STEPHY: A Graph Neural Inference Framework for Rapid Estimation of Regional Epidemic Dynamics from Large Viral Phylogenies

Abstract Genomic surveillance has become a cornerstone of epidemic response, yet translating the large phylogenies generated by modern sequencing efforts into actionable regional epidemic intelligence remains beyond the reach of most public health systems. Existing phylogeographic methods are too computationally demanding for routine deployment, creating a bottleneck between genomic data generation and epidemiological insight. Here, we introduce STEPHY, a readily deployable graph neural inference toolkit that enables rapid estimation of regional epidemic dynamics directly from large viral phylogenies. By representing phylogenies as region-level graphs and integrating information across connected regional outbreaks, STEPHY jointly estimates region-specific reproduction numbers, recovery rates, source-sink roles, and the most likely index location — epidemiological quantities that directly inform geographically targeted surveillance and intervention. STEPHY is adaptable to diverse pathogens and surveillance settings through locally tailored simulations. We validate STEPHY through simulation studies across complex metapopulation systems, demonstrating accurate recovery of the regional epidemiological quantities and showing that jointly modeling cross-regional dependencies substantially outperforms independent region-by-region analysis. We apply STEPHY to five densely sampled SARS-CoV-2 lineage phylogenies from Denmark, one of the most comprehensively sequenced national epidemics on record, spanning the Alpha, Delta, and Omicron waves under markedly different epidemic conditions. STEPHY consistently identifies Hovedstaden as the dominant source of viral dissemination, with its source role becoming increasingly pronounced from Alpha through Omicron, revealing a progressive strengthening of spatial transmission heterogeneity across successive variant waves. Inference across phylogenies containing 571 to 7,177 sequences requires only milliseconds, confirming that STEPHY scales to the demands of real-world genomic surveillance. Together, these results establish STEPHY as a scalable and readily deployable toolkit for regional epidemic inference and response using routine genomic surveillance data.