Omicron has been anticipated by the model. The shape of it, not the name, the historical frame, or the particular nation where it would initially be discovered. A computational model trained on the evolutionary history of coronaviruses and the structural constraints of their spike protein had already identified the cluster of mutations that would make something like Omicron particularly dangerous months before the variant that would change the course of the pandemic appeared in southern Africa in November 2021. It had classified those mutations according to how likely they were to provide structural fitness and immunological escape. The majority of them were carried by Omicron. Since the validation was done after the fact, the model was not awarded credit for the prediction made in real time. However, the retrospective analysis, which was published in Nature in late 2023, revealed something that truly surprised the researchers: the instrument had been effective.
One of the more tangible illustrations of what AI-driven viral prediction actually looks like in practice is the system known as EVEscape, which was created by a partnership between academics at Harvard and Oxford. The model integrates two different data streams: structural data from protein folding tools, which determines which mutations are physically acceptable given the geometry of the protein, and evolutionary data from millions of sequenced viral genomes, which demonstrates which mutations have occurred and survived throughout the virus’s history. When you combine those two inputs, you obtain a prioritized list of mutations that are both structurally feasible and consistent with what evolutionary selection has previously allowed. The most likely mutations to emerge next are those that score strongly on both dimensions.
Participating in the pathogen genomics consortia that exchange sequencing data via GISAID’s global database, contributing to protein language model applications for virology, and conducting their own analyzes of circulating viral lineages, researchers affiliated with the University of Toronto’s computational biology programs in Toronto are part of the larger international network engaged in this work. The work is not done in a vacuum. Data sharing at a scale that would have been nearly unthinkable twenty years ago is essential to modern viral prediction. More than 16 million SARS-CoV-2 genome sequences that have been exchanged by researchers from more than 100 countries are currently stored in GISAID. The AI models learn from the evolutionary signal in that dataset, which includes which mutations emerged, which persisted, and which were selected against.
The stakes are currently highest when using these methods to combat H5N1 avian influenza. Since early 2024, H5N1 has been spreading thru dairy cattle herds in the United States, causing spillover infections in farm workers at a rate that worried epidemiologists more than the public discourse. In order to determine which changes will most effectively enable the virus to spread between mammals rather than requiring bird-to-human contact in every instance, several groups have been performing EVEscape-style analysis to the spike protein equivalent of H5N1. Planners for pandemic preparedness should start developing countermeasures against the mutations that the models identify as having the highest chance before they manifest, not after.
This technology has a serious and unresolved biosecurity component. A group at the Arc Institute released research on Evo, a genomic foundation model that was trained on 2.7 million prokaryotic and viral genomes, in February 2024. The model might produce new DNA sequences, such as those for functioning bacteriophages, which are viruses that infect bacteria and are completely artificial intelligence (AI)-designed rather than naturally occurring. The study was published as a scientific breakthrough, and it was. Additionally, it was the first time an AI system producing viable viral genomes from scratch was shown to the public. The international biosecurity community is aggressively debating the gap between that capability and a system that might produce something more harmful than a bacteriophage, but there are currently no established governance structures.
In a 2023 report on biotechnology governance, the National Academies of Sciences specifically brought up the dual-use issue: the same tools that allow vaccine designers to predict which mutations to include in a universal flu vaccine could be used to create enhanced pathogens in different hands and with different intentions. Rather than taking the place of the real advantages, that risk coexists with them. AI-driven pathogen surveillance was specifically identified in the WHO pandemic accord negotiations in 2024 and 2025 as a capability that needed to be part of the infrastructure for international pandemic preparedness. It also required governance frameworks, which are currently lacking in comprehensive form.
