The authors analysed 753 vertebrate and 481 invertebrate genomes to examine how methylation-associated deamination influences genome dinucleotide composition. Across vertebrates, the primary pattern of variation in dinucleotide frequencies was a pronounced depletion of CpG. This dominant axis indicates that long-term, context-dependent mutational processes linked to methylation have substantially reorganised vertebrate genome composition over evolutionary time.
Although methylation-associated deamination of 5-methylcytosine at CpG sites produces TpG/CpA as immediate mutational products, the strongest between-genome correlate of CpG loss was enrichment of AG/CT, not TpG/CpA. The dataset nevertheless showed the expected dataset-wide mass balance for TpG/CpA, but those dinucleotides varied little among genomes, whereas AG/CT displayed the largest genome-to-genome variation correlated with CpG depletion. This result positions AG/CT enrichment as a more informative marker of the long-term mutational axis than the direct deamination products.
To probe causal mechanisms, the authors implemented a forward-evolution model driven by measured seven-nucleotide human germline substitution rates. When the model excluded a CpG-specific increase in mutability, it produced neither the observed CpG depletion nor AG/CT enrichment. Introducing a single CpG-specific mutability term — calibrated solely to reproduce the mammalian CpG ratio — recreated both empirical features. This finding supports the interpretation that CpG-specific mutability, likely driven by methylation-associated deamination, is necessary to generate the broad compositional patterns seen across vertebrate genomes.
The modelling outcome also implies that AG/CT enrichment arises as a second-order consequence of a context-dependent network of substitutions, rather than as a direct, primary product of CpG deamination. The authors emphasise that context (here represented by seven-nucleotide substitution rates) is critical: without a CpG-specific term, the measured context-dependent rates are insufficient to reproduce the vertebrate dinucleotide axis.
Within individual genomes, CpG depletion was not uniform. The strongest depletion signal was found in transposable elements, and CpG depletion weakened with increasing distance from these elements. This spatial pattern indicates that regions enriched for mobile elements—where methylation levels and turnover may differ—contribute disproportionately to the genome-wide CpG landscape.
When comparing major clades, the methylation-associated dinucleotide axis tracked Amniota more closely than it did traits such as endothermy. The axis was also associated with an expanded GC-rich isochore compartment in genomes that exhibited stronger CpG depletion and AG/CT enrichment. These associations suggest that the long-term action of methylation-linked mutational processes contributed to large-scale compositional partitioning in vertebrate genomes.
A Machine Learning analysis incorporated dinucleotide features across the surveyed genomes and identified CpG depletion and AG/CT enrichment as the principal features distinguishing vertebrates from invertebrates. Invertebrate genomes showed markedly attenuated signals for both CpG loss and AG/CT enrichment, implying that the methylation-associated mutational axis is a defining characteristic of vertebrate genome evolution.
The study synthesises comparative genome analysis, context-aware forward modelling, and machine learning to argue that a methylation-associated, CpG-specific increase in mutability organised vertebrate dinucleotide composition over roughly 500 million years. Although TpG/CpA are the immediate mutational products of CpG deamination, the clearest long-term marker of this axis across genomes is AG/CT enrichment, a second-order consequence of context-dependent substitution dynamics. Within genomes, transposable elements are hotspots of CpG depletion, and across clades the axis aligns with Amniota and expansion of GC-rich isochores.
Limitations and source reporting
This summary is based solely on the content presented in the preprint abstract and article metadata. Detailed methods, full model parameters, quantitative effect sizes, and additional results were not reported in the provided source text; readers should consult the full preprint for those specifics and note the article is a preprint and has not been peer reviewed.