DIMPLE: A Tree-Guided Divide-and-Conquer Framework for ILS-Aware Phylogenetic Network Inference on Large-Scale Data

Abstract

Inferring evolutionary histories of species provides a foundation for comparative genomics and evolutionary analyses. Extensive evidence of reticulate evolution motivates the inference of phylogenetic networks, which represent relationships that species trees cannot capture. However, accurate reconstruction of large phylogenetic networks remains a major challenge because of the vast space of possible network topologies. Here, we propose DIMPLE, a divide-and-conquer framework for phylogenetic network inference without an explicit restriction on network level. DIMPLE uses a tree-of-blobs (TOB), which represents the tree-like relationships between regions of potential reticulation while leaving relationships within those regions unresolved. Within each unresolved region, DIMPLE partitions the set of taxa into pairwise-disjoint subsets, each containing at most a user-specified number of taxa. Multiple division runs are allowed, and subsets from different runs may overlap. DIMPLE infers a subnetwork on each subset, merges the subnetworks within each region, and uses the TOB to assemble the resulting networks into a network on the full taxon set. Experiments with true inputs demonstrate accurate merging, with improved reticulation recovery over existing methods on higher-level networks. With estimated inputs, DIMPLE achieves competitive topological accuracy and more consistently recovers hybrid ancestry than competing methods across varying levels of network complexity. Its computational requirements support analyses of hundreds of taxa. Applications to fern and gymnosperm datasets recover documented hybridization and previously reported phylogenetic conflict.