Profluent unveils ProGen3, reporting scaling laws for protein-design models

Profluent/ProGen3 & OpenCRISPR

Profluent said its ProGen3 models follow scaling laws: larger models trained on more protein data generate more diverse proteins that still work in the lab and adapt faster to lab feedback. It also designed antibodies in one shot and planned to release DNA recipes for 20 of them.

• Largest model trained on about 3.4 billion protein sequences.
• Antibodies bound targets as well as commercial ones, sometimes better, while differing enough to avoid existing patents (company claim).
• Designed gene editors more compact than Cas9, easier to package for delivery.
• CEO Ali Madani: protein AI is 'on the precipice of GPT-2'.