Detection of local recombination and mutation hotspots in bacterial genomes using neural posterior estimation
summary
The gist
Detection of local recombination and mutation hotspots in bacterial genomes using neural posterior estimation presents a new statistical framework to infer gene-specific recombination and mutation…
In short
The method uses neural posterior estimation (NPE) trained on simulations to infer gene-specific recombination and mutation rates across bacterial genomes. It identifies hotspots using a circular posterior-propagated Difference-of-Gaussians detector, allowing for the joint inference of both recombination and mutation hotspots.
Key concepts
- Neural Posterior Estimation (NPE)
- NPE is a statistical framework trained on simulations conditioned on a fixed clonal genealogy. It learns to recover true local recombination and mutation rates by using gene length as an observed statistic, effectively learning how gene structure influences these evolutionary parameters.
- Circular Posterior-Propagated Difference-of-Gaussians (ppDoG)
- This detector identifies hotspots by contrasting narrow and broad-scale smoothed log rates. It treats the genome circularly, allowing it to detect local enrichment signals across the entire chromosome, providing a spatial contrast between local rates and regional background rates.
- Summary Statistics
- These are mathematical measures derived from genetic data like Watterson’s estimator or Tajima’s D. The paper uses these statistics to capture information about recombination and mutation while significantly reducing the input dimensionality for the neural network, making the analysis more efficient.
Terminology used across episodes
This episode discusses
- Detection of local recombination and mutation hotspots in bacterial genomes using neural posterior estimation · Paper Radio
- Validating Bayesian Inference Algorithms with Simulation-Based Calibration
- Collective Fluctuations in models of adaptation
The paper
Detection of local recombination and mutation hotspots in bacterial genomes using neural posterior estimation · Read on arXiv
Haochu Liu, Richard G. Everitt, Xavier Didelot
Department of Statistics, University of Warwick · Zeeman Institute for Systems Biology and Infectious Disease Epidemiology Research, University of Warwick · School of Life Sciences, University of Warwick
Motivation: Recombination and mutation rates in bacteria vary along the genome, and regions of elevated rate (hotspots) often reflect selection on specific genes. Detecting them is difficult because the two processes leave overlapping signals in sequence data, clonal relationships act as a confounder, and inference must scale to whole genomes. Existing methods assume a constant mutation rate and/or rely on a small set of summary statistics. Results: We present a simulation-based framework that infers gene-specific recombination and mutation rates under a well-specified model of bacterial recombination. Neural posterior estimation (NPE) is trained on simulations conditioned on a fixed clonal genealogy, using a wide range of summary statistics. Gene length is also used as both a parameter and an observed statistic, so that the model learns about its effect on other statistics and thus becomes applicable across different gene lengths. To identify hotspots, we introduce a circular posterior-propagated Difference-of-Gaussians detector that contrasts local and regional smoothers of the posterior log rates, carrying NPE uncertainty into hotspot probabilities. On simulated data, the method recovers the true rates, with credible intervals that narrow as gene length increases. We also apply the new methodology to three real datasets in Bacillus cereus, Staphylococcus aureus and Escherichia coli to showcase its usefulness in a range of situations. Availability and implementation: The Python implementation is freely available from https://github.com/haochu-liu/bacnpe.
Transcript
Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.
Ines: We're looking at the paper: "Detection of local recombination and mutation hotspots in bacterial genomes using neural posterior estimation".
Marcus: Here's the gist of the paper. Detection of local recombination and mutation hotspots in bacterial genomes using neural posterior estimation presents a new statistical framework to infer gene-specific recombination and mutation rates across bacterial genomes and identify hotspots through a circular posterior-propagated Difference-of-Gaussians detector The method uses neural posterior estimation (NPE) trained on simulations conditioned on a fixed clonal genealogy to recover true rates, with credible intervals narrowing as gene length increases Inference Framework The methodology is based on a simulation-based approach that exploits the ClonalOrigin model to approximate the ancestral recombination graph conditional on a fixed clonal genealogy T This model allows for control over the clonal genealogy, which is the tree obtained by following the clonal ancestral relationships between the genomes The framework extends this model by considering that several regions along the genome each have its own recombination and mutation rate, calling these regions genes even if they may not represent biological gene units Training Signal Construction The neural posterior estimation (NPE) is trained on simulations conditioned on a fixed clonal genealogy using a wide range of summary statistics The model learns about the effect that gene length has on the summary statistics and their relation with local recombination and mutation rates by using gene length as both a parameter and an observed statistic The compression network represents each gene by a vector z = (s1,..., sd, l) where s1,..., sd are the summary statistics and l is the gene length Hotspot Detection Mechanism To identify hotspots, a circular posterior-propagated Difference-of-Gaussians detector is introduced that contrasts local and regional smoothers of the posterior log rates The score is calculated as D(m)i = N(m)i − B(m)i, where N(m)i and B(m)i denote the narrow- and broad-scale smoothed log rates for gene i in posterior draw m The posterior hotspot probability is estimated as Pbhot i = 1/M X M m=1 I h D(m)i > log Fmini, where Di = Ni − Bi is the score computed from the log rates Empirical Validation and Findings The method was tested on simulated datasets where true parameter values are known, showing that it recovers the correct values with good correlation for both recombination and mutation rates In applications to real datasets like Bacillus cereus, Staphylococcus aureus, and Escherichia coli, candidate hotspots were identified for both recombination and mutation The results suggest that in the B.
Ines: cereus group recombination is the driving evolutionary force of diversification as a result of its commensal lifestyle Furthermore, the analysis revealed that a mutation hotspot in E.
Marcus: coli was previously misclassified as a recombination hotspot because previous methods assumed a constant rate of mutation throughout the genome Benchmarking Performance The comparison between NPE, NLE, and NRE showed that NPE generally requires less simulation budgets to achieve good testing results than NLE and NRE The empirical coverage of the 95% posterior credible intervals was close to the nominal level for all methods Euclidean errors between posterior medians and true parameter values were generally small, with a median of 0.41 The final NPE configuration selected was a compact model consisting of one hidden layer with 128 units and an output dimension of 2 + 1 = 3 Application Specific Hotspots For the Bacillus cereus group, three real-data applications identified seventeen recombination hotspots and no mutation hotspots For Staphylococcus aureus, ten mutation and six recombination hotspots were found, with some mutation hotspots close to mobile element insertion sites For Escherichia coli, only two recombination hotspots and two mutation hotspots were found The ppDoG detector successfully identified candidate regions in all three cases, with the minimum enrichment thresholds being adjusted based on the observed contrasts The choice of bandwidths is a practical strategy where hB is fixed at 1,000,000 bp and hN and Fmin are adjusted for each dataset to ensure clear local signals The method's robustness is demonstrated by the agreement between inferred gene lengths and observed values across all tested datasets The paper concludes that the joint inference of mutation and recombination hotspots potentially indicates that some of the excess homoplasy described in previous analyses is attributable to high mutation rather than recombination The method is implemented into open source software which should be useful for many scientific researchers working on bacterial genomics Limitation and Future Work A limitation of the approach is its dependence on the assumptions encoded by the simulator used for training, specifically the ClonalOrigin model Future work could build on this framework by developing more comprehensive simulation models that impose different restrictions on recombination events The simulation framework could also be extended to models with biased recombination, natural selection, changes in population size, or migrations between structured populations Training NPE models on simulations spanning these alternative mechanisms would make it possible to estimate additional evolutionary parameters More flexible simulators combined with validation across multiple data-generating models could ultimately improve the robustness and biological interpretability of simulation-based inference The present analysis primarily resolves variation in recombination and mutation parameters at the gene level, with a posterior-propagated Difference-of-Gaussians method applied as a postprocessing step in order to detect hotspots at a wider genomic level Diagnostic Results The diagnostics showed that the NPE-C posterior was well calibrated and provided accurate point estimates, with empirical coverage of 93.6% for ρs and 95.1% for θs Euclidean errors between the posterior medians and true parameter values were generally small, with a median of 0.41 The SBC rank histograms showed no pronounced departures from uniformity, while TARP expected coverage closely followed the diagonal The inferred gene lengths almost perfectly matched their true values, confirming that gene-length information was preserved by the inference pipeline The paper provides a Python implementation freely available from https://github.com/haochu-liu/bacnpe Hotspot Identification Details The ppDoG detector treats the posterior log rates along a chromosome as an irregularly sampled spatial signal, using circular genomic distances to treat genes near the origin and end of a circular chromosome as neighbours The score D(m)i = N(m)i − B(m)i is calculated on the log scale, and exp(D(m)i) is the corresponding local-to-background fold enrichment Genes with Pbhot i ≥ 0.9 were designated candidate hotspot genes The bandwidths hN and hB set the genomic scales to which the detector responds, and the narrow bandwidth should be comparable to the expected width of a localized enrichment The broad bandwidth hB should be substantially larger than hN so that B(m)i follows regional variation without following the local peak The choice of Fmin is adjusted for each dataset and rate type so that the narrow smoother provided a clear local signal distinct from the background The paper recommends checking that the main candidate regions remain stable under moderate changes of hN and Fmin The results show that in E.
Ines: coli, a mutation hotspot was previously misclassified as a recombination hotspot because previous methods assumed a constant rate of mutation throughout the genome The method's ability to jointly infer mutation and recombination hotspots potentially indicates that some of the excess homoplasy described in previous analyses is attributable to high mutation rather than recombination The summary statistics computed include Watterson’s estimator for the number of segregating sites, Tajima’s estimator for Tajima’s D, and homoplasy index HI Linkage disequilibrium (LD) between two segregating sites is captured by statistics such as D, D′, and r2 The four-gamete (G4) test is used to indicate that at least one recombination event has occurred between two sites These statistics are summarized as exact and proportional LD and G4 features Kelly’s statistics, Hudson’s estimator RM, and Wall’s statistics are also defined for summary statistics The paper uses these statistics to retain information relevant to recombination and mutation while substantially reducing the dimensionality of the inputs Implementation Details The full NPE procedure is summarized in Algorithm 2, which combines the posterior-estimation objective with the length-preserving compression network The compression network uses a fully connected multilayer perceptron (MLP) gψe to compress the input vector z into m learned nonlinear features ze = gψe(z) The final NPE configuration selected was one hidden layer with 128 units and an output dimension of 2 + 1 = 3 This architecture was chosen after a staged tuning procedure that reduced the computational cost while allowing major architectural choices to be evaluated systematically The training used a learning rate of 10−4, a batch size of 100, and an early-stopping patience of 20 epochs The paper provides the Python implementation freely available from https://github.com/haochu-liu/bacnpe The paper is an arXiv preprint with the identifier arXiv:2611.13763v1
q-bio.GN: The results confirm that the method is able to recover the correct values for recombination and mutation rates in simulated datasets The application to Bacillus cereus identified seventeen recombination hotspots and no mutation hotspots, while Staphylococcus aureus found ten mutation and six recombination hotspots The application to Escherichia coli found two recombination hotspots and two mutation hotspots, with one region clearly coinciding with the first The results highlight the importance of considering both the mutation and recombination rates as variable along the genome The paper demonstrates that NPE is generally more computationally efficient when posterior distributions were required for many observations The paper is an arXiv preprint with the identifier arXiv:2611.13763v1
q-bio.GN: Discussion and Future Directions A limitation of our approach is its dependence on the assumptions encoded by the simulator used for training, which could be improved by investigating the use of the full ARG model as a simulator or another approximation Future work could build on this framework by developing more comprehensive simulation models that impose different restrictions on recombination events The simulation framework could also be extended to models with biased recombination, natural selection, changes in population size, or migrations between structured populations This would make it possible to estimate additional evolutionary parameters and reduce bias when analysing empirical genomes The present analysis primarily resolves variation in recombination and mutation parameters at the gene level, with a posterior-propagated Difference-of-Gaussians method applied as a postprocessing step in order to detect hotspots at a wider genomic level The paper is an arXiv preprint with the identifier arXiv:2611.13763v1
q-bio.GN: Validation Diagnostics The diagnostics indicated that the NPE-C posterior was well calibrated and provided accurate point estimates, with empirical coverage of 95.0% for gene length Euclidean errors between the posterior medians and true parameter values were generally small, with a median of 0.41 The SBC rank histograms showed no pronounced departures from uniformity, while TARP expected coverage closely followed the diagonal The inferred gene lengths almost perfectly matched their true values, confirming that gene-length information was preserved by the inference pipeline The paper is an arXiv preprint with the identifier arXiv:2611.13763v1
q-bio.GN: Hotspot Interpretation The results show that in E.
Title and authors: Marcus: Now we get into the actual summary of "Detection of local recombination and mutation hotspots in bacterial genomes using neural posterior estimation." The authors present their simulation-based framework designed to infer gene-specific recombination and mutation rates under a well-specified model of bacterial recombination.
Ines: They use neural posterior estimation, or NPE, trained on simulations conditioned on a fixed clonal genealogy. This is the engine that recovers the true rates. And to find those hotspots, they introduce a circular posterior propagated Difference-of-Gaussians detector.
Yuki: I’m thinking about the practical application of that detector—it contrasts local and regional smoothers of the posterior log rates, which means it carries that neural posterior estimation uncertainty directly into the hotspot probabilities.
Marcus: That uncertainty propagation is what makes it different; instead of just picking genes with the highest absolute rate, they are contrasting how much a region deviates from both its immediate neighborhood and its broader context. The score they calculate is D of m i equals N of m i minus B of m i, where N and B are narrow- and broad-scale smoothed log rates.
Ines: So, the main implication is that they are not just identifying high rates; they are detecting relative spatial enrichment in the posterior log rates. This contrasts with methods that might only look at a single point estimate for a rate.
Yuki: That sounds like it could help us see if we’re looking at true biological variation or just noise within the model, which is crucial when you're working with complex genomic data.
Marcus: They show empirical validation on simulated data where they know the true parameter values, and they report good correlation for both recombination and mutation rates. They also test it on real datasets like Bacillus cereus, Staphylococcus aureus, and Escherichia coli to identify candidate hotspots for both processes.
Ines: The finding in *Bacillus cereus* is interesting because it suggests recombination is the main driving force of diversification there due to its lifestyle as a commensal organism. But wait, they also found a mutation hotspot in *E. coli* that was previously misclassified as a recombination hotspot because older methods assumed constant mutation rates across the genome.
Yuki: That’s a big finding for me because it shows how easily prior assumptions can lead to misinterpretations, and it highlights the need to consider both mutation and recombination rates when analyzing these genomes.
The paper's summary: Ines: The paper outlines several improvements they made to this approach. First, they used gene length as both a parameter and an observed statistic in the training signal construction, which lets the model learn how gene length affects other statistics.
Marcus: That addresses one of the limitations I see in older methods where you often only use a fixed set of summary statistics. By incorporating gene length into this way, they make it more applicable across different gene lengths.
Yuki: And on top of that, they are improving how they identify hotspots by using that circular posterior propagated Difference-of-Gaussians detector to carry the NPE uncertainty into the final probability estimation. That’s a big step in making the detection more robust.
Ines: They also compared their neural posterior estimation method against other approaches like NLE and NRE, showing that NPE generally requires less simulation budget to get good testing results than those methods.
Marcus: And they achieved good empirical coverage for the ninety-five percent posterior credible intervals, with Euclidean errors between the posterior medians and true parameter values being generally small, with a median of zero point four one.
Yuki: That small error rate suggests the method is quite reliable when you look at the precision of its estimates compared to what we actually know from simulations. It’s good to see that when we test it on real data, these results hold up reasonably well across different species like *Staphylococcus aureus* and *E. coli*.
The paper's improvements: Ines: So wrapping up the paper "Detection of local recombination and mutation hotspots in bacterial genomes using neural posterior estimation." The main implication is that by jointly inferring both processes, we can potentially attribute some of the excess homoplasy seen in previous analyses to high mutation rates rather than just recombination.
Marcus: That’s a big statement about how we interpret evolutionary signals; it suggests our previous understanding of which process dominates homoplasy might be incomplete. They found specific hotspots across *Bacillus cereus*, *Staphylococcus aureus*, and *Escherichia coli* for both recombination and mutation.
Yuki: For the population genetics side, this means that when we analyze these bacterial genomes, we can no longer treat mutation and recombination as separate things; they are intertwined in a way that needs joint inference. It suggests we need to be more careful about what sample size supports which evolutionary story.
Ines: Exactly. The method itself is computationally efficient because it uses neural posterior estimation, and it’s implemented in open source software, which should help researchers use this framework for gene-level inference across many different genomic analyses.
Marcus: It’s a lot of work to build that simulation model to condition the NPE on the clonal genealogy, but the results show that when you get those posterior distributions right, they can provide accurate point estimates and preserve important information like gene lengths.
Yuki: I just think this paper shows us how combining different statistical approaches—simulation-based inference with neural networks—can actually yield a more nuanced picture of bacterial evolution at the gene level.
Ines: That’s the gist of it. We'll keep an eye on this framework as researchers try to build more comprehensive simulation models to test its limits, because that dependence on the simulator is a thing they flag as a limitation.
Marcus: Yeah, so we’re looking forward to seeing how this framework evolves when they build in more complex evolutionary scenarios or selection pressures into those training simulations.
Conclusion: Ines: So we’ve seen how they used neural posterior estimation to infer gene-specific recombination and mutation rates in bacterial genomes, which is pretty cool because they built a way to find those hotspots using a circular detector that carries the uncertainty into the final probability estimate.
Marcus: Yeah, and what I liked was how they trained that model on simulations conditioned on a fixed clonal genealogy, so they could recover the true rates even when looking at real data like *Bacillus cereus* or *Staphylococcus aureus*.
Yuki: For me, it’s interesting that in those studies, recombination seemed to be the main driver for diversification in some groups, which suggests we should look at recombination more closely when thinking about how these bacteria evolve their niches.
Ines: Right, and they even found something surprising—a mutation hotspot in *E. coli* that was previously misidentified as a recombination one because older methods assumed constant mutation rates across the whole genome.
Marcus: That’s a key point for me, because it shows how those old assumptions about uniform rates can totally skew what we think is happening in the data before we even look at the statistics.
Yuki: And from a population genetic standpoint, if you can disentangle mutation and recombination hotspots like this, it gives us much better tools to understand the actual evolutionary pressures acting on those bacterial populations.
Ines: Overall, it’s a solid statistical framework that uses neural posterior estimation to handle those complex relationships between gene length and the observed summary statistics.
Marcus: The comparison against other methods showed that this neural approach was generally more computationally efficient when you needed those posterior distributions for many observations, which is important for large cohort studies.
Yuki: It really shows how these simulation-based approaches can provide a way to push beyond just picking the highest rates and actually figure out where the significant localized signals are.
Ines: So that’s "Detection of local recombination and mutation hotspots in bacterial genomes using neural posterior estimation" for this week, showing us how joint inference helps uncover hidden evolutionary forces in bacteria.
Marcus: Definitely something worth checking out if you’re working with sequence data or looking at those specific summary statistics.
Yuki: We’ll be back next time to look at how trajectory inference helps us understand human aging from DNA methylation data.
More episodes
- 2610.33881-Sensorimotor behavior from an untrained brain-and-nerve-cord connectome
- 2610.40344-Morphological Features of Murine Pulmonary Arterial Networks for Diagnosing Vascular Disease
- Daily Summary for 2026-11-19
- 2609.10193-Are You Learning Biological Signal or Shortcuts? Auditing and Mitigating Bias in Protein-Protein Interaction Datasets
- 2610.38236-Generative discovery of chemical reaction networks
- 2507.11176-An Interpretable AI framework Quantifying Traditional Chinese Medicine Principles Towards Enhancing and Integrating with Modern Biomedicine
- The day in conversation for 2026-11-19
- Daily Summary for 2026-11-19
- 2504.03590-Towards a Unified Framework for Determining Conformational Ensembles of Disordered Proteins
- 2603.15217-A multiscale discrete-to-continuum framework for structured population models