AI-Designed Peptides Outperform GLP-1 in Lab
June 3, 2025
- 67% of AI-designed peptides outperformed GLP-1 in binding assays vs. ~0.1% for traditional methods.
- Peptides with 10+ changes delivered the biggest gains showcasing the power of AI creativity.
Breakthrough: 62 of 93 Beat GLP-1
Scientists at 310.ai challenged our generative protein model MP4 to reinvent GLP-1. We selected 93 of the most diverse AI-designed candidates for testing. In experimental assays, 62 of 93 (67%) matched or exceeded native GLP-1’s activation signal, a dramatic leap over the ~0.1% hit rate typical of structure-based designs. The top hit reached striking 19.4% improvement.
Unexpected Creativity: Bold Edits Win
Even more striking: the most altered sequences often performed best. Designs with 10+ amino acid changes consistently showed the highest activity gains. MP4 didn’t just tweak GLP-1—it reimagined it. Its most unconventional outputs proved to be the most potent, further validating the predictive strength of our AI-driven approach.
We are now advancing these molecules through additional functional assays and pursuing opportunities to patent and license the most promising candidates.
Method: From Code to Lab in Days
All designs were created fully in silico—no structural templates, no manual edits. We first generated 150,000 candidates with MP4, then filtered them down to 96 using our proprietary computational pipeline. These finalists hit the bench within days, validating our rapid, scalable design-to-data loop.
MP4 Today—and Tomorrow
MP4 is the flagship foundation model for 310.ai. It was demoed at NVIDIA’s GTC keynote and covered in Nature for its breakthrough in protein design. Launched in June 2024, MP4 has 1.5 billion parameters, was trained on 3.2 billion biology data points, and consumed over 3,800 MI250 GPU days. In April 2025, one AI-designed protein demonstrated experimental ATP interaction, signaling a breakthrough for programmable biology. Its successor, MP5, will bring deeper multimodal and multidisciplinary capabilities when it releases in July 2025.
___
Special thanks to the Terasaki Institute and Matterworks for their support and collaboration towards this project.
Note: These results reflect enhanced functional response in vitro and do not imply greater clinical efficacy or safety.