Google DeepMind releases SynthID Bio for research use

SynthID Bio embeds an imperceptible signature directly into biological code, according to Google DeepMind.[2] The method can subtly guide amino-acid choices in protein sequences and adjust atomic coordinates in predicted 3D structures.[2] Nature reported that the hidden clue is encoded at the sequence level and can be detected only with a secret key, which would be shared with trusted partners such as DNA-sequence makers.[3] Google DeepMind says verification can be performed on a synthesized physical protein, not just a digital model.[2]

The designs tested were made with AlphaProteo and a SynthID Bio version of ProteinMPNN.[4] Laboratory testing covered VEGF-A, the SARS-CoV-2 spike protein RBD and PD-L1, with Adaptyv Bio assisting.[4] The watermarked and unwatermarked designs matched on hit rate, binding affinity and sequence diversity in those tests.[4] Google DeepMind also said the watermark preserved biological function in laboratory testing.[2] These findings describe the reported tests; they do not establish performance across all proteins or research settings.[4]

Google DeepMind said it was releasing the tools on an open basis for research use.[1] It also said the code and laboratory data were open source and that it was releasing model weights to researchers.[4] The company describes the watermark as a digital ID card that could help DNA labs verify synthetic proteins.[1] It says verification could help keep sequence databases accurate and help labs automate and strengthen biosecurity.[1] Google DeepMind named the Protein Data Bank, UniProt and GenBank in discussing the databases.[4]

Google DeepMind said it still needed to make the watermark harder to remove deliberately.[4] Nature reported that a watermarked protein could be redesigned to retain its structure and function while obscuring its synthetic origin.[3] The Next Web reported that very short proteins may have too few marked amino acids for detection, and that joining a marked protein to an unmarked one could dilute the signal.[4] University of Sheffield researcher Haiping Lu said it remained unclear whether the method would work broadly or be widely adopted.[5]