Google DeepMind releases SynthID Bio code, weights and Nature paper

Google DeepMind introduced SynthID Bio on September 30, 2026. It extends the company's SynthID watermarking work to protein design tools. The goal is provenance: data produced by a model can later be traced back to it. DeepMind published a paper in Nature the same day, put the code on GitHub under Apache 2.0, and released model weights.

What SynthID Bio is

SynthID Bio is a family of watermarking methods, not a product with a user interface. DeepMind describes it as a proof of concept. A signal is embedded while a model generates its output, and a detector can later check whether that signal is present. The SynthID Bio announcement (opens in a new tab) splits the work in two parts.

One part, SynthID Bio-sequence, hooks into the sampling step of an existing open model, ProteinMPNN. The other, SynthID Bio-structure, is a fine-tuned version of AlphaFold 3. There the marking behavior sits in the model weights, so output from that checkpoint carries the signal whoever runs it.

DeepMind reports that AlphaFold 3's prediction accuracy is preserved in the fine-tuned version, with near-perfect detectability, maintained structural feature distributions, and resistance to digital noise or minor coordinate changes. For the sequence side, DeepMind says laboratory testing found no loss of biological function compared with unmarked output.

What was actually released

Three artifacts came out together:

• The paper, "Function-preserving watermarking of AI-generated proteins", in Nature. It was received on 29 September 2025, accepted on 29 July 2026 and published on 30 September 2026, under CC BY 4.0.
• The google-deepmind/synthidbio repository (opens in a new tab), created on 2026-09-29, with the sequence-side code and validation data. Software is Apache 2.0. ProteinMPNN parameters bundled there stay under MIT, and the data directory is CC BY 4.0.
• Weights for the fine-tuned AlphaFold 3 variant. The AlphaFold 3 repository (opens in a new tab) points to a separate parameter file and keeps those weights under the AlphaFold 3 Model Parameters Terms of Use, not an open-source license.

The repository README calls the sequence work a drop-in replacement for existing pipelines, enabled through flags rather than a rewrite.

How the detector side works for a developer

Marking and detection use a shared parameter set, and the README is blunt that detection fails if they do not match. The documented parameters include --ngram_len (default 4), --watermark_keys (default 0 through 24), --context_history_size (default 1024), --watermark_temperature (default 1.0) and --watermark_top_k (default 21, the vocabulary size).

Detection produces a g-value. The README's examples show unmarked output near 0.5 and marked output above it, for instance 0.8902 in one sample run. A standalone script, compute_g_values.py, scores an existing FASTA file without regenerating anything. Install instructions use Python 3.9 with uv or venv on Linux, and the repository ships a pytest suite covering generation, detection accuracy and the standalone calculator.

The method is meant to be model-agnostic. The README describes wrapping SynthIDLogitsProcessor from the synthid-text package around another autoregressive decoder. Two constraints apply. Keep the full alphabet by disabling top-k filtering, and decode strictly left to right, because each position's signal depends on the positions before it.

Where SynthID Bio is meant to fit

DeepMind frames watermarking as one layer among several, not a complete answer. The announcement quotes two outside voices. Sarah Carter, a biosecurity policy expert and Principal at Science Policy Consulting, who reviewed the work, calls it a piece of the puzzle for tracking provenance. James Diggans, Vice President, Policy and Biosecurity at Twist Bioscience, who gave early feedback on the paper, describes it as a promising addition to existing screening practice.

DeepMind also suggests the marks could help keep public databases such as the Protein Data Bank, UniProt and GenBank accurate, by labeling or flagging entries that came from a model.

Limits and open questions

DeepMind names tamper resistance as the main open problem: making the mark hold up when someone deliberately tries to strip it. It also points to pairing watermarks with provenance metadata, along the lines of C2PA for digital media, or with central registries.

Other things the published pages do not settle:

• Whether any model provider other than DeepMind has agreed to turn this on.
• Whether a synthesis provider or database operator has committed to checking for the marks.
• Detection rates are reported qualitatively in the blog post. The numbers sit in the Nature paper rather than the announcement.
• Work extending the approach to larger genetic sequences, with the Hie lab at Stanford University and Arc Institute, has early laboratory results but no technical manuscript yet.

DeepMind asks organizations interested in partnering to write to [email protected] with a high-level proposal, and says not to include confidential information in that first contact.