Protein design, what exactly is it?

If the functions of life are carried out by proteins, then protein design is humanity's ability to actively 'write' the code of life—of course, here 'code' is just a metaphor. Proteins cannot really be programmed line by line; rather, we customize their 3D structures and functions by designing their amino acid sequences.

From the earliest point mutations on natural proteins to today's AI-driven de novo design of entirely new proteins that have never existed in nature, the boundaries of protein design capabilities have been continuously pushed for half a century. It is not only a cutting-edge topic in basic biology but also the underlying technology for trillion-dollar industries like biopharmaceuticals, industrial enzymes, and synthetic biology.

1. What is protein design?

Proteins are long-chain molecules made up of 20 kinds of amino acids linked in a specific order. This chain folds into a specific 3D structure, and structure determines function—antibodies can recognize antigens, enzymes can catalyze reactions, collagen can support tissues, all because of their structure.

Protein design has two core levels: first is the inverse folding problem—given a target 3D structure, deduce the optimal amino acid sequence that can fold into it; second is function-driven de novo design—start directly from the desired function (like binding a target or catalyzing a reaction), while generating a completely new protein backbone and corresponding amino acid sequence.

This is exactly the reverse of protein structure prediction:

Structure prediction: Sequence → Structure (given the sequence, tells you what it looks like)

Protein design: Function/Structure → Sequence (tells you what function you want and gives you the corresponding sequence)

Structure prediction is 'understanding,' protein design is 'creating.' The difficulty is on a totally different scale—a protein with 300 amino acids has possible sequence combinations of 20³⁰⁰, far more than the number of atoms in the observable universe.

2. Evolution of three generations of methods

 Flowchart illustrating the evolution of third-generation protein design technologies

The development of protein design has roughly gone through three generations of methods. Each generation doesn’t completely replace the previous one but rather builds on it with new capabilities.

First Generation: Rational Design – Experts’ “Handmade Work”

Time: 1980s to early 2000s

Based on the known protein structures and functional knowledge, a few amino acid sites are targeted for mutation. The advantage is precision and controllability, but it heavily relies on expert experience, with limited sites that can be modified. It can only optimize existing proteins and finding entirely new proteins from scratch is difficult.

Second Generation: Directed Evolution – Simulating Natural Selection

Time: 1990s to present, Nobel Prize in Chemistry 2018

Random mutations to build libraries → high-throughput screening → select better mutants → iterate multiple rounds. The idea is simple: since we don’t know how to design better proteins, we let “natural selection” help us find them.

Directed evolution is very powerful. As long as there’s a sensitive screening method, you can optimize proteins for better performance without deeply understanding the mechanism. But it has obvious limits: high screening costs, low search efficiency, and difficulty in optimizing multiple traits simultaneously.

In actual projects, the two methods are often combined – “rational design for choosing sites, directed evolution for library screening,” also called “semi-rational design.”

Third Generation: AI-driven Design – From “Prediction” to “Generation”

Timeframe: Accelerated in the mid-to-late 2010s, entering a boom period after 2020

AI entering the protein design field roughly went through two stages:

The first stage is 'AI-assisted' — using machine learning models to predict the effects of mutations on protein function and providing direction for directed evolution, reducing the screening from thousands to just dozens.

The second stage is 'AI-generated' — models design entirely new protein sequences from scratch without relying on natural proteins as a starting point. Specify a target, and the model directly generates a protein that can bind to it; specify an active site, and the model directly generates an enzyme scaffold to catalyze that reaction.

This is a qualitative leap — before, we could only 'tweak' natural proteins, now we can directly 'create' proteins that have never existed in nature.

Key milestones:

2021: AlphaFold2 made breakthroughs in structure prediction, providing a reliable tool for design verification.

2022–2023: The David Baker team successively released ProteinMPNN (sequence design) and RFdiffusion (scaffold diffusion generation), forming the standard two-step AI-from-scratch design process.

2024: Nobel Prize in Chemistry awarded for computational protein design and structure prediction.

Since 2025: Continuous breakthroughs in AI antibody from-scratch design and AI enzyme design.

April 2026: AI-driven protein R&D platform company Matwings Technology officially launched its conversational protein R&D agent MatwingsVenus™ (Xiaowu™). By deeply integrating AI design, automated wet lab experiments, and expert knowledge, this platform turns complex R&D capabilities that were previously limited to large institutions into a 'shared lab' that individual developers can easily access, enabling rapid transformation from ideas to products.

July 2026: Nobel laureate Jennifer Doudna’s team published research in Science on AI-designed synthetic RNA-guided nucleases, with some variants reaching or even surpassing the activity of natural enzymes.

3. four core directions

 Distribution Chart of Application Scenarios for the Four Major Proteins

Protein design isn’t just a single technology; it’s a huge technical ecosystem with a wide range of applications. Based on the application scenarios, it can be divided into four categories:

3.1. Therapeutic protein design — the “precision manufacturing” of drug molecules

This is currently the most popular area, covering antibodies, cytokines, fusion proteins, vaccine antigens, and more. Given a disease target, proteins can be designed from scratch to specifically bind to it. Traditional antibody discovery can take months or even years, but AI design can shrink this timeline to just a few weeks.

3.2. Industrial enzyme design — the “catalyst engine” of bio-manufacturing

AI optimization of industrial enzymes is the fastest path to industrialization. With short validation cycles, low regulatory requirements, and a direct commercial value from improved performance — like higher temperature or alkaline tolerance and better catalytic efficiency — these improvements can directly reduce production costs.

3.3. Diagnostic and detection protein design — highly specific “probes”

Fields like in vitro diagnostics, biosensors, and molecular testing need proteins that are highly specific and stable. AI design allows precise control over protein target affinity, thermal stability, and storage stability, quickly responding to new target detection needs.

3.4. Novel functional material protein design — exploring the “protein universe”

This is the most cutting-edge and imaginative area: designing completely new proteins from scratch, without relying on natural protein scaffolds, with entirely new folding patterns and functions — like protein nanoparticles, protein switches, and biomaterials. On April 24, the AI-driven protein R&D platform company Matwings Technology officially launched its conversational protein R&D intelligent agent MatwingsVenus™ (Xiaowu™). By deeply integrating AI design, automated wet lab experiments, and expert insights, the platform turns the complex R&D capabilities once limited to large institutions into a “shared lab” that even individual developers can easily access, enabling rapid transformation from ideas to products.

4. Dry-Wet Loop: The 'Last Mile' of AI Protein Design

The value of AI protein design ultimately depends on experimental validation. No matter how perfect the design looks on a computer, if it can’t be verified experimentally, it has no industrial value.

This is why the industry is increasingly emphasizing the 'dry-wet loop'—tight integration of dry experiments (computational design) and wet experiments (laboratory verification). AI design, experimental validation, data feedback, and model iteration form a complete cycle.

Why is this loop so important?

First, AI design isn’t 100% successful yet, so experimental feedback is needed for verification and optimization. The more data, the more accurate the model—this creates a positive cycle.

Second, the industry needs 'functional proteins,' not just 'sequences.' There are many steps from sequence to function, including expression, purification, and testing.

Third, whoever has high-quality wet lab feedback data will have increasingly accurate models.

Domestic platforms are already exploring this. Take Shanghai Matwings Technology’s MatwingsVenus™ (Xiaowu™) platform as an example: it has built a closed loop of AI design and automated experimental validation. After a user proposes a functional requirement, the platform designs the protein sequence, and candidate sequences are directly sent to the automated wet lab platform for expression, purification, and testing. The experimental results automatically feed back to drive the next AI iteration. Traditional protein R&D projects that would take 2–5 years can be compressed to 2–6 months under this AI loop mode.

This isn’t 'AI replacing scientists'; it’s AI freeing scientists from a lot of repetitive trial and error, letting them focus on more creative scientific problems.

5. Industrial Applications and Market Prospects

The industrial implementation of protein design is moving from 'proof of concept' to 'scaled application.'

Innovative drug field: Therapeutic proteins like AI-designed antibodies, cytokines, and fusion proteins are advancing rapidly. Some AI-designed protein drugs have already reached late-stage clinical trials. In 2026, a breakthrough in the de novo design of nanobodies marks the leap of AI macromolecule drug design from 'theoretical prediction' to 'industrial usability.'

Industrial enzyme field: This could be the earliest track for AI protein design to achieve large-scale commercialization. Validation cycles are short, regulatory requirements are low, and performance improvements directly translate to commercial value. In July 2026, the Doudna team’s AI enzyme design achievements further proved that AI can create functional enzymes beyond natural evolution.

Synthetic biology field: The application of AI-designed enzymes in metabolic pathways is making many previously unfeasible biosynthesis pathways possible. From bulk chemicals to high-value natural products, AI enzyme design is expanding the boundaries of bio-manufacturing.

Looking at market data, according to market research reports, the AI protein design market is estimated at $610.3 million in 2025 and is expected to grow to $718.97 million in 2026, reaching $2.0133 billion by 2032, with a CAGR of 18.59%. The overall protein engineering market is about $4.55 billion in 2025, expected to reach $4.98 billion in 2026, and potentially $9.51 billion by 2032.

And this is just the beginning. Once AI-designed proteins enter large-scale commercial production, the market space will be much larger than it is now—not only because it replaces traditional R&D methods, but more importantly, because it can create things that traditional methods simply cannot, thereby creating entirely new markets.

6.Industry outlook:

Currently, the global protein design industry is enjoying dual benefits from both a technological boom and industrial implementation. With upgrades in AI large-model computing power, continuous iteration of databases, and improved automated experimental systems, protein design will move from 'single-point technology breakthroughs' to 'full-scale industry applications.'

In the future, the core development trends of protein design will focus on two directions: First, extreme intelligence, using AI agents to achieve fully automated end-to-end R&D, further lowering the technical barrier and making customized protein development more widespread; second, deep scenario specialization, developing more highly specific, highly stable, and highly active customized proteins targeting industry-specific challenges, covering more cutting-edge fields.

Domestic AI protein design platforms like MatwingsVenus™ (Xiaowu™) are breaking the overseas tech monopoly thanks to their full-chain, closed-loop, and easy-to-implement advantages. They provide low-cost, high-efficiency, and practical protein R&D solutions for local research institutions, biomedicine companies, and bio-manufacturing firms, helping China's synthetic biology and bio-industry achieve original innovation and even get ahead in the game.

From passively adapting to nature to actively designing life molecules, protein design technology is redefining the boundaries of bio-manufacturing. With AI and synthetic biology empowering each other, this microscopic tech revolution is set to become a key driver of high-quality growth in the global bio-economy.