DeepMind watermarks AI proteins with SynthIDBio

AI protein design tools have produced successes such as enzymes that digest plastics or block venom proteins, but the same tools could be used to make toxins or alter viral proteins. Existing software for screening DNA sequences does not pick out AI-designed proteins, because nobody has characterized them well enough to know they are threats, and nearly a year after that risk was flagged it still was not clear what anyone could do about it.

The DeepMind work started from ProteinMPNN, a popular AI design tool developed by the Baker Lab, and from Google's SynthID, which adds a subtle watermark to AI-generated digital material. The team's SynthIDBio variant steps in as ProteinMPNN builds a protein one amino acid at a time, suggesting choices consistent with a watermark only when they still fit the protein's structure and interactions. Detection requires scanning the whole sequence with the key and measuring how often the suggested amino acids appear, and Google has also developed the software to do this.

Google envisions DNA synthesizers holding keys from trusted organizations, such as universities or major biotech companies, so they can quickly tell whether an unknown protein is an AI design from a trusted source and focus resources on evaluating untrusted ones. Whether that helps in practice hinges on key distribution, on proteins long enough to carry a readable signal, and on whether design tools beyond ProteinMPNN can be brought into the same scheme, and it is not clear the approach will be especially useful in its original form.