// Applications · GenHub2

GenHub2:
the first application on PQCF.

Virus classification for genetic laboratories, on a federation of lab nodes where each lab's data stays on its own hardware. Phase A (one node) and phase B (a lab's two nodes) are done; its first tasks have run through the federation.

// Methods and results

What GenHub2 does,
and how we measured it.

1 · The problem. A genetic laboratory sequences a sample and must say, quickly, which virus it holds, which lineage, and which mutations it carries; for an outbreak, also how it relates to other genomes. GenHub2 does this on the lab's own hardware: the lab's raw data does not leave its node. It is research software, not a validated diagnostic product.

2 · The data. A held-out test set: 600 GenBank genomes requested from NCBI Virus, 20 each of 30 viruses that matter to Bulgarian labs (among them Crimean-Congo haemorrhagic fever, West Nile, tick-borne encephalitis and Dobrava-Belgrade viruses), chosen evenly from the first 1,000 accessions of each. They are classified against all 19,607 complete viral genomes in NCBI RefSeq (the snapshot of 13 November 2026). Every accession is published, so the test reruns on the same genomes.

How they are held out. Nothing is trained: the reference set is RefSeq, and the test genomes are GenBank records that are not in RefSeq. Genomes whose sequence is identical to a RefSeq genome (3) were removed, and 4 that NCBI did not return were left out, so 593 were classified. Close relatives of the reference genomes were not removed: the test measures how well a genome is placed against the reference set, as a laboratory's sample would be, not how well entirely novel viruses are found. Near-identical genomes among the 593 themselves were not removed either.

3 · The method. Each genome's nearest RefSeq genome by MinHash distance (Mash, Ondov et al. 2016, doi:10.1186/s13059-016-0997-x; K = 21, 1,000 hashes a genome), labelled with that genome's species, genus and family from NCBI Taxonomy. The call's confidence comes from the estimated identity: high at 0.95 or more, medium from 0.85, low below.

4 · "With proteins". Nucleotide similarity fails for divergent viruses. So every genome is also searched, by DIAMOND blastx (Buchfink et al. 2021, doi:10.1038/s41592-021-01101-x), against all 720,495 proteins of the RefSeq viral genomes (and 1,462 coding sequences without a protein record, translated); each region votes for its best hit's species by bit score. GenHub2 uses the protein call only where MinHash's call is low or missing: alone, proteins are worse at high confidence (255 against 274 of 275 right), because related species share near-identical proteins.

5 · The evaluation and its baseline. The truth is NCBI's taxonomy of each GenBank record, and a call is right only if its species name is exactly NCBI's (likewise genus and family). The baseline is MinHash alone.

593 held-out genomesspecies rightgenusfamily
MinHash alone (the baseline)564 (95.1%)572 of 573 (99.8%)591 (99.7%)
MinHash, high-confidence calls only (275)274 (99.6%)100%100%
proteins alone547 (92.2%)566 of 573586
combined: proteins for MinHash's weak calls572 (96.5%)573 of 573593 (100%)

What the protein search changed. Its call was used for 126 of MinHash's 133 low-confidence or missing calls, and differed from MinHash's in 13 of them: it corrected 10, turned 2 right calls wrong, and left 1 wrong. Net, 8 more species right: 564 to 572.

By virus (MinHash alone; species right of the genomes classified):

virusrightvirusright
25 of the 30 virusesevery genome (18–20 each)Rotavirus A16 of 20
Hepatitis E virus2 of 20 (naming, below)Rabies virus19 of 20
Tick-borne encephalitis virus15 of 20Human metapneumovirus19 of 20

Of the 29 species misses, 18 are naming: NCBI files these hepatitis E genomes under an older species name than the reference genome's; GenHub2 named the right virus, and we count them as misses (the family is right for 19 of the 20). The others are real limits of nucleotide matching: tick-borne encephalitis against a close relative, rotavirus segments, a rabies genome, and one metapneumovirus genome that shared no hashes. Every result per virus is in the results file. Mutations of 1,000 SARS-CoV-2 genomes were compared with Nextclade's: identical for 885; every other difference lies at a genome end or next to an insertion or deletion, where the two aligners differ.

6 · Two nodes (phase B), 16 November. Two Erlang nodes in their own containers on one DGX Spark, five efficiency cores (Cortex-A725) each, against one of them alone; every workload run four times, alternating, and the best of each kind reported. The results were identical in every run.

workloadtasksone node, 5 corestwo nodes, 5 + 5speed-up
MinHash classification, 1,593 genomes323.4 s1.8 s1.91
protein classification, 593 genomes1082.0 s43.9 s1.87
mutations, 1,000 SARS-CoV-2 genomes202.3 s1.2 s1.97
lineages of the 1,000208.4 s4.4 s1.92
tree of the 1,000, 100 bootstrap replicates14861.3 s31.3 s1.96

One machine stands in for two here; the times vary between runs because the machine is shared (the first one-node tree took 86.1 s). After a correction, the rerun gave the same results and speed-ups of 1.80 to 1.95 (below).

7 · The federation experiment, and its limits. On 17 November, 4 distance tiles and 10 bootstrap replicates of a tree of 1,000 genomes went from one DGX Spark (DGX-1) through the FVLLMP federation to another (DGX-2), as GA4GH TES tasks, and came back byte for byte identical, checked by hash. This establishes correctness, not speed: through the federation they took 36.9 s, against 15.2 s on the lab's own 10 cores, because each lab may run one task at a time on a peer for now. Only tree tasks have run this way so far, and the executor that sent them is a prototype in GenHub2.

8 · How to reproduce it. Phase A's acceptance test, 20 checks, about 20 minutes on 10 cores (REPRODUCE.md; without the human genome, a reduced run of 14 checks); phase B's, 20 checks, about 10 minutes more (REPRODUCE-PHASE-B.md); the federation check needs a federation node and a token (FEDERATION.md); the reference data, its snapshot and how to check a rebuilt copy (REFERENCE-DATA.md).

// Announcement · 17 November 2026 · GenHub2

GenHub2 runs across machines: on a lab's own nodes, and through the FVLLMP federation.
The same results, byte for byte, wherever a task runs.

A lab's work now spreads over more than one machine, in two ways:

  • Over the lab's own nodes (phase B): through PQCF's executor for one owner's machines.
  • Through the FVLLMP federation: GenHub2 sends tasks to the federation node it runs on, as GA4GH TES tasks. That node runs each task itself, or sends it to a peer node that accepts its tasks. The first tasks went from one DGX Spark (DGX-1) to another (DGX-2) on 17 November.

The rule that holds it together: a task is plain data, and its result depends only on it. So the same task gives the same bytes on any node, and every result carries a record of where it ran.

Phase B: two nodes, one owner. Two Erlang nodes on one DGX Spark, five cores each, against one node alone (16 November):

workloadone nodetwo nodesspeed-up
classification of 1,593 genomes by MinHash3.4 s1.8 s1.9
classification of 593 held-out genomes by proteins (DIAMOND)82.0 s43.9 s1.9
mutations of 1,000 SARS-CoV-2 genomes (minimap2)2.3 s1.2 s2.0
lineages of the 1,000 (Nextclade)8.4 s4.4 s1.9
a tree of the 1,000 with 100 bootstrap replicates61.3 s31.3 s2.0
  • Identical results: every result is the same on two nodes as on one. The neighbour-joining tree, now built from 36 independent tiles of its distance matrix, is byte for byte the tree one task built in phase A.
  • Trees are spread: the distance matrix in tiles, and bootstrap replicates as independent tasks. Every branch of a tree now shows its support.
  • Data stays at the lab: the steps that read a lab's raw data (validation, removal of human reads, the consensus genome) run only on the lab's own node.
  • References by hash: a new node fetches the 1.5 GB of reference data it needs from the lab's node by hash, and checks it (18 s).

The federation: GenHub2's tasks on another lab's machine.

  • GenHub2 talks only to its own node, with its own token. Where a task runs is that node's decision, under its owner's policy. GenHub2 sees the same task either way, and downloads the results from its own node.
  • Results are checked: every input and output is named by its SHA-256, the way GenHub2's own store names data.
  • A task's container has no network, a read-only file system and fixed limits. Everything it needs comes in as an input.
  • Task images are never public and pass through no registry. GenHub2 hands an image to its own node, and the federation carries it to the peers. Every node that runs an image can read it, its recipe and its purpose, and checks them against the federation's charter before allowing it.
sent from DGX-1, run on DGX-2result
4 distance tiles and 10 bootstrap replicates of 1,000 SARS-CoV-2 genomesall 11 tasks ran on DGX-2, every result byte for byte DGX-1's own; the tree from DGX-2's tiles is phase A's tree

What did not go perfectly, also written down.

  • A bug our first two-node run missed. A task run on the second node left its output there, and the lab's node did not fetch it. The test had run every workload on one node first, so the same results were already at the lab's node and nothing was missing. Batching the bootstrap replicates made new files, and a tree job failed. Now every result's files are fetched and checked, and the test runs two nodes first. The results are unchanged; the speed-ups of the rerun are 1.80 to 1.95.
  • Task size matters. The protein search loads a database of 720,495 proteins on every call. In 30 small tasks it took 229 s on one node; in one task per core, 82 s.
  • The federation is not faster yet. Each lab may run one task at a time on a peer for now: the 11 tree tasks took 36.9 s through the federation, against 15.2 s on the lab's own 10 cores. The first goal is identical results; capacity comes next.
  • So far only tree tasks have run on the federation. Classification, mutations and whole jobs through it are next.
  • One machine stands in for two in phase B: two Erlang nodes on one DGX Spark. Across the federation, the second machine is real.

Next: more of GenHub2's work as federation tasks (classification, proteins, mutations, lineages, whole read sets), more task slots once measured, and phase C: two labs, with sharing, accounts and roles.

Code and results: github.com/quanterall/genhub2, with the phase B results and the federation results.

How to reproduce the results: phase B's test from a fresh clone (20 checks, about 10 minutes after phase A's). The federation check needs a node of the federation and a token from its owner; it is in the same repository (FEDERATION.md).

// Announcement · 14 November 2026 · GenHub2

GenHub2 starts: virus classification for genetic labs, on a federation of lab nodes.
The first application built on PQCF. Open source, Apache 2.0.

GenHub2 classifies DNA and RNA viruses from sequencing data for government and private genetic labs, first in Bulgaria and then with international collaborators. It does not need a central supercomputer:

  • Every lab runs at least one node of the FVLLMP federation, an NVIDIA DGX Spark on the lab's own premises.
  • The nodes compute for each other through a secured gate (mutual TLS), with no central point.
  • A lab's data stays on its own hardware. Results, aggregated counts and the datasets a lab approves are what travel.

Three layers:

layerwhatrole
infrastructure: FVLLMPthe labs' nodes, the secured gates between them, a credit ledgermachines, identity, transport, accounting
framework: PQCFtasks as plain data, executors for one node, one owner's nodes and federation peers, a provenance record for every resultthe distributed computing under every heavy step
application: GenHub2uploads, analyses, trees, monitoring, sharing, the web interfacewhat the labs use

The goals:

  • Rapid classification of DNA and RNA viruses for epidemic response.
  • Data sharing between labs under the GA4GH standards: htsget, DRS, Beacon and Passports. Tasks will pass between nodes as TES tasks.
  • Epidemic monitoring with alerts, computed at every lab from counts per lineage, region and week. Patient-linked data is never pooled.
  • Phylogenetic trees spread over the federation, beyond the 1,000-genome trees one machine builds in seconds.
  • Links to related treatment knowledge (records, registered trials, literature), marked as information, not clinical advice. Software that recommends treatment is a medical device under the EU MDR (Regulation 2017/745) and a high-risk AI system under the AI Act (Regulation 2024/1689).
  • Learned models only where they earn their place: a model is adopted only if it beats the established methods on held-out data, in a published test.
  • Open source under Apache 2.0, built in Erlang and Rust, with an EMQX 5.8 MQTT broker (Apache 2.0) inside each lab.

Why it is PQCF's first application.

  • Every analysis is a PQCF task: plain data, with a result that is the same whichever node runs it.
  • A node runs only the task types it has allowed.
  • Every result carries its provenance: the task's hash, the code's checksum, the node and the time. The labs' audit and reproducibility requirements come from that directly.
  • Nothing genome-specific goes into PQCF itself, so the next application on PQCF needs nothing from GenHub2.

The methods in GenHub2 are classical. The federation and distribution layer is the part we are building for PQCF, and the same layer will run quantum simulation workloads later. GenHub2 is its proving ground.

What the distribution layer does today:

  • Phase A: runs on one node, through PQCF's single-machine executor.
  • Phase B: PQCF's executor for one owner's machines exists and is tested; GenHub2 uses it from phase B.
  • Phase C: the executor for federation peers was not built on 14 November. A prototype now runs GenHub2's tasks through the federation (17 November).

Phase A, one node, is done and measured on one DGX Spark (10 cores), with SARS-CoV-2 as the first acceptance test:

testresult
SARS-CoV-2 reference genome uploadedspecies right, no mutations against itself, lineage B (clade 19A), in 1.4 s
reads simulated from a GenBank SARS-CoV-2 genome (PX916490.1), with human reads mixed inevery human read pair removed and every viral pair kept; the genome rebuilt from the reads equals the original at every base; lineage BQ.1.1, as the original
593 held-out genomes from NCBI Virus, 30 viruses (among them Crimean-Congo haemorrhagic fever, West Nile, tick-borne encephalitis, Dobrava-Belgrade), against all 19,607 complete viral genomes in RefSeqspecies right for 95.1% (99.6% of the calls made with high confidence); genus 99.8%; family 99.7%
1,000 SARS-CoV-2 genomes (138 of them with GenBank records that name Bulgaria), mutations compared with Nextclade'sidentical for 885; every other difference lies at a genome end or next to an insertion or deletion, where the two aligners differ
the specification's targetsa 10 MB upload in 4.5 s (target 5 s); 1,000 genomes classified in 0.7 s (target 60 s); a neighbour-joining tree of 1,000 genomes in 4 s (target 5 minutes)

The methods are established ones: MinHash distances, minimap2, neighbour-joining and FastTree. The lab uses it through a web page and an API. The site's events reach every service on the lab's MQTT broker and the browser live, and the results export as CSV, PDF, JSON and SVG.

Seven more methods (14 November), each measured against what it adds to, on the same data:

methodaddsmeasured
DIAMOND protein search, for MinHash's weak callsdivergent virusesspecies right for 96.5% of the 593 held-out genomes, up from 95.1%; family 100%
VeryFastTreemaximum-likelihood trees within the target1,000 genomes in 250 s (FastTree: 700 s), with 96–100% of FastTree's splits
UShERa new genome placed in an existing tree100 left-out genomes in 0.8 s; 69 into exactly their lineage, 84 into it or a parent or child of it
22 Nextclade datasets, chosen automaticallylineages beyond SARS-CoV-2: influenza, RSV, mpox, measles, dengue and more178 held-out genomes reached a dataset, none another virus's
Hostile, now the default host-read removalhuman viruses kept99.5% of Epstein-Barr virus reads kept, which removal against GRCh38 deletes; 2 of 10,000 human read pairs left
iVar primer trimmingamplicon sequencing (ARTIC)the consensus exact; untrimmed, 9 bases are left unknown at primer sites
Kraken 2 + Brackenmixed samplesa 60/30/10 mixture of SARS-CoV-2, influenza A and RSV estimated as 59.6/29.8/9.9

What did not go perfectly, also written down.

  • 18 of the 29 species misses are naming. NCBI files these hepatitis E genomes under an older species name than the reference genome's: the right virus, under a different name. We count them as misses.
  • The other misses are real limits of nucleotide matching for divergent viruses: tick-borne encephalitis against a close relative, rotavirus segments, a rabies genome.
  • A maximum-likelihood tree of 1,000 whole SARS-CoV-2 genomes took about 12 minutes with FastTree on 10 cores, beyond the 5-minute target. VeryFastTree, added since, does it in about 4 minutes.
  • Proteins alone are worse than MinHash at high confidence (255 against 274 of 275 right): related species share near-identical proteins. So the protein call only replaces MinHash's weak calls.
  • Lineage datasets cover part of the list: the influenza datasets cover only the HA segment, and 18 of the 30 viruses tested have none.

Its results, in TechUPGRADE's simulation engine and 3D visualiser. GenHub2 is the first external application integrated with the running TechUPGRADE Digital Twin Platform: it operates in its own tenant and uses TechUPGRADE's simulation engine and 3D visualiser through the platform's external integration API; the genome and structure viewer opens from GenHub2's own results by a signed hand-off (gh_viewer.erl). The viewer shows a genome's amino-acid changes, in the form GenHub2's Nextclade step writes them (S:N501Y, S:H69-), along the genome, down to the single nucleotide, and on 26 3D structures: 24 measured, from the RCSB Protein Data Bank, and 2 AlphaFold predictions, marked as such. Selecting a region lists, residue by residue, what it touches within 4 Å: the human ACE2 receptor, a drug in an active site, RNA, or an antibody. For RSV, for example, it shows which residues of the fusion protein the infant antibody nirsevimab grips, the question a laboratory must answer quickly for a new strain. Viruses so far: SARS-CoV-2, RSV-A and RSV-B.

What it is not, yet: the lineages in the film are demo data from Nextclade's public reference trees, not samples; it visualises and does not diagnose; it runs behind a sign-in, and the film is its only public view; and its code, written with an AI coding assistant, has not yet been reviewed by an independent virologist (a review dataset by dataset, SARS-CoV-2 and RSV first, is planned). Watch the film of GenHub2 in its TechUPGRADE tenant (8 minutes), with the platform's other films on demo.quanterall.com.

Next:

  • Phase B: two nodes of one owner (done on 16 November, announced above).
  • Phase C: two labs through the federation's gate, with sharing, accounts and roles.
  • Phase D: monitoring across labs, from counts only.
  • Phases E and F: learned models where they beat the baselines, and scale.

The code, the specification's tests and every result: github.com/quanterall/genhub2, with the phase A results and the accessions of every genome used.

How to reproduce the results: a step-by-step page: the three images, the human reference genome from NCBI with its checksum and indexes, the tests, and the acceptance test (20 checks, about 20 minutes on 10 cores); every result must equal the published run. Without the human genome, a reduced run makes the other 14 checks and says it skipped the 6 that need it.