TL;DR
Mantis 2 is a foundation model for infectious disease forecasting that combines mechanistic simulations with real-world observational data to enable zero-shot predictions across diverse pathogens and locations. By treating simulations as a "scientific prior," the model significantly outperforms traditional statistical and machine learning baselines, particularly in data-sparse environments like emerging outbreaks.
What Problem Does This Paper Address?
Forecasting infectious diseases is notoriously difficult. Public health agencies need to know when an outbreak will peak and how long it will last to allocate resources effectively. However, traditional models are often "siloed"—built for one specific disease in one specific location. Furthermore, during the early stages of an emerging outbreak, historical data is scarce, making it nearly impossible for purely statistical models to learn the underlying dynamics.
While general-purpose time-series foundation models (like Chronos or TimesFM) have shown promise, they often lack the domain-specific "mechanistic" knowledge—the biological rules of how diseases spread—required for high-stakes public health decisions. Mantis 2 aims to bridge this gap by integrating scientific knowledge from simulations with the messy reality of surveillance data.
Methods
Mantis 2 builds upon the Chronos-2 transformer backbone, but introduces several domain-specific innovations:
- Mixed Training Curriculum: The model is trained on a blend of synthetic and real data. The training starts with 70% simulated trajectories to teach the model "biological plausibility" and gradually shifts to 100% observational data to fine-tune it on real-world noise and reporting delays.
- Latent Feedback Recurrence: The authors implemented a "chain-of-thought" style mechanism for time series. The model processes hidden representations through multiple recurrences before generating a forecast, allowing for deeper computation on complex patterns.
- Multivariate & Multi-horizon: Unlike its predecessor, Mantis 2 can jointly forecast up to six variables (e.g., cases, hospitalizations, wastewater levels, and deaths) simultaneously. It extends the forecast horizon to 16 weeks and can utilize up to 10 years (512 weeks) of historical context.
Datasets
The model was trained on a massive, diverse corpus:
- Simulated Data: 20,000 scenarios generated by the "Reservoir" simulator, covering human, vector-borne, and environmental transmission under varied epidemiological parameters.
- Real-World Data: Over 1,000 scenarios from sources like the Thai Ministry of Public Health, Project Tycho (historical US data), WHO FluNet, and the National Notifiable Diseases Surveillance System (NNDSS).
- Evaluation: The model was tested on 30 distinct disease-location tasks, including "out-of-domain" pathogens like COVID-19, Ebola, and Mpox to test its zero-shot generalization capabilities.
Key Insights and Ideas
- Simulation as a Scientific Prior: The most striking finding is that simulation-rich training acts as a powerful prior. It is most beneficial when local history is short (8–26 weeks). As more local data becomes available, the model relies more on observed patterns, and the relative advantage of simulation diminishes.
- Mitigating "Catastrophic" Errors: Mechanistic training significantly reduced the frequency of exceptionally large forecast errors. By learning the "rules" of transmission, the model is less likely to predict biologically impossible outcomes, even when the input data is noisy.
- The Value of Context: Increasing the usable historical context from 2 years to nearly 10 years allowed the model to better interpret seasonal recurrences and rare epidemic patterns.
Key Results
Mantis 2 achieved the best aggregate probabilistic accuracy (measured by Weighted Interval Score, or WIS) across short, medium, and long horizons compared to six baselines, including SARIMA, LSTMs, and Chronos-2.
- Peak Anticipation: It missed roughly 30% fewer outbreak peaks than the next best model (Chronos-2).
- Lead Time: Successfully anticipated peaks were identified more than six weeks in advance on average.
- Zero-Shot Performance: In a direct comparison on 27 datasets, Mantis 2 reduced pooled WIS by 31.42% relative to the original Mantis model.
Limitations and Future Work
While Mantis 2 represents a significant step forward, the authors note several areas for improvement:
- Retrospective Bias: The current evaluation is retrospective. Prospective, real-time testing is needed to confirm these benefits in active public health crises.
- Data Coverage: While diverse, the training data cannot guarantee performance for every possible emerging pathogen or unique surveillance system.
- Internal Fusion: It remains unclear exactly how the transformer internally combines mechanistic priors with statistical observations. Future work could investigate these internal representations to improve model interpretability.
- Optimization of the Mix: The authors used a 35% average synthetic blend, but the "optimal" ratio likely varies by disease type and data availability. Refining this balance remains an open research question.
Summary updated