Watermarking AI-Designed Proteins

Google DeepMind's SynthID Bio system embeds traceable watermarks in AI-generated proteins without compromising function, addressing biosecurity risks.

Google DeepMind developed SynthID Bio, a watermarking system for AI-designed proteins that embeds subtle signals during design by altering amino acid sequences and atomic positions. Lab tests confirm these watermarked proteins maintain biological functionality for targets including immune regulation, blood vessel growth, and viral proteins. The system works with AlphaFold3-generated proteins without requiring external modification and was tested against synthetic toxins and bacteriophages created by open-source AI tools.

SynthID Bio’s mechanism relies on embedding watermarks that remain detectable even after proteins undergo reprocessing through other design tools. However, it cannot identify creators, trace illicit use, or prevent AI proteins from evading DNA sequence databases. Watermarking may also alter protein structure sufficiently to reduce functionality, and researchers working on non-sensitive projects face added complexity.

This system directly mitigates biosecurity risks by enabling traceability of AI-generated proteins—a critical step for preventing malicious deployment. While not a complete solution, it addresses gaps in tracking AI-designed biological molecules and provides a foundation for future systems like FoldMark, which could distinguish among up to one million hypothetical users. The next phase requires resolving implementation challenges while balancing traceability with biological utility.

Source: Singularity Hub