How can protein structure prediction become a business engine in the AI era?

In 1958, British biochemist John Kendrew solved the first three-dimensional protein structure — myoglobin — using X-ray crystallography. Sixty-eight years later, the global Protein Data Bank (PDB) has archived over 240,000 experimentally resolved structures, and AI-predicted protein structure databases have ballooned to over 200 million, covering almost all known organisms’ proteomes.

In the 2020 CASP14 competition — the 'Olympics' of global protein structure prediction — AI methods achieved a median Global Distance Test (GDT) score of 92.4 for 87 target domains. GDT is scored out of 100, and a score above 90 means the predicted structure deviates from the experimental structure by less than roughly 2 Å, considered 'experimental-level accuracy.' This breakthrough is seen as essentially solving the challenge of predicting the static three-dimensional structure of single-chain monomer proteins and directly contributed to the awarding of the 2024 Nobel Prize in Chemistry for the groundbreaking contribution of AI in protein structure prediction.

According to different market research organizations, the global market for protein structure prediction tools is estimated at $400–500 million in 2025 and is expected to grow to $2.4–3 billion by 2032, with a compound annual growth rate of nearly 30%. This article clarifies three things: why protein structure prediction matters, how AI cracked it, and how this capability is turning into a business.

1. Why do we need protein structure prediction?

Proteins are the core molecular machines that carry out life’s functions. Whether an antibody can precisely target a cancer cell, an enzyme can efficiently catalyze reactions in industrial reactors, or whether a receptor can be activated by a drug molecule — all these 'can it or can’t it' questions are written in the protein’s three-dimensional folding. Knowing the structure is like having the blueprint to this 'molecular lock' — drug design, enzyme engineering, and antibody optimization all get a solid anchor.

For decades, scientists have relied on three experimental methods to resolve protein structures — X-ray crystallography (requires high-quality crystals, which is very difficult for membrane and flexible proteins), nuclear magnetic resonance (limited to small proteins under ~30 kDa), and cryo-electron microscopy (whose applicability has greatly expanded with recent resolution breakthroughs, though very small proteins still pose challenges). From protein expression and purification to structure resolution, each structure takes months to a year on average, costing tens of thousands to hundreds of thousands of dollars.

As of 2025, the PDB has about 240,000 experimental structures — although this number sounds large, it covers less than 0.01% of the hundreds of millions of known protein sequences. This is a huge information gap: the cost of gene sequencing has dropped from hundreds of millions in 2000 to a few hundred dollars today, causing an explosion in sequence data, while the throughput of structure resolution cannot keep up. Filling this gap is protein structure prediction — directly calculating the three-dimensional structure from the amino acid sequence.

2, the Three Traditional Prediction Approaches: Each Has Its Limits, Hard to Bridge the Structural Data Gap

Protein structure prediction algorithms are mainly divided into three types.

Homology modeling is the classic method—when the target protein has high sequence similarity with a protein of known structure, it uses the known structure as a template for prediction. It's quite accurate but heavily relies on templates, and it can't do much for proteins that lack a suitable template.

Ab initio physical modeling relies on molecular force fields and thermodynamic and physical parameters to simulate the spontaneous folding process of proteins. In theory, it applies broadly, but the computation is enormous and the accuracy usually doesn't match template-based methods.

Machine learning-based modeling is the most revolutionary direction in recent years. Deep neural networks learn the mapping between amino acid sequences and 3D structures by training on large-scale sequence-structure pairing data—including capturing co-evolution signals (which positions' amino acids 'change together,' suggesting they're close in space). These methods not only get close to experimental accuracy in homology modeling scenarios but also show unprecedented ability in predicting proteins without templates.

What really made this path 'work' was that competition in 2020.

 Technological evolution

3. The AI Era: A Paradigm Shift in Structure Prediction

In the 2020 CASP14 competition, a brand-new deep learning method achieved a median GDT score of 92.4 across 87 evaluation domains—"A breakthrough in predicting the static structures of single-chain monomer proteins, solving a problem that has puzzled the scientific community for over fifty years." (Current Protein & Peptide Science, 2025)

The essence of this breakthrough is the shift from "physical simulation" to "data-driven" approaches. Looking back from the initial glimpses in CASP13 to the historic leap in CASP14, and then to the structural database expanding to about 214 million predicted structures, the impact of this paradigm shift goes far beyond academia. Its core significance is that structural information is transforming from a "scarce resource" into "scalable infrastructure"—no longer the exclusive weapon of a few structural biology labs.

But solving a scientific problem is just the beginning of the story. What's really important is: what value can this capability create?

4. From 'Predicted Structures' to 'Creating Value': Three Dimensions of Commercial Transformation

 Three Major Commercial Applications

The protein structure forecasting market is expected to reach a scale of several billion dollars by 2032, with a compound annual growth rate approaching 30%. Behind this growth rate is the strong demand for structural information in fields such as biomedicine, synthetic biology, and enzyme engineering. Specific business conversion focuses on three dimensions.

Dimension One: Let drug development "look at blueprints first, then start construction"

Small molecule drug design has traditionally relied on high-throughput screening—casting a net among hundreds of thousands to millions of compounds. If the three-dimensional structure of the target protein is known, structure-based drug design (SBDD) can be conducted: precisely analyzing the shape and key residues of the active pocket, and targeting or screening complementary compounds.

Breakthroughs in protein structure prediction mean that many targets previously unable to enter the SBDD process—especially hard-to-crystallize membrane proteins (GPCRs, ion channels, etc.)—now have high-precision prediction models for "visual-based design."

The deeper impact lies in the field of antibody drugs. A 2025 review published in mAbs noted that AI methods have made breakthroughs in antigen-conditioned antibody design—the compute-generated antibodies have been experimentally validated to bind to specific targets. A 2024 review published in Briefings in Bioinformatics further points out that deep learning has been applied to the skull-based skeleton generation and binding interface optimization of antibodies. As of August 2025, 144 antibody drugs have been approved by the FDA worldwide, 1,516 candidate molecules are in clinical development, and global antibody therapy market sales are expected to exceed $267 billion in 2024. Every week of research and development time saved by accelerating antibody discovery with protein structure prediction corresponds to tens of millions in capital costs and market opportunities.

Dimension Two: Transforming Enzyme Engineering from "Blind Men Touching an Elephant" to "Precision Surgery"

One of the core challenges in industrial enzyme development is stability—natural enzymes often lose their inactivation rapidly at high temperatures, extreme acids and bases, or organic solvents. Traditional methods use random mutations and directed evolution to screen for more stable variants, but the vast majority of random mutations are non-functional, with a very low true positive rate.

When protein structure prediction can accurately present the enzyme's three-dimensional structure, rational design becomes possible: structural models tell you which residues are located in the hydrophobic core (mutations directly affect folding stability), which form salt bridges or hydrogen bond networks (which introduce new stabilizing interactions), and which are located at the active site entry points (mutations may alter substrate specificity). AI goes a step further—not only predicting the structure of the current sequence, but also searching in sequence space for variants that "fold into similar structures but with better performance," directly outputting a list of candidate mutations and prioritizing them, compressing traditional periods in months into days.

Dimension Three: Making Protein-Protein Interactions 'Designable' Interfaces

The core pathological mechanisms of many diseases involve abnormal protein-protein interactions (PPIs). But PPI interfaces are usually large and flat (1500–3000 Ų, compared to small molecule binding pockets which are typically only 300–500 Ų), making them extremely difficult to target with small molecule drugs.

A 2025 review published in *Trends in Biotechnology* pointed out that machine learning methods based on co-folding and atomic graphs can now predict "de novo" PPIs that don’t exist in nature – including antibody-antigen complex structure prediction and molecular glue-induced PPI modeling. When PPIs can be predicted and designed, a large batch of targets traditionally considered "undruggable" might find new inhibition strategies through AI-designed binding proteins.

5. Beyond Prediction: Closed-Loop Design from Structure to Function

 Outlook for the Future

When structure prediction is no longer a bottleneck, the industry's next question naturally shifts to: who can take this capability out of the server and directly generate proteins with real industrial value?

On April 24, 2026, Matwings Technology officially released the conversational protein research and development AI agent — MatwingsVenus™ (Xiaowu™). Its positioning is not just as a 'prediction tool,' but as an agent-centered, one-stop protein R&D platform.

At the structure prediction level, the platform integrates over 200 protein design tools and supports retrieval of tens of billions of real-labeled protein data. Users only need to input their task objectives in natural language, and the system can automatically break down the task, coordinating structure prediction, functional analysis, sequence design, and more, completing the full workflow from deep research, enzyme mining, and directed evolution to de novo design.

Even more groundbreaking, MatwingsVenus™ (Xiaowu™) directly links AI design capabilities to automated shared laboratories. Once the agent completes a design, the platform can automatically feed the results into plasmid ordering and experimental planning workflows, guiding robots to carry out sample preparation, protein purification, and functional testing. Finally, the experimental results flow back into the next round of AI design — forming a 'computationally driven wet experiment, wet experiment feeding back into computation' closed-loop iteration. This model shifts protein R&D from being 'driven by expert experience and massive experiments' to 'AI-assisted, data-closed-loop driven.'

Currently, the platform has completed full-process validation in multiple real projects. In a de novo design project targeting an immune regulatory receptor, MatwingsVenus™ (Xiaowu™) used the target structure and functional requirements as input to automatically handle backbone screening, interface design, and sequence optimization. Verified by automated experiments, it successfully generated dozens of new binding molecules with in vitro cell-blocking activity, completing the full closed-loop process of de novo binding molecule design. In another directed evolution project for a protein sweetener, the platform used continuous iterations of 'AI design — automated experiment — AI feedback — redesign,' producing candidate variants with sweetness several times higher than the wild type while maintaining high heat resistance around 75°C.

If AI structure prediction solved the 'seeing' problem, MatwingsVenus™ (Xiaowu™) represents connecting 'seeing' with 'design' and 'validation' into a closed loop — not only calculating what the structure looks like, but generating molecules that actually do something, making them in the lab, testing them, and then optimizing them again. This is the full-fledged form of protein structure prediction as a commercial infrastructure.

6. Looking Ahead: The Next Frontier of Structure Prediction

The story of protein structure prediction is far from over. Right now, two frontiers are redefining the field: first, moving from 'taking a static photo' to 'shooting a dynamic video'—predicting how proteins switch between different states, rather than just freezing one conformation; second, shifting from 'modifying natural proteins' to 'designing from scratch on demand'—given a target, directly creating proteins that don’t exist in nature but can bind to it precisely.

As researchers in the field have pointed out, dynamic ensemble prediction and programmable protein design could represent two 'AlphaFold moments' happening in biology.

The ultimate goal of protein structure prediction has never been 'prediction' itself—it’s about using the understanding of structure to ultimately achieve precise design and modification of function. From 'seeing' to 'creating,' this technological revolution built over sixty-plus years and breakthrough in a single moment has only just set sail on its commercial journey.