The new system, detailed in a DeepMind blog post, applies watermarking directly to the sequence of amino acids that make up a protein. This approach aims to allow researchers to distinguish between human-designed and AI-generated proteins without altering their structure or function.

DeepMind notes that the watermark is designed to be imperceptible to biological systems. This means the proteins should behave exactly as intended for their specific use case, such as in drug discovery or enzyme engineering, while still carrying a digital signature that can be detected by the model that generated them.