We’re introducing SynthID Bio, bringing our watermarking technology to synthetic biology.

Google unveiled SynthID Bio, a new watermarking technology designed to embed imperceptible, verifiable identifiers into AI-generated protein sequences and 3D structures without altering their biological function. The tool was tested in laboratory settings across multiple target proteins, demonstrating that watermarked designs performed comparably to unwatermarked versions while preserving natural diversity. Researchers highlighted the importance of provenance layers for biosecurity, particularly in open scientific databases where tracking the origin of biological designs is critical. The announcement includes visualizations and comparative data, such as binding affinity measurements and structural predictions, available on the Google DeepMind website.

SynthID Bio embeds watermarks directly into protein sequences and predicted structures, ensuring that the biological function remains intact while enabling verification of the design’s origin. Laboratory tests showed that watermarked proteins matched the performance of unwatermarked versions across multiple targets, with binding affinity measurements (KD) indicating no significant loss in function. Visual comparisons of predicted and ground truth structures, including a watermarked VEGF-A protein binder, were provided to illustrate the technology’s precision. The research and additional details are accessible via the Google DeepMind website.

The technology aims to address biosecurity concerns by creating a verifiable provenance layer for AI-designed proteins, which can be crucial for tracking and validating biological designs in open databases. By embedding watermarks that do not alter biological function, SynthID Bio seeks to maintain the integrity of scientific research while enabling traceability. The announcement emphasizes the potential for this tool to strengthen oversight in synthetic biology, where the rapid advancement of AI-generated designs poses new challenges. Visual and structural data from the research are available for further review.

Google DeepMind’s research on SynthID Bio demonstrates the feasibility of watermarking AI-designed proteins without compromising their functionality or natural diversity. The technology was tested across several target proteins, with results showing that watermarked designs performed indistinguishably from unwatermarked versions in laboratory settings. Structural visualizations, including predicted and ground truth models of a watermarked VEGF-A protein binder, were included to highlight the tool’s accuracy. Full details and research findings are published on the Google DeepMind website.