#25 · Genomics

Neural and hybrid models infer population history from LD, ARGs, and mtDNA distances

The three preprints share a common thread: recasting population-genetic and phylogenetic inference as a learning or optimization problem over genealogical or distance-based representations. LinkedNN uses a neural network to map linkage disequilibrium decay to recent effective population size, and is reported to outperform prior methods on sparse data. ARGformer applies a transformer to ancestral recombination graphs, learning self-supervised embeddings used for population structure and ancestry retrieval. Hyb-Adam-UM addresses mtDNA phylogeny reconstruction by completing missing pairwise distance matrices with an alignment-limited Needleman-Wunsch backbone and a robust treatment of ultrametric violations. What is new is that the modeled objects are demographic and genealogical quantities themselves: effective population size, recombination graph structure, and tree-consistent distances, rather than sequence features.