AI De Novo Protein Design System ProteusAI: Generating Novel Functional Proteins from Scratch Without Natural Templates

AI De Novo Protein Design System ProteusAI: Generating Novel Functional Proteins from Scratch Without Natural Templates

The protein design field is undergoing a paradigm shift from "modifying nature" to "creating nature." On November 27, 2030, the Artificial Intelligence Biology Laboratory at ETH Zurich published a paper in Nature Biotechnology officially releasing the ProteusAI de novo protein design system.

Unlike previous systems such as AlphaFold and RoseTTAFold that predict natural protein structures, ProteusAI aims to design entirely new proteins that have never existed in nature. The system accepts functional descriptions as input—for example, "an enzyme that catalyzes cellulose hydrolysis at 80 degrees Celsius"—and then generates amino acid sequences satisfying that function along with their three-dimensional folding structures.

Technical Architecture

ProteusAI's core is a three-layer generation architecture. The first layer, a "function encoder," converts functional requirements expressed in natural language into mathematical representations. The second layer, a "structure generator," creates protein backbones using diffusion models. The third layer, a "sequence optimizer," finds optimal amino acid sequences while satisfying structural constraints.

Research team leader Professor Anna Mueller revealed that ProteusAI absorbed all 210,000 resolved structures from the Protein Data Bank (PDB) during training, along with over 250 million protein sequences from UniProt. "The key breakthrough is in the third layer," Mueller explained. "Previous systems generated sequences that were theoretically feasible but often could not fold in practice. Our optimizer increased the experimental validation success rate from under 5% to 67%."

Validated Design Results

The paper reports 12 novel protein designs validated through wet-lab experiments, covering three application directions:

Industrial Enzymes: Two thermostable cellulases showing activity three times higher than natural enzymes at 80 degrees Celsius, applicable to second-generation biofuel production.

Drug Target Binding Proteins: Three mini-proteins that precisely bind the PD-L1 protein, with molecular weight only one-tenth that of antibodies, holding promise for tumor immunotherapy.

Self-Assembling Nanostructures: Seven protein cage structures that spontaneously assemble into specified geometric shapes, useful for drug delivery and vaccine carriers.

In the drug target binding protein direction, the ETH team has signed a collaborative research agreement with Novartis. Thomas Weber, Novartis head of protein engineering, stated: "The PD-L1 binding protein designed by ProteusAI showed comparable affinity to existing antibodies in animal experiments, but with much smaller molecular weight and better tissue penetration."

Limitations

The system's main limitation currently lies in membrane protein design. Membrane proteins account for 30% of the human proteome and represent the most important drug target category, yet ProteusAI's design accuracy for transmembrane regions is only 28%. "The membrane protein environment is extremely complex—lipid bilayers, aqueous phases, and interfacial regions each have different physicochemical rules," Mueller acknowledged. "This is our next major focus."

Regarding ethics, critics worry that de novo designed proteins could be used to develop biological weapons. Mueller responded that the system has a built-in function screening module that rejects design requests related to known toxins or pathogens, but admits this screening mechanism is not foolproof.

ProteusAI's academic version code is open-sourced on GitHub, with commercial licensing under negotiation. The ETH team expects to release a specialized optimization version targeting membrane protein design by mid-2031.