AI‑designed viruses are a test of whether biosecurity can keep pace

Scientists have pioneered a new frontier in biological engineering by utilizing artificial intelligence (AI) to craft genetic instructions for bacteriophages, viruses that infect bacteria. These AI-designed viruses, when tested in the laboratory, have proven functional, infecting E. coli bacteria rather than humans. The study employed a bacteriophage model, Evo 2, specifically engineered with safety constraints, excluding viruses infecting animals, plants, and humans from its training data.

However, the research underscores a critical question: as AI becomes more adept at designing biological systems, are our existing safeguards sufficient to keep pace? This work is seen as beneficial, given the potential of bacteriophages for treating antibiotic-resistant bacterial infections. Nonetheless, the surge in biological design capabilities also presents new challenges for biosecurity.

Genome language models, akin to AI language models, analyze genetic code patterns, enabling researchers to study biology and generate DNA sequences through computational means. Evo 2, for example, was trained on trillions of DNA building blocks, providing a potent new avenue for biological exploration while introducing potential risks.

These genetic sequences must undergo multiple transformations before becoming tangible biological threats, offering opportunities to mitigate risks at each stage. Safeguards can be embedded within AI models, as Evo 2's developers omitted human-infecting viruses during training, observing the model's poor performance with human viral proteins.

Such precautions must evolve as AI tools in biology grow more powerful. Another checkpoint arises when digital DNA transforms into physical DNA, with DNA manufacturing companies tasked to screen the requested genetic sequence and the requester for potential security concerns. International protocols for DNA synthesis screening are under consideration, with researchers advocating for standardized approaches as this technology gains wider adoption.

Laboratories and research institutions act as additional layers of protection, evaluating potentially risky research beforehand, including containment methods, access controls, and benefit-risk assessments. Funding bodies, like the UK Research and Innovation (UKRI), have established teams to identify and manage security risks in collaborative research.

Public health preparedness is crucial as well; as biological design accelerates, health authorities must swiftly detect and investigate unusual outbreaks. The Metagenomics Surveillance Collaboration and Analysis Program (mSCAPE) in the UK, led by the Health Security Agency, analyzes genetic material from samples to pinpoint and monitor emerging pathogens, a role that remains pertinent as AI technologies advance.

Preparedness also entails rapid development of diagnostic tests, treatments, and vaccines upon the emergence of new threats. However, a significant global biosecurity challenge lies in the uneven preparedness among countries; some possess robust systems for identifying and responding to biological risks, while others are lagging.

Similarly, vaccine regulation exhibits disparities in countries' capacities to assess new products. Given the transnational nature of biological threats, strengthening these capabilities on an international scale is paramount. Delicate trade-offs emerge, as stringent restrictions could impede useful research, including efforts to comprehend diseases or develop treatments, while lax safeguards might allow scientific advancements to outpace preventive mechanisms.

Furthermore, the open accessibility of AI tools, such as Evo 2, facilitates discovery and expands access to potent research instruments, yet raises concerns regarding the control over their usage once widely disseminated. No single safeguard can guarantee comprehensive protection, emphasizing the necessity for a multifaceted, collaborative approach to ensure the responsible development and application of AI in biological sciences.