Researchers at the University of Maryland have adapted Google DeepMind’s SynthID technology to embed invisible watermarks into AI-designed protein sequences. The method utilizes an unbiased Gumbel sampling approach to modify the probability distribution of amino acids during the generation process, specifically within the ProteinMPNN model released in 2022.
Tests confirmed that the watermarking process does not affect the pLDDT scores, which measure the predicted folding quality of proteins. The watermark is integrated directly into the sequence at the time of generation, making it resistant to file format changes or copy-paste actions. Detection requires access to a specific private key.
This proof-of-concept aims to enhance biosecurity by allowing organizations to verify if a protein sequence originated from a specific AI model. The technology could potentially integrate with existing screening frameworks, such as those maintained by the International Gene Synthesis Consortium.
Currently, there is no commercial product named SynthID Bio available. The research remains in the proof-of-concept stage, serving as an extension of the existing SynthID system used for text, images, audio, and video.
