SynthID Bio

Properties

category: watermarking
maturity: emerging
approach: Tournament-sampling watermarking applied to AI-designed biological molecules: SynthID Bio-sequence integrates into ProteinMPNN during autoregressive protein sequence generation, and SynthID Bio-structure is a fine-tuned AlphaFold 3 model that embeds a signal into predicted 3D structures; detection uses g-value scoring.
access: open-source
owns: A Google DeepMind family of function-preserving watermarks for AI-generated proteins (sequences and structures), published in Nature in 2026, with an Apache-2.0 reference repository, aimed at provenance for biological databases and DNA-synthesis screening.
distinctFrom: Closest is synth-id (Google DeepMind SynthID for text, image, audio, video). SynthID Bio is a separate modality (protein sequences and 3D structures) with a function-preservation goal, released as its own repository and Nature paper; none of the tracked watermarking rows cover biological sequences.
source: Google blog: SynthID Bio
notes: Announced 2026-09-30 on Google's blog; the opening sentence quoted is confirmed on that page. GitHub repo google-deepmind/synthidbio was created 2026-09-29 and is Apache-2.0 for software; its README states the AlphaFold 3 model parameters are available only under the separate AlphaFold 3 Model Parameters Terms of Use, so the weights are not open-source. The repo README cites the Nature paper 'Function-preserving watermarking of AI-generated proteins' (2026). Edge of the anti-slop scope: this is provenance and biosecurity watermarking of AI-designed biological sequences and structures rather than content slop. Wet-lab functional-binder claims were not verified against the paper.

Provenance

Source: https://blog.google/innovation-and-ai/models-and-research/google-deepmind/synthid-bio/

Properties

category

watermarking

maturity

emerging

approach

Tournament-sampling watermarking applied to AI-designed biological molecules: SynthID Bio-sequence integrates into ProteinMPNN during autoregressive protein sequence generation, and SynthID Bio-structure is a fine-tuned AlphaFold 3 model that embeds a signal into predicted 3D structures; detection uses g-value scoring.

access

open-source

owns

A Google DeepMind family of function-preserving watermarks for AI-generated proteins (sequences and structures), published in Nature in 2026, with an Apache-2.0 reference repository, aimed at provenance for biological databases and DNA-synthesis screening.

distinctFrom

Closest is synth-id (Google DeepMind SynthID for text, image, audio, video). SynthID Bio is a separate modality (protein sequences and 3D structures) with a function-preservation goal, released as its own repository and Nature paper; none of the tracked watermarking rows cover biological sequences.

source

Google blog: SynthID Bio

notes

Announced 2026-09-30 on Google's blog; the opening sentence quoted is confirmed on that page. GitHub repo google-deepmind/synthidbio was created 2026-09-29 and is Apache-2.0 for software; its README states the AlphaFold 3 model parameters are available only under the separate AlphaFold 3 Model Parameters Terms of Use, so the weights are not open-source. The repo README cites the Nature paper 'Function-preserving watermarking of AI-generated proteins' (2026). Edge of the anti-slop scope: this is provenance and biosecurity watermarking of AI-designed biological sequences and structures rather than content slop. Wet-lab functional-binder claims were not verified against the paper.