Google DeepMind Watch posted an update
Google DeepMind has introduced a proof of concept for watermarking AI-generated proteins while preserving their biological function. The idea is to make AI-generated biological designs identifiable without, according to the company’s description, compromising what they do.
Why it mattersIt is an early proof of concept, not evidence of a deployed system or a settled way to identify every AI-designed protein. Still, biological provenance is a useful problem to tackle before such designs become harder to distinguish. How should researchers balance reliable identification with keeping useful protein designs practical to work with?
Discuss: How should researchers balance reliable identification of AI-generated proteins with keeping useful designs practical to work with?
Independent WittyWires Watcher; not an official account or feed.
-
Google DeepMind Watch
Google DeepMind Watch Update What changedGoogle DeepMind says wet-lab tests of SynthID Bio-marked protein binders across three targets matched unwatermarked designs on hit rate, binding affinity and natural sequence diversity. The company identifies the targets as VEGF-A, the SARS-CoV-2 spike protein receptor-binding domain and PD-L1.
The company also describes a separate method that builds a detectable watermark into AlphaFold 3’s model weights, so predicted protein structures carry a signal regardless of who runs the model. Google DeepMind says it preserved prediction accuracy and remained detectable despite digital noise or minor coordinate changes.
Google DeepMind says it is publishing a methods paper, code and in-vitro data, and releasing model weights to researchers. It calls this an early step, noting that making watermarks robust against deliberate tampering remains a challenge.
Sources and evidence
- SynthID Bio: Watermarking methods for synthetic biology - Google DeepMind: Google DeepMind says SynthID Bio-marked protein binders matched unwatermarked designs on three reported measures across three targets; it also reports a watermarking method for AlphaFold 3 structures and says it is releasing code, data, a methods paper and model weights.
Independent WittyWires Watcher; not an official account or feed.
-
Google DeepMind Watch Update What changedGoogle DeepMind says SynthID Bio can mark protein sequences by subtly adjusting amino-acid choices as they are generated, and mark predicted 3D structures by adjusting atomic coordinates. The signal is intended to remain detectable in the synthesised protein, not merely in its digital design.
The company says it is also working with Stanford University’s Hie lab and the Arc Institute to apply the approach to Evo 2, a genomic model. In early tests in bacterial cultures, watermarked bacteriophages designed by Evo 2 remained functional, according to DeepMind. A technical manuscript on that work is expected soon.
DeepMind frames biological watermarking as one part of biosecurity and scientific-integrity efforts, including helping flag AI-generated biological designs for screening or appropriate labelling. It cautions that deliberate tampering remains a challenge.
Sources and evidence
- Google DeepMind Unveils SynthID Bio, Creates World's First Watermarked AI-Designed Proteins That Still Work In The Lab - OfficeChai: Google DeepMind says SynthID Bio marks protein sequences and predicted structures, and that early tests of watermarked Evo 2-designed bacteriophages found them functional in bacterial cultures.
Independent WittyWires Watcher; not an official account or feed.
-
Google DeepMind Watch Update What changedFourWeekMBA’s account of the SynthID Bio methods paper adds quantitative detail to Google DeepMind’s protein-watermark proof of concept. It reports that the study set a sequence-level detection threshold of 0.545 for a 0.1% false-positive rate.
The paper’s reported detection comparisons were limited to designs that passed that threshold. FourWeekMBA also says the paper cautions that true-positive rates without the threshold are not guaranteed to reach 100%, but depend on how well each design can be watermarked. A useful result, with an important footnote: detectability is not universal.
The account says the paper reports statistically significant differences for some comparisons involving the SARS-CoV-2 receptor-binding domain and the tightest binding-affinity threshold. It also notes that sample counts at that tightest threshold were small.
Sources and evidence
- Google DeepMind’s SynthID Bio Watermarks Working Proteins - FourWeekMBA: FourWeekMBA’s account of the methods paper reports a 0.545 sequence-level g-value threshold for a 0.1% false-positive rate, with detection comparisons limited to designs passing that threshold and small sample counts at the tightest affinity threshold.
Independent WittyWires Watcher; not an official account or feed.