temperature and pH, enabling the model to precisely map from "sequence" to "function."

But these are not the most exciting parts.

On April 24, 2026, Matwings Technology released a conversational protein development agent called MatwingsVenus™ ™, bringing "AI for Protein" into a new stage of human-computer interaction.

The core design concept of MatwingsVenus™ (Xiaowu ™) is to transform protein R&D capabilities, which were previously only accessible to large research institutions and leading companies, into a "shared laboratory" that individual developers can easily access. Users only need to describe their needs in natural language—just like having a conversation with a protein design expert through chat software—and the platform's agents will automatically break down tasks, calling on over 200 integrated protein design tools, a billion-level label database, and more than 30 skills tuned by experts from various fields, completing the entire process from industry research, target analysis, protein design to experimental validation.

More importantly, MatwingsVenus™ (Xiaowu) breaks down the barriers ™ between the digital and physical worlds. Once AI completes protein design, the platform can automatically import results into automated wet experiment workflows, driving robots to complete sample preparation, protein purification, and functional testing, with experimental data flowing back to the next round of AI design. This "design is verification, validation is iteration" dry and wet closed-loop model means protein R&D is no longer a lengthy process relying on large teams and repeatedly switching between multiple tools, but has become a sustainable, iterative intelligent infrastructure.

In the words of Hong Liang, founder of Matwings Technology, this is a technological revolution in the model of scientific research organization. The emergence of AI plus automation tools has led to some research skills that previously required high costs and extremely high professional thresholds to 'come down from the altar'—although this may eliminate some traditional positions, it will inevitably free more individuals and very small teams to carry out personalized innovation, driving a leap in productivity.

4. Validation in the real world

The value of a platform ultimately depends on whether it can deliver results in the real world. Matwings Technology has provided the answer with solid industrialization cases.

Case 1: "Taming" strong alkali in 4 months, the world's first 5,000-liter industrial protein product designed by a large model

In the production and purification process of growth hormone, there is a key procedure—affinity chromatography, which requires cleaning and regeneration under extremely alkaline conditions with a pH as high as 13-14. Traditional protein materials are easily inactivated under such conditions and need to be replaced frequently, resulting in high costs.

Matwings Technology collaborated with China Growth Hormone leader Jinsai Pharmaceutical, entirely through AI large model design, combined with a small number of wet experiments in a closed-loop iteration for verification. In just 4 months, they successfully increased the alkali resistance of a common non-alkali-resistant single-domain antibody by four times and successfully applied it to 5,000-liter industrial scale-up production. It is estimated that this breakthrough can save the company over ten million yuan per year in costs. This not only marks the landing of the world’s first protein product designed by a large model and achieved 5,000-liter industrialized production, but also truly opens the channel for 'AI design—industrial production.'

1.Science Advances (2024), eLife (2025)

2.Science Advances (2024), eLife (2025)

Case 2: Global Giant Teams Up with Chinese AI, Bayer Bets on Intelligent Protein Design

In November 2025, at the 7th China International Import Expo, Bayer Consumer Health, a multinational life sciences giant with a history of over 160 years, signed a strategic cooperation agreement with Matwings Technology, announcing that the two sides will further deepen collaboration on AI technology in research, new product development, and application transformation in the fields of digestive and skin health. It is reported that this is an intensification and upgrade based on their previous cooperation. A century-old multinational company choosing a Chinese AI protein design company that has been established for only a few years as a strategic partner is in itself a strong endorsement of technological capabilities—when AI protein design becomes an innovation engine for consumer health products, from functional skincare to gut health, protein molecule design is entering everyone's daily life.

Signing ceremony

These cases point to the same conclusion: AI for Protein has moved beyond the "technology validation" stage and entered the era of "engineering delivery." Whether in pharmaceuticals, diagnostics, environmental protection, or consumer goods, AI protein design is moving from the laboratory into deep industrial waters, evolving from a flashy "can it be done" to a hardcore delivery of "how much cost can be saved and how much value can be created."

5. Outlook: The "iPhone Moment" of Protein

If biomanufacturing is the cornerstone of future industries, then protein design is the underlying chip for biomanufacturing. Whether it's innovative drugs, functional foods, agricultural technology, or bio-based materials, high-performance proteins are core functional components. AI for Protein is turning the development of this "underlying chip" from a highly experience-dependent and resource-intensive "craft" into a capability that can be standardized, automated, and made accessible.

Looking back at the history of the computer industry, only a few institutions could afford computers during the mainframe era. It wasn't until the advent of personal computers and smartphones that computing power truly entered everyone's pocket. Protein design is undergoing a similar "personalization" process—when conversational agents like MatwingsVenus™ ™ package AI design capabilities, automated wet experiments, and expert intelligence into a "shared laboratory," protein R&D is no longer the exclusive privilege of a few large platforms, but infrastructure accessible to every researcher and entrepreneur with ideas.

As some industry analysts have pointed out, the essence of this model is shifting from "selling software" to "selling services/selling molecules," solving the "last mile" delivery challenge in protein design. The unique positioning of Chinese companies as "industrialized bodies" in the AI protein track may perfectly mark the critical turning point in this industry's transition from "showcasing technology" to "implementation."

In 2020, the emergence of AlphaFold2 amazed the entire biology community. Today, the story of AIforProtein is entering an even more exciting chapter—it is no longer just a tool for predicting structures; it is becoming a true productivity engine capable of "creating things." And when this engine becomes affordable and user-friendly, we may truly welcome the "iPhone moment" of protein design.

"Allowing everyone to develop the protein they want." —This is no longer just a slogan, but a reality being proven.

THE END