Google figures out how to watermark AI-designed proteins
Google DeepMind's SynthIDBio watermarks AI-designed proteins without breaking function, aiding biosecurity checks.
Google DeepMind published SynthIDBio, which adapts its SynthID watermarking system to proteins designed with ProteinMPNN. A key biases amino-acid choices only when they remain compatible with a functional protein, spreading a detectable signal along the chain. Existing screens often miss novel AI-designed proteins, so marked sequences from trusted researchers can be identified and unmarked designs scrutinized. The work responds to concerns that the same tools used for beneficial enzymes could also design toxins or alter viral proteins.
- DeepMind adapted SynthID into SynthIDBio for ProteinMPNN sequence design.
- Suggested amino acids are kept only when they still support a functional protein.
- Detection scans the full sequence with a secret key, not a simple yes-or-no.
- Marked designs from trusted researchers can be distinguished for closer biosecurity review.
- Vendors
- Products
- SynthIDSynthIDBio
- AI models
- ProteinMPNN
From the original article
7 min read at arstechnica.com
Intended to help with biosecurity, it works with a popular AI protein design tool.
By John Timmer · arstechnica.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.