Bacterial genome hotspots no longer rest on a few summary statistics: neural posterior estimation separates recombination from mutation rates and assigns hotspot probabilities
Synopsis
Recombination and mutation rates vary along bacterial genomes, the two signals overlap, and clonal relationships confound inference, while earlier methods often assumed a constant mutation rate or relied on a small set of summary statistics. This work trains a simulator-based neural posterior estimator conditioned on a fixed clonal genealogy to infer local recombination and mutation rates for each gene segment, then applies a circular posterior-propagated Difference-of-Gaussians detector that carries posterior uncertainty into hotspot calls. On simulated data the inferred rates track the true values and credible intervals narrow as gene length grows; candidate hotspots are reported for three real datasets in the Bacillus cereus group, Staphylococcus aureus and Escherichia coli.
Interpretation
Recombination and mutation rates can now be estimated separately for each gene segment of a bacterial genome, and hotspot calls come with a probability. The earlier comparable inference used approximate Bayesian computation with a small set of summary statistics, and recombination-hotspot detection assumed a constant mutation rate across the genome. On simulated data with known true values, inferred recombination and mutation rates correlate well with the correct values, empirical coverage of posterior credible intervals is close to nominal for both rates, and the median Euclidean error between posterior medians and true values is small.
The approach replaces an intractable likelihood with a neural density estimator trained on simulations, so one training run yields posteriors for every segment. Neural posterior estimation directly approximates the posterior distribution, and after training new observations need neither fresh simulations nor a separate Markov chain Monte Carlo run per segment. Training data come from the ClonalOrigin simulator conditioned on a fixed clonal genealogy, with compressed summary statistics plus gene length as input; for the Bacillus cereus group, 50,000 simulations took about 96 hours of wall-clock time on a single compute node and training took about 56 minutes.
Hotspots come from a Difference-of-Gaussians detector applied to posterior draws rather than from picking the genes with the highest rates. On the circular chromosome the detector contrasts a narrow-scale local smoother with a broad-scale regional background smoother of log rates, so it responds to relative enrichment rather than absolute rate. In the Bacillus cereus group data no mutation hotspots and 17 recombination hotspots were detected; in Staphylococcus aureus ten mutation and six recombination hotspots; in Escherichia coli two recombination and two mutation hotspots.
Perspective
The method infers local recombination and mutation rates per gene segment under a fixed clonal genealogy and reports candidate hotspots from the contrast of two smoothing bandwidths on a circular chromosome; the demonstrated scope is simulated data plus three published datasets in the Bacillus cereus group, Staphylococcus aureus and Escherichia coli.
Inference depends on the assumptions encoded by the training simulator, and neural posterior estimation does not correct for discrepancies between that simulator and the true data-generating process; hotspot calls are scale-dependent evidence of relative enrichment rather than direct evidence of a specific hotspot mechanism, and bandwidths and enrichment thresholds must be adjusted per dataset.