June 18 News: The prestigious international AI conference ICML 2025 has announced its paper review results. A major breakthrough was achieved through a joint study by MoleculeMind, a leading AI-driven protein design company founded by Professor Jinbo Xu, known as a pioneer in AI protein folding, and The Hong Kong Polytechnic University. Their research paper, titled Retrieval-Augmented Zero-Shot Enzyme Generation for Targeted Substrates, has been accepted by the conference. This achievement marks a significant milestone in the field of AI-driven enzyme design.
AI-Designed Enzymes: A Critical Step to Break Bottlenecks in the Bioeconomy
Enzymes, nature’s efficient and eco-friendly molecular machines, are central to the trillion-dollar bioeconomy, powering advancements in biomedicine, green chemistry, environmental degradation, and sustainable agriculture. However, naturally evolved enzymes, shaped by billions of years of evolution, are limited to reactions found in nature. As new synthetic chemicals such as advanced plastics, specific pharmaceutical intermediates, and persistent pollutants continue to emerge, the limitations of natural enzymes are becoming increasingly evident. Overcoming these constraints is now a critical challenge in the field of biomanufacturing, making AI-driven enzyme design an urgent necessity.
McKinsey research shows that the lack of ideal biocatalysts is one of the main barriers to scaling up production in the bioindustry. In the pharmaceutical, chemical, and agricultural sectors alone, enzyme limitations lead to annual production losses exceeding $100 billion. Traditional methods for enzyme discovery and optimization—such as directed evolution or rational design—are time-consuming, costly, and heavily reliant on expert intuition. With a success rate below 1%, these approaches are nearly powerless when faced with entirely new substrates, falling far short of the industry’s urgent needs.
In recent years, AI-driven protein design has brought new hope for precise enzyme generation. By learning the structure-function relationships of known enzymes, AI models can generate new catalysts from scratch, enabling the design of enzymes with specific catalytic functions. However, AI training is highly dependent on existing enzyme-substrate pairing data. When it comes to novel synthetic molecules, AI models often struggle due to a lack of relevant training data.
Pioneering Zero-Shot Enzyme Design with AI: A New Chapter in Enzyme Generation
To solve the critical challenge of generating new enzymes without direct catalytic data, MoleculeMind and The Hong Kong Polytechnic University jointly developed an innovative AI-driven enzyme design method called SENZ. By cleverly combining large-scale bioinformatics retrieval with generative AI, this approach enables enzyme generation even in zero-shot settings. The research paper has been officially accepted by ICML.
SENZ abandons the mainstream protein similarity-based generation approach and instead retrieves and designs functionally relevant enzymes based on substrate structural similarity, creating a direct path from “unknown substrates” to “super enzymes.” This method guides AI to extract catalytic patterns from “similar molecules” and reassemble them into “molecular keys” capable of targeting new substrates. It features three major innovations: First, it uses massive global enzyme databases to identify known enzymes that may not directly catalyze the target molecule but whose substrates are structurally similar, providing a “structural blueprint” for design. Second, it introduces a proprietary enzyme-substrate classifier to distill the underlying logic of biological reactions and build a “bioreaction map.” Third, it employs generative AI to design novel enzymes capable of efficiently and precisely catalyzing the target substrate.
To validate SENZ’s capabilities, the research team designed a theoretical enzyme to target a pollutant not found in nature. Methylphosphonic acid is a persistent environmental pollutant for which no effective degradation method or efficient natural enzyme exists. The team compared SENZ’s output with several leading enzyme design approaches, including unconditional generation based on Transformer models and structure-based methods. The results showed that SENZ-generated enzymes significantly outperformed both baselines and natural enzymes.
According to the team at The Hong Kong Polytechnic University, SENZ’s success is akin to giving every chemical compound a custom-made key. This innovation could help biomanufacturing move beyond its dependence on natural evolution. In pharmaceuticals, it enables the rapid design of catalytic pathways for complex or hard-to-synthesize drug molecules, dramatically reducing production costs and speeding up drug development. In environmental protection, it makes it possible to create tailor-made “super enzymes” that efficiently degrade persistent pollutants such as plastics. In biomanufacturing, it supports the sustainable and efficient production of bio-based materials, fine chemicals, and food additives.
Ongoing Innovation: Expanding On-Demand Design Capabilities
SENZ is just one of MoleculeMind’s “molecular key” technologies. The company has developed over a dozen original, industrially valuable AI protein design methods, including NewOrigin, the world’s first multimodal AI foundational model for proteins, and MoleculeOS, the industry’s first full-featured platform for protein prediction, optimization, and design.
Powered by these latest technologies, MoleculeMind has formed deep partnerships with leading companies like Cathay Biotech across biomanufacturing and biopharma sectors to develop urgently needed “super proteins.” In one project, a key enzyme optimized by AI increased microbial yield by 5x compared to the wild-type strain, enhancing commercial output. In another, MoleculeMind helped a pharmaceutical company tackle protein vaccine stability, producing dozens of strong candidate proteins in just three days. These candidates showed improved antibody titers and immune response, while also bypassing existing vaccine stability patents.
Professor Jinbo Xu, founder of MoleculeMind, revealed that over the next three years, the company will expand its on-demand design capabilities across antibodies, vaccines, and industrial enzymes. The goal is to build a next-generation AI engine for molecular design that delivers powerful biological solutions to meet the challenges of health, environment, and sustainable development.
Leave a comment