Talk Title: Using machine learning to accommodate complexity in phylogenetics and population genetics

Date: Friday, December 6, 2025. 11:00 am Eastern Time (EST)

Speaker: Megan L. Smith, Mississippi State University

Join the webinar at https://iscb.junolive.co/Nucleus/live/mainstage/iscbacademycosi99

Non-ISCB members, please first register here: https://iscb.swoogo.com/ISCBnucleus-registration

Talk Abstract: As the availability of genomic data from across the tree of life has increased, the extent of heterogeneity in phylogenomic datasets has become increasingly evident. The diverse processes shaping genomic variation necessitate increasingly complex models that cannot always be accommodated in standard likelihood or Bayesian frameworks. In light of this heterogeneity, machine learning has emerged as a particularly promising approach. First, I’ll describe our applications of supervised machine learning to infer phylogenetic relationships and demographic histories. While promising, these approaches rely on the use of data simulated under the models of interest to train machine learning algorithms. When the models used to simulate these data do not include processes important in shaping genetic variation in our focal systems, it leads to a mismatch between training data and empirical data. Our results indicate that such model violations can mislead inferences of introgression. However, domain adaptation approaches aim to overcome this limitation of supervised machine learning. Using domain adaptation, we demonstrate that accurate inferences of introgression are possible, even in the presence of complex processes not modelled in the training data.