Breaking Down the Walls of the Laboratory: How AI Reshapes the Platform Logic of Biological Research

If we were to condense the past hundred years of biotechnology into a single painting, the most striking background color might not be data, but trial and error. From the completion of the human genome draft to the revolution in protein structure prediction, new technologies have repeatedly knocked on the door of scientific research, yet the most fundamental paradigm—identifying problems, conducting experiments, observing results, and then redesigning—has hardly changed.

From Traditional Lab to AI-Powered Biological Research

Slowness is an innate characteristic of biological research.

For a drug target protein, going from the initial concept to obtaining the first candidate molecule via traditional processes can take months or even years, involving countless rounds of wet lab screening and failures. Even in the most "predictable" field within the entire chain—protein engineering—training a researcher capable of independently modifying proteins takes five years, and successfully designing a protein suitable for industrial application often takes another five years. Beyond time costs, bottlenecks in efficiency and success rates have long been long-standing pain points in the industry: with traditional methods, finding mutants that meet requirements through trial and error experiments has a success rate of only about 1%, with experimental samples often numbering in the thousands.

If the first wave of AI represented by AlphaFold solved the problem of "seeing what a protein looks like," today we are entering a new phase—where AI can not only see clearly but also directly address the functional question of "whether this protein can be used."

It is against this background that a quietly occurring structural change is reshaping the infrastructure of biological research: the "biological research platform" is evolving from a "dedicated tool" into "shared infrastructure."

1. Where is the 'last mile' of biological research?

To understand this transformation, we first need to return to a fundamental question: why has biological research always been so difficult?

Proteins are the core executors of life activities. They are composed of 20 types of amino acids arranged in combinations. A medium-sized protein contains hundreds of amino acid residues, and its sequence combination space expands exponentially—finding a protein sequence that meets a specific functional requirement is no easier than finding a planet suitable for human habitation in the vast universe.

Traditional protein engineering roughly follows two approaches: one is rational design, which involves purposeful site-directed modification based on the understanding of protein structure and function; the other is directed evolution, which mimics the natural 'mutation-selection' cycle in the laboratory and approaches the target function through multiple iterations. The bottleneck of the former lies in our still limited understanding of proteins; the pain point of the latter is that it is too time-consuming and costly, and high-quality mutants are extremely rare in the vast mutational space.

A more crucial problem is the connection. Computational design produces a batch of candidate sequences, but who will verify them? The traditional process is like this: after scientists complete the design on a computer, they manually send the sequence information to an outsourced laboratory or an internal wet-lab team. The process in between involves ordering plasmids, sample preparation, purification, functional testing, and other time-consuming steps with very low tolerance for error before the results are manually entered into the system for the next iteration. If any link in this chain fails, the entire cycle gets infinitely prolonged.

‘Design is for design, verification is for verification’—the invisible wall between computational scientists and wet-lab scientists is the real barrier that needs to be broken to improve the efficiency of biological research.

Digital Design Stream

2. When the platform becomes a new entry point: from "exclusive to big companies" to "usable by individuals"

With this background, let's look at Matwings Technology's recently launched MatwingsVenus™ ™ platform, which gives us a clear logic.

The most fundamental change of this platform is not a more precise AI protein design model, but the integration of the entire biological research workflow into a single conversational interface.

In the words of Liu Hao, founder of Matwings Technology, "Protein design used to be like 'looking for a needle in a haystack,' but now it has been rewritten by AI." MatwingsVenus™ ™ supports retrieval-level real protein labeling data at the tens of billions, integrating over 200 protein design tools, 50 platform-certified domain experts, and more than 30 expert-optimized skills. Users only need to input the task goal described in natural language—for example, "I need a single-domain antibody that remains stable under pH 13"—and the system will automatically break down the task, call the corresponding tools and models, and complete everything from market research and in-depth research to molecular design, prediction, screening, and experimental collaboration.

However, what truly sets this platform beyond the essence of "another AI tool" is its "conversational dry-wet closed-loop" design.

Once the AI completes the protein design, users can choose to place orders directly on the platform. The platform uses its self-developed communication mechanism to automatically import the design sequence into plasmid ordering and experiment scheduling processes, driving the robot to complete a series of wet experiments such as sample preparation, protein purification, and functional testing. Experimental results will be automatically fed back to the AI model as a basis for the next round of optimization. Computation-driven wet experiments, wet experiments feed back into computation—thus forming a truly executable, feedback-free, and iterable R&D closed loop.

Matwings Technology has a vivid description of this: MatwingsVenus™ ™ aims to transform complex R&D capabilities that only large enterprises and research institutes could previously master, into a "shared laboratory" that every individual developer can easily access. This model not only breaks the traditional CRO "fee-per-project" heavy labor model, but more importantly, it achieves standardized, high-throughput protein molecule delivery through an agent platform of "AI design and automated experiments."

Capital feedback supports this judgment: just this March, Matwings Technology completed a Series A financing round of over 200 million yuan, jointly led by PetroChina Kunlun Capital, Yonghua Investment, and Shanghai Future Industry Fund.

3. Not just theoretical talk: What are these real cases saying?

No matter how beautiful the concept is, in the end, it depends on the outcome.

 The MatwingsVenus™ ™ platform has already completed multiple real-world cases in innovative drugs and synthetic biology, with at least two of them being highly convincing.

Case 1: A "targetless" target.

Immune regulatory receptors are an important direction in innovative biopharmaceutical R&D, widely involved in oncology, autoimmune diseases, and inflammatory diseases, with extremely high clinical value and market potential. However, a specific receptor de novo design project faces three challenges: first, the target is novel and lacks similar drug molecules to reference; Second, the target surface is mainly polar, lacking typical "highly drug-like" binding hotspots; Third, natural ligands already possess nanomolar (nM) level affinity, meaning artificially designed binders must "snatch food from the tiger's mouth" to snatch binding sites from highly competitive natural ligands.

Based on the MatwingsVenus™ platform ™, AI agents automatically complete the entire computation process, including skeleton screening, interface design, sequence optimization, and druggability prediction, quickly outputting dozens of high-quality binder design sequences. Most importantly, binder samples prepared by automated experimental platforms performed outstandingly in vitro cell activity tests—dozens of molecules showed clear cell blocking activity and high affinity potential. A target starting from scratch achieves a full integration from design to validation on a single platform.

Case 2: Making sweet protein a "snack" lasts longer.

The protein-type sweetener Monellin is naturally quite sweet, but stability remains a persistent challenge—it is highly sensitive to changes in pH, temperature fluctuations, and long-term storage conditions. The MatwingsTechnology team adopted a multi-round iterative strategy of "agent design—automated experiments—AI feedback—agent redesign," narrowing the search space step by step and gradually optimizing candidate sets.

The final automated experiment results showed that the sweetness of several modified Monellin variants was more than ten times higher than that of the wild type, and their heat tolerance remained at a high range of about 75°C. From the sweetness to high-temperature tolerance, multiple performance indicators optimized simultaneously—this is exactly what traditional directional evolution requires multiple rounds of experimentation to achieve, but the AI platform accomplishes it in a shorter cycle.

4. Faster, More Accurate, Lower Threshold: A New Paradigm of Scientific Research Reshaped by AI

Returning to the question at the beginning of the article: what does a “biological research platform” really mean today?

If you asked a biologist five years ago, “What is the ideal research tool?” the answer would most likely have been: a tool that allows me to fail fewer experiments. But today, as AI protein design technology continues to mature, as automated experimental equipment gradually connects to the cloud, and as hundreds of specialized tools and expert knowledge are integrated into a single interface accessible via conversation—the new answer is emerging:

The ideal research platform is one that frees scientists from tedious repetitive work, allowing them to focus on the creativity and judgment that truly require human intelligence.

Multiple driving forces resonate behind this trend.

At the policy level, the “14th Five-Year Plan” clearly includes biomanufacturing as a key development area, and the State Council’s “Artificial Intelligence Action Opinions” specifically propose empowering biomanufacturing. Eight departments, including the Ministry of Industry and Information Technology, have issued transformation guidelines across four dimensions: strain design, process prediction, and others. As Yu Xuejun, Chairman of the China Biotech Fermentation Industry Association, puts it, the Party Central Committee “requires breakthroughs in key core technologies across the entire chain.”

At the technological level, protein engineering driven by deep learning is accelerating in maturity. This is a common trend across the industry.

At the industry level, the business model itself is being redefined—from traditional CROs charging “per head” to platform-based companies delivering “per molecule.” It took MatwingsTechnology just over three years to successfully deliver more than 30 industrial projects, with clients spanning Fortune 500 companies and multiple domestic listed companies. The application scenarios of their technical services range from innovative drug development to over ten industries including in vitro diagnostics, food and beverages, beauty and skincare, and textiles and detergents.

A more critical milestone is Matwings's collaboration with GenScript Pharmaceuticals, where AI large models increased the alkali resistance of a non-alkali-tolerant single-domain antibody by 4 times, successfully applying it to 5,000-liter industrial-scale production—the world’s first protein product designed by a large AI model and scaled up to 5,000 liters of industrial production.

This marks that the industrial validation of AI protein design has moved beyond the “can it be done” stage and into a new stage of “can it be delivered at scale.”

Industrialized Achievements

5. When Everyone Can "Develop" Their Own Proteins

A few years ago, if you told a graduate student that you could, just like writing a document, use natural language to tell an AI, "I need an enzyme that works at high temperatures," and then receive an experimentally validated candidate molecule within a few weeks, they would most likely think you were talking about science fiction.

Today, this is no longer science fiction.

As Matwings Technology stated when releasing MatwingsVenus™ (Xiaowu™), this platform represents not just the iteration of a tool, but the transformation of R&D capabilities that were previously accessible only to large companies and major research institutes into infrastructure that individuals can also use. From cleaning test tubes and running gels to designing proteins and coordinating automated robots to complete the entire process—the interface of scientific research work is being completely rewritten by AI and automation.

The essence of a biological research platform is evolving from a collection of "an expensive instrument" or "a set of exclusive tools" into a full-stack system that integrates computation, experimentation, and expert knowledge. As more and more APIs are opened, and as more knowledge is condensed into reusable skills, the productivity ceiling in this field will be pushed higher and higher.

For Matwings Technology, this transformation has only just begun. But for any team engaged in protein-related research, an important question to pay attention to is: what kind of platform will you use to design your next protein?