Today, we announced SynthID Bio, a method to watermark and detect AI-designed proteins, the building blocks of life.
We believe this is the first time humanity has successfully synthesised AI-designed proteins that are both functional and watermarked. It is a spectacular engineering achievement from our @GoogleDeepMind research teams, but also a vital safeguard as we enter a transformational era of biological research.
Generative AI is helping the scientific community design novel biological artefacts, such as functional proteins, with systems like AlphaProteo and RFDiffusion, and even bacteriophages. This powerful technology could advance how we develop effective disease treatments, but it also has the potential to be misused. For instance, AI-generated proteins could circumvent biosecurity screening measures used by gene synthesis companies, as well as threaten to pollute shared scientific databases.
For many years, my team and I at Google DeepMind have worked to ensure AI models are safe, building solutions that help track the provenance of generative AI outputs. Our research led us to develop SynthID, the state-of-the-art solution that embeds imperceptible watermarks into AI-generated artefacts like images, videos, audio and text without affecting their quality. Already, SynthID has been used to watermark more than 100 billion images and videos and over 60,000 years of audio, and is also used by a number of our key industry partners, including OpenAI, NVIDIA and Kakao.
Adapting SynthID to the biological world was a significant challenge. In the world of images, videos and audio, imperceptibility has to consider human perception and avoid visual artefacts or acoustic distortions. Biology is different. A protein sequence isn’t just a string of text: it is a physical entity that must function in the natural world, with real biological consequences. Watermarks only work if the resulting molecules retain their function. For a protein, this means folding, engaging with cellular receptors and interfacing with molecules exactly like its unwatermarked equivalent.
With SynthID Bio, we achieved effective watermarking while preserving biological activity. We developed techniques to watermark both protein sequences as well as predictions of their three-dimensional protein structures. Our extensive testing of SynthID Bio showed that watermarked molecules functioned in both biological environments and against real therapeutic targets. In wet-lab experiments, our watermarked binders also matched the success of their unwatermarked equivalents, confirming that our system works as intended.
I hope SynthID Bio will become a key verification layer in tracing AI-generated designs. By automatically flagging AI-generated proteins, it can help synthesis providers streamline the slow, manual security reviews that currently stall vital research.
SynthID Bio will also help preserve the scientific commons. Breakthroughs like AlphaFold rely on clean public databases, like the Protein Data Bank, which risk being polluted by vast volumes of AI-generated predictions. By filtering out these misclassified contributions, SynthID Bio keeps our shared databases and future algorithms accurate and free of inadvertently mislabelled predictions.
We are continuing to improve SynthID and extend it to even more biological modalities. As an early example, I am excited to share that in an ongoing collaboration with the Hie lab at Stanford University and Arc Institute, we extended SynthID Bio to watermark the genome of an Evo 2-designed bacteriophage. Early laboratory testing in bacterial cultures has shown that our watermarking methods can preserve the bacteriophage’s function. This is an early milestone and a promising technical proof-of-concept, and we look forward to sharing more details with the scientific community soon.
Developing effective standards for biological provenance should be a collaborative process. Like our work with the AlphaFold database, we are partnering with the scientific and biosecurity communities to develop these technologies together. As a first step, we are publishing our methods paper, open-sourcing the code and in vitro data, and releasing the weights to the research community, so we can realise the full biosecurity benefits of this work.
We are entering an era in which biology is changing from an observational science to one we can responsibly engineer. Watermarking won’t solve biosecurity by itself, but it is an important step towards a safer and more resilient world. It is my belief that we should not measure progress solely by the complexity of the biological structures we can model, or their potential impact, but also by the foresight and responsibility with which we guide these advancing capabilities.
Read our blog: https://deepmind.google/blog/introducing-synthid-bio/ and our technical paper: https://www.nature.com/articles/s41586-026-10965-y
