Massively scalable inference of level-1 phylogenetic networks

Recent advancements in sequencing technologies have enabled large-scale
phylogenomic analyses. While these analyses often rely on phylogenetic
trees, increasing evidence suggests that non-treelike evolutionary events,
such as hybridization and horizontal gene transfer, are prevalent in the
evolutionary histories of many species, and in such cases, tree-based
models are insufficient. Phylogenetic networks can capture such complex
evolutionary histories, but current methods for accurately inferring them
lack scalability. For instance, state-of-the-art model-based approaches
are limited to around 30 taxa. Implicit network inference methods like
NeighborNet and Consensus Networks are fast but lack biological
interpretability. Here, we introduce a novel method called InPhyNet that
merges a set of non-overlapping, independently inferred networks into a
unified topology, achieving linear scalability while maintaining high
accuracy under the multispecies network coalescent model. Our simulations
show that InPhyNet matches the accuracy of SNaQ on datasets with 30 taxa
while drastically decreasing the overall network inference time. InPhyNet
is also more accurate than implicit network methods on large datasets
while maintaining computational feasibility. Re-analyzing a phylogeny of
1,158 land plants with InPhyNet, we recover known reticulate events and
provide evidence for the controversial placement of Order Gnetales within
gymnosperms. These results demonstrate that InPhyNet enables biologically
meaningful network inference at previously unprecedented scales.