Delphy: Bayesian Phylogenetics at Outbreak Speed

A researcher wanting a proper Bayesian reconstruction of how H5N1 moved through American dairy cattle has faced an unattractive choice: feed 3,339 genomes to BEAST X, the field's reference implementation, and wait. When the authors of a paper published in Nature on 16 September 2026 tried precisely that, they stopped the run after 21 days without it finishing. Their own tool completed the same analysis in 42 minutes [1].

That gap is the story. Not because 42 minutes is impressive in isolation, but because the difference between three weeks and three quarters of an hour is the difference between a retrospective paper and a decision an epidemiologist can act on while the outbreak is still moving.

Delphy, released alongside the paper by Patrick Varilly, Pardis Sabeti and colleagues, rebuilds Bayesian phylogenetic inference around a different internal representation of a tree. It reports two to three orders of magnitude speed gains over BEAST 2 and BEAST X while, the authors argue, sampling the same posterior [1]. It runs in a web browser. The code is MIT-licensed. And in a week dominated by AI agents and foundation models, none of its performance comes from machine learning.

The paper, "Scalable near-real-time Bayesian phylogenetics for outbreaks with Delphy", appeared in Nature on 16 September 2026 as an open access article [1]. The first author is Patrick Varilly and the senior author is Pardis Sabeti of the Howard Hughes Medical Institute, with co-authors spread across more than a dozen affiliations [1].

The technical move is a data structure the authors call an explicit mutation-annotated tree, or EMAT: a timed bifurcating tree carrying an explicit root sequence, with individual mutation and missing-data events placed on branches at explicit times [1]. Conventional Bayesian tools compute the Felsenstein tree likelihood, integrating over every possible history at every variable site. The authors' complaint is that this is mostly wasted work on outbreak data, where "an overwhelming effort goes to vanishingly rare ones" [1]. Delphy reformulates the calculation around local mutations rather than all mutated sites, and proposes large tree rearrangements using mutation-directed subtree pruning and regrafting, which needs no hand-tuning [1].

The benchmarks span five real pathogens and four simulated datasets: SARS-CoV-2, Zika, Ebola, human mpox, the H5N1 cattle outbreak, and synthetic sets of up to 100,000 sequences [1]. On 757 SARS-CoV-2 genomes from early 2020 Boston, Delphy and BEAST X produced maximum clade credibility trees the authors describe as being in "excellent agreement", and they argue the results are "statistically indistinguishable" from conventional Bayesian methods given the shared Bayesian foundation [1]. The 100,000-sequence simulation finished within a day on a 96-vCPU cloud machine, reaching an effective sample size of 248 in under 15 hours [1].

The delivery mechanism deserves as much attention as the algorithm. Delphy ships as a client-side web application: a TypeScript interface talking to a C++ core compiled to WebAssembly, with the explicit property that data never leave the user's machine [1]. A live analysis of US H5N1 sequences is updated daily at delphy.bio. The project was relicensed under MIT in December 2025, and the team withdrew its patent application, dedicating the intellectual property to the public domain [2].

Why This Was Slow in the First Place#

Phylogenetics reconstructs the family tree of a set of sequences. For pathogens, adding dates turns that tree into something more useful: an estimate of when lineages diverged, how fast the population grew, and where introductions came from. This is the substrate of modern outbreak investigation.

There have long been two ways to get there. Approximate methods build one tree quickly. Maximum parsimony tools such as UShER were designed for exactly the low-diversity data outbreaks generate, and can place new SARS-CoV-2 genomes onto an existing tree in seconds [3]. Maximum likelihood tools, and the Nextstrain pipeline built on them, have been the visible face of real-time pathogen tracking since 2018 [4]. Their limitation is structural: they yield a single topology, and any statement about uncertainty has to be bolted on afterwards [1].

Bayesian methods take the opposite approach. BEAST 2 and BEAST X sample thousands of plausible trees and return a posterior distribution, so a claim like this lineage most likely entered the herd in late March arrives with a credible interval attached, and population growth or geographic spread are inferred jointly rather than post hoc [5], [6]. This is the right framework for outbreak questions. It is also expensive enough that, in the authors' words, Bayesian phylogenetics remains "inaccessible to most epidemiologists and public health bodies" [1].

Delphy's claim is that this trade-off was an artefact of implementation rather than a law of nature. Once explicit mutation events are projected out, the authors state, Delphy samples the same tree distribution as the existing tools [1]. They validated the roughly 8,000 lines of new C++ against the Well-Calibrated Simulation Study framework, which is the appropriate way to check that a Bayesian sampler is actually hitting its target [1].

Why This Matters#

The obvious beneficiaries are public health laboratories without computing infrastructure. A browser-based tool needing no cluster, no queue, no installation and no data upload removes several barriers at once [1], and the privacy property is not cosmetic: many national agencies cannot send pathogen sequences to a third-party server before they are ready to share them. Running locally sidesteps the question.

Speed also changes what questions are worth asking. A three-week analysis answers a question you had three weeks ago. A 42-minute analysis [1] can be re-run when 40 new sequences arrive, which is what surveillance actually looks like.

There is a quieter implication for how we think about AI in the life sciences. This fortnight also brought DeepMind's AlphaGenome Atlas, a petabyte of predictions for nine billion possible single-nucleotide variants [7], and a Nature paper on turning research papers into executable AI agents [8]. Both matter. But the largest practical gain for outbreak response came from rethinking a data structure. Machine learning is not the only route to a hundredfold improvement, and treating it as the default directs attention away from problems where engineering pays better.

Critical Analysis#

Narrow by design#

Delphy's efficiency depends on low sequence diversity. For thinly sampled data spanning long evolutionary timescales, the authors note, the explicit-mutation approach stays exact but "may offer smaller efficiency gains, or even a loss" [1]. This is a tool for outbreaks, not deep phylogenomics.

The current release implements a single-partition HKY substitution model with a coalescent prior allowing flexible population growth, and fine-scale phylogeography is not yet supported beyond basic parsimony-based ancestral state reconstruction [1]. Researchers needing structured coalescent models, relaxed clocks across partitions or sophisticated discrete trait analysis will still reach for the older tools.

Input quality is now the bottleneck#

Data handling introduces real losses. Partially ambiguous nucleotides, the Y and R codes that pervade consensus sequences from clinical samples, must be treated as missing, and tips with date uncertainty beyond a year should be excluded [1]. Most tellingly, roughly 18% of recent public H5N1 sequences lacked precise dates or locations and had to be dropped [1]. That figure is a comment on surveillance metadata quality rather than on Delphy, and a reminder that faster inference does not fix an impoverished input.

Rough edges at scale#

The authors call their parallelisation a proof of concept, noting the 100,000-sequence run used under 25% of the machine's capacity at any moment [1]. They observed numerical inaccuracy in the parallelised coalescent prior when too many branches were active, and small discontinuities in mutation-count traces during the largest H5N1 analyses, judged negligible for public health quantities [1]. They recommend corroborating any conclusion depending on exact tip placement through multiple independent runs [1], which is sound practice and also an admission that rugged tree landscapes remain hard.

A final caveat belongs to the reader. Comparisons between MCMC implementations are sensitive to configuration, and agreement demonstrated on a 757-sequence dataset is a claim independent groups should test on their own data before treating Delphy as a drop-in replacement. The MIT licence and public Zenodo archive make that straightforward [1], [2].

Expert Perspective#

The lineage is clear enough. UShER showed in 2021 that mutation-annotated trees could make phylogenetic placement tractable at pandemic scale [3]. Nextstrain showed that real-time visualisation changes how findings travel [4]. Neither offered full Bayesian uncertainty. Delphy's contribution is to bring the mutation-annotated representation into a genuinely Bayesian sampler, a harder problem than it sounds, because tree and mutation placements have to be sampled jointly.

The comparison that matters is therefore with BEAST rather than with UShER. BEAST X, published in Nature Methods in 2025, remains the more general instrument, with a modelling range Delphy does not approach [6]. Delphy is not a replacement; it is a faster, narrower tool aimed squarely at the outbreak case, and most groups will end up using both.

What makes the release unusual is the combination of choices around it: browser-based execution, local-only data, an MIT licence, a withdrawn patent, a live public analysis. Those are decisions about who gets to do this kind of work. In a week when Nature's editorial page argued that AI companies must work with the research community on attribution and consent [9], and when reporting described a protein-ligand structure model trained on proprietary pharmaceutical data that reportedly outperforms public equivalents but is not available to anyone outside the consortium [10], a fully open release with the patent dropped reads as a position rather than an afterthought.

Key Takeaways#

1. Delphy reports two to three orders of magnitude speed gains over BEAST 2 and BEAST X on outbreak data, and its authors argue it samples the same posterior, through explicit mutation-annotated trees rather than machine learning [1].
2. The headline benchmark: an H5N1 analysis of 3,339 genomes in 42 minutes, against a BEAST X run abandoned after 21 days [1].
3. It runs client-side in a browser via WebAssembly, so sequence data never leave the analyst's machine, and it is MIT-licensed with the patent application withdrawn [[1]](https://www.nature.com/articles/s41586-026-11012-6), [[2]](https://github.com/broadinstitute/delphy).
4. Scope is deliberately narrow: single-partition HKY, a coalescent prior, no fine-scale phylogeography, reduced benefit on diverse or thinly sampled data [1].
5. The binding constraint on real-time genomic epidemiology may now be metadata rather than compute; about 18% of recent public H5N1 sequences were unusable for want of precise dates or locations [1].

Frequently Asked Questions#

Is Delphy an AI tool? No. It is a Bayesian Markov chain Monte Carlo sampler. The speed comes from a redesigned data structure and new tree-rearrangement proposals, not from a learned model [1].

Does it give the same answers as BEAST? On the datasets tested, the authors report indistinguishable maximum clade credibility trees and compatible posterior supports against BEAST X, and argue the two sample the same tree distribution once explicit mutation events are projected out [1]. Independent replication would strengthen that case.

Can it replace Nextstrain? They answer different questions. Nextstrain provides continuously updated maximum likelihood phylogenies and visualisation across many pathogens [4]. Delphy provides Bayesian posteriors with quantified uncertainty for one dataset. Many groups will use both.

What data does it need? A multiple sequence alignment in FASTA or MAPLE format with dates, which may be imprecise, encoded in the sequence identifiers, plus optional metadata [1]. Tips with date uncertainty greater than a year should be excluded.

Is it safe to use with unreleased sequences? The browser version computes locally and the authors state that data never leave the user's machine [1]. Institutions should still verify this against their own information governance requirements rather than relying on a published description.

What is the current state of the H5N1 outbreak it was tested on? The US CDC reports 71 total human A(H5) cases since February 2024, 64 of them detected through monitoring of exposed people, and assesses current risk to the general population as low; that page was last updated on 6 March 2026 [11]. Consult current official guidance.

Where can I get it? The application and live H5N1 analysis are at delphy.bio, the source at github.com/broadinstitute/delphy under an MIT licence, with the archived code carrying a Zenodo DOI [1], [2].

References#

1. Varilly, P. et al. "Scalable near-real-time Bayesian phylogenetics for outbreaks with Delphy." Nature, published online 16 September 2026. Open access. DOI: 10.1038/s41586-026-11012-6. Archived code: Zenodo, DOI 10.5281/zenodo.20717608. Volume and page numbers not yet assigned at time of writing.

2. Delphy source repository, Broad Institute. github.com/broadinstitute/delphy. MIT licence; relicensed and patent application withdrawn 17 December 2025.

3. Turakhia, Y., Thornlow, B., Hinrichs, A. S., De Maio, N., Gozashti, L., Lanfear, R., Haussler, D. & Corbett-Detig, R. "Ultrafast Sample placement on Existing tRees (UShER) enables real-time phylogenetics for the SARS-CoV-2 pandemic." Nature Genetics 53, issue 6 (June 2021). DOI: 10.1038/s41588-021-00862-7.

4. Hadfield, J., Megill, C., Bell, S. M., Huddleston, J., Potter, B., Callender, C., Sagulenko, P., Bedford, T. & Neher, R. A. "Nextstrain: real-time tracking of pathogen evolution." Bioinformatics 34(23), 4121–4123 (2018). DOI: 10.1093/bioinformatics/bty407.

5. Bouckaert, R., Vaughan, T. G., Barido-Sottani, J. et al. "BEAST 2.5: An advanced software platform for Bayesian evolutionary analysis." PLOS Computational Biology e1006650 (2019). DOI: 10.1371/journal.pcbi.1006650.

6. Baele, G., Ji, X., Hassler, G. W., McCrone, J. T., Shao, Y., Zhang, Z., Holbrook, A. J., Lemey, P., Drummond, A. J., Rambaut, A. & Suchard, M. A. "BEAST X for Bayesian phylogenetic, phylogeographic and phylodynamic inference." Nature Methods 22, 1653–1656 (2025). DOI: 10.1038/s41592-025-02751-x.

7. Google DeepMind. "AlphaGenome Atlas: A predictive map of every possible DNA letter change in the human genome." 8 September 2026. deepmind.google/blog. Company announcement, not peer reviewed.

8. Miao, J., Davis, J. R., Zhang, Y., Pritchard, J. K. & Zou, J. "Reimagining research papers as interactive and reliable AI agents." Nature, 16 September 2026. DOI: 10.1038/s41586-026-11044-y.

9. Nature Editorial. "AI companies must work with the research community to protect attribution." 16 September 2026. nature.com/articles/d41586-026-02886-7.

10. Artificial Science "Pharma's Locked Structures Just Trained an AI". 16 September 2026. Artificial Science

11. US Centers for Disease Control and Prevention. "H5 Bird Flu: Current Situation." Page last updated 6 March 2026. cdc.gov/bird-flu/situation-summary.

Nothing in this article constitutes medical, clinical or public health advice. Statements attributed to the Delphy authors are their reported findings; interpretation and criticism in the analysis sections are editorial.