AI Protein Research: From 'Reading Heavenly Books' to 'Writing Code,' What Is Life Science Going Through?

If we compare life to a precise machine, proteins are the core parts in the machine that each perform their own roles: structural scaffolds, catalytic enzymes, signal switches, and carriers for transporting substances. The human genome encodes about 20,000 protein-coding genes, and with post-translational modifications and alternative splicing, the actual types of functional proteins far exceed this number. Each type has its own unique three-dimensional structure and biological function.

Over the past few decades, scientists around the world have been "decoding" the book of life that is proteins—gene sequencing, structural analysis, functional verification—all through hands-on experiments. The involvement of artificial intelligence has now pushed protein research from an age of reading into an era of programmable writing. This is also currently one of the most transformative fields in life sciences: AI-driven protein research.

I . So, what exactly is AI protein research studying?

Three Key Research Areas in AI Proteomics

AI protein research relies on big AI models, deep learning algorithms, multiple sequence alignment, molecular dynamics simulation, and other technologies to create an intelligent research system covering the full process of protein structure analysis, functional optimization, de novo design, and scenario validation. Unlike the traditional R&D approach that depends on trial-and-error experiments, this method achieves a brand-new research logic: "algorithm-driven predictions first, precise experimental validation, and iterative data feedback."

The industry usually divides AI protein research into three core directions:

1. Protein structure prediction — "seeing" the protein’s shape

This was the first breakthrough AI achieved in the protein field. In 2021, AlphaFold 2 predicted single-protein structures with near cryo-EM accuracy in the CASP competition, solving the protein folding problem that had troubled structural biology for nearly 50 years. In 2024, the Nobel Prize in Chemistry was awarded to David Baker and Minkyung Baek for protein design, and Demis Hassabis and John Jumper for protein structure prediction, officially marking this field’s top recognition in science.

Today, structure prediction has moved from a "single breakthrough" to "comprehensive coverage": from single-chain proteins to protein complexes, protein-ligand complexes, protein-nucleic acid complexes, and even protein conformational dynamics — AI is gradually revealing the complete picture of protein structures. A 2026 review published in Communications Biology points out that structural biology is entering a "second wave": full-scope protein conformation prediction and routine de novo design of high-affinity protein complexes are emerging as two entirely new frontier directions.

2. Protein function prediction — "understanding" what the protein does

Knowing the structure is just step one; you also need to know what it actually does.

AI function prediction aims to answer questions like: Who does this protein bind with? What reactions does it catalyze? Under what conditions is it stable? What side effects might it have? Traditional methods rely on expensive and time-consuming experiments, while AI can quickly infer protein functional properties from sequence and structural data — from enzyme catalytic activity to antibody affinity, and even protein thermal stability and solubility.

The rise of Protein Language Models has led to breakthrough progress in this area. These models are trained on billions of protein sequences, learning the "grammar rules" of amino acid sequences. Just like large language models understand text, they can understand proteins — predicting functions, stability, and even interactions without explicitly knowing the structure.

3. De Novo Protein Design — 'Creating' Entirely New Proteins

This is the highest level of AI protein research and also the most imaginative direction: designing completely new proteins from scratch that have never existed in nature but possess specific functions, without relying on any known natural protein templates. In the industry, this is called De Novo Protein Design. The logic is reversed: first define the function you want, and AI directly generates the amino acid sequences that fulfill that function.

Since 2025, progress in this area has been particularly rapid. Research teams from Stanford University and SLAC National Accelerator Laboratory have developed a fast protein design method, deriving two completely different new proteins from a single designed protein — one of which is the most active artificially designed enzyme to date. Meanwhile, a team at Nanjing University used AI to design a super-stable protein called 'SuperMyo,' whose mechanical unfolding strength is more than five times that of its natural counterparts, with a melting temperature above 100°C, remaining structurally and functionally intact even at the extreme high temperature of 150°C, with their work published in Nature Chemistry.

These advances mean that AI protein research has moved from 'analyzing the existing' to 'creating the unprecedented.'

II. Three Disruptive Industrial Changes Brought by AI Protein Research

Technological progress ultimately manifests in industrial value. The significance of AI protein research isn't just the number of papers published; it's how it truly improves the efficiency of protein R&D, reduces costs, and expands possibilities.

1. Screening Logic Shift: From 'Needle in a Haystack' to 'Directed Design'

The core pain point in traditional protein engineering is the 'needle in a haystack' problem. A protein of 300 amino acids has 20³⁰⁰ possible sequence combinations — a number far beyond the total atoms in the observable universe. Natural random mutations are like searching for a specific star in the universe, with an extremely low success probability.

AI fundamentally changes the search logic. It doesn’t 'try randomly'; it 'searches with direction' — by deeply understanding the relationship between protein sequence, structure, and function, it directly pinpoints 'promising' regions in the huge sequence space, reducing the number of experimental candidates from hundreds of thousands to just dozens or even a few.

Industry estimates suggest that the experimental success rate for traditional protein modification is roughly one in a thousand. In AI-assisted design projects, the effective hit rate of candidate proteins can increase by one to two orders of magnitude, reducing the R&D cycle from years to months and achieving a leap in efficiency.

2. Expanding performance boundaries: breaking the limits of natural protein evolution

Natural proteins are the result of billions of years of evolution, but evolution’s goal is 'survival,' not 'industrial production' or 'human disease treatment.' Natural proteins are usually only suited to mild conditions like normal temperature and physiological pH, while industrial manufacturing and innovative drugs often require proteins to withstand extreme conditions like high temperature, strong alkalinity, or high organic solvent concentrations.

AI-driven de novo design breaks this limitation. Since it doesn’t rely on natural templates, it can create stable structures and catalytic functions that simply don’t exist in nature. For drug targets that are hard to develop with traditional methods—like the 'smooth proteins without binding pockets,' which have smooth surfaces and lack typical binding sites—AI can directly design proteins with high affinity, expanding the range of targets accessible for drug development.

3. R&D model: from 'expert-dependent' to 'platform-based production'

Traditional protein engineering relies heavily on the experience of senior experts. A skilled protein engineer often needs more than ten years of training, and the supply of talent has long limited industry expansion.

AI industrial platforms are changing this. By embedding general industry expertise and mature algorithm workflows into the platform system, researchers can use AI tools to independently handle most protein modification and design tasks. This lowers the industry's entry barriers and accelerates the pipeline expansion for biotech, pharmaceuticals, and synthetic biology companies.

III. The technical foundation of AI protein research: not just one model, but a whole system

AI Protein Three-Layer Technology Platform

People generally misunderstand that AI protein research is the same as a single structure prediction model. In fact, a complete industry-level AI protein system is divided into a three-layer progressive architecture:

First layer: foundational model layer. This includes protein structure prediction models, protein language models, diffusion generation models, and so on. These models serve as the 'base,' trained on massive amounts of data to gain a general understanding of proteins. Currently, industry-leading technical approaches include structure generation based on diffusion models, sequence generation based on large language models, and hybrid solutions that combine both.

Second layer: specialized tool layer. On top of the foundational models, there are numerous professional tools optimized for specific tasks—affinity prediction, stability prediction, druggability assessment, signal peptide design, codon optimization, expression prediction, and more. Every subproblem has its own corresponding algorithm and tool.

Third layer: closed-loop system layer. This is the most critical and also the hardest layer. Individual models and tools can only solve computational problems, while actual protein research requires a complete 'design–synthesis–expression–purification–detection–feedback' loop. Whether an AI-designed protein can be expressed, is active, or stable enough ultimately needs to be verified in wet lab experiments.

A 2025 review in Nature Reviews Bioengineering clearly states that AI is transforming protein engineering from random trial-and-error into a predictable and reproducible engineering discipline, with the core support being the 'AI design–experimental validation' closed-loop system.

According to various industry research institutions, the global AI protein design market is estimated to be around $500 million to $1.5 billion in 2025 (the difference comes from statistical scope: a narrow scope includes only AI design software and services, while a broad scope covers the entire protein engineering-related market), and it is expected to maintain a 15–20% compound annual growth rate over the next few years. The underlying driver of this rapid market growth is the strong demand from the industry for a complete dry-wet closed-loop R&D system.

IV. Industry Practice: Exploring the Dry-Wet Closed Loop on the MatwingsVenus™ (Xiaowu™) Platform

Amid the global wave of AI-driven protein research, Chinese companies are swiftly catching up and developing their own unique approaches. Taking the MatwingsVenus™ (Xiaowu™) platform from Shanghai Matwings Technology as an example, domestic companies are following a full-chain path of 'AI design, automated experiments, and industrial application.'

The core idea of the MatwingsVenus™ (Xiaowu™) platform is to upgrade AI protein research from being a 'tool-based approach' to becoming a 'dry-wet closed-loop R&D infrastructure.' Specifically, this is reflected in three levels:

First, conversational interaction lowers the entry barrier. The platform uses a conversational intelligent agent structure, allowing users to describe their needs in natural language. The agent automatically handles task breakdown, tool calls, solution design, and result presentation. This enables experimental staff who are not familiar with computational biology to quickly use AI for protein design without having to learn complex algorithms or software operations.

Second, an integrated toolchain covers the entire workflow. The platform combines over 200 specialized protein design tools and large-scale functional annotation databases. From structure prediction and sequence design to property assessment, mutation optimization, and druggability analysis, it covers the main aspects of protein research. Users can complete the entire design workflow on a single platform without switching between different tools.

Third, the dry-wet closed loop ensures practical application. More importantly, the MatwingsVenus™ (Xiaowu™) platform seamlessly connects cloud-based AI design with automated wet-lab experiments—the AI-designed sequences can directly drive synthesis, expression, purification, and functional testing in automated labs. Experimental results are fed back into the model to optimize the next round of design. Each experiment improves AI accuracy, and each iteration boosts R&D efficiency.

So far, the platform has proven this model effective in multiple directions: In a de novo binder design project targeting immune regulatory receptors, dozens of AI-designed molecules showed clear cell-blocking activity in vitro. In a project modifying the sweet protein Monellin, multiple rounds of iteration increased the sweetness of several samples by more than tenfold compared to the wild type, while maintaining heat resistance around 75°C.

These cases show that AI protein research is not just a concept in academic papers—it’s already generating tangible industrial value in the practice of Chinese companies.

V. The Next Stop for AI Protein Research: From Tool to Infrastructure

Looking back at the development of AI protein research over the past decade, you can see a clear evolution path:

Late 2010s: Deep learning started being used for protein structure prediction, with accuracy steadily improving

2021: AlphaFold 2 achieved a breakthrough in structure prediction — "AI can see proteins"

2023–2024: Diffusion models and generative AI entered the protein design field — "AI can design proteins"

2025–2026: The closed-loop system of AI design and automated experiments gradually matures — "AI can deliver functional proteins"

Each step moves forward, gradually pushing AI protein research from "paper achievements" toward "industry infrastructure."

Of course, challenges still exist. A perspective article published in Nature Machine Intelligence in 2026 pointed out that the interpretability of protein language models urgently needs improvement — it cannot be just an afterthought. In addition, issues like data quality, experimental costs, and talent cultivation also require the industry to face them together.

But the direction is clear. As AI gets better at understanding proteins, as experimental validation becomes more automated, and as the design-validation loop becomes more efficient — protein R&D is transforming from a craft dependent on experience and luck into an engineering discipline that is predictable, reproducible, and scalable.

For China's life sciences industry, this is not only an opportunity for technological upgrade but also an important lever for achieving self-controlled supply chains. From antibody drugs to industrial enzymes, from vaccines to synthetic biology — every field that relies on protein performance could see systemic efficiency gains thanks to advances in AI protein research.

The story of AI protein research has only written its first chapter, but its final chapter may be far grander than we can imagine today.