Google DeepMind’s breakthrough watermarks AI-designed proteins

Google DeepMind has developed a method called SynthID Bio to watermark AI-designed proteins, ensuring their biological function remains intact. The system embeds a detectable signature into either the protein’s amino acid sequence or its predicted three-dimensional structure, addressing a unique challenge compared to watermarking images. Unlike altering pixels in an image, changing a protein’s structure can affect its folding and functionality.

The research, published in Nature, demonstrated that watermarked protein binders retained their function in lab tests across three targets: VEGF-A, PD-L1, and the SARS-CoV-2 spike protein. The structural watermark achieved a detection rate above 99.8% with a 0.1% false-positive rate, while the sequence watermark showed no significant differences in binding affinity compared to non-watermarked proteins.

Despite these successes, SynthID Bio is still a proof of concept. The sequence watermark can be removed through resequencing, and the structure watermark is vulnerable to certain transformations. The study also noted that current watermarks are zero-bit, meaning they can only indicate presence or absence without carrying additional provenance data.

The technology aims to address biosecurity concerns by providing a way to track the origin of AI-generated biological sequences. This could help screen novel sequences against known threats and maintain the integrity of scientific databases. However, the researchers emphasized that watermarking alone won’t solve biosecurity issues but is a step toward a safer future.