Mpox Is Still Spreading. Can AI Get Ahead of the Virus This Time?

Why the mpox outbreak is a test case for AI in public health#

In April 2026, the world recorded 1,066 confirmed mpox cases and three deaths. That is a small number next to the peaks of 2024 and 2025, and it is easy to read it as the end of the story. It is not. Across the 16 months from January 2025 to the end of April 2026, 59,709 cases and 241 deaths were reported, and nine countries outside Central and East Africa now report community transmission of the clade Ib strain, from Colombia to Thailand.

Then, in February, the World Health Organization reported something virologists had been quietly dreading: a recombinant virus carrying genetic material from both clade Ib and clade IIb. Two viruses had swapped pieces of their genomes inside a single patient. It was spotted only because someone sequenced the sample.

That detail is the reason this topic matters to anyone interested in artificial intelligence. Sequencing generates enormous quantities of genetic text. Making sense of it quickly, across dozens of countries, is exactly the sort of job people hope machine learning can do. So can it? The honest answer is "partly, and not in the way the headlines suggest". This article walks through what the data show, what the algorithms are really doing, and where the evidence is thin.

Background: what genomic surveillance actually measures#

Clades, lineages and why they matter#

A virus's genome is its instruction manual, written in a four-letter chemical alphabet. When scientists sequence it, they read that manual letter by letter. Every time the virus copies itself, small typos creep in. Viruses with a shared ancestor carry shared typos, and that lets researchers draw a family tree, called a phylogenetic tree.

Mpox viruses are grouped into broad branches called clades. The 2024 Nature Medicine analysis of global genomic surveillance from the WHO used 10,546 high-quality genomes from 65 countries, collected between 1958 and 2024, and divided the virus into clade I, clade IIa, clade IIb (including the lineage B.1 behind the 2022 global outbreak) and sub-lineages within them. Clade Ib is the newer branch that emerged in the eastern Democratic Republic of the Congo and has since travelled.

The APOBEC3 fingerprint#

Here is the detail that makes mpox genomics unusually informative. Humans carry enzymes, called APOBEC3 enzymes, that defend cells by deliberately damaging viral DNA. When a virus passes from person to person, it picks up a characteristic pattern of these edits. Viruses that jump from animals to people lack the pattern, or have it less often. A 2023 Science paper used this signal to argue that mpox had been transmitting between humans since at least 2016.

In other words, a genome carries a record of who has been passing it along. That record is what algorithms are trained to read.

Sequencing, defined plainly#

Whole-genome sequencing reads every letter of the virus. Amplicon sequencing, a cheaper shortcut, copies selected stretches and reads only those; a 2025 paper in Emerging Infectious Diseases describes a partial-genome amplicon approach designed to improve surveillance. Portable nanopore sequencers read DNA as it threads through a tiny pore, and they have taken genomics out of central laboratories and into regional hospitals.

What sequencing has already revealed about the mpox outbreak#

A new clade learns to spread between people#

In eastern Congo, researchers sequenced 58 viral genomes from 54 patients and found APOBEC3-type mutations in 28 of 35 linking mutations and 55 of 82 unique mutations, the pattern expected when a virus moves person to person. Their analysis suggested sustained human transmission of clade Ib since at least September 2023. The same study recorded 670 hospitalised cases between September 2023 and June 2024, most linked to sexual contact.

The signature was not limited to clade Ib. A 39-year-old man returned to Dublin from the Democratic Republic of the Congo in February 2025 with clade Ia mpox, the first confirmed clade Ia case outside Africa to be reported internationally. Of his virus's unique mutations, 83% matched APOBEC3 editing patterns, consistent with sustained human spread. He recovered and nobody else was infected.

Three clades under one roof#

In the Republic of Congo, passive surveillance identified 16 clade Ia cases, 32 clade Ib cases and one introduction of clade IIb by August 2025. Nanopore sequencing placed the clade IIb case in lineage A.2.2, which has been emerging in West Africa, with a common ancestor estimated at mid-2023. The authors describe the first time three clades have been seen circulating together in one confined area, which raises the odds of recombination.

The recombinant#

That risk became concrete in the WHO's February 2026 notice. Two confirmed cases of a Ib/IIb recombinant were reported: one in the United Kingdom, detected in December 2025, and one in India, detected in September 2025 and reclassified in January 2026. The two genomes were more than 99.9% similar. Both patients had mild disease and recovered, no onward transmission was detected, and the WHO left its risk assessment unchanged.

What the numbers show at a glance#

The chart below converts the APOBEC3 counts reported in the two studies above into percentages. It is a simple way to see how consistently the human-transmission signature turns up, though note that each bar rests on a small number of mutations, and the two studies counted different things.

Where does AI actually fit into mpox genomic surveillance?#

Pattern-finding in genomes, mostly without the hype#

When people say "AI" in this setting, they usually mean one of two things. The first is unsupervised machine learning: algorithms that sort data into groups without being told what the groups should be. A 2024 study in Frontiers in Genetics used techniques called t-SNE, principal component analysis and hierarchical clustering on 165 gene regions from mpox genomes. Its authors reported that sequences from after 2022 showed a linear pattern of adaptation to the human host, and then used AlphaFold 2 to predict how the proteins differed in shape.

The second is supervised prediction, where a model learns from labelled examples. A 2025 study in Communications Biology built models that combined host ecology with viral genome features to predict which animals might harbour orthopoxviruses, and flagged hotspots in southeast Asia, equatorial Africa and the Amazon that overlap with low smallpox vaccination coverage.

Neither is a chatbot. Both are statistical tools that need clean data, and both produce hypotheses that still require laboratory or epidemiological confirmation.

Dashboards and tooling that automate the routine#

A great deal of the practical gain comes from automation rather than cleverness. MpoxRadar, built at the Robert Koch Institute, downloads mpox sequences from GenBank daily and lets users filter by gene, country, time and sequencing platform, with trend analysis through linear regression. Its developers noted that the virus mutates at an exceptional rate for a DNA virus, which is why a tool that never sleeps is useful. Nextstrain also maintains an mpox dataset for its Nextclade tool, which assigns new sequences to clades automatically.

Forecasting cases, with caveats#

A separate strand of work predicts case counts rather than reading genomes. One recent example pairs machine-learning time-series forecasting with hotspot mapping to provide early warning of mpox in the United States. Forecasts like these depend heavily on how many cases are tested and reported, which brings us to the real bottleneck.

Large language models: promise, but little proof#

Genome language models, which treat DNA the way chatbots treat sentences, are advancing quickly. A 2026 review in Frontiers in Genetics notes that one such model, Evo 2, was trained on 9.3 trillion DNA base pairs. Yet the same review does not describe pathogen surveillance or outbreak detection applications. No peer-reviewed evidence that such a model has so far flagged an mpox lineage before conventional methods did.

The real bottleneck is samples, not algorithms#

An algorithm can only learn from sequences that exist. This is where the world's response has been uneven. The WHO recommends that countries with community transmission sequence at least 10% of confirmed cases, and prioritise imported, unusual or severe cases. That is a modest target, and many places struggle to meet it.

There has been progress. When Africa CDC ended its continental mpox emergency on 23 January 2026, it reported that laboratory and sequencing capacity had expanded more than tenfold during the response, that confirmed cases had fallen by 60% across 2025, and that the case fatality rate had dropped from 2.6% to 0.6%. It also stressed that lifting the emergency does not signal the end of mpox in Africa.

The WHO's own testing picture is still patchy. A multi-country evaluation of point-of-care tests is under way in Belgium, the Democratic Republic of the Congo, Ghana and Madagascar, and twelve mpox nucleic acid tests have been listed for emergency use. Fast, local testing feeds fast, local sequencing, and that is the foundation on which any AI layer must sit. A smart model fed by a thin trickle of sequences will still be late.

What it means for vaccines and response#

Surveillance exists to change decisions. If a new variant spreads in a way that evades vaccine protection or diagnostic tests, public health teams need to know early. The WHO's February notice recommends strengthening vaccination access for at-risk populations and integrating mpox services with HIV and sexually transmitted infection care. By the end of May 2026, all MVA-BN vaccine doses allocated through nine rounds had been delivered to 19 countries, with about 135,000 doses still available through the allocation mechanism.

Genomic data help here in two ways. They reveal whether mutations land in regions that matter for tests and vaccines, and they show whether outbreaks are linked, which tells teams where to target vaccination. Machine learning can speed up the sorting. It cannot replace the clinicians, laboratory staff and epidemiologists who collect the samples and act on the results.

So can AI get ahead of mpox?#

The evidence is cautious but not gloomy. Genomic surveillance has already done what it promised: it showed that clade Ib was spreading between people, caught a clade Ia case in Dublin, exposed three-clade co-circulation, and found a recombinant. Machine learning has added useful ways to cluster genomes, predict reservoirs and automate clade assignment. What it has not yet shown is that it can predict, rather than detect, the next consequential change.

Prediction would need a model that connects specific mutations to specific behaviour, such as transmissibility or immune escape, and that link has to be tested in laboratories. For now, the biggest wins are likely to come from unglamorous work: more sequencing, faster sharing of data, and open tools that anyone can run. AI can be a good assistant on that path, but it is not a shortcut.

Frequently asked questions#

Q: Is the mpox outbreak over?

A: No. The WHO reported 1,066 confirmed cases and three deaths for April 2026, far below earlier peaks, but its report states the outbreak remains a graded health emergency. Africa CDC ended its continental emergency in January 2026 while stating that the disease remains endemic.

Q: What is clade Ib, and why do scientists watch it so closely?

A: Clade Ib is a branch of the mpox virus first described in eastern Congo. Genomic work there found sustained human-to-human transmission, and it has since been reported with community transmission in nine countries outside Central and East Africa.

Q: What is a recombinant virus, and should I be worried about the clade Ib/IIb one?

A: Recombination happens when two related viruses infect the same cell and swap stretches of genetic material. The WHO reported two cases of a Ib/IIb recombinant, both mild with no onward transmission, and said its risk assessment was unchanged. Two cases is too few to draw firm conclusions, so monitoring continues.

Q: What does AI actually do in genomic surveillance?

A: Today it mostly sorts and compares. Examples include clustering gene regions to study adaptation, predicting likely animal hosts and forecasting case counts. Large genome language models are promising, but no published evidence that they have yet led an outbreak response.

Q: How much of the virus needs sequencing for surveillance to work?

A: The WHO recommends sequencing representative samples, with a minimum of 10% of confirmed cases in settings with community transmission, and prioritising imported, unusual or severe cases.

Q: Does the mpox vaccine still work against new clades?

A: The sources reviewed here do not report that vaccine protection has been lost to a new variant, and the WHO continues to recommend strengthening vaccination access for at-risk groups. Laboratory testing of any new lineage remains the way to check.

Q: What is APOBEC3, and why does it appear in so many mpox studies?

A: APOBEC3 enzymes are part of the human antiviral defence and leave a characteristic pattern of edits in viral DNA. When that pattern is abundant, it indicates the virus has been passing between people, as in the 2023 Science analysis pointing to human transmission since at least 2016.

References#

1. World Health Organization. Mpox: Multi-country External Situation Report no. 66, 31 May 2026. Official organisation.
2. World Health Organization. Mpox: recombinant virus with genomic elements of clades Ib and IIb, Disease Outbreak News, 14 February 2026. Official organisation.
3. Otieno JR, Subissi L, et al. Global genomic surveillance of monkeypox virus. Nature Medicine, 2024. Peer reviewed.
4. O'Toole Á, et al. APOBEC3 deaminase editing in mpox virus as evidence for sustained human transmission since at least 2016. Science, 2023. Peer reviewed.
5. Masirika LM, et al. Epidemiological and genomic evolution of the ongoing outbreak of clade Ib mpox virus in the eastern Democratic Republic of the Congo. Nature Medicine, 2025. Peer reviewed.
6. Dowling-Cullen C, Fahey L, Igoe D, et al. Mpox due to monkeypox virus clade Ia infection detected outside of Africa: a case study. Communications Medicine, 2025. Peer reviewed.
7. Koukouikila-Koussounda F, Yinda CK, et al. Evidence of monkeypox virus clade IIb lineage A.2.2 in the Republic of the Congo and co-circulation of clade Ia, Ib and clade IIb. Nature Medicine, 2026. Peer reviewed.
8. Africa CDC. Africa CDC Declares End to Mpox As Continental Health Emergency (reported by allAfrica), 23 January 2026. News report of an official announcement.
9. Nasri F, Kongkitimanon K, Wittig A, Fuchs S, et al. MpoxRadar: a worldwide MPXV genomic surveillance dashboard. Nucleic Acids Research, 2023. Peer reviewed.
10. Zhang S, et al. Compositional features analysis by machine learning in genome represents linear adaptation of monkeypox virus. Frontiers in Genetics, 2024. Peer reviewed. (Full text: PMC.)
11. Tseng KK, et al. Viral genomic features predict Orthopoxvirus reservoir hosts. Communications Biology, 2025. Peer reviewed. (Full text: PMC.) Retrieved via PubMed.
12. Machine learning-based time series forecasting with polygon-based emerging hotspot analysis for early warning of mpox in the United States. Taylor & Francis, 2026. Peer reviewed (only the title and abstract-level description were accessible to me).
13. Meng Z, Yang Z, Zhu M, Xu J. Survey of large language models in bioinformatics. Frontiers in Genetics, 2026. Peer reviewed.
14. Nextstrain. nextclade_data repository, mpox datasets. Research lab software resource.
15. Monkeypox Virus Partial-Genome Amplicon Sequencing for Improvement of Genomic Surveillance during Mpox Outbreaks. Emerging Infectious Diseases, 2025. Peer reviewed.