Deep Learning and Likelihood Approaches for Viral Phylogeography Converge on the Same Answers Whether the Inference...

Area of Science:

• Computational Biology
• Epidemiology
• Machine Learning

Background:

• Phylogenetic tree analysis is crucial for understanding viral transmission dynamics and epidemiology.
• Traditional likelihood-based phylogenetic methods are computationally intensive, limiting real-time outbreak analysis.
• Deep learning offers a promising avenue for computationally efficient, likelihood-free inference from phylogenetic data.

Purpose of the Study:

• To extend, compare, and contrast a deep learning-based likelihood-free inference method for phylogenetic trees.
• To evaluate the accuracy, robustness, and computational efficiency of deep neural networks against traditional methods.
• To apply the developed deep learning method to real-world phylogeographic data, such as the SARS-CoV-2 pandemic.

Main Methods:

• Trained multiple deep neural networks using simulated outbreak phylogenies across five locations.
• Compared the performance and robustness to model misspecification of neural networks against Bayesian inference.
• Implemented conformalized quantile regression for uncertainty quantification and applied the method to SARS-CoV-2 phylogeographic data.

Main Results:

• Deep neural networks achieved accuracy comparable to Bayesian inference on simulated data.
• Both neural networks and Bayesian methods exhibited similar biases when faced with model misspecification.
• The deep learning approach was over 3 orders of magnitude faster than traditional methods after training, with comparable robustness and accuracy on SARS-CoV-2 data.

Conclusions:

• Deep learning provides a computationally efficient and accurate alternative for phylogenetic tree analysis in epidemiology.
• Trained neural networks can effectively mimic the statistical properties of likelihood-based phylogenetic models.
• This approach holds significant potential for real-time policy-making during rapidly evolving outbreaks.