Google DeepMind researchers have developed a method to add an identifiable watermark to proteins created with artificial intelligence, potentially helping scientists trace the origin of biological designs while supporting biosecurity efforts. The technique, described in a study published in Nature, subtly changes the probability with which amino acids are selected during protein design. These small changes create a statistical pattern that can be detected later without significantly affecting the protein’s structure or function. Researchers tested the approach using SynthIDBio alongside ProteinMPNN, a widely used artificial intelligence system for protein design. Their experiments focused on protein binders, including designs intended to attach to specific biological targets. The researchers reported that watermarked binders performed similarly to unwatermarked versions in laboratory testing. Scientists said the approach could create a record showing whether a protein originated from a particular laboratory, artificial intelligence model or manufacturer. However, experts noted that the method may be harder to apply to smaller or highly complex proteins because their sequences have fewer options for modification. DeepMind is also exploring the approach for other AI-designed biomolecules, including DNA. The researchers acknowledged that current watermarks can be removed relatively easily and are working to make the technology more resistant to such changes.
Read more here
