NEW YORK — What happened: Bio-AI researchers released an open-weights generative model trained on billions of metagenomic sequences and structural conformations. The system successfully designed, synthesized, and tested functional enzymes that do not exist in nature.
Why it matters: Enzyme engineering previously relied on evolutionary trial and error over months or years. This framework generates viable, high-stability protein structures for industrial chemistry in minutes.
Deep dive: The model combines equivariant graph neural networks with diffusion processes operating directly in 3D coordinate space. It scores candidates not just on static folding energy, but on predicted catalytic turnover rates under fluctuating pH and thermal conditions.
Report check (claims vs what is verified vs still rumor): The developers claim a 70% success rate in synthesizing functional enzymes on the first try. Wet-lab replication by independent university groups has verified high activity for plastic-degrading variants, though commercial yield scalability is still unproven.
Open questions: How will regulatory bodies handle genetically engineered organisms and proteins designed entirely by automated machine learning pipelines?