David Baker’s AI Protein Design Transforms Drug Discovery and Therapeutics

David Baker’s AI Protein Design Transforms Drug Discovery and Therapeutics

The Shift to Generative Protein Design

Generative AI has moved the field beyond mere structure prediction—the process of determining how existing proteins fold—into the realm of de novo design, where entirely new proteins are built to meet specific functional requirements. Traditional physics-based methods once required years to design proteins, but modern tools such as RFdiffusion and ProteinMPNN can now generate candidate binders and enzymes that reach laboratory testing within weeks. This process mimics image-generating diffusion models: a network begins with random noise and iteratively refines it into a structure with plausible bond geometry and a foldable shape. The design process typically involves a multi-step pipeline: * RFdiffusion: Generates a protein backbone by adapting a structure-prediction network into a diffusion model. It can produce backbones that satisfy specific constraints, such as target binding sites or enzyme active sites. * ProteinMPNN: A graph-based neural network that assigns the optimal amino acid sequence to the generated backbone. * Validation: Because a plausible in silico structure does not guarantee function, researchers use self-consistency checks, often re-predicting the structure of the generated sequence to ensure it remains stable before committing to synthesis via X-ray crystallography or cryo-electron microscopy.

Applications and Therapeutic Potential

The ability to design proteins with atomic precision has significant implications for the antibody drug market, which is valued at hundreds of billions of dollars. Historically, designing antibody loops—the regions responsible for binding—has been challenging due to their flexibility. AI-guided design now allows researchers to construct these loops to bind user-specified epitopes, bypassing time-consuming experimental screens. Research at the IPD has demonstrated the efficacy of these tools. For example, in scaffolding the p53 helix that binds the MDM2 protein, the strongest AI-designed binder bound approximately three orders of magnitude more tightly than the natural peptide it was built to replace. Furthermore, RFdiffusion2 has successfully scaffolded enzyme active sites, producing working retroaldolase and hydrolase enzymes after screening fewer than 96 candidate sequences per reaction.

Industry Integration and the Future of Medicine

The impact of this technology is reflected in the rapid growth of the biotech sector. In 2024, Xaira Therapeutics launched with over $1 billion in funding. The company features a leadership team that includes David Baker as a scientific advisor, former Genentech CSO Marc Tessier-Lavigne as CEO, and a board of directors that includes Nobel laureate Carolyn Bertozzi, former FDA head Scott Gottlieb, and former Johnson & Johnson CEO Alex Gorsky. Despite this progress, Baker notes that the field must remain grounded. While the ability to design proteins on a computer is a reality, he cautions that revolutionizing medicine will require a deeper understanding of biology. “The hype is that for therapeutics, there’s a lot more than the basic activity of a protein binding or catalyzing a reaction,” Baker said.

Beyond Proteins: The Emergence of Virtual Cells

As protein design matures, the industry is looking toward “virtual cell” models to predict complex biological behavior. Experts like Tom Sercu of Biohub emphasize that transformative AI in biology requires large-scale, high-quality datasets. To support this, Biohub has announced a $500 million commitment to the Virtual Biology Initiative. Current efforts to model cellular behavior include: * Xaira Therapeutics: Unveiled X-Cell, a virtual cell model with 4.9 billion parameters, trained on the X-Atlas/Pisces dataset containing 25.6 million cells. * Arc Institute: Developed the STATE model to predict how immune, cancer, and stem cells respond to genetic or drug perturbations. * Ginkgo Bioworks: Through its Virtual Cell Pharmacology Initiative, the company is using high-throughput automation to profile small molecules and generate diverse biological datasets. While current models often focus on transcriptomics, researchers like Hani Goodarzi of the Arc Institute note that cells are complex systems defined by multiple layers, including metabolism, spatial organization, and post-translational regulation. The ultimate goal, according to Sercu, is to build a foundational resource for cellular biology equivalent to the Protein Data Bank (PDB), which has hosted over 253,000 molecular structures over the last five decades.

Find more reporting in our Health section.