AI BioDesign operates a runaway feedback loop: an AI designs sequences, the lab tests them in parallel in one or a few tubes, and the results are fed back to the model to generate the next batch, enabling rapid learning from vast data.

Founder Baker targets tangible milestones within 18–24 months and open access within five years, including therapies for new diseases, plastics-degrading enzymes, and bio-computers that consume less power than silicon-based ones.

Open data and tools are a stated goal, with models, datasets, methods, reagents, and benchmarks slated for public release, while acknowledging ecological and misuse risks in synthetic DNA.

David Baker, a Nobel laureate, has long pursued de novo protein design, moving from prediction (as in AlphaFold) to designing proteins with new folds, exemplified by his 2003 Top7 protein.

The project favors building small, specialized AI models trained on massive, experimentally generated datasets rather than a single general-purpose model, arguing the experimental infrastructure is the bottleneck and key to progress.

Experiments are framed as learning opportunities for the model, with success measured by information gained per round and the model’s improved confidence on challenging sequences.

AI BioDesign, a Seattle accelerator backed with $94.6 million from the Paul G. Allen Estate, aims to design proteins and molecules that do not exist in nature by combining AI with high-throughput in-tube testing.

Researchers collect millions of design sequences in pooled experiments; sequencing then reveals which designs work, accelerating training data beyond traditional lab workflows.