Google figures out how to watermark AI-designed proteins

Key Points:

• Google DeepMind researchers have developed a protein watermarking system, SynthIDBio, that embeds identifiable watermarks into AI-designed protein sequences without compromising their function, aiding biosecurity efforts.
• The watermarking process integrates with ProteinMPNN, an AI protein design tool, selectively incorporating amino acids consistent with both the watermark and functional protein structure, allowing detection through statistical analysis of the sequence.
• Watermarked proteins were experimentally shown to retain their biological activity, such as binding target proteins, indicating the method does not impair protein functionality.
• This watermarking approach could help DNA synthesis companies identify AI-designed proteins from trusted sources, streamlining threat screening by focusing on untrusted or unknown sequences, although it does not guarantee complete security.
• Limitations include reliance on secure key distribution, challenges with short proteins, potential dilution of watermarks through sequence fusion, and partial compatibility with other protein design tools, meaning the system is an initial but promising step toward mitigating AI-related biosecurity risks.