Delphy Brings Near-Real-Time Bayesian Phylogenetics to Outbreak Response

When a novel pathogen begins to spread, genomic sequencing is often the fastest way to answer the questions that matter most: where did it come from, how quickly is it evolving, and which lineages are driving transmission? For more than a decade, Bayesian phylogenetic inference has been the gold standard for answering these questions, because it does more than draw a tree of relationships among viral genomes. It explicitly quantifies uncertainty, estimating probability distributions over divergence times, transmission dates, and evolutionary rates rather than offering a single best guess. Yet the computational cost of Bayesian methods has always been their Achilles’ heel. As sequencing output exploded during the COVID-19 pandemic, the classic Bayesian phylogenetics pipelines simply could not keep pace with the flood of tens of thousands of new genomes per week, and many public health teams fell back on faster but less rigorous maximum-likelihood approaches that discard much of the temporal and uncertainty information that makes phylogenetics genuinely useful in an outbreak.

A new study published in Nature introduces Delphy, a scalable Bayesian phylogenetics platform designed from the ground up for near-real-time analysis of outbreak-scale datasets. The work addresses a bottleneck that has frustrated the genomics community since 2020: how to preserve the statistical rigor of Bayesian inference while making it fast enough to run continuously on datasets containing hundreds of thousands of genomes. The authors demonstrate that Delphy can produce dated, uncertainty-aware phylogenies from massive SARS-CoV-2 datasets on timescales of hours rather than days or weeks, opening the door to a style of genomic epidemiology in which phylogenetic estimates are refreshed as routinely as case counts.

The core challenge Delphy tackles is one of combinatorial explosion. Bayesian phylogenetic inference requires exploring an astronomically large space of possible trees, branch lengths, and substitution model parameters, typically via Markov chain Monte Carlo (MCMC) sampling. Each step of an MCMC chain proposes a small change to the current state of the tree and evaluates how well the resulting configuration explains the sequence data, a calculation whose cost scales with both the number of taxa and the length of the genome alignment. With a few hundred genomes, a desktop workstation can handle the workload. With hundreds of thousands, the likelihood evaluations become prohibitively expensive, and the mixing of the chain—the efficiency with which it explores the space of trees—deteriorates because the vast majority of proposed moves are rejected. Traditional implementations such as BEAST, which has served the field admirably for years, were architected for datasets orders of magnitude smaller than those generated during the pandemic.

Delphy’s designers approached the problem by rethinking the inference engine rather than simply optimizing existing code. The platform restructures the likelihood computation to take advantage of modern hardware accelerators, particularly graphics processing units (GPUs), which excel at the kind of massively parallel arithmetic that underlies phylogenetic likelihood calculations. In a phylogenetic likelihood evaluation, the probability of observing the sequence data at each site of the genome must be computed for every node of the tree, and these computations are largely independent of one another. By mapping them onto thousands of parallel processing cores, Delphy amortizes the cost of each likelihood evaluation across hardware that consumer laptops and cloud instances now routinely carry. The result is a dramatic speedup in the inner loop of Bayesian inference, the very component that dominated runtime in earlier tools.

Hardware alone, however, would not have been sufficient. The study also describes algorithmic innovations in how the chain explores tree space and how the analysis is partitioned. Rather than requiring the entire dataset to be re-analyzed from scratch each time new genomes arrive, Delphy is built to support incremental updates, in which an existing posterior sample of trees is extended and refined as additional sequences become available. This design mirrors the operational rhythm of outbreak response, where sequencing data arrive in a continuous stream and analysts need updated estimates on a daily or even hourly cadence. The ability to warm-start an analysis from a previous posterior, rather than beginning each run cold, converts what was historically a batch process into something closer to a living model of an epidemic.

Another distinguishing feature of the platform is its treatment of time. During the COVID-19 pandemic, the field converged on the value of time-calibrated phylogenies, in which branch lengths are expressed in units of calendar time using sampling dates as calibration points. These dated trees allow epidemiologists to estimate when lineages diverged, when a variant likely entered a region, and how quickly lineages were growing or declining in frequency. Delphy incorporates temporal structure directly into its Bayesian model, jointly estimating the evolutionary rate and the timing of divergence events alongside the tree topology. Crucially, it retains full posterior distributions over these quantities, so that an analyst can report not just a point estimate of when a lineage emerged but a credible interval that honestly reflects the limits of the data. In fast-moving outbreak situations, where decisions about border measures, vaccine updates, and public communications hinge on timing estimates, that quantified uncertainty is not a luxury; it is the difference between a defensible inference and an overconfident one.

The authors validate the platform on datasets drawn from the SARS-CoV-2 pandemic, the largest real-world stress test genomic epidemiology has ever faced. Global repositories accumulated many millions of viral genomes during that outbreak, and even curated subsets routinely exceeded the practical limits of legacy Bayesian tools. Delphy’s benchmarks show that analyses on datasets of this scale, which would have been effectively impossible with prior software, complete in hours on appropriate hardware, with posterior estimates that are consistent with those obtained by slower reference methods on smaller subsets. This kind of cross-validation matters: speed is worthless if the fast answer is wrong. By demonstrating agreement between its accelerated inference and established approaches on tractable data, the study builds the case that Delphy’s approximations and engineering choices preserve the statistical integrity that Bayesian methods are meant to guarantee.

The implications for public health practice extend well beyond coronaviruses. The same architecture applies to any rapidly evolving pathogen for which dense genomic sampling is available, including influenza, respiratory syncytial virus, dengue, mpox, and foodborne bacterial outbreaks. In each of these settings, the operational question is similar: can a public health laboratory turn a week’s worth of new genomes into an updated picture of transmission before the picture changes again? A tool that delivers dated phylogenies with quantified uncertainty on a near-real-time cadence shifts genomic epidemiology from retrospective analysis toward genuine situational awareness. It also lowers the barrier for smaller laboratories and public health agencies in low-resource settings, since the computational demands, while substantial, are within reach of cloud computing budgets that are far smaller than the supercomputing clusters previously imagined necessary for pandemic-scale Bayesian work.

The study also speaks to a broader lesson about scientific software. The pandemic exposed a gap between the methods developed in academic statistics and systematics communities and the operational needs of public health. Tools that were exemplary for studies of dozens or hundreds of taxa could not simply be scaled by throwing more processors at them; they required rethinking data structures, sampling strategies, and the software engineering practices needed for continuous, reliable operation. Delphy represents the kind of ground-up redesign that gap demands, combining modern statistical machinery with the performance engineering of contemporary high-performance computing. Its open availability means that the wider community can scrutinize, extend, and build upon the platform, which is essential if it is to become trusted infrastructure rather than a one-off demonstration.

Challenges remain, of course. Bayesian inference on enormous datasets still requires careful attention to model adequacy, because a fast and precise answer to a poorly specified model is of limited value. Recombination, which is substantial in coronaviruses, complicates the strictly tree-based assumptions underlying most phylogenetic software, and integrating recombination-aware inference at pandemic scale remains an open problem. Sampling bias—where genomes are sequenced unevenly across geography, time, and severity of illness—continues to shape what any phylogenetic method can honestly conclude. None of these caveats diminishes the significance of the advance, but they frame the work realistically: Delphy removes a computational barrier that has constrained the field for years, leaving the scientific community freer to focus on the biological and epidemiological subtleties that no amount of computing power can resolve on its own.

If the trajectory of genomic surveillance continues, the coming years will see sequencing embedded ever more deeply in routine public health, from wastewater monitoring to hospital diagnostics. In that world, the demand for phylogenetic analyses that are simultaneously rigorous, current, and scalable will only grow. Delphy’s demonstration that Bayesian phylogenetics can operate at near-real-time speed on outbreak-scale data marks a turning point in that trajectory, promising that the statistical rigor that made phylogenetics indispensable to evolutionary biology can finally be delivered at the tempo at which epidemics actually unfold.

Subject of Research: Scalable near-real-time Bayesian phylogenetic inference for infectious disease outbreak genomics

Article Title: Scalable near-real-time Bayesian phylogenetics for outbreaks with Delphy

Article References: Varilly, P., Schifferli, M., Yang, K., Cronan, P., Specht, I., Burcham, T., Glennon, O., Jacks, O., Laning, E., Marrs, L., Oba, K., Yeung, S., Zhao, K. W., Parker, E., Omah, I., Pekar, J. E., Luebbert, L., Andersen, K. G., Park, D. J., … Sabeti, P. C. (2026). Scalable near-real-time Bayesian phylogenetics for outbreaks with Delphy. Nature. https://doi.org/10.1038/s41586-026-11012-6

Image Credits: AI Generated

DOI: 10.1038/s41586-026-11012-6

Keywords: Bayesian phylogenetics, Delphy, outbreak genomics, viral evolution, SARS-CoV-2, genomic epidemiology, GPU computing, Markov chain Monte Carlo, phylodynamics, public health surveillance, time-calibrated phylogenies, real-time inference