Protein research enters the 'programmable era': How is AI cracking the life's code unsolved for 50 years?

Proteins are the main performers of life’s functions. From enzymes that catalyze biochemical reactions to hemoglobin that carries oxygen, to antibodies that fight off pathogens, every protein carries out precise and indispensable tasks in living organisms. Understanding proteins, redesigning proteins, creating proteins—this is the ultimate mission of protein research, and it’s also a key engine in fields like biomedicine, green manufacturing, and new materials.

However, protein research has long faced a huge gap: the complex mapping between sequence, structure, and function, often called the 'second genetic code.' Today, this gap is being bridged by AI in ways never seen before, and China's contribution plays an important role in this profound transformation.

1. Massive sequences and scarce structures: the core dilemma of protein research

  Massive Sequences and Sparse Structures

Proteins are mainly composed of 20 standard amino acids connected in specific sequences, folded into unique conformations in three-dimensional space to function effectively. Nobel laureate Christian Anfinsen once proposed that under in vitro buffering conditions, the primary amino acid sequence of a protein contains all the information that determines its natural three-dimensional conformation. However, accurately predicting three-dimensional structures from one-dimensional sequences and how structure determines function has been a challenge that has puzzled academia for more than half a century.

Currently, the global public database UniProt has recorded over 250 million protein sequences, and the number continues to grow exponentially. However, through experimental methods such as X-ray crystallography and cryo-EM microscopy, about 247,000 high-resolution three-dimensional proteins have been resolved in the PDB (Protein Structure Database). Even in the highly regarded human proteome, the proportion of proteins with complete experimental high-resolution structures was less than 35% in 2021. The vast sequence-structure gap limits our understanding and utilization of protein functions.

2. From trial and error to design: a paradigm leap in protein research

In May this year, the open-source ESMFold2 model released the "ESM Atlas," which contains 1.1 billion predicted protein structures and 6.8 billion protein sequences, with over 800 million more predicted protein structures than the AlphaFold protein structure database. This tool is based on a protein language model, trained on billions of protein data points from the "Tree of Life." In determining the correct structure of interacting protein complexes, its performance surpasses existing methods.

Meanwhile, a research institute developed the atomic interaction generation model Void-X, which uses a bottom-up strategy to directly generate the atomic arrangement at the protein interaction interface. The model has 170 million parameters and achieves an accuracy of 78.3% in intra-chain protein cluster prediction tasks and 68.2% in inter-protein chain atomic cluster prediction tasks. This achievement has been published in the Proceedings of the National Academy of Sciences (PNAS).

More notably, protein research is moving from "predicting structure" to "designing functions." In June this year, a Chinese research team achieved a breakthrough in AI macromolecule drug design, achieving over 90% target success rate under ultra-low-throughput experimental validation (each target only needs to test 14-50 candidate molecules) (data sourced from Tsinghua University). This means protein research is moving from "observation and trial and error" to an era of "prediction and design."

3. The Rise of Chinese Platforms: How MatwingsVenus™ (Xiaowu ™) Reshapes Protein Design

Breakthroughs in cutting-edge technologies are certainly exciting, but for most researchers and enterprise R&D personnel, the real challenge is: how can these advanced technologies truly work for us?

Traditional protein R&D faces multiple challenges: AI design requires computing power and algorithms, while wet experimental verification demands equipment and manpower, often creating a huge gap between the two. Researchers have to rush between different platforms and switch between different tools, greatly reducing research efficiency.

Amid this wave of intelligent protein research, Shanghai Matwings Technology's independently developed intellectual property platform—MatwingsVenus™ (Xiaowu ™)—is driving AI protein design implementation with strong engineering capabilities.

The MatwingsVenus™ ™ platform deeply integrates protein pre-trained large models trained on tens of billions of data with multi-objective optimization algorithms, building end-to-end prediction and generation capabilities from sequences to functions. The platform can simultaneously accurately evaluate and intelligently design dozens of properties such as protein stability, activity, immunogenicity, and expression levels, transforming the previous R&D model that relied heavily on "expert experience and extensive experiments" into a new paradigm of "data-driven, intelligent generation."

In real-world industrial projects, the value of this capability has been validated:

In an industrial transaminase transformation project for a partner, the MatwingsVenus™ ™ platform accurately predicted and designed key mutations, increasing the enzyme's half-life under industrial production conditions by 42 times, while expanding the optimal reaction temperature from 40°C to 65°C, significantly improving catalytic efficiency and process economics.

In antibody drug discovery scenarios, the platform can simultaneously optimize multidimensional attributes such as affinity, stability, and developability, significantly shortening candidate molecular screening cycles and shortening the optimization process from over a year to just a few months.

Compared to traditional methods, the MatwingsVenus™ ™ platform shortens the protein design cycle from the traditional 2~5 years to 2~6 months, significantly reducing the cost of industrialization trial and error. More importantly, it marks a key step toward multi-objective collaborative optimization. Moreover, the platform relies on local data and computing power deployment to ensure the security and autonomous control of R&D data.

4. The Future Is Here: The "Programmable Era" of Protein Research

The Programmable Era of Protein Research

When the relationship between a protein's sequence, structure, and function can be precisely decoded by AI models, protein research enters a whole new 'programmable' era. We can:

- Design more efficient biocatalysts, allowing industrial production to move away from high-energy, high-pollution chemical processes;

- Create more precise antibody drugs, targeting previously undruggable sites while reducing immunogenicity;

- Build entirely new biomaterials and synthetic biology systems, enabling carbon neutrality and sustainable manufacturing.

Shanghai Matwings Technology's MatwingsVenus™ (Xiaowu™) platform is a practitioner of this vision. Through continuously iterated AI models combined with wet-lab verification loops, MatwingsVenus™ (Xiaowu™) is helping an increasing number of biopharma, fine chemical, and synthetic biology companies quickly turn their protein design ideas into real productivity.

Protein research is at a historic turning point, transitioning from 'understanding life' to 'designing life.' AI-powered protein design technology is opening unprecedented possibilities for the bioeconomy.