A research team at Google DeepMind has introduced a method for embedding detectable markers into artificially designed proteins, aiming to improve how laboratories screen synthetic genetic material. The approach, detailed in a paper published Wednesday in the journal Nature, builds on earlier work that explored embedding subtle signatures in AI-generated content such as text and images.

The system, called SynthIDBio, modifies the protein design process by influencing the selection of amino acids during construction. Rather than altering the final protein structure, the technique introduces specific amino acid choices that collectively form a pattern identifiable only with knowledge of the encoding key. This pattern persists through standard laboratory processing and does not affect the protein’s biological function, according to the researchers.

The technique was tested on proteins designed to interact with known biological targets. In these trials, the watermarked proteins retained their intended binding capabilities, suggesting compatibility with complex design workflows. The method relies on tools like ProteinMPNN, which generates viable protein structures, and integrates the watermarking step within its sequential assembly process.

Because the watermark is distributed across the sequence, detection requires scanning the full length of the protein and comparing amino acid choices against expected patterns using the shared key. This statistical identification method means that thresholds for confirmation can affect the balance between false positives and false negatives.

Google suggests the technology could help genetic synthesis facilities assess whether unfamiliar protein sequences originate from trusted design sources. Currently, such facilities screen for known hazardous genetic elements, but they lack reliable ways to verify the origin of novel designs that do not match existing natural sequences. By providing a verification mechanism tied to authorized keyholders—such as academic institutions or established biotechnology firms—the system could streamline this evaluation process.

Limitations and concerns have been noted. Very short proteins may not carry enough watermark signals to be reliably detected, and the approach depends on secure key management to prevent misuse. Additionally, the method applies only to design platforms that follow the specific sequential amino acid selection process used by ProteinMPNN, leaving other AI-driven design tools outside its current scope.

The researchers acknowledge that the system does not eliminate biosecurity risks but may enhance existing screening protocols. They also highlight potential vulnerabilities, such as the possibility of masking watermarked sequences through fusion with unrelated proteins.

The work represents one of the first attempts to apply digital watermarking principles to biological sequences in a way that preserves function. While practical implementation challenges remain, the development contributes to ongoing efforts to address security considerations in rapidly advancing AI-driven biological engineering.