International Edition
Latest News
Technology

Google DeepMind SynthID Bio: Watermarking AI-Designed Proteins for Biosecurity

Google DeepMind has introduced SynthID Bio, a new method designed to sneak a subtle signal into proteins during the design process as a way to address biosecurity concerns and biological AI slop. While proteins were once created only…

Google DeepMind SynthID Bio: Watermarking AI-Designed Proteins for Biosecurity

Google DeepMind has introduced SynthID Bio, a new method designed to sneak a subtle signal into proteins during the design process as a way to address biosecurity concerns and biological AI slop.

While proteins were once created only by nature, AI now allows scientists to design completely new proteins from scratch. These AI-generated structures create a risk of polluting databases that scientists rely on, potentially forcing them to build on a rotten foundation without a way to distinguish experimentally validated structures from dreamt-up ones. Beyond database pollution, biosecurity is a major concern, as AI-designed proteins can complicate screening efforts by DNA synthesis companies that check genetic blueprints against databases of controlled sequences.

Google DeepMind Introduces SynthID Bio for Watermarking Proteins

How SynthID Bio Works

Adapted from technology that labels AI-generated text, images, audio, and video, SynthID Bio can be built directly into protein-design systems such as AlphaFold and RFdiffusion. When the system encounters a location where a set of chemically related amino acids will all work—such as leucine, isoleucine, valine, serine, or threonine—it uses one that is consistent with the watermark when possible. As another approach, the method can scan the domain of functional proteins to find the subset possessing an adequate frequency of watermark amino acids.

As a result, the watermark is randomly distributed across the entire length of the protein. Detecting the signal is not a simple yes-or-no question; users must scan the whole sequence using the key and measure how often the amino acids suggested by the tool appear. Google has also developed the necessary software to perform this detection.

A multi-colored set of linked ribbons, coils, and swirls, meant to represent the backbone of a protein
Photo: Ars Technica

Watermarked Proteins Function as Intended in Tests

For functionality assessment purposes, the researchers employed the platform to design proteins that physically interact with key natural proteins previously targeted with AI designs. The watermarked versions worked just fine and bound their intended targets, suggesting no reason to expect serious problems in more complicated design tasks as long as a protein is long enough for the system to detect a watermark.

About the author: Anika Shah - Technology

MSc in Computer Science, senior reporter. Anika focuses on AI ethics, cybersecurity, and emerging hardware—frequently moderating panels at CES and Web Summit. “Anika Shah decodes tech breakthroughs and startup disruption shaping tomorrow’s digital landscape.”