DNAMark: AI-Generated DNA Watermarking for Biosecurity and Traceability Across the Central Dogma

Background

The rapidly advancing field of synthetic biology, particularly with the advent of AI language models capable of designing functional genetic sequences, presents significant challenges. There is a critical need for robust mechanisms to ensure biosecurity and protect intellectual property. These powerful AI tools, while enabling innovation in therapeutics and research, also pose dual-use risks, including the potential for designing harmful biological agents. Establishing accountability and traceability for AI-generated DNA sequences is essential for responsible innovation and addressing growing regulatory concerns.

Current approaches to securing digital content, such as those used for natural language, are ill-suited for biological sequences. DNA watermarking faces unique and complex hurdles due to the limited four-nucleotide alphabet and the critical requirement to preserve biological function, especially protein integrity. Existing methods often struggle to maintain traceability across the central dogma (DNA to RNA to protein) and lack the necessary robustness against biological mutations or adversarial modifications. This results in reduced statistical power and an inability to proactively embed verifiable signatures during the generation process.

Technology

This technology introduces two novel watermarking techniques, DNAMark and CentralMark, designed to embed robust, inheritable digital signatures into synthetic DNA sequences generated by AI language models. DNAMark achieves function-preserving watermarking by targeting the third base of codons, utilizing synonymous substitutions without altering the encoded protein’s amino acid sequence. CentralMark creates an inheritable watermark detectable in both DNA and the translated protein by targeting the second base of codons, influencing amino acid identity. Both methods employ a trained watermark model, leveraging DNA or protein embeddings to guide modifications, ensuring robustness and statistical unbiasedness with minimal impact on biological function.

This technology is uniquely positioned to address critical biosecurity and intellectual property challenges in synthetic biology. Unlike generic AI watermarks, these methods are specifically tailored for biological sequences, accounting for genetic code degeneracy and crucial protein function constraints. It provides proactive security by embedding digital signatures during DNA generation, rather than relying on reactive screening. CentralMark further differentiates by enabling watermark detection across the central dogma, from DNA to protein, ensuring traceability throughout biological workflows. Its use of semantic embeddings ensures robustness against mutations, while adaptive strategies preserve sequence quality and biological function, mitigating dual-use risks posed by advanced DNA language models.

Benefits

• Enhances biosecurity and intellectual property protection for AI-generated DNA.
• Enables secure tracing and accountability of synthetic DNA sequences.
• Preserves biological function and sequence quality during watermarking.
• Provides inheritable watermarks detectable in both DNA and translated proteins.
• Offers robust detection against various genetic modifications and attacks.
• Facilitates regulatory compliance and responsible AI use in synthetic biology.

Applications

• Monitor AI-designed pathogens
• Protect synthetic DNA IP
• Secure therapeutic DNA design
• Ensure regulatory compliance
• Integrate AI design tools

Additional info

This system embeds robust, inheritable digital signatures into AI-generated synthetic DNA. DNAMark uses synonymous codon substitutions for function-preserving embedding, while CentralMark targets the second codon base for protein-level traceability. These methods leverage biological language models and adaptive strategies to ensure high detection rates, biosecurity, and intellectual property protection with minimal impact on biological function.