DeepMind Introduces SynthID Bio to Watermark AI-Designed Proteins for Traceability

DNA synthesis companies currently perform safety screening on sequences submitted by customers. One important method is comparing new orders against known pathogens, toxins, and other potentially risky biological sequences, and when highly similar content is found, further verifying the customer's identity and purpose. The premise of this process is that an unfamiliar sequence either can be matched in a database or has identifiable similarity to a known dangerous sequence.

Generative AI is shaking this premise. AI can actively design new proteins that have never existed in nature, and even if two proteins differ greatly in amino acid sequence, they may still form similar structures and achieve similar functions. This means a protein redesigned by AI may show almost no obvious sequence connection to a known dangerous protein, yet still have similar functions. Faced with such an unfamiliar sequence, relying only on database comparison makes it increasingly difficult to determine whether it comes from a natural organism, normal scientific design, or an artificially designed potentially risky sequence.

What SynthID Bio aims to provide is precisely this additional piece of provenance information. It cannot directly prove a protein is safe or dangerous; rather, it tells screeners where this design came from.

SynthID Bio includes two methods. SynthIDBio-sequence is used to watermark a protein's amino acid sequence, and it is integrated into ProteinMPNN—a model that selects suitable amino acids for different positions based on a given protein backbone and ultimately generates a complete sequence. The watermark is not forcibly adding a special sequence after protein design is complete; instead, it adds a set of hidden rules into ProteinMPNN's original amino acid selection process: when multiple suitable amino acids exist at a position, the system slightly changes the tendency of different candidate amino acids to be selected based on a secret key. Looking at any single position alone, this difference is hard to notice, but after similar choices continually appear throughout the entire sequence, they form a detectable statistical pattern. The party holding the corresponding detection key can use this to determine whether the sequence was generated by a model with SynthID Bio.

SynthIDBio-structure targets AI-predicted three-dimensional protein structures. Rather than waiting for AlphaFold 3 to complete structure prediction and then modifying the positions of a few atoms, the researchers directly fine-tune part of the AlphaFold 3 network, so that the model automatically leaves hidden watermark signals when generating the three-dimensional structure. These signals are embedded into very subtle geometric relationships such as distances and angles between atoms. For researchers using protein structures normally, these differences are almost imperceptible, but a specially trained detector can identify them.