Speaker
Description
Bayesian phylogenetic inference is commonly used to generate distributions of phylogenies and evaluate evolutionary models from genetic data, concluding with a visual inspection of randomness in parameter sampling. Visual inspection impedes full automation of the algorithm making it inaccessible to non-experts, obstructing cloud-based runs, and impeding the running of multiple replicates. One weakness of stochastic algorithms such as this is the possibility of a result being a local maximum of likelihood rather than a true global maximum, which can only be discerned with copious replicates randomly seeded that better explore probability space and agree on a result. Here we develop a machine learning model to flag run convergence, thus replacing the visual inspection step. Three algorithms for multivariate time series classification were implemented in Python v3.12 with SKtime v0.35: Rocket Classifier (F1 score of 1), Temporal Dictionary Ensemble (F1 score of 0.997), and K Neighbors (F1 score of 0.998). Training datasets of bayesian phylogenetic inference run states were generated by running BEAST v1.10.4 on a set of 200 simulated intra-host HIV sequences inferred with SANTA-SIM v1.0, then sectioning the data into multivariate time series of length 16 states, with each section labeled according to run convergence as evaluated by a sufficiently small phylogenetic path distance between sampled trees and the true simulated phylogeny. Successful detection of run convergence, even of local likelihood maxima, may empower phylogenetics to produce better and faster results.
| Expedited Notification | No thanks, I do not require Expedited Notification |
|---|
