Google introduces SynthID Bio protein-watermarking proof of concept
Google says laboratory tests found watermarked protein designs matched unwatermarked versions, while DeepMind describes the technology as a proof of concept.
AI illustrationWatermarks embedded in artificial intelligence protein designs are the focus of SynthID Bio, introduced by Google on September 30. The company says the technology can mark biological designs while preserving their function.
Google DeepMind describes the work as a proof of concept in a post by Pushmeet Kohli, David Stutz, Ali Cowen-Rivers, and Jeremy Ratcliff. Its description concerns watermarking AI-generated proteins without losing biological function.
According to Google, the marks are imperceptible and verifiable, and are placed directly in designs including protein sequences and predicted three-dimensional structures. The company presents the approach as a way to establish provenance, strengthen biosecurity, and protect open scientific databases.
Google said laboratory tests across target proteins found that marked designs matched the performance and natural diversity of unmarked versions. Its accompanying material compares binding affinity across three targets and shows a watermarked protein binder’s predicted structure. Those test results are reported by Google.