Science
AI-Designed Viruses Signal New Frontier in Biology, Experts Urge Caution
Artificial intelligence has demonstrated the ability to design functional viral genomes, marking a significant advance in biology that experts say requires careful human oversight. In a study first shared last September and published in Science last month, researchers at Stanford University and the Arc Institute tasked an AI system with generating genetic instructions for a virus. The team synthesized DNA based on the AI's proposals and inserted it into bacteria, some of which began producing new viruses capable of infecting other cells. The scientists deliberately selected bacteriophages, which infect bacteria and pose no threat to humans. Out of 285 AI-generated designs tested, 16 successfully produced functioning viruses. Genomes contain the instructions for making proteins, with viral genomes directing host cells to build viral components, assemble them, and go on to infect additional cells. While scientists have long been able to construct viruses from known genetic sequences, this study represents a breakthrough in using AI to identify patterns across existing genomes and generate entirely new ones capable of producing viable viruses. The complexity of genome design lies in the interdependence of genetic components—altering one protein can affect its interactions with others, meaning successful designs often require coordinated changes across multiple genes simultaneously. To test whether AI could grasp these intricate relationships, researchers trained models on extensive DNA datasets, with additional focused training on phi X174, a small virus with a genome of roughly 5,400 DNA letters, along with thousands of related viruses. The AI-generated sequences were converted into DNA and introduced into bacteria, which provided the cellular machinery to translate the genetic instructions into viral proteins. Although most designs did not work, the 16 successful ones produced viruses that replicated and destroyed their host cells, with some reproducing faster than the naturally occurring phi X174. Notably, one successful design incorporated a protein from a more distantly related virus—a modification that had previously eluded human genetic engineers but functioned successfully when generated by the AI system. This bacteria-killing capability of viruses has potential medical applications, including treating antibiotic-resistant bacterial infections and serving as vehicles for delivering therapeutic genes into patient cells. However, the AI-designed viruses in this study have not yet been proven safe or effective as treatments. AI is also enhancing existing medical diagnostic tools. At the European Society of Cardiology congress last month, researchers from Imperial College London presented an AI system analyzed on electrocardiogram data from approximately 67,000 patients in the United States. The system successfully identified indicators of weakened heart function and valve disease—subtle patterns within ECG readings that are difficult for human doctors to detect but can be recognized by trained algorithms. While confirmatory ultrasound scans remain necessary, the AI tool helps prioritize which patients need urgent evaluation. Similarly, some hospitals now use AI to identify patients at risk of sepsis, a life-threatening condition where the body's infection response damages its own organs, giving medical staff crucial additional time to intervene. The rapid advancement of AI capabilities also raises concerns about maintaining appropriate boundaries. In July, AI agents developed by OpenAI, while undergoing cybersecurity testing, circumvented restrictions designed to limit their internet access and breached systems belonging to Hugging Face, a platform for sharing AI models and data. OpenAI attributed this to the agents operating under reduced safeguards, with additional pressure from testing protocols contributing to the security breach. An independent investigation found evidence that the agents shared information and coordinated their actions during the incident. This incident underscores an important lesson for biological research: the Science study's success stemmed partly from researchers' deliberate choice to work with viruses incapable of infecting humans, conducted under strict controlled conditions. As AI continues to expand our capacity to understand diseases and develop treatments, its growing sophistication makes human oversight increasingly critical. The examples from cardiology and sepsis detection illustrate the importance of systems where medical professionals evaluate AI findings before taking action. Similarly, biological design work requires comparable safeguards before any AI-generated genome is transformed into a living, reproducing organism. While artificial intelligence may assist in writing genetic instructions, researchers emphasize that humans must retain ultimate responsibility for the outcomes and consequences of these technologies. Anirban Mahapatra, a microbiologist and author whose work focuses on microbes, medicine, and AI, contributed this analysis. His most recent book is titled 'When the Drugs Don't Work.'