Nanopore reads to isolate results on a laptop
Can a lab with a MinION and an ordinary laptop go from nanopore reads to the results an isolate
paper reports: assembly, completeness, sequence type, resistance genes and GTDB species? This page
measures that through irit-mcp (the new flye and medaka tools) on one public Klebsiella
pneumoniae isolate. It compares the result with the same isolate’s Illumina reads and with a
published hybrid analysis.
Everything marked measured was run on the project laptop on 2026-11-13. Cited numbers come from the sources linked next to them.
Summary
Section titled “Summary”- The ONT-only route fits in 8 GB. Flye then Medaka ran in a kernel-enforced 8 GB cgroup with
swap off, and nothing was OOM-killed (
memory.events:max 0,oom_kill 0). Flye peaked at 3.85 GiB and Medaka at 7.36 GiB, the closest step to the cap (measured). - Time. Flye took 22 min and Medaka 9.4 min on 6 threads, on a 12-core machine shared with other builds (load average 15–55). The whole run, with an Illumina comparison and all typing, took 74 min of wall clock (measured).
- Same answers as Illumina. The ONT-only assembly (Flye + Medaka) and the Illumina assembly
(Shovill) agree:
- ST101, all 7 alleles exact.
- The same 49 AMRFinderPlus elements, with identical method, identity and coverage for every one. These include blaKPC-3, blaOXA-232, blaCTX-M-15, armA and the gyrA/parC quinolone mutations.
- CheckM2: 100 % complete; contamination 0.44 % (ONT) vs 0.38 % (Illumina).
- Species:
s__Klebsiella pneumoniae, ANI 98.92 % (ONT) vs 98.96 % (Illumina). - The unpolished Flye draft gave the same ST and the same 49 elements as well.
- Far more contiguous. The ONT assembly has 11 contigs, including a 5.49 Mb chromosome contig (not closed into a circle), against 127 contigs and N50 174 kb from Illumina.
- Matches the published analysis. Our assemblies contain every element the study’s Autocycler assembly of these reads has (45/45). The four extra elements show up in our Illumina assembly too, so they come from the newer AMRFinderPlus database, not from nanopore errors.
- Three more isolates with small resistance plasmids (below):
- Same ST as the study for all three, and every resistance element of the hybrid assemblies, except one.
- The exception: default Flye missed AHB29’s 4,999 bp blaSHV plasmid, as the study’s
assemblers did. Flye
--metafound it. - Deep reads (450x) had to be subsampled to fit 8 GB.
- Flye closed several small plasmids as two or three copies in one circle, which multiplies gene counts. irit-mcp now warns about it.
- What the laptop doesn’t do: basecalling. These reads arrived already basecalled with the
supmodel. Basecalling is GPU work, and on a CPU it is roughly 30× slower (see basecalling).
Why Flye + Medaka (cited)
Section titled “Why Flye + Medaka (cited)”- Consensus assembly is the most accurate long-read-only method, but too heavy for a laptop.
Autocycler (Wick, Howden & Stinear, Bioinformatics 2025,
doi:10.1093/bioinformatics/btaf474; bioRxiv
10.1101/2025.05.12.653612) builds a consensus from
many input assemblies. Each Autocycler assembly typically had fewer than 10 errors, mostly
homopolymer lengths, and fewer with R10.4.1 and
supbasecalls. - The trade-off in numbers (Nagy et al., Microbial Genomics 2026, 12(2):001631,
doi:10.1099/mgen.0.001631). This study compared
assemblers on 92 Enterobacterales isolates, mostly E. coli and K. pneumoniae, with R10.4.1
supreads and Illumina reads:- Autocycler + Medaka was the most accurate long-read-only route: median 0 errors per genome, and 51/92 genomes error-free. Its consensus combined 20 assemblies (4 each from Canu, Flye, Raven, Miniasm and Hybracter).
- Flye was median Q61 unpolished and Q67 after Medaka. Medaka mainly removed indels.
- Plasmids: Flye recovered 56 % of the reference plasmids, the other assemblers 93–96 %.
- Typing: no significant difference between long-read-only and hybrid assemblies in resistance, virulence or stress gene counts, or in MLST.
- Compute: the whole comparison took 13,428 CPU-hours.
- So irit starts with Flye + Medaka. Each is one program with a pinned Bioconda version, and
both fit 8 GB. Autocycler’s 20 assemblies per isolate would take hours on a laptop, and Canu
alone is slow. The
flyetool returns each contig’s circularity and depth, so a missing small plasmid is visible rather than silent.
| Isolate | K. pneumoniae AF47, carbapenemase-producing, from a rectal screening swab (NEKSUS study, England, 2024) |
| ONT | ENA ERR15424148: PromethION, R10.4.1, Rapid Barcoding Kit V14, Dorado sup methylation-aware v5.0.0 model (cited, Nagy et al.); 84,759 reads, 385.8 Mb (64× of the 6.05 Mb assembly), 309 MB gzipped; leanqc Q20 84.8 % of bases (measured) |
| Illumina | ENA ERR15422825: NovaSeq X, 2×150, 9.36 M reads, 730 MB gzipped |
| Published results | BioProject PRJEB93885. Per-assembler AMRFinderPlus, MLST, CheckM2, contig and accuracy tables on FigShare, doi:10.6084/m9.figshare.29584931 |
Why AF47:
- It is the second-smallest K. pneumoniae download in the set (1.04 GB for both read sets).
- Its ONT depth (64×) is close to the study’s 60× subsample, and to what a multiplexed MinION run gives.
- It is a hard genome. In the study, no assembler closed its chromosome: Flye gave 13 contigs (N50 3.03 Mb), and Autocycler 171.
- Its resistance profile (KPC-3, OXA-232, CTX-M-15, ArmA, QRDR mutations) tests whether long-read-only assemblies get the genes and point mutations right.
The ENA FASTQ headers are bare (@ERR15424148.1 <uuid>). Dorado’s basecalling-model tag is gone,
so Medaka’s automatic model choice and --bacteria can’t work, and the model has to be named. We
used r1041_e82_400bps_sup_v5.0.0, which matches the basecaller model. The bacterial methylation
model is the alternative.
Setup (measured)
Section titled “Setup (measured)”| irit | nanopore-assembly branch, irit-mcp debug build, run by a scripted MCP client (one stdio session, calls in order) |
| Limits | systemd-run --user -p MemoryMax=8G -p MemorySwapMax=0; the client sampled the unit’s memory.current every 0.2 s for each call’s peak |
| Tools | Flye 2.9.6 (2.9.6-b1802), Medaka 2.2.2 (CPU, PyTorch), fastp 1.4.0, Shovill 1.4.2 (SPAdes, depth 100), mlst 2.35.0, AMRFinderPlus 4.2.7 (DB 2026-10-29.1, --plus, organism Klebsiella_pneumoniae), CheckM2 1.1.0 (DB v3, lowmem), gtdbprep 0.3.5 ANI screen (GTDB R232, species_only, sketches from GTDB’s server) |
| Threads | 6 for Flye, Medaka, fastp, Shovill, CheckM2 and the ANI screen; 4 for AMRFinderPlus |
| Machine | Arch Linux, 12 cores, 30 GB RAM, shared with other builds (load average 15–55 during the run) |
Steps (measured)
Section titled “Steps (measured)”| Step | Wall time | Peak memory (GiB) |
|---|---|---|
| leanqc, ONT reads | 2.7 min | 0.31 |
flye (nano-hq, genome_size 5.5m) |
22.2 min | 3.85 |
medaka (r1041_e82_400bps_sup_v5.0.0, default batch_size 100) |
9.4 min | 7.36 |
medaka, batch_size 50 (a later run) |
4.8 min | 4.78 |
medaka, batch_size 25 (a later run) |
4.9 min | 3.66 |
| fastp, Illumina | 2.4 min | 2.77 |
| shovill, Illumina (SPAdes, 100×) | 16.8 min | 3.77 |
| mlst (ONT + Illumina; Flye draft) | 13 s; 9 s | 0.33 |
| amrfinder: Medaka / Flye draft / Illumina | 4.2 / 4.6 / 0.6 min | 0.39 |
| checkm2 (2 genomes) | 6.3 min | 4.56 |
| gtdbprep ANI screen (2 genomes) | 4.3 min | 1.33 |
| isotable | 31 s | 1.12 |
The whole unit peaked at 7.3 GiB with no OOM events.
-
Medaka’s 7.36 GiB at the default
batch_sizeis most of the cap. A smaller batch fixes it at no cost here:- 50: peaked at 4.78 GiB.
- 25: peaked at 3.66 GiB.
Both gave the same consensus sequence as the default run (11 contigs, every base equal) and weren’t slower. The default run’s 9.4 minutes was probably slowed by memory pressure near the cap. On 8 GB machines, use
batch_size50, or 25 for more headroom. -
AMRFinderPlus took 7× longer on the long-read assemblies than on the Illumina one. Both are about 6 Mb.
Results (measured)
Section titled “Results (measured)”| ONT only (Flye + Medaka) | Flye draft | Illumina (Shovill) | Study (cited) | |
|---|---|---|---|---|
| Contigs | 11 | 11 | 127 (≥500 bp) | Flye 13; Autocycler 171 |
| Length | 6,050,636 | 6,050,890 | 5,920,807 (≥500 bp) | Flye 6,051,272; Autocycler 6,022,885 |
| Longest contig / N50 | 5,490,603 | 5,490,840 | N50 173,801 | Flye N50 3,032,324 |
| Circular contigs | 5 (139.7, 17.7, 10.8, 6.1 and 3.9 kb) | same | — | Flye 5 |
| GC % | 56.41 | 56 | ||
| MLST | ST101 (PERFECT) | ST101 | ST101 | ST101, all assemblers |
| AMRFinderPlus elements | 49 | 49 | 49 | Autocycler 45 (AMRFinderPlus 4.0.3) |
| CheckM2 completeness / contamination | 100 / 0.44 | 100 / 0.38 | Flye 100 / 0.47; Autocycler 100 / 0.42 | |
| GTDB R232 species (ANI, AF) | K. pneumoniae (98.92, 0.87) | K. pneumoniae (98.96, 0.87) | — |
Notes:
- The 49 elements are the same in all three of our assemblies. Gene symbol, method (e.g.
ALLELEX,POINTX), % identity and % coverage are identical for every one. - Against the study’s Autocycler assembly, ours have all 45 of its elements, plus
mrx(A),terA,traTandompK35_G62AfsTer1. Those four are in our Illumina assembly too, so they come from the newer AMRFinderPlus and database, not from nanopore errors. - The study’s own Flye run reported a partial
blaOXAand lacked the gyrA/parC/ompK36 mutations that its Autocycler assembly had. On our full read set, Flye + Medaka has all of them. - Base-level accuracy (errors per Mb against an Illumina-polished reference) was not measured
here: irit has no read-mapping tool yet. The study measured AF47’s Flye assembly at 4
substitutions and 6 indels unpolished (Q57.8), and 0 and 0 after Medaka with all reads (Q100)
(cited,
assembly_nucleotide_accuracy_cleaned.csv).
Basecalling
Section titled “Basecalling”Basecalling turns the raw electrical signal into bases. It is the step before these reads, and the laptop didn’t do it.
- On a GPU (cited): Ryan Wick measured
sup@v5.2.0at 2.69×10⁷ samples/s on one H100, which basecalled a 132 Gbp PromethION run in 18.75 h (blog, 2025). At that rate, this isolate’s 386 Mb would take about 3–4 min.hacwas 6–10× faster thansup. - On a CPU (cited, older R9.4.1 models): Dorado
hactook 13 h 31 min for an E. coli read set on a 16-core Xeon, and 27 min on an RTX A6000 GPU, about 30× slower (Rawsamble, arXiv 2410.17801).supis another 6–10× slower thanhac(the H100 numbers above). - Apple silicon: Dorado runs on it, but no published bacterial-scale numbers were found.
Dorado 2.0.1 added an Apple Silicon implementation for the v6.0.0
hacmodel. - A practical route for a lab without a GPU: basecall
haclive on the MinION’s computer, or send the raw data to a sequencing service or GPU machine forsup. Then do everything from here on a laptop. Wick foundsup@v5.2.0clearly better thanhac@v6.0.0for assembly (blog, 2026).
Three more isolates: small plasmids (measured)
Section titled “Three more isolates: small plasmids (measured)”Measured 2026-11-14 on the same laptop, each run in an 8 GB, no-swap cgroup, through irit-mcp.
The three isolates come from the same study as AF47 (Nagy et al.), so the study’s hybrid
assemblies (Unicycler with Illumina, unicycler_bold in its FigShare tables) give the expected
plasmids and genes. They were chosen because a resistance gene sits on a small plasmid:
| Isolate | Reads (ENA) | Gene on a small plasmid (hybrid assembly) | The study’s Flye + Medaka |
|---|---|---|---|
| K. pneumoniae AF38 | ERR15423950, 2.58 Gb (about 450x) | blaOXA-232 on 6,141 bp | found |
| E. coli AHB29 | ERR15424439, 0.71 Gb (about 135x) | blaSHV on 4,999 bp | missed (so did its Autocycler + Medaka) |
| E. coli AF39 | ERR15423954, 1.89 Gb (about 340x) | one of two blaTEM-1 copies on 3,731 bp | missed that copy |
Route: leanqc, then rasusa to 100x, then Flye
(genome_size set), Medaka (model r1041_e82_400bps_sup_v5.0.0, batch_size 50), AMRFinderPlus
--plus and mlst. For AHB29 and AF39, Flye also ran in --meta mode, which keeps
uneven coverage.
Deep reads need subsampling first. The first AF38 run gave Flye all 2.58 Gb with
asm_coverage 50. That option limits only Flye’s first stage, and the consensus stage, which
maps every read, went over 8 GB after 27 minutes and was killed. irit-mcp got a
rasusa tool for this, and Flye’s description now says so.
| Isolate | Flye | Medaka | Contigs (circular) | Chromosome | ST (study’s) |
|---|---|---|---|---|---|
| AF38 | 24 min, 4.54 GiB | 8.9 min, 5.28 GiB | 7 (7) | closed, 5.51 Mb | 231 (231) |
| AHB29 | 20 min, 5.15 GiB | 6.9 min, 4.87 GiB | 3 (3) | closed, 5.19 Mb | 73 (73) |
AHB29 --meta |
21 min, 4.98 GiB | 5.8 min, 5.15 GiB | 13 (6) | closed | 73 |
| AF39 | 13 min, 6.96 GiB | 4.2 min, 6.58 GiB | 7 (6) | closed, 5.15 Mb (the hybrid’s is in two pieces) | 3032 (3032) |
AF39 --meta |
30 min, 6.46 GiB | 4.3 min, 6.49 GiB | 6 (6) | closed | 3032 |
Rasusa took 53–145 s and at most 1.85 GiB. Nothing was OOM-killed once the reads were subsampled.
Resistance genes against the hybrid assembly (AMRFinderPlus elements of type AMR, including point mutations; the study ran an older AMRFinderPlus database):
| Assembly | Hybrid’s elements found | Missing | Extra |
|---|---|---|---|
| AF38 | 21 of 21 | none | mrx(A), ompK35 truncation (newer database) |
| AHB29 | 5 of 6 | blaSHV | none |
AHB29 --meta |
6 of 6 | none (blaSHV-1: the newer database names the allele) | none |
| AF39 | 24 of 24 | none | mrx(A), cirA Y277* |
AF39 --meta |
24 of 24 | none | same |
- AHB29’s 4,999 bp plasmid was missed by default Flye, as by the study’s assemblers. In these
reads it sits at about 11x against 98x for the chromosome. That is low even though these reads
come from a rapid kit, which recovers small plasmids better than ligation (Wick et al.,
Microbial Genomics 2021, doi:10.1099/mgen.0.000631).
--metarecovered it, along with ten extra contigs at 3–18x depth. Long-read assemblers in general miss or distort small plasmids (Long read genome assemblers struggle with small plasmids, Microbial Genomics 2023, doi:10.1099/mgen.0.001024). - AF39’s 3,731 bp plasmid was found by both modes this time, inside an 11,269 bp circle at 20x.
Small plasmids come out as several copies. Flye closed several small plasmids as one circle holding two or three copies:
- AF38’s 6,141 bp blaOXA-232 plasmid is an 18,377 bp contig with three copies, so AMRFinderPlus lists blaOXA-232 three times.
- AHB29’s plasmid in the
--metaassembly is a 14,997 bp circle: three copies of 4,999 bp, three blaSHV-1 hits. - AF39’s 3,731 bp plasmid is in an 11,269 bp circle with two blaTEM-1 copies.
- Two contigs of AF47 (10,848 and 17,738 bp) have the same shape.
A gene’s presence is right, but its copy count and the plasmid’s length are not. Medaka keeps the copies.
irit-mcp now warns about these. For each circular Flye contig up to 300 kb, it measures how much of the contig’s sequence (21-mers) occurs more than once in it. Every multi-copy contig above scored 0.92–1.0; every other contig of these assemblies scored at most 0.2. Contigs above 0.5 get a warning naming the copy count where the contig is a clean multiple.
Limits and open questions
Section titled “Limits and open questions”- Four isolates, one study. AF47’s chromosome didn’t close; the other three closed. All four
come from one study’s PromethION runs with the rapid barcoding kit and
supbasecalls. Ligation kits (which recover small plasmids less well, Wick et al. 2021) and other labs are untested. - Small plasmids: check them with
--metaas well as the default. On AHB29 only--metafound the resistance plasmid; it also adds low-depth fragments, so read its extra contigs with care. - Corrupt downloads. Four ENA downloads (AHB29 and AF39, twice each) arrived the right size but corrupt, with no error from curl; a third copy matched ENA’s MD5. Flye stopped at an “invalid sequence symbol” deep into the file. Check reads against ENA’s MD5s before a run; irit can’t do that yet.
- isotable needed a long-read read-QC row (fixed, #21).
It marked the ONT sample
incompletefor “no read count / no Q30 (fastp)”, and leanqc’s FastQC thresholds failed long reads on per-position quality. isotable now takes reads, read N50 and median read quality from leanqc’s long-read numbers, and leaves the short-read checks ungraded. The ONT sample passes. - The Medaka model is downloaded on first use.
r1041_e82_400bps_sup_v5.0.0is not in the Bioconda package; Medaka fetched it into the environment during the call. irit-mcp now hashes the model file Medaka logged into the run record (#21). - Not measured: Autocycler on a laptop, and Medaka on GPU or Apple silicon.