New watermarking tool for synthetic biology tracks ai-generated proteins

Google DeepMind has introduced SynthID Bio, a watermarking technology designed to track AI-generated proteins while preserving their biological function. The tool embeds an imperceptible signature into the biological code, making it verifiable both digitally and in physical proteins. This innovation addresses concerns about biosecurity and scientific integrity, as AI-generated proteins can bypass traditional DNA synthesis screening and mislead research. SynthID Bio was tested on three proteins, including VEGF-A and the SARS-CoV-2 spike protein RBD, and showed no compromise in binding affinity or natural sequence diversity. It also integrates with AlphaFold 3, allowing for detectable signatures in predicted 3D structures. Sarah Carter, a biosecurity policy expert, praised the tool for linking designs to model developers, enhancing safety and streamlining screening processes. The project, led by Pushmeet Kohli and Alexander I. Cowen-Rivers, aims to strengthen biosecurity by enabling automated verification of synthetic sequences. It also seeks to improve the integrity of biological databases like the Protein Data Bank. The team plans to expand the technology to watermark bacteriophages and is open-sourcing its methods and data to foster collaboration in biosecurity research.