Improved ancestral genome reconstruction using a learned gene-content grammar
Abstract
Ancestral gene content inferences allow inferring the set of genes - and by extension, the cellular features and metabolic capabilities - of ancestral organisms, based on data from modern genomes. Current methods differ in their approach to ancestral inferences and the kinds of errors they make: reconciliation methods map gene trees onto species trees, and tend to under-estimate ancestral contents due to phylogenetic noise; profile methods model the evolution of phylogenetic profiles (presence-absence or count data) on the species tree, and tend to return inflated ancestors because they ignore gene trees and as a result can only account for horizontal gene transfer (HGT) in a limited manner. For reasons of tractability, both approaches also share a core limitation: neither uses the fact that genes do not act alone but belong to operons, protein complexes, and metabolic pathways that may be gained and lost together or experience shared selective constraints. Here, we show that this context - the grammar of gene content - provides a rich source of information that can be used to greatly improve ancestral gene content inference and metabolic reconstruction under both the reconciliation- and profile-based approaches. We model this structure as an Ising model and infer its parameters from 113,104 bacterial and archaeal genomes (one per species representative in GTDB). We validate the model on extant taxa using phylum-level holdout (i.e. using test data from different prokaryotic phyla than training data), showing that it can accurately "denoise", i.e., reconstruct gene repertoires from highly fragmented and noisy input data, learning about protein-protein interactions and gene essentiality during the training process. When applied to ancestral reconstructions, the denoiser fills gaps in conservative reconstructions and removes excess genes from overly-generous ones, such that different reconstruction methods converge to broadly concordant conclusions. By using this gene content grammar, patchy method-dependent ancestral reconstructions can be turned into organism-like ones, and yield agreement on the gene families, cell-biological features, and metabolic capabilities of the deepest nodes in the tree of life.
Related
- Ancestral Karyotype Reconstruction Reveals Chromosome Rearrangements and Large Inversions in Oryza
- Ancestral protein reconstruction uncovers unanticipated properties of the ancestral Arf protein and of modern Arf1
- A Specific Ancestral Sequence Reconstruction of Flagellin Can Drive Motility in Salmonella enterica but not Escherichia coli
- Reconstruction of ancestral plant genomes for inter-crop translational research
- Reconstruction of the gastropod ancestral karyotype reveals lineage-specific chromosomal evolution and adaptations during habitat transitions
BibTeX
@misc{szollosi2026improved,
title = {{Improved ancestral genome reconstruction using a learned gene-content grammar}},
author = {G. J. Szöllosi and A. Spang and B. Boussau and T. A. Williams},
howpublished = {bioRxiv},
year = {2026},
month = sep,
doi = {10.64898/2026.09.03.749268},
url = {https://www.biorxiv.org/content/10.64898/2026.09.03.749268v1},
note = {CC BY-NC},
}
RIS
TY - UNPB
TI - Improved ancestral genome reconstruction using a learned gene-content grammar
AU - G. J. Szöllosi
AU - A. Spang
AU - B. Boussau
AU - T. A. Williams
PY - 2026
DA - 2026/09/03
PB - bioRxiv
DO - 10.64898/2026.09.03.749268
UR - https://www.biorxiv.org/content/10.64898/2026.09.03.749268v1
AB - Ancestral gene content inferences allow inferring the set of genes - and by extension, the cellular features and metabolic capabilities - of ancestral organisms, based on data from modern genomes. Current methods differ in their approach to ancestral inferences and the kinds of errors they make: reconciliation methods map gene trees onto species trees, and tend to under-estimate ancestral contents due to phylogenetic noise; profile methods model the evolution of phylogenetic profiles (presence-absence or count data) on the species tree, and tend to return inflated ancestors because they ignore gene trees and as a result can only account for horizontal gene transfer (HGT) in a limited manner. For reasons of tractability, both approaches also share a core limitation: neither uses the fact that genes do not act alone but belong to operons, protein complexes, and metabolic pathways that may be gained and lost together or experience shared selective constraints. Here, we show that this context - the grammar of gene content - provides a rich source of information that can be used to greatly improve ancestral gene content inference and metabolic reconstruction under both the reconciliation- and profile-based approaches. We model this structure as an Ising model and infer its parameters from 113,104 bacterial and archaeal genomes (one per species representative in GTDB). We validate the model on extant taxa using phylum-level holdout (i.e. using test data from different prokaryotic phyla than training data), showing that it can accurately "denoise", i.e., reconstruct gene repertoires from highly fragmented and noisy input data, learning about protein-protein interactions and gene essentiality during the training process. When applied to ancestral reconstructions, the denoiser fills gaps in conservative reconstructions and removes excess genes from overly-generous ones, such that different reconstruction methods converge to broadly concordant conclusions. By using this gene content grammar, patchy method-dependent ancestral reconstructions can be turned into organism-like ones, and yield agreement on the gene families, cell-biological features, and metabolic capabilities of the deepest nodes in the tree of life.
ER -