Google DeepMind Embeds Watermarks into AI-Designed Proteins Without Compromising Function

Google DeepMind has developed a way to embed watermarks into AI-designed proteins without compromising their function. The technology, called SynthID Bio, uses a proof-of-concept approach that subtly influences the selection of amino acids in protein sequences and predicted 3D structures.

According to researchers, SynthID Bio tackles a growing biosecurity problem: AI-generated biological designs can differ significantly from known threats, making it difficult to determine their origins using conventional screening methods. The technology has been integrated into existing tools such as AlphaProteo and ProteinMPNN, which design protein binders.

Laboratory tests have shown promising results, with watermarked proteins matching their unmarked counterparts in terms of hit rate, binding affinity, and natural sequence diversity. However, researchers acknowledge that making watermarks resistant to deliberate removal remains a major challenge.