Google develops watermarking for AI-designed proteins

The challenge in watermarking proteins lies in their fundamental structure. Unlike digital images with millions of pixels, proteins are composed of only 20 amino acids, with limited room for subtle alterations. Any change must not disrupt the protein's structural integrity or catalytic activity. Google's approach, detailed in a recent research paper, integrates with popular AI protein design tools like ProteinMPNN. During the design process, SynthIDBio, a variant of SynthID, suggests amino acids that are consistent with the watermark, provided they do not inactivate the protein. This ensures that the resulting watermarked proteins remain functional.

The system works by using a cryptographic key to influence the AI's amino acid selection. When the AI encounters a position where multiple chemically similar amino acids would function, it prioritizes one that aligns with the watermark. This process distributes the watermark randomly throughout the protein sequence, making it detectable only with the correct key and difficult to remove. The team demonstrated that watermarked proteins successfully bind to their intended targets, suggesting the method is viable for complex design tasks.

This technology is intended to assist DNA synthesizers, which currently screen ordered sequences for known threats. However, AI-designed proteins can be novel and uncharacterized, making their potential threat level difficult to assess. By using keys provided by trusted organizations, synthesizers can quickly identify AI-designed proteins from known sources, allowing them to focus resources on evaluating truly unknown sequences.