SynthIDBio watermarks AI-designed proteins while preserving tested binding function

Selected AI-designed protein binders retained their measured binding performance after researchers embedded a machine-detectable signature in their sequences. The result is part of SynthIDBio, Google DeepMind’s proof of concept for tracking the origin of AI-generated biological designs.[1]

Published in Nature on 30 September, the work examines two separate forms of watermarking: one for amino-acid sequences and another for computationally predicted molecular structures. The functional experiments address a demanding requirement for the sequence approach—a mark must coexist with the biological job the molecule was designed to do.[1]

One watermark lives in the sequence; another in the shape

A protein sequence is a chain of amino-acid instructions. A structure prediction represents the molecule’s expected three-dimensional arrangement. SynthIDBio-sequence introduces a detectable pattern as a sequence-design model samples possible amino acids. SynthIDBio-structure instead fine-tunes a structure-prediction model related to AlphaFold 3 so that certain properties of its output carry a hidden detectable signature. These are two different media and two different provenance claims; one need not imply the other.[1]

The researchers prepared watermarked and unwatermarked binders against selected targets and compared their interactions in vitro. Preserved binding provides a physical check beyond a computational quality score. The finding applies to the tested designs and assays; other protein functions require their own evaluation.[1]

In the structure-watermark evaluation, the reported detection rate exceeded 99.8% at the specified false-positive threshold. That measures the detector’s performance on the evaluated structure outputs. It is separate from the question of whether a marked sequence produces a functional protein.[1]

An origin signal that can be altered

A detected mark can support attribution to a compatible generating system. It does not establish a design’s safety or approval status. Equally, absence of a mark cannot establish natural origin: another system may not use watermarking, or subsequent modification may have weakened the signal.[1]

The authors examine the possibility of deliberate removal and acknowledge that changes can erase or weaken the marks. A provenance record would therefore need more than the detector alone—for example, the associated design record and a documented method for checking disputed results.[1]

Potential uses include recording AI-assisted designs in biological databases and helping laboratories track experimental material. To serve those purposes, different users would need consistent detection practices and clear handling of false positives, modified outputs and unmarked generators.[1]

Function and detectability must be tested together

A stronger mark can be easier to detect while imposing more change on the generated output. SynthIDBio explores that trade-off among detectability, structural accuracy and biological function. Measuring all three is essential because a detectable signature has little scientific value if it undermines the experiment.[1]

The study establishes that a provenance signal and the tested binding function can coexist in selected designs. The next task is to map where that result holds: across more functions, more modifications and independently tested detectors.[1]

Report an error or suggest a correction