AI protein design
Generative AI invents new proteins, enzymes and even antibodies that never existed in nature.
Open in the interactive tree →David Baker's lab at the University of Washington released RFdiffusion in 2023, a diffusion model that designs proteins around a target, followed by a series of upgrades. AI-designed proteins now serve as binders, enzymes, vaccines and drug candidates. The goal is a protein for any function on demand.
As of October 2026
In November 2025 Nature published antibodies designed from scratch with a fine-tuned RFdiffusion, tested against targets including bacterial toxins and influenza and coronavirus proteins, with electron microscopy confirming that designs bound as predicted; the RFantibody software was released freely. RFdiffusion3 (a preprint from September 2025) designs structures around ligands, DNA and other non-protein atoms at about one tenth of the computational cost of earlier approaches. Whether AI-designed proteins will succeed as medicines is still unproven.
Open steps
- Binders for hard targets High AI leverageHit rates are improving, but most designs still fail, especially for membrane proteins, flexible loops and sites that need very high affinity.
- Fast, efficient designed enzymes High AI leverageDesigned enzymes now match an average natural enzyme for simple hydrolysis, but harder or non-natural chemistry stays far less efficient; AI must design active sites and motion.
- Safe drugs, not just binders High AI leverageA binder must also be stable, makeable and not trigger anti-drug antibodies; predicting immunogenicity and aggregation from sequence is still weak.
- Screening designed sequences for harm Medium AI leverageAI can redesign toxins so that DNA-order screening misses them; screening must check predicted function, not just sequence similarity.
Where AI could help
High AI leverage. Design and ranking are computation and reported hit rates have climbed, but medicines still need developability tests and trials.
- Zero-shot design of antibodies and binders against new targets
- Predicting stability, immunogenicity and manufacturability early
- Designing enzymes and binders around small molecules and DNA
- Closed-loop design with automated wet-lab testing
Shown so far
- In November 2025 Nature published RFdiffusion-designed antibodies whose binding poses were confirmed by cryo-EM against targets including influenza hemagglutinin. source
- In June 2025 Chai Discovery reported in a company technical report (not peer reviewed) that Chai-2 designed antibodies with a 16-20% hit rate on 52 novel targets, testing about 20 designs per target in a two-week cycle. source
- A Bits to Binders competition tested about 12,000 AI-designed binders in functional CAR-T screens, with team success rates from 0.6% to 38.4% (webinar announced for June 2026). source
Prerequisites
- Transformer Architecture2017Protein language and generative design models are built on transformer networks
- AI protein structure prediction2020