Google DeepMind Adds Verifiable Watermarks to AI-Designed Proteins

AI-designed proteins can carry a verifiable signal of their origin within the design itself, Google DeepMind says, without sacrificing biological function. The company introduced SynthID Bio on September 30, 2026, bringing its watermarking technology to protein sequences and predicted three-dimensional structures.

The signal sits inside the biological design

SynthID Bio embeds what DeepMind calls an imperceptible, verifiable watermark directly into biological designs. The signal is therefore part of the design, rather than simply a description attached to the announcement.

One example is a predicted structure of a watermarked VEGF-A protein binder. DeepMind’s illustration uses color to show the watermark signal for each amino acid. That makes the embedded signal visible in the explanatory graphic, while the company describes the watermark itself as imperceptible.

Keeping the protein’s job intact

DeepMind says laboratory tests across target proteins found that watermarked designs matched the performance and natural diversity of unwatermarked versions. That is the company’s reported result, not an independent assessment. It addresses a central requirement of this approach: adding a detectable origin signal while preserving the biological function of the design.

An origin record for open science

DeepMind wants the watermark to provide a provenance layer for biological designs, including those held in open scientific databases. The stated benefit is traceability: a verifiable signal embedded in a design can help establish its origin. The company presents that capability as a contribution to both database integrity and biosecurity.

Those are intended benefits, distinct from the laboratory finding.