Function-preserving watermarking of AI-generated proteins (www.nature.com)

Researchers have introduced SynthIDBio, a groundbreaking method for watermarking AI-generated proteins, crucial for establishing their provenance and addressing biosecurity and information authenticity concerns. This approach embeds function-preserving watermarks into protein sequences and structures, allowing for effective tracking without compromising their biological functions. SynthIDBio-sequence, which integrates watermarking into the widely used ProteinMPNN model, has demonstrated that it can produce watermarked protein binders with binding affinities comparable to non-watermarked versions and boasts high watermark detection accuracy. SynthIDBio-structure, fine-tuned from AlphaFold 3, embeds watermarks into biomolecular structures while maintaining structural integrity. The significance of SynthIDBio lies in its ability to address the challenges posed by the rapid advancement of AI in biology, particularly in ensuring the responsible use of generated data. This watermarking technique not only offers a privacy-preserving alternative to traditional methods but also promises to maintain the essential qualities of the proteins involved. With potential applications in fields such as therapeutics and diagnostics, SynthIDBio represents a pivotal step toward safeguarding scientific integrity and biosecurity in an era increasingly influenced by AI-driven biological engineering.