Google DeepMind is testing a new approach to identifying proteins designed with artificial intelligence, introducing a proof-of-concept watermarking system intended to help DNA suppliers trace unfamiliar biological designs while maintaining existing safety screening processes.
The technology, called SynthID Bio, was introduced on Sept. 30, 2026. It extends Google’s SynthID watermarking research into protein design, an increasingly important field as AI systems become more capable of generating novel biological structures.
SynthID Bio Aims to Trace AI-Designed Proteins
Companies that manufacture custom DNA routinely screen orders to identify sequences associated with known biological threats. Novel sequences can be more difficult to assess because they may not match material already contained in screening databases.
Google says these unfamiliar designs can require additional manual investigation, potentially slowing legitimate research.
SynthID Bio is designed to provide an additional signal indicating that a protein originated from an AI-assisted design process. The watermark does not replace biological safety screening or determine whether a protein is harmful.
Instead, researchers envision it as another piece of information DNA suppliers could consider when reviewing unusual orders.
Two Watermarking Methods Under Development
Google DeepMind’s project documentation describes two approaches.
The first works with ProteinMPNN, an AI model used in protein design, and embeds a watermark directly into amino acid sequences.
The second modifies AlphaFold 3 so that a hidden pattern is incorporated into the predicted three-dimensional coordinates of a protein structure.
According to research published in Nature, the structural watermark detector identified more than 99.8 per cent of watermarked predictions in the AlphaFold 3 evaluation set while operating at a 0.1 per cent false-positive rate.
Researchers said the recommended configuration maintained the structural accuracy measures evaluated during testing.
Laboratory experiments were also conducted on the sequence-based approach. Google reported comparable binding strength and experimental success rates between watermarked and non-watermarked proteins tested against three biological targets.
However, the experiments began with structures derived from proteins already known to bind successfully. That leaves questions about how the technology would perform across a broader range of novel protein designs.
Tampering Remains a Challenge for Protein Watermarks
SynthID Bio currently provides a relatively simple signal showing that a design has been watermarked rather than a detailed record identifying its creator or history.
Researchers also demonstrated that watermarks could be removed by rewriting protein sequences or conducting additional processing on protein structures.
That limitation raises an important issue for any future safety application: a watermark that can be intentionally removed cannot function as a complete security mechanism.
False positives and missed detections could also influence which DNA orders receive additional scrutiny. Any operational use would therefore require testing to determine how the technology performs alongside existing screening procedures.
AI Protein Design Creates New Opportunities and Risks
The development comes as AI-powered protein design continues to advance in biomedical research.
In September 2024, Google reported that its AlphaProteo system generated proteins that successfully bound to seven targets associated with infection and disease research. The system was unsuccessful with an eighth target, and Google noted that strong binding represents only an early stage in developing practical biological or medical applications.
For Canadian researchers working in biotechnology, pharmaceutical development and academic life sciences, advances in AI-designed proteins could support faster experimentation and new approaches to drug discovery. At the same time, international DNA synthesis and screening standards can affect laboratories regardless of where they operate because biological research frequently relies on global suppliers.
DNA Screening Systems Face More Sophisticated Designs
Concerns about AI-generated biological sequences are not entirely theoretical.
In October 2025, Twist Bioscience reported research showing that AI-designed variants of toxin and viral protein sequences were able to bypass standard screening software.
Those results involved digital sequences, and researchers did not establish whether the resulting proteins would retain harmful biological activity if manufactured. Twist said the findings nevertheless contributed to software patches and improved detection methods.
SynthID Bio could potentially complement such screening systems by giving suppliers additional information about a sequence’s origin.
If reliable, that information could allow reviewers to spend less time determining where unfamiliar designs originated and focus more attention on evaluating potential biological risks. Whether the approach can meaningfully reduce delays in real-world DNA ordering systems has yet to be demonstrated.
Google Plans Further Testing of Biological AI Safeguards
Google DeepMind and Isomorphic Labs have placed the watermarking project within a broader biosecurity framework that includes threat assessment, evaluations, safeguards and monitoring.
SynthID Bio remains a proof of concept rather than a standalone safety solution. Researchers will need to examine how effectively it integrates with DNA order screening, how false positives and missed detections affect review processes, and whether future versions can better withstand attempts to remove embedded watermarks.
As AI becomes more capable of designing proteins, technologies such as SynthID Bio could provide suppliers and researchers with additional tools for identifying AI-generated biological material. Its practical value, however, will depend on whether the watermark can remain reliable, difficult to manipulate and useful alongside established biological risk assessments.
Sophie MacKenzie is a contributor to CA News Ottawa, covering news, politics, business, technology, sport, entertainment, and lifestyle. She focuses on clear, accurate reporting and useful information that helps readers stay informed about current affairs and developments that shape their communities. Her work highlights relevant stories, emerging trends, and important issues, presenting them in a balanced, accessible, and reader-friendly manner.
