AI models still can't design proteins based on 'vibe' alone

Researchers have created a new benchmark to test how well AI models can design proteins using natural language descriptions. It turns out no current AI model can consistently perform well across all stages of protein design, from recognizing existing proteins to generating new ones based on open-ended instructions.

Protein design is a core task in drug discovery and biotechnology, but it often relies on highly structured, technical inputs. The idea of 'vibe protein design' is that scientists could simply describe what they want in plain language, and an AI would create it. This benchmark shows that this vision is still far off, meaning human experts remain essential for translating broad ideas into precise biological instructions.