Delphy maps 100,000 viral sequences in a day

When a new virus spreads, epidemiologists want to know how it arrived, how many times it was introduced, and which cases are connected. The standard analytical approach, Bayesian phylogenetics, builds not one viral family tree but many plausible ones, preserving genuine uncertainty about an outbreak's history. The problem is that running such an analysis on thousands of outbreak sequences has been painfully slow. A peer-reviewed study published in Nature on 16 September describes Delphy, a tool designed to compress that bottleneck by orders of magnitude and run inside an ordinary web browser.

Storing the differences

Most phylogenetic tools treat every genome as a full string of genetic letters, comparing them all from scratch at each analytical step. During an outbreak, though, circulating virus copies are genetically very similar. Delphy exploits that similarity. It uses a structure called an explicit mutation-annotated tree, which works something like a version-history file: instead of saving every complete document, it records only what changed between versions. A related technical paper published in Genetics reports that this representation can yield as much as 1,000-fold gains in statistical efficiency when sequences differ at few positions. That efficiency is one part of Delphy's design, working alongside optimised computational routines to deliver the speed gains described in the Nature paper.

To understand why speed matters here, consider what Bayesian analysis actually provides. A conventional method returns its single best-supported tree. Bayesian inference retains many alternative trees and shows how uncertain their timing and relationships remain. An analyst can report not just that two clusters are probably related, but how confident that conclusion is. That honesty matters when public-health decisions depend on whether cases share a recent ancestor or arrived independently. The trade-off has always been computational cost: the number of possible tree shapes grows enormously with the number of sequences, and established platforms such as BEAST 2 require long runs and specialist judgement.

The authors report that Delphy processed a simulated dataset of 100,000 sequences within a day, delivering results they describe as comparable with established Bayesian analyses (Nature). That benchmark has not been independently reproduced. Delphy currently supports a narrower range of population models than a general-purpose platform like BEAST 2, making it a purpose-built outbreak tool rather than a universal replacement for evolutionary analysis.

From simulation to an active outbreak

Delphy has moved beyond test datasets. A Virological.org post documents researchers applying Delphy's web interface during the 2026 Bundibugyo-virus outbreak in the Democratic Republic of the Congo and Uganda. That confirms use in a live outbreak setting, though no evidence yet shows it changed a specific public-health decision or shortened an intervention.

The browser-based design is relevant for accessibility. The Broad Institute, which played a central role in the work, describes this as part of Delphy's intended value for outbreak response, potentially letting health agencies start an analysis without a high-performance computing cluster. Public availability, licensing terms and practical hardware requirements were not established in the investigation.

Speed also cannot repair what happens before sequences reach the software. A separate Bundibugyo analysis illustrates the upstream reality: researchers had to remove outliers caused by sequencing errors and incorrectly recorded dates across 626 genomes before any tree-building could begin. Faster analysis cannot fix delayed sampling, poor sequence quality, or missing surveillance.

The peer-reviewed result shows that Bayesian outbreak reconstruction can scale much further than previously demonstrated, at least in the authors' tests. The next meaningful evidence is whether outbreak teams can reproduce that speed with messy real-world data and use the time saved to inform an actual decision.