A Paradigm Shift in Protein Design – From 'Fishing in the Sea' to 'Precise Customization'

When you think of proteins, you might think of eggs, milk, or protein powder at the gym. But to scientists, proteins are the most intricate nanomachines in cells—enzymes that catalyze reactions, hemoglobin that carries oxygen, antibodies that fight viruses, and even keratin that makes up our skin and hair. All of these are made from polypeptide chains folded from 20 different amino acids.

In the past, humans approached protein modification like a game of "blind trial and error": random mutations, massive screenings, relying on time and luck to find improved versions. Today, a brand-new paradigm—"protein design"—is moving us from luck-based approaches to rational strategies, even allowing us to create proteins that have never existed in nature.

1. So, what exactly is protein design?

Protein design

Simply put, proteins are high-molecular chains made up of 20 amino acids, and these chains twist and fold into unique three-dimensional structures. It’s this unique structure that gives a specific protein its biological function. 

The essence of protein design is figuring out a two-way pathway. 

In nature, the information flow is: gene sequence → amino acid sequence → three-dimensional structure → biological function. Traditional structural biology and genomics focus on predicting structure and function from sequences, which is a “forward analysis” approach. Protein design, on the other hand, follows a reverse-engineering path: starting from the desired human function, you first build the ideal 3D structure, then work backwards to determine the amino acid sequence that can fold stably into that structure, and finally synthesize a protein with the specified function. 

The core challenge here is called the “inverse folding problem”: a given protein backbone structure could correspond to an astronomical number of amino acid sequence combinations, but only a tiny fraction of them will fold correctly, remain stable, and carry out their function. The key question protein design aims to answer is how to efficiently and accurately find that ‘needle in the haystack’ in the vast sequence space. 

Creating a high-performing protein product usually requires a lot of expert experience combined with tens of thousands of trial-and-error experiments. With traditional methods, protein design can take 2–5 years, involve tens of thousands of experimental samples, cost tens of millions of yuan, and the success rate in R&D is only about 0.1%–1%.

2. From Physical Equations to Generative AI: The Three Major Technological Steps in Protein Design

Protein design isn’t something that happens overnight; its tech stack has evolved step by step.

Step One: Physical energy functions and classical computation. Early design relied on precise calculations of molecular force fields, simulating van der Waals forces, electrostatic interactions, hydrogen bonds, and hydrophobic effects between atoms. Tools like the Rosetta suite used methods such as Monte Carlo sampling to explore sequence and conformation spaces, achieving the feat of designing entirely new folded proteins from scratch. Its strength lies in having an interpretable physical model, but the computational cost is extremely high. When dealing with large proteins, the sequence space explodes, making convergence difficult.

Step Two: Statistical potentials and co-evolution analysis. To reduce computation, scientists turned to the protein sequence and structure databases already existing in nature. By analyzing co-evolution information of amino acids in homologous proteins and calculating the likelihood of amino acids appearing in particular structural contexts (statistical potentials), faster and more successful sequence design became possible. At this point, design no longer relied solely on first principles but began to learn from the “big data” of evolution.

Step Three: The explosion of deep learning and AI models. Over the past five years, protein structure prediction models like AlphaFold2 have marked AI as a core driving force at the frontier of protein science. Subsequently, diffusion models, protein language models, graph neural networks, and others have been applied to protein design. They can directly learn distribution patterns from multimodal data of sequence, structure, and function. Now, given a specific functional site or target pocket, algorithms can act like a “molecular-level Midjourney,” generating entirely new protein backbones and sequences that nature hasn’t evolved over millions of years, and experimental validation success rates are rapidly climbing.

3. How is AI reshaping protein design?

 Searching Vast Protein Databases

In recent years, the integration of artificial intelligence technology is completely changing this situation.

Data is the core resource driving technological progress. Take Matwings Technology as an example: the protein sequence dataset they built contains nearly 9 billion sequences, with about 500 million sequences labeled with functional tags, clearly indicating how proteins perform under specific temperatures, pH levels, and pressures. These data not only cover regular biological protein sequences but also include special functional sequences from extreme environments like volcanoes and deep-sea trenches, such as heat-resistant or acid- and alkaline-resistant proteins. This huge dataset helps models better understand the relationship between protein sequences, structures, and functions.

Large models trained on this data are enabling a two-way protein design framework, from 'sequence → structure' to 'structure → sequence.' Researchers no longer rely solely on traditional physical calculations and sequence optimization algorithms—methods that can achieve high accuracy in individual systems but are costly and limited in design space. Nowadays, AI models can directly target the core goal of 'function prediction,' turning complex protein design into a streamlined, demand-oriented process that requires only minimal experimental verification.

4. From Theory to Practice: The Diverse Applications of Protein Design

Protein design isn't just some theoretical exercise in the clouds—it’s already showing transformative productivity across multiple fields.

l Custom Industrial Enzymes: Traditional detergent enzymes and feed enzymes need to resist high temperatures, strong acids or bases, and protease degradation. Through protein design, the core structure of enzymes can be globally stabilized while maintaining or even enhancing the active site's efficiency, increasing the enzyme’s half-life under extreme conditions by hundreds or even thousands of times, freeing industrial production from reliance on mild conditions.

l Synthetic Biology Components: Designing signal proteins, switch proteins, and metabolic pathway enzymes that don’t exist in nature provides standardized, predictable building blocks for artificial synthetic cells and bio-manufacturing. This means future bioengineers will be able to 'program' life like stacking blocks.

l Smart Antibodies and Drugs: Proteins can be designed from scratch to bind predetermined epitopes, or existing antibodies can be optimized for affinity, specificity, and immunogenicity in a multi-objective way. Protein design makes it possible to respond quickly to emerging infectious diseases and design dual-antibody or even multi-antibody logic gates.

AI-driven protein design is also thriving across multiple areas.

Take Matwings Technology’s general-purpose protein design AI model AIACCLBIO® as an example: the model gradually masters the complex mapping between protein sequences and functions, enabling rapid and efficient optimization and modification of proteins. The results are impressive: protein design cycles have been shortened from the traditional 2-5 years to 2-6 months, the number of experimental samples reduced from tens of thousands to around 100, project costs dropped from tens of millions to around a million yuan, and the R&D success rate increased to 30%.

In the field of biomedicine, de novo protein design aims to create protein molecules with entirely new structures and functions that don’t exist in nature. The MatwingsVenus™ platform has a case study for a de novo design project targeting an immune regulatory receptor. The platform takes the target's structure and functional requirements as input, then the AI automatically completes all computational steps, including backbone selection, interface design, sequence optimization, and druggability prediction, quickly producing high-quality binder design sequences. In the end, it successfully generated dozens of completely new binder molecules with in vitro cell-blocking activity, completing full verification of the de novo binder design process. Achieving this wasn’t easy—the target was novel, there were no similar drug molecules for reference, and the target surface lacked typical high-druggability binding hotspots, all of which made the design extremely challenging.

In the field of industrial enzyme design, AI-assisted enzyme protein design combines AI technology and automation to analyze the deep relationships between enzyme sequences, structures, and functions, achieving efficient targeted design and performance optimization of enzyme molecules. For example, Matwings Technology and its partners developed enzymes for pancreatitis detection reagents, cutting costs to one-tenth of imported products.

Regarding ultra-stable protein design, Chinese researchers combined single-molecule force spectroscopy, molecular dynamics simulations, and AI protein design methods to successfully create de novo ultra-stable "SuperMyo" protein modules with controllable mechanical properties. The best-performing variant had a mechanical unfolding force exceeding 1000 pN, five times that of the natural titin Ig domain, and a melting temperature above 100°C.

5. All-in-One Platform: Making Protein Design Accessible to Everyone

The technical barriers to protein design are being significantly lowered by the next generation of platforms.

MatwingsVenus™ (Xiaowu™), released by Matwings Technology in April 2026, is a conversational AI platform for protein research and development. Users can simply enter their task goals in natural language, and the system will automatically break down the tasks, deploying the necessary design, prediction, analysis, and screening capabilities to carry out in-depth research, enzyme discovery, directed evolution, de novo design, and automated wet-lab collaboration.

The platform supports searches across tens of billions of real labeled protein data points, integrates over 200 protein design tools, more than 50 experts certified by the platform, and over 30 skill sets fine-tuned by specialists in various fields. Most importantly, the platform implements an intelligent R&D model of 'design as validation, validation as iteration' — deeply coordinating cloud-based design with physical experiments. Once design is completed, the task seamlessly transitions to experimentation, and the results feed back into the next round of AI-driven design, creating a closed-loop cycle where computation drives wet experiments and wet experiments refine computation.

In the platform update of June 2026, MatwingsVenus™ (Xiaowu™) added three core models: BoltzGen, LigandMPNN, and Protenix. With the integration of BoltzGen, the platform can now design binders for multiple types of biological targets, generating candidate binding molecules like proteins, peptides, nanobodies, and cyclic peptides according to protein, peptide, small molecule, and other targets. Notably, related studies show that BoltzGen achieved nanomolar affinity for 66% of targets in nanobody-protein complex design.

6. Outlook

Prospects for Protein Design

The '15th Five-Year Plan' proposal has already listed biomanufacturing as a forward-looking future industry. With the deep integration of AI and biotechnology, protein design is shifting from being an 'art in the lab' to a 'programmable science.' From new drug development to industrial enzyme modification, from synthetic biology to green biomanufacturing, protein design is reshaping how we understand and use life's molecules.

We have reason to believe that protein design is becoming a foundational enabling technology in the bioeconomy era. Just like chip design tools drove exponential growth in the semiconductor industry, mature protein design platforms will completely change the starting point of 'biomanufacturing'—we are no longer just readers of life, but gradually becoming programmers and architects of life.

From 'slow trial and error' to 'efficient and precise design,' the next decade of protein design is just getting started.