FoldMark: Safeguarding Protein Structure Generative Models with Distributional and Evolutionary Watermarking

FoldMark achieves over 95% watermark accuracy at 32 bits in protein generative models without compromising structural quality.

• Why it matters: Protecting protein design from misuse is critical as AI-driven methods accelerate breakthroughs but also pose biosecurity risks, such as creating pandemic-capable proteins.
• What they did: FoldMark employs distributional and evolutionary principles to embed watermarks into models like AlphaFold3 and RFDiffusion, enabling traceability of up to 1 million users.
• The result: The watermarking maintains high structural integrity (>0.9 scTM scores) and demonstrates practical effectiveness in wet lab applications, including EGFP and CRISPR-Cas13, with over 90% detection success.