Date
13
Feb 2026
Time
14:00:00
28 Rue du Docteur Roux, Paris, France
Address
Building: Lwoff Room: Salle RETROVIRUS (RdC 14)
Location
2026-02-13 14:00:00
2026-02-13 15:00:00
Europe/Paris
Dept of Computational Biology | Seminar “Real-time estimation of outbreak transmission dynamics using deep-learning-based phylodynamics” by Ruopeng Xie
Real-time estimation of outbreak transmission dynamics using deep-learning-based phylodynamics Reconstructing epidemic dynamics in real-time has become crucial for effective disease management, as demonstrated by the COVID-19 pandemic. Most methods rely on epidemiological time series […]
28 Rue du Docteur Roux, Paris, France
Anna Zhukova
anna.zhukova@pasteur.fr
About
Real-time estimation of outbreak transmission dynamics using deep-learning-based phylodynamics
Reconstructing epidemic dynamics in real-time has become crucial for effective disease management, as demonstrated by the COVID-19 pandemic. Most methods rely on epidemiological time series data (e.g., reported cases), which can be biased or incomplete due to variable testing policies, particularly in resource-limited settings. Phylodynamics, which leverages viral phylogenies, offers an independent, alternative data source but can be computationally intensive and suited mainly for retrospective analysis. Recent deep-learning approaches (e.g., PhyloDeep, PhyloCNN) have improved scalability, yet challenges remain due to poorly-resolved phylogenies caused by limited genetic divergence during outbreaks. We present a deep-learning framework that integrates pathogen-specific evolutionary and epidemiological signals to enable real-time phylodynamic inference from imperfect genomic data.
To assess the impact of realistic genomic conditions, we simulated transmission trees using multi-type birth-death models for major respiratory viruses (SARS-CoV-2, influenza, and RSV), and transformed them into poorly resolved phylogenies with virus-specific levels of molecular resolution. PhyloCNN models were retrained and performance was evaluated. Significant discrepancies emerged when models trained on one type of phylogeny were applied to another, underscoring the importance of jointly aligning the birth-death model and phylogenetic resolution with the characteristics of each pathogen.
To account for temporal variation in transmission rates (e.g. driven by public health interventions, population immunity, or the emergence of variants with increased transmissibility) we introduced a scalable subtree decomposition strategy to estimate Rt from large and temporally-structured phylogenies, overcoming the limitations of tree-level averaging in deep learning models. Finally, we conducted incremental simulations that mimic realistic outbreak scenarios to assess the sensitivity of real-time inference and determine optimal genomic sampling proportions for accurate epidemic tracking.
Deep-learning–based phylodynamics extends the scalability of birth–death inference to phylogenies with over thousands of tips. Our findings highlight its potential to reliably reconstruct epidemic dynamics from genomic data in near real-time, supporting timely and informed public health decision-making.
Location
Building: Lwoff
Room: Salle RETROVIRUS (RdC 14)
Address: 28 Rue du Docteur Roux, Paris, France