Prediction of pathogen transmissibility and virulence from sequence
Predicting how a changed genome will behave in a human population - how contagious, how lethal, how immune-evasive - without having to test it.
State of the art
A RAND Delphi study (2025) concluded that AI-enabled design can plausibly assist optimization of existing pathogens but that prediction of population-level transmissibility remains out of reach; AI protein design can produce functional variants whose phenotype is still only knowable by experiment.
- Transmissibility is an emergent property of host behavior, immunity and ecology, not of sequence alone
- Animal and organoid models correlate poorly with human epidemic potential
- The experiments that would validate such predictions are the ones that must not be run
Who’s working on it
- Debora MarksResearcherHarvard Medical School
Built EVEscape for antibody-escape forecasting (Nature, 2023) and the EVEREST viral variant-effect benchmark (2026), which deliberately excludes virulence endpoints
- Trevor BedfordResearcherNextstrain, previously Fred Hutchinson Cancer Center and HHMI
His 2024 to 2026 frequency-dynamics framework separates intrinsic transmissibility from immune escape using sequence data alone, the closest thing to a population-level predictor
- Nextstrain ↗Program
Open real-time phylogenetics platform, with its automation paper published September 2026, supplying the surveillance data any predictive model is judged against
- RAND Global and Emerging RisksProgram
Ran the 2025 Delphi elicitation finding no method maps sequence to human transmissibility, and a 2026 assessment of whether LLM agents can drive biological design tools
- CEPI ↗Program
Funded the EVEREST benchmark and supplied the biosecurity review that kept harm-enhancing phenotypes out of the released datasets in 2026
- Latch BioStartup
Released the BioSecBench benchmarks in 2026, measuring whether AI agents can infer transmissibility, immune escape and virulence from experimental data
What to invest in
| Kind | Investment | Scale |
|---|---|---|
| research | Predictive modelling benchmarked only on historical natural variants and attenuated systems, with structured hazard review | $$ $10M – $100M |
| policy | Governance of dual-use predictive tools, including access controls and audited evaluation environments | $ < $10M |
Depends on
Required by
No other capability builds on this directly.