Researchers introduced SynthIDBio, a method to watermark AI-generated protein sequences and structures. SynthIDBio-sequence embeds watermarks into protein sequences while keeping their function intact. It produced functional protein binders with binding strength matching non-watermarked versions. SynthIDBio-structure uses a fine-tuned AlphaFold3 model to add imperceptible watermarks to biomolecular structures. The watermarks are nearly always detectable. The work is a proof-of-concept for provenance tracking in AI-driven biology. It targets biosecurity and information integrity challenges. The approach preserves experimental function in designed proteins. It also maintains structural prediction accuracy. The paper appears in Nature.
What changed
Existing watermarking methods reduced protein function or structural accuracy.
What it unlocks
Embedding undetectable watermarks in AI-generated proteins without losing biological function.
- near-perfect watermark detection accuracy
- low nanomolar binders for SC2RBD
- subnanomolar binders for VEGF-A and PD-L1
Sources