Google DeepMind publishes SynthIDBio in Nature, invisible function-preserving watermarks for AI-generated proteins

A Google DeepMind team published the SynthIDBio method family in Nature, embedding detectable, hard-to-remove and imperceptible invisible watermarks into AI-generated protein sequences and structures. The method integrates SynthID's tournament sampling strategy into ProteinMPNN; calibrated at a 0.1% false positive rate, sequence watermark detection (TPR) reached 100%, and watermarked binders showed no significant difference in binding affinity to SC2RBD, VEGF-A and PD-L1 versus non-watermarked controls.