Protein AI: A Key Step Towards Industrialization

 The Technological Revolution in Proteins

Proteins are the main executors of life’s functions. From enzymes that catalyze metabolism, to antibodies that fight disease, to fibers that make up muscles, each protein carries out its unique role with a precise three-dimensional structure. However, obtaining an ideal protein that meets industrial or medical needs has long been a lengthy, highly unpredictable trial-and-error process.

When artificial intelligence deeply entered this field, everything changed. Protein AI—a technological revolution powered by computing, algorithms, and biological data—is pushing our understanding of proteins from passive "decoding" to active "creation." And the hybrid platforms connecting virtual design with real-world modification are becoming the key to breaking new ground.

1. From sequence to structure to function: How AI fully deciphers proteins

 Number of PDB testimonies per year

However, faced with the reality that there are over 254 million known protein sequences worldwide, scientists still know very little about the three-dimensional structures of these molecular machines. As of January 2025, the globally authoritative biological structure database, the Protein Data Bank (PDB), has only about 220,000 experimentally determined protein structures, less than one-thousandth of the known sequences. In 2025, the total number of structures submitted to the PDB by the global research community historically surpassed 20,000, but only about 17,600 were ultimately released to the public that year, while the AlphaFold database already had over 241 million predicted structures. This massive gap highlights a core bottleneck: traditional experimental methods are not only time-consuming and costly but also can never keep up with the speed at which nature designs proteins.

What really breaks this deadlock is artificial intelligence. In 2020, AlphaFold2 achieved a revolutionary breakthrough in structure prediction at the CASP14 international assessment, with a median GDT_TS score of 92.4 and main-chain precision reaching sub-angstrom level (0.96 Å), marking the arrival of atomic-level accuracy in protein structure prediction, comparable to experimental methods. Beyond structure prediction, protein language models have also opened a new front in another dimension. Large-scale pre-trained protein language models (like the ESM series, Tranception, etc.) achieve ROC-AUC scores of 0.85-0.95 in predicting mutation effects across multiple deep mutational scanning (DMS) benchmark datasets, significantly outperforming traditional statistical potential methods. The rapidly evolving computational paradigm is pushing protein engineering from a time-consuming trial-and-error process to a predictable, designable, rational engineering era.

Protein Language Models (PLMs). Scientists have found that protein amino acid sequences bear an astonishing structural similarity to human languages: the 20 standard amino acids are like letters, which arrange into local secondary structures (α-helices, β-sheets, etc., similar to phrases), further folding into complete proteins with specific three-dimensional structures and biological functions (similar to chapters).

By self-supervised pre-training on large-scale protein sequence databases, protein language models can understand the "grammar" and "semantics" of proteins, uncovering statistical rules and evolutionary patterns between amino acid sequences, protein structures, and functions. These models can not only predict protein function and three-dimensional structure but also "design" entirely new amino acid sequences based on functional requirements.

2. From Understanding to Creation: Protein Design Enters the Generative Era

Being able to understand structures is just the first step. The more disruptive capability of AI in protein science lies in creating something out of nothing.

By using cutting-edge generative AI methods like diffusion models and protein language models, researchers can now directly "design" 3D structural scaffolds that meet specific functional requirements, and then reverse-engineer brand-new amino acid sequences to encode them — sequences that might never have existed in nature. Whether it’s highly efficient enzymes that can break down plastics, or smart antibodies that precisely target the tumor microenvironment, AI is transforming protein design from screening limited combinations of mutations into an efficient, intelligent search and creation process across a vast sequence space.

However, a core bottleneck lies between the virtual and the real: perfect structures in a computer may not necessarily express efficiently or work stably in test tubes or cells. Relying purely on computational predictions often can't accurately capture thermal stability, solubility, and catalytic activity in real biological environments, causing many beautifully designed "digital proteins" to remain at the conceptual stage.

3. Dry-Wet Closed Loop: How MatwingsVenus™ Turns AI Design into Reality

Connecting the last mile between "silicon-based design" and "carbon-based implementation" requires a completely new way of working. This is exactly the core problem that Shanghai Matwings Technology aims to solve with its MatwingsVenus™ (XiaoWu™) platform.

MatwingsVenus™ (XiaoWu™) is not just a simple prediction software; it’s a "Design-Build-Test-Learn" (DBTL) closed-loop system that deeply integrates AI algorithms with automated high-throughput experiments. Its workflow is clear and efficient:

AI Design Engine: Integrates multiple in-house deep generative models and physics-based energy calculation modules to produce a large number of candidate protein sequences in parallel for specific targets (like improving enzyme activity, enhancing thermal stability, or modifying substrate specificity), intelligently exploring a vast sequence space.

High-Throughput Build and Test: The automated wet lab module constructs thousands of designs in expression systems in a very short time and completes multi-dimensional functional tests such as activity and stability, generating a throughput of real screening data that was unimaginable with manual operations.

Data Feedback and Iterative Evolution: Every high-throughput test result is fed back to the AI model in real time, continuously evolving its understanding of the intrinsic "sequence-structure-function" relationships. After several iterations, the platform can quickly converge on the optimal solution from a huge pool of possibilities.

This seamless connection between the "dry" and "wet" sides completely changes the trial-and-error logic of traditional protein engineering. Projects that used to rely on expert experience and months or even over a year of manual experiments could potentially be shortened to just a few weeks on the MatwingsVenus™ (XiaoWu™) platform, showing a clear success advantage in complex tasks with multiple targets and constraints. Whether it’s reshaping active sites of industrial enzymes or systemically optimizing antibody fragments for developability, MatwingsVenus™ (XiaoWu™) provides one-stop empowerment from concept design to the delivery of high-performance mutants.

4. Heading Towards the Vast Universe of Bio-Intelligence Manufacturing

Protein AI Agent

The ultimate goal of AI in protein science goes far beyond just tweaking individual molecules. What it really brings is industrial enzymes with catalytic efficiency far beyond traditional chemical catalysts, biomaterials that can withstand extreme environments and be reused, protein drugs that can accurately identify and eliminate pathogens, and entirely new metabolic pathways in engineered microbes. 

Based on the MatwingsVenus™ (Xiaowu™) Agent, Matwings Technology is working with partners in pharmaceuticals, fine chemicals, diagnostics, and other fields to turn these visions into reality. When rational design is backed by the power of high-throughput validation, what we’re really kicking off is a new era of “bio-manufacturing” where designing life components can be as easy as writing code.