Research
Google hides a signature in AI-designed proteins, for now only in tests
Google DeepMind has unveiled SynthID Bio, a hidden signature that makes it possible to recognise a protein created with AI. It says the signature does not change how the protein works, but for now this is a lab test.
The essentials
3 confirmed facts · 1 according to the source · 3 open questions
- On 30 September, Google DeepMind unveiled SynthID Bio, a technique that leaves a hidden signature in AI-designed proteins.12
- According to the sourceGoogle DeepMind says that, in lab tests with three proteins, the marked designs worked the same as the unmarked ones.12
- The company itself describes SynthID Bio as a proof of concept.2
- The checks that screen DNA orders do not detect AI-designed proteins, according to Ars Technica and Google DeepMind itself.23
Why it matters
To make a protein designed on a computer, you have to order DNA from a company, which compares the order against a list of known threats2. A protein invented by an AI does not look like anything on that list. If the mark is adopted, those companies could automatically tell which program an order comes from. Everything suggests it would only work for the programs that build it in.
The details
According to Google DeepMind, SynthID Bio works in two ways. In a protein's sequence, meaning the list of building blocks (amino acids) that make it up, it subtly steers which blocks are chosen. In the three-dimensional shapes that a program predicts, it adjusts the position of the atoms. The company maintains that the signature can also be checked in the protein once it has been made, not just in the file2.
The company tested it with molecules designed to bind to another specific protein: VEGF-A, a part of the SARS-CoV-2 spike and PD-L1. It says the marked versions bound just as strongly as the normal ones2. It has also built the mark into AlphaFold 3, its program for predicting the shape of proteins: according to Google DeepMind, whatever it generates comes out signed, no matter who runs it2.
What we don't know
- When it will be used outside the lab, and whether the companies that make DNA and other makers of protein design programs will adopt it: the sources give no dates or commitments.
- Whether anyone independent has checked the results. The two outside voices that assess the work appear in Google DeepMind's announcement: one reviewed it and the other gave an opinion on the study.
- Whether someone who deliberately wants to erase the signature can do so. The company only says that it withstands small changes to the three-dimensional shapes.
Background
According to Ars Technica, a warning came almost a year ago that the programs that look for dangerous DNA sequences do not recognise AI-designed proteins, because nobody has studied them enough to know whether they are a threat. Since then, it had not been clear what could be done3.
Sources
- 1We’re introducing SynthID Bio, bringing our watermarking technology to synthetic biology.
- 2Introducing SynthID Bio
- 3Google figures out how to watermark AI-designed proteins
Was this story useful? Yes Not really
This story was written in Spanish by an artificial intelligence system from the sources listed above, and translated automatically. Before publication, a program checks that every figure and every name appears in the sources, and that the translation keeps the same facts, figures and sources. It has not been reviewed by a person. Spanish original · How we work (in Spanish)