Google DeepMind proposes watermarking for AI-designed proteins

In short: Google DeepMind published research on a way to “watermark” AI-designed proteins so they can be identified later.

What happened

Google’s DeepMind team published a paper describing “protein watermarking.” It is a method for adding a hidden signature to a protein’s amino acid sequence, without stopping the protein from working.

Proteins are built from 20 amino acids, like a long string of beads with only 20 bead types. AI tools can design new protein sequences for helpful tasks, but the same tools could also be misused to create harmful toxins or tweak viral proteins.

The new method is based on Google’s SynthID, which is used to watermark AI-made text and images. For proteins, the team created SynthIDBio and used it with a popular protein design tool called ProteinMPNN. As the AI picks amino acids one by one, SynthIDBio quietly nudges some choices, but only when those choices still fit the protein’s structure and function.

To check that this worked, the researchers designed proteins meant to bind to specific natural proteins. The watermarked versions still bound to their targets, which suggests the watermark did not break the designs.

Why it matters

Companies that manufacture DNA often screen orders for known dangerous sequences. The problem is that AI-designed proteins can look new and unfamiliar, so existing checks may not flag them. Google’s idea is that trusted labs could watermark their AI-designed proteins using secret “keys” (like a password used to prove something is authentic), so DNA makers can quickly recognize them and spend more time reviewing the unwatermarked, unknown designs.

The paper also notes limits. Short proteins may not carry a detectable watermark, key handling could be a weak point, and mixing a watermarked protein with other sequences could dilute the signal.

Source: Arstechnica