Reconstructing ancient genomes from gene counts: A robust likelihood framework with sampling bias correction.

Deducing the makeup of ancient genomes is a fundamental challenge in evolutionary biology. While vast genomic datasets exist that span the entire tree of life, current methods for ancestral reconstructions struggle to resolve the inherent ambiguities of gene-sequence evolution at scale. Here, we present a numerically robust computational framework that overcomes the topological uncertainty of gene trees. Instead of tracking every single event, our phylogenetic gain-loss-duplication (GLD) model is based on birth-death processes over gene copies along the species tree. We show that the likelihood and its gradient can be computed efficiently under an adjustable observation bias of minimum gene family size. The framework facilitates unconstrained numerical likelihood maximization and ancestral inference by posterior probabilities. We apply this framework to kingdom-level reconstructions over a 269-genome archaeal dataset and demonstrate that GLD recovers ancestral states with high accuracy. We compare GLD inference with phylogenetic reconciliation from gene sequences (ALE method) and show that implausibly frequent horizontal gene transfer inferred by ALE are often statistical artifacts of collapsing phylogenetic signals in large alignments. In contrast, GLD inferences reveal how two layers of opposing evolutionary mechanics shape microbial genomes: a high-frequency tension between genome streamlining and the pervasive influx of transient genes, complemented by an adaptive counterbalance of recurrent, modular losses, and punctuated massive gains. The GLD framework provides a statistically sound foundation for hypothesizing about gene content evolution across the diversity of entire kingdoms.