DeepMind watermarks proteins designed by AI
Published in Nature, SynthID Bio hides a signature in a protein’s sequence and 3D structure. Watermarked proteins kept their function in lab tests.
Method
The sequence watermark is a tiny bias in amino acid choices; a second method writes the signature into the atomic coordinates of the predicted 3D structure. Detection is 100% for sequences and above 99.8% for structures at a 0.1% false positive rate.
Lab results
The team designed binders for VEGF-A, PD-L1 and the SARS-CoV-2 binding domain. Watermarked designs matched unmarked ones on hit rate, affinity and sequence diversity.
What it is for
The target is DNA synthesis screening: a watermark would tell providers automatically that an order came from a model with safeguards. Resistance to deliberate tampering is still unsolved, and the tools are being open-sourced.