DeepMind adds detectable marks to AI-designed proteins
Google DeepMind introduced two SynthID Bio methods that place detectable marks in AI-designed protein sequences and predicted structures. The tested protein binders retained the functions examined in laboratory experiments. Such marks could help identify where a biological design came from, but resistance to deliberate alteration remains unfinished work. The result is a demonstration of provenance marking, not a ready-made safety certification for biological designs.
Artificial Intelligence··Night
Two kinds of design can carry a mark
Google DeepMind introduced SynthID Bio on September 30 as a way to place detectable marks in AI-generated protein sequences and predicted three-dimensional structures. NeoTeo also reported the announcement and distinguished the separate methods for sequences and structures. The sequence method steers amino-acid selection in the ProteinMPNN design tool so that a detectable signal can be left in the resulting design. For predicted structures, DeepMind adjusts a small part of AlphaFold 3 and embeds a signal in the calculated atomic coordinates. These are two different biological objects and two different routes for marking them. A mark in a sequence should not be assumed to serve the same purpose at the same stage as one in a predicted structure.[1], [2]
Laboratory tests compared function
The team combined AlphaProteo-designed binders with a SynthID Bio-enabled version of ProteinMPNN and tested them against three targets: VEGF-A, the receptor-binding domain of the SARS-CoV-2 spike protein, and PD-L1. It compared marked and unmarked designs for hit rate, binding strength and sequence diversity. DeepMind reports comparable results on those measurements. That supports preservation of the functions examined in these binders; it does not establish the same outcome for every protein or design method. On the structure side, the company says prediction accuracy was maintained and the signal survived small digital changes. Those findings also belong to the model and test conditions described, rather than a general field result. No broad deployment of the method in biological production was announced.[1]
A provenance mark is not a safety approval
DeepMind proposes that the mark could give DNA synthesis providers one signal about whether an unfamiliar design came from a trusted model. Marking synthetic entries in databases such as Protein Data Bank, UniProt and GenBank is another proposed use. Safety screening would still require a separate assessment of the protein and its intended use. The company says resilience against deliberate tampering still needs work. NeoTeo describes tests in which particular redesign operations removed a mark, reinforcing its role as an additional trace rather than a complete screening system. DeepMind says it is releasing the methods paper, code, laboratory data and research weights. Those materials can let researchers examine how well the signals survive changes beyond the initial demonstrations.[1], [2]
