This repository measures how the calls of PCAn v1.0, a tool that predicts point centromeres in budding yeast genome assemblies, depend on the assembly they are made from. Installed from its pinned commit with MEME suite 4.11.2, PCAn ran on 134 Saccharomycetaceae assemblies from 24 genera and made 1,519 calls, 2 to 26 per assembly. It called all 16 centromeres of the S288C reference, each with SGD's boundary between CDEII and CDEIII, and five runs gave identical output. In a simulation that cut ten chromosome-level assemblies into contigs (1,600 replicates holding 15,520 unfragmented calls), 518 calls were lost and every one had a breakpoint inside its extraction window; none was lost with its window intact, and all 391 calls gained were candidates that PCAn's one-call-per-contig rule had suppressed before the cuts. At a contig N50 of 5 kb, mean recall relative to the unfragmented call was 0.93 with uniform breakpoints and 0.79 when breakpoints favoured AT-rich sequence. Planted CDEII changes of -10 to +20 bp were reported at the planted length in 207 of 210 edits; changes of +30 and +40 bp mostly fell outside PCAn's 30 bp length anchor. In 69 Saccharomyces cerevisiae strains with a published short-read and a long-read assembly, 975 of the 1,103 PCAn calls on the long-read assemblies were intact and called in the short-read assembly: 603 of 608 in homozygous strains and 372 of 495 in heterozygous ones. Two intact regions went uncalled. Calls were broken at about half the rate of AT-matched null windows (ratio 0.51, 0.35 to 0.66), and the breakpoints of the short-read assemblies fell in AT-rich windows more often than uniform placement predicts (maximum-likelihood gamma 5.0, 4.5 to 5.6). In 33 homozygous strains whose Illumina reads were assembled at 5x to 80x with SPAdes and MEGAHIT (462 assemblies), the fraction of long-read calls found intact and called rose from 0.09 (SPAdes) and 0.01 (MEGAHIT) at 5x, where contigs were shorter than the flanks the comparison needs, to 0.98 and 0.89 at 20x, where both settled, and 0.99 and 0.92 at 80x. Intact calls had the long-read CDEII length in at least 98.9 per cent of cases at every depth, and from 10x on breaks fell at centromeres less often than in AT-matched windows.
In budding yeasts of the Saccharomycetaceae, the centromere of each
chromosome is a short DNA sequence on which the kinetochore assembles. This is
a point centromere. It has three parts. CDEI is a short motif bound by the
protein Cbf1. CDEII is an AT-rich stretch wrapped around the centromeric
nucleosome, which carries the histone H3 variant Cse4; in the 16 centromeres
of the S. cerevisiae reference, SGD's CDEII features are 76 to 85 bp long
(results/arm0/s288c_vs_sgd.tsv). CDEIII is a motif of about 25 bp bound by
the CBF3 complex. PCAn's 1,519 calls on the arm 0 assemblies have a median
length of 118 bp from the start of CDEI to the end of CDEIII, and 98 per cent
of them lie between 65 and 201 bp (results/arm0/calls.tsv).
PCAn v1.0 (software record: Zenodo 10.5281/zenodo.17293587) scans an assembly with FIMO for a CDEIII motif, takes a window upstream of every hit, scans each window for a CDEI motif, and scores every pair of CDEI and CDEIII hits from the two motif scores and the AT content of the CDEII between them. It keeps the 50 best pairs and takes the median CDEII length of the top five as an anchor. It then removes pairs whose CDEII length is more than 30 bp from the anchor or whose CDEII AT content is 70 per cent or less, keeps one pair per CDEIII hit, per sequence and per contig, and removes pairs that are outliers in both length and AT content. What survives is the call set. Motifs and thresholds are set per genus. The window runs from 249 bp before the CDEIII hit's start to its end, and the CDEII AT percentage is computed as the released code computes it.
Contig N50 is the length L such that contigs of length L or longer hold half of the assembly. It measures contiguity and nothing else. An assembly can have a high N50 and be wrong, because a misjoin makes a contig longer, or a low N50 and be accurate base for base. Two assemblies with the same N50 can have very different length distributions, which is why every arm 1 replicate also records L50 and the contig count. Here contig N50 is always computed after splitting sequences at every run of N, so that published scaffolds, cut replicates and new contig assemblies sit on one axis.
A short-read assembler joins reads that overlap. At low depth some stretches of the genome have no reads at all and the assembly breaks there. More depth closes those gaps until repeats longer than the reads become the limit. Illumina libraries tend to under-cover very AT-rich sequence, and CDEII is among the most AT-rich sequence in a yeast genome, so a centromere is a plausible place for a short-read assembly to break. Heterozygosity adds a second effect: where the two copies of a region differ, a short-read assembler may keep both as separate contigs.
When PCAn reports no centromere where one is expected, four things can be true. The sequence may be absent from the assembly. It may be present but split across two contigs. It may be present and whole, with PCAn's filters removing it. Or the strain may carry a centromere that PCAn's motifs do not match. Only the last says something about the genome. Arms 2 and 3 tell the first three apart by finding the sequence on either side of each centromere in the other assembly.
Arm 1 compares PCAn's calls on an altered assembly with its calls on the same assembly unaltered, and arms 2 and 3 compare PCAn's calls on two assemblies of one strain. Agreement in either case means PCAn's output is stable across the change; it does not mean the calls are right. The only external truth in this repository is the S288C reference against SGD's centromere annotation.
Every file except the arm 3 reads was downloaded on 2026-10-09 (UTC); the
reads of each arm 3 strain were downloaded, checked against ENA's MD5 and
deleted when that strain was processed (logs/arm3_strains.tsv). Sizes,
SHA-256 checksums, source checksums where the source publishes one, and
download times are in logs/downloads.tsv, logs/assemblies_arm0.tsv and
logs/assemblies_arm2.tsv.
PCAn is installed from its GitHub repository at commit
a5fa46f0cb971e7d499fd52b38b0c29166b33ef8, the v1.0 release archived on Zenodo
(10.5281/zenodo.17293587); scripts/01_install.sh holds the address. One
patch is applied so that Arxiozyma assemblies run
(patches/0001_arxiozyma_cdeiii_motif.patch): the Arxiozyma entry of the
motif table names a CDEIII motif file that is not in the release, and the
patch points it at CDEIII_Arxiozyma_MEME.txt, which ships with PCAn. Nothing
else is changed.
The species and their assembly accessions come from the species table of the
Figshare collection 10.6084/m9.figshare.c.7630151 (Supplementary Data 5,
MD5-checked). Of its 166 rows, 138 are Saccharomycetaceae species. PCAn v1.0
has no motif entry for Naumovozyma, so its 3 species were not run, and NCBI
had suppressed the assembly listed for Maudiozyma bulderi (GCA_933962305.1,
now GCA_933962305.2), so nothing was run in its place
(results/arm0/excluded_assemblies.tsv). That leaves 134 assemblies from 134
species and 24 genera, downloaded by exact accession and version, 0.51 GB
compressed. The role of every row is in config/species_arm0.tsv, and what
was run for each Saccharomycetaceae species in
results/arm0/species_counts.tsv.
The S288C reference is SGD release R64-5-1
(S288C_reference_genome_R64-5-1_20240529.tgz); its 16 chromosomes are
identical, base for base, to NCBI's GCA_000146045.2
(results/arm0/s288c_sequence_check.tsv).
Of the 33 chromosome-level or complete assemblies among the arm 0 assemblies, 26
have as many PCAn calls as nuclear chromosome sequences; the other 7 do not
and are ineligible (results/arm1/eligibility.tsv). The rule in
config/arm1_design.md, written before any replicate ran, took one per genus
by highest contig N50: ten genomes from Ashbya, Eremothecium,
Huiozyma, Kazachstania, Kluyveromyces, Lachancea, Nakaseomyces,
Saccharomyces, Torulaspora and Zygosaccharomyces, holding 6 to 16 calls
each and 97 in all (results/arm1/genomes.tsv). For the planted variants,
the same file's rule took S288C, the genome with the shortest median CDEII
(Kazachstania africana, 49 bp) and the one with the longest (Eremothecium
gossypii, 167 bp) (results/arm1/variant_genomes.tsv).
The long-read assemblies are the telomere-to-telomere panel of O'Donnell et
al. 2023 (ScRAP), taken from NCBI: 69 accessions, sequenced with Oxford
Nanopore and Illumina, 36 at contig and 33 at scaffold level, with contig N50
of 438 to 952 kb (median 910 kb; results/arm2/assembly_stats.tsv). The
short-read assemblies are those of
Peter et al. 2018, distributed as one archive by the 1002 Yeast Genomes
project (1011Assemblies.tar.gz, 3,999,623,201 bytes, MD5
10cd6ced9cd9a1064ee35583d644c458, matching the server's md5.txt). Their
contig N50 ranges from 3.4 to 128 kb (median 84 kb;
results/arm2/short_read_assembly_stats.tsv). Peter et al. built them with
ABySS from Illumina reads, as their Methods describe.
Strains were matched by identifier, never by name alone. ScRAP's
supplementary table gives each of its 142 strains a standardized name, and
for 69 of them that name is a strain code of Peter et al.'s Table S1; those 69
are the pairs. The other 73 were dropped because their standardized name is
not in Table S1 (results/arm2/pair_candidates.tsv). Of the 69 long-read
assemblies, 68 are haploid or collapsed assemblies and one is ScRAP's diploid
assembly of a heterozygous strain (config/arm2_pairs.tsv). Peter et al.'s
Table S1 lists 38 of the strains as homozygous and 31 as heterozygous.
The reads are the paired-end Illumina HiSeq 2000 runs of ENA study PRJEB13017,
the 1002 Yeast Genomes project. Each run was linked to its strain through the
reads-file name in Peter et al.'s Table S17, which matches the submitted file
names of the run in ENA. Of the 69 arm 2 strains, 33 qualify under the
registered rule (homozygous, euploid, a haploid or collapsed long-read
assembly, and a linked run); 36 do not: 28 are not homozygous, 5 are not
euploid, 2 are neither, and 1 has a diploid long-read assembly and is
heterozygous. The 33 runs, with base counts, read counts, file MD5s, the
linking evidence and the depth available (104x to 668x, median 301x), are in
config/arm3_runs.tsv.
Bash runs every stage; Python is called from bash. Each stage skips work it
has already done and says so, measures itself through
scripts/lib/measure.py (elapsed time, peak resident memory, the data
directory's footprint, the drive's free space; logs/resources.tsv), and
refuses to start, with the shortfall, when its projected time or disk exceeds
the budget in project.conf. The results here came from WSL 2 on Windows,
with 16 threads (logs/resources.tsv) and less RAM visible to Linux than the
16 GB the design assumed; SPAdes was capped at 4 GB accordingly.
bash scripts/00_configure.sh --disk 35 --data-dir <path> --yes writes
project.conf. The working budget was set to 35 GB, with the hard ceiling
10 per cent above it. The data directory sat on the Linux filesystem, because
creating small files on the Windows drive mounted in WSL was far slower.
bash scripts/01_install.sh builds two conda environments. PCAn's is built
from the exact pins of its pcan_specs.yml (Python 3.8.17, MEME suite
4.11.2, BLAST+ 2.16.0, Biopython 1.83, pandas 2.0.3), because a later FIMO
can compute p-values differently at a threshold boundary. The pins solve only
with flexible channel priority, which the script sets for that environment
alone. Everything else goes in a second environment (logs/versions.tsv).
Explicit lock files are in env/. conda clean then freed 0.80 GB
(logs/conda_clean.tsv). Measured: 622 s.
bash scripts/02_fetch_tables.sh fetches the supplementary tables, the
Figshare copies, the Figshare listing and the SGD release, records checksums,
and converts them to TSV. SGD's HTTPS host reset the connection repeatedly,
so the release comes from SGD's archive bucket on Amazon S3, by path.
bash scripts/03_fetch_assemblies.sh --set arm0 checks each accession
version's status with NCBI before downloading it with the NCBI datasets
command-line tool; a replaced or suppressed version is listed and nothing is
substituted. Measured: 906 s for 134 assemblies.
bash scripts/04_run_pcan.sh --assembly <fasta.gz> --genus <genus> --out <tsv>
is the only place PCAn runs. PCAn asks four questions on standard input (the
genome path, the genus, the Kazachstania species, and whether to run the
BLAST synteny step); the wrapper answers them from its arguments and answers
no to BLAST, because synteny here is judged geometrically, as whether 10 kb
flanks survive on one contig. PCAn runs FIMO twice into the same output
directory, overwriting the first pass with the second; a pass-through FIMO
wrapper keeps both passes, gzip, with their command lines. The wrapper then
replays PCAn's candidate building and filters in PCAn's own environment
(scripts/lib/pcan_trace.py), so that pandas sorts tied scores exactly as
PCAn does, and records the filter that removed each candidate. The replay
reproduced PCAn's call table in every one of the 2,342 runs made for arms 0
to 2: 134 in arm 0, 2,070 in arm 1 and 138 in arm 2 (logs/arm0_runs.tsv,
results/arm1/fragmentation_runs.tsv, results/arm1/variant_runs.tsv,
results/arm2/pcan_runs.tsv).
Every species is run with the motifs and thresholds of PCAn's own entry for its genus, unchanged in every arm: CDEIII thresholds of 1e-5 to 1e-7 for the species here, and CDEI thresholds of 1e-2, or 1e-3 for Arxiozyma, Kluyveromyces and Zygosaccharomyces. A contig carries at most one call.
python scripts/05_reproduce.py --jobs 8 runs PCAn on every arm 0
assembly, compares the S288C calls with SGD's centromere annotation, and
runs S288C five times to check that the output is deterministic. Measured:
948 s with 8 parallel PCAn runs; one PCAn run took a median of 21.9 s (10.2
to 389.6 s; logs/arm0_runs.tsv).
python scripts/06_fragment.py --jobs 8 cuts each arm 1 genome at a Poisson
rate found by bisection to reach a target contig N50 of 1,000, 500, 200, 100,
50, 20, 10 or 5 kb, under two models: uniform, and AT-weighted, where the
probability of a cut in a 500 bp window is proportional to exp(gamma (a -
a_genome)), with a the window's AT fraction and gamma = ln(10) / 0.30, a
tenfold rise per 30 percentage points of AT. Ten seeded replicates per genome,
model and level; clean cuts, no sequence removed. Each replicate's FASTA is
deleted after PCAn has run, since its seed regenerates it. Measured: 4,520 s
for 1,900 replicates (1,600 for the two models and 300 for the gamma
sensitivity analysis) with 8 parallel runs, in two runs, because the first
was stopped on request after 358 replicates and the second skipped them. No
replicate had to be cut to fit the 24 hour budget (results/arm1/cuts.tsv).
Of the replicates, 140 needed no cut, because the assembly was already at or
below the target, and were scored without a PCAn run.
python scripts/07_plant_variants.py --jobs 8 edits CDEII by microhomology:
an insertion duplicates an adjacent stretch of
CDEII in tandem, and a deletion removes a stretch from inside it, never
touching CDEI or CDEIII. Random insertions would shift AT content in a way
real variants do not, and PCAn scores AT content. Single edits of -10, -5, +5,
+10, +15, +20, +30 and +40 bp at every centromere of the three genomes (280
runs), and, in S288C, +10 bp at 1, 2, 4, 8, 12 and all 16 centromeres, five
seeded replicates per level (30 runs). Measured: 840 s.
python scripts/08_score_perturbations.py scores every unfragmented call in
every replicate as retained exactly, retained with other boundaries, lost with
a cut in its window, or lost with the window intact, and gives the filter
responsible for the last. Measured: 45 s.
python scripts/09_build_pairs.py builds the arm 2 pairs and the arm 3 runs
from the ScRAP and Peter et al. tables, the Peter et al. archive and ENA's run
report for PRJEB13017, with every exclusion and its reason. Measured: 5,043 s,
nearly all of it the 4.0 GB archive.
bash scripts/03_fetch_assemblies.sh --set arm2 fetches the 69 long-read
assemblies (622 s). python scripts/10_compare_pairs.py --arm 2 --jobs 4 then
runs PCAn on both assemblies of every strain and compares them by the method
registered in config/analysis_plan.md before any arm 2 data were examined
(commit cbf81a5, 2026-10-09 04:41 UTC). The 2 kb flanks on either side of each
long-read call are aligned to the short-read assembly with minimap2
(-x asm10); a flank is placed only if exactly one primary alignment has
identity of at least 0.95, covers at least 90 per cent of the flank and has
mapping quality of at least 20. Nothing is forced. Each call gets one of five
statuses: intact and called, intact but not called, split, containing an N
run, or absent. Twenty null windows per call, matched on length and on AT
content within 2.5 percentage points, go through exactly the same steps,
because a null that skipped the liftover would measure the liftover. Measured:
1,982 s with 4 parallel jobs, peak 1,840 MB. That first run stopped in its
final aggregation: pandas had read the null windows' kind, the word "null",
as a missing value, and the line summarising C5 failed. Both were fixed
before any confirmatory result was written; the aggregation alone took 97
to 163 s in the runs after the fix (logs/resources.tsv).
bash scripts/11_assemble_reads.sh processes the 33 arm 3 strains one at a
time in the registered order. Each read file is downloaded from ENA,
resuming any transfer that breaks, checked against ENA's MD5, and fanned out
in one pass through named pipes to seqtk, which draws every depth and seed;
then it is deleted. The registered design streamed the reads instead, but on
this link the first strain's streams broke repeatedly, once with a TLS
record that failed its integrity check, and a broken stream must start again
from its first byte; the change and its date are at the end of
config/analysis_plan.md. While one strain assembles, the next strain's
reads download, because the link and the processors would otherwise sit idle
in turn; assemblies never overlap. Reads and working directories are deleted
when a strain is done, or by the cleanup trap if it is interrupted. SPAdes
runs under the 4 GB cap from project.conf, which limits its address space
rather than its resident memory. On 16 threads it ran out of address space
for the first three strains, at 20x and above for CCN and at 40x and above
for AEG and BAM, while holding about 1.5 GB;
from the fourth strain on it runs on one thread per GB of the cap, 4 here,
which on BAM's 40x reads gave contigs identical to a 16-thread run without a
cap. The failed runs of the first three strains stay failed, as registered.
ENA's server also wrote error messages of its own into some transfers, in
place of the data at that point: an XML fragment beginning
<Error><Code>ConnectionClosedException</Code> that reports a premature end
of the message body. That leaves a file of the right size and a wrong MD5,
different on every try. Before this was found, downloads were moved to
Windows' own curl on the assumption that the damage arose inside WSL; it did
not help and was reverted. Since 20:11 UTC on 10 October the script finds each
such message in a downloaded file, fetches the 4 MiB on either side of it
again as a byte range, writes them back in place and then checks the MD5 as
before. Three runs of a strain ended without intact reads and three were
stopped part way; each is a row of logs/arm3_strains.tsv, and every one of
the 33 strains was assembled in the end. The stage was interrupted five times,
once when the Linux VM shut down with the session that had started it and four
times by hand to change the download method or the SPAdes threads
(logs/resources.tsv).
The registered time budget was 48 h. At 02:19 UTC on 11 October, with 14
strains complete and none cut, it was raised to 96 h so that all 33 strains
would be assembled. The running stage kept the 48 h it had read at its start
and, at 17:11 UTC, refused the 32nd strain, whose projected end, from the
slowest download and assembly seen so far, lay 0.1 h past it; a second run
with the new budget assembled the last two. Counted once where they
overlap, the strain intervals in logs/arm3_strains.tsv cover 47.3 h, as do
the assemble_reads rows of logs/resources.tsv.
python scripts/10_compare_pairs.py --arm 3 then compares every assembly
with the strain's long-read assembly by the arm 2 method.
python scripts/12_analyse.py writes the reporting summary, with Holm's
correction across the fourteen registered tests, and decides for every
sensitivity analysis whether it changes a conclusion by the rule set before
any sensitivity result was read. python scripts/13_figures.py draws the
figures.
Changes made after the analysis plan was registered are recorded, dated, at
the end of config/analysis_plan.md: the overlap of download and assembly in
arm 3, the download of reads to a file in place of a stream, the SPAdes
thread count, the detour through Windows' curl and its correction, the repair
of the server's error messages, the raised arm 3 budget, two exploratory
additions to arm 2, and two points of implementation that change no arm 2
result.
PCAn ran on all 134 assemblies and called all 16 centromeres of the S288C reference at SGD's boundary between CDEII and CDEIII, with identical output in five runs.
PCAn completed on every one of the 134 assemblies and made 1,519 calls, a
median of 11 per assembly (2 to 26; logs/arm0_runs.tsv,
results/arm0/calls.tsv). Call lengths cluster by genus: the median call is
83 or 84 bp in eight genera, among them Kazachstania, 119 bp in
Saccharomyces, 168 bp in Zygosaccharomyces and 198 bp in Kluyveromyces
(results/arm0/calls.tsv, genus from config/species_arm0.tsv).
On S288C, PCAn called all 16 SGD centromeres and nothing else. Each call
covers SGD's centromere feature to within 1 bp at either end, and the boundary
between CDEII and CDEIII matches SGD's in all 16. PCAn's CDEII is 2 bp longer
than SGD's in every case; SGD's CDEI features are 10 bp long and PCAn's CDEI
motif is 8 bp (results/arm0/s288c_vs_sgd.tsv).
FIMO never reached its cap of 100,000 stored matches: the largest first pass
held 2,927 hits (results/arm0/fimo_cap.tsv), so no call depends on genome
size through that cap.
In the simulation, every lost call had a breakpoint inside its extraction window, and mean recall at a contig N50 of 5 kb was 0.93 with uniform breakpoints and 0.79 with AT-weighted ones.
Recall here is relative to the unfragmented call; it is self-consistency, not
accuracy. Across 1,600 replicates, 15,520 unfragmented calls were scored:
15,002 retained, 518 lost, every lost call with a cut in its window, and none
lost with the window intact (results/arm1/fragmentation_replicates.tsv,
results/arm1/pipeline_losses_by_filter.tsv). The geometric rule predicted
520 losses. Two of those calls survived: one on a fragment of 170 bp, and one
recovered with a different CDEI that shortened its CDEII from 49 to 23 bp
(results/arm1/fragmentation_calls.tsv). Mean recall over genomes, with
bootstrap intervals over the ten genomes (results/arm1/recall_pooled.tsv;
the figure shows the median over replicates instead, which is higher at 5 kb
because a few genomes lose many calls):
| Target contig N50 | Uniform | AT-weighted |
|---|---|---|
| 50 kb | 0.996 (0.992 to 0.999) | 0.973 (0.952 to 0.989) |
| 20 kb | 0.979 (0.972 to 0.985) | 0.938 (0.910 to 0.963) |
| 10 kb | 0.967 (0.952 to 0.980) | 0.865 (0.797 to 0.924) |
| 5 kb | 0.926 (0.901 to 0.946) | 0.786 (0.696 to 0.861) |
At 200 kb and above, recall is 0.994 or more under both models. The
AT-weighted model costs more calls because CDEII is AT-rich, which puts cuts
where calls are. The calls gained, 391 in all, are a pipeline effect of
another kind: each was a candidate that PCAn's one-call-per-contig rule had
removed from the unfragmented assembly, which became the best candidate on a
contig of its own once the cuts separated it from the unfragmented call
(results/arm1/gained_calls_by_origin.tsv). A call that survives may still
lose the evidence used to confirm it: the median fraction of calls whose
contig extends 10 kb beyond both ends was 0.50 (uniform) and 0.44
(AT-weighted) at 50 kb, and zero at 10 kb and below. The arm 0
assemblies have contig N50 of 16 kb to 1.75 Mb (median 220 kb;
results/arm0/assembly_stats.tsv), and the arm 2 short-read assemblies 3.4 to
128 kb; both sets are marked under the top panels.
In the simulation, PCAn reported planted CDEII changes of -10 to +20 bp at the planted length in 207 of 210 edits, and did not recover most changes of +30 bp or more, which fall outside its 30 bp length anchor.
For each size, 35 centromeres were edited one at a time across the three
genomes (results/arm1/detection_envelope.tsv). At -10, -5, +5 and +20 bp
all 35 were called at the planted length; at +10 bp, 34; at +15 bp, 33. The
three misses were removed by the filter for outliers in both length and AT
content (results/arm1/planted_single.tsv). At +30 bp, 15 edited centromeres
were still called and 13 at the planted length; at +40 bp, 3 were called and
none at the planted length. The 52 edits not recovered at those two sizes were
all removed because their CDEII length fell more than 30 bp from the anchor;
the 5 called at another length took a different CDEI. The interval for these
fractions comes from three genomes and is rough.
The progressive design asks whether PCAn can count a genome part of the way
through a transition. With +10 bp edits at 1, 2, 4, 8, 12 or 16 of S288C's
centromeres, PCAn called a median of 16 centromeres at every level, and every
edited centromere at the planted length in 29 of 30 replicates; in one
replicate at 4 edits it called 3 of them and 15 in all. Unedited centromeres
were unchanged in every replicate (results/arm1/planted_progressive.tsv).
For a change of 10 bp, the median anchor does not stop PCAn from counting a
mixed genome.
In 69 strains, 975 of 1,103 calls on the long-read assembly were intact and called in the published short-read assembly; 126 were broken there and 2 were intact but not called.
These are the confirmatory outcomes registered in config/analysis_plan.md
(results/arm2/confirmatory.tsv). PCAn made 15 to 17 calls on each long-read
assembly (16 in 66 strains).
C1. The two assemblies gave the same call count in 44 of 69 strains (0.638, Wilson interval 0.520 to 0.741). Where they differed, the short-read assembly usually had more calls, by up to 13.
C2. Of 1,103 calls on the long-read assemblies, all liftable, 975 were intact
and called in the short-read assembly (0.884, bootstrap interval over strains
0.834 to 0.928), 2 intact but not called (0.002), 62 split (0.056), 27 in an N
run (0.024) and 37 absent (0.034). Both calls that were intact but not called
were removed by the one-call-per-contig rule: in each case the short-read
contig joins the region with another call that lies on a different contig of
the long-read assembly, and PCAn kept the other one
(results/arm2/centromere_status.tsv; the filter, from the trace, in
results/arm2/exploratory_centromere_flanks.tsv).
C3. Among the 975 intact calls, 909 had the same CDEII length in both assemblies (0.932, 0.904 to 0.956). The other 66 differed by -8 to +6 bp; 62 of them are in heterozygous strains.
C4. Calls were broken in the short-read assembly at about half the rate of
AT-matched null windows: 126 of 1,103 calls against 4,926 of 22,060 windows,
a break-rate ratio of 0.51 (0.35 to 0.66) and a difference of -0.109 (-0.135
to -0.081). The stratified permutation test gave p = 0.0001 before
correction, the smallest value 10,000 permutations allow; after Holm's
correction it is 0.0011 (results/reporting_summary.md). Short-read
assemblies of these strains do not break at centromeres more often than in
other sequence of the same AT content; they break there less often.
Zygosity divides the strains sharply. This comparison is exploratory, though
excluding heterozygous strains is a registered sensitivity analysis. In the 38
homozygous strains, 603 of 608 calls were intact and called, 3 were split,
CDEII lengths agreed for 599 of 603, and call counts agreed in 33 strains. In
the 31 heterozygous strains, 372 of 495 were intact and called, CDEII lengths
agreed for 310 of 372, call counts agreed in 11, and 195 of the 566
short-read calls had no long-read counterpart, against 6 of 609 in
homozygous strains (results/arm2/call_counts.tsv,
results/arm2/short_calls_without_long_counterpart.tsv,
figures/supplementary/figS2_zygosity.png). The heterozygous short-read
assemblies are also longer and more fragmented: median total length 14.2 Mb
against 12.0 Mb, and median contig N50 34 kb against 99 kb
(results/arm2/short_read_assembly_stats.tsv).
Part of the "broken" count in heterozygous strains is not a break. Aligning
the flanks again and counting every full-length match, secondary ones
included, shows that for 46 of the 123 broken calls in heterozygous strains a
flank matches two or more places in the short-read assembly, as when both
haplotypes are kept as separate contigs; the registered rule leaves such a
flank unplaced, so the call counts as split or absent. For 34 a flank matches
nowhere, 27 have an N run between placed flanks, and 16 have both flanks
placed but too far apart, on different contigs or in opposite orientations.
The 3 broken calls in homozygous strains all have a flank that matches
nowhere. In heterozygous strains the doubly matched flank strikes calls and
null windows at a similar rate, 46 of 495 calls and 980 of 9,900 windows, so
it inflates both sides of C4 alike
(results/arm2/exploratory_flank_summary.tsv, exploratory). Supplementary
Data 4 of the Figshare collection lists two different centromere sequences for
14 of the broken calls and 3 of the 66 calls with a different CDEII length, so
those may be allelic rather than errors of either assembly
(results/arm2/centromere_status.tsv, column possibly_allelic).
A call that survives may lose its synteny evidence: for 819 of the 975 intact
calls (0.84) the short-read contig extends 10 kb beyond both ends of the call,
529 of 603 in homozygous strains and 290 of 372 in heterozygous ones
(results/arm2/exploratory_centromere_flanks.tsv, exploratory).
Matching calls by sequence instead of by flanks is stricter, because a
single base of difference breaks the match. For 874 of the 1,103 long-read
calls a short-read call has an identical sequence from CDEI to CDEIII: 601 of
608 in homozygous strains, 273 of 495 in heterozygous ones
(results/arm2/call_counts.tsv, exploratory).
Recall relative to the long-read calls settled at 20x with both assemblers, at least 98.9 per cent of intact calls kept the long-read CDEII length at every depth, and from 10x on breaks fell at centromeres less often than in AT-matched windows.
All 33 qualifying strains were assembled with SPAdes and MEGAHIT at 5x, 10x
(seeds 11, 22 and 33), 20x, 40x and 80x, the full run capped at 80x: 462
assemblies, of which 455 completed (results/arm3/assemblies.tsv). The seven
failures are SPAdes runs of the first three strains, CCN at 20x, 40x and 80x
and AEG and BAM at 40x and 80x, which ran out of address space under the 4 GB
cap before SPAdes was limited to four threads; as registered, they stay
missing. Every subsample came within 0.4 per cent of its target depth, and
mate names agreed read for read in all of them. Median contig N50 rose from
2.1 kb at 5x to 29 kb at 10x, 95 kb at 20x and 135 kb at 80x with SPAdes, and
from 1.3 kb to 14 kb, 62 kb and 72 kb with MEGAHIT.
Recall (D1), the mean over strains of the fraction of long-read calls that
are intact and called, with its bootstrap interval over strains
(results/arm3/confirmatory.tsv):
| Target depth | SPAdes | MEGAHIT |
|---|---|---|
| 5x | 0.089 (0.064 to 0.114), 33 strains | 0.009 (0.002 to 0.017), 33 strains |
| 10x, seed 11 | 0.845 (0.786 to 0.892), 33 strains | 0.661 (0.604 to 0.710), 33 strains |
| 20x | 0.982 (0.969 to 0.994), 32 strains | 0.890 (0.862 to 0.915), 33 strains |
| 40x | 0.992 (0.981 to 1.000), 30 strains | 0.909 (0.894 to 0.924), 33 strains |
| 80x | 0.990 (0.981 to 0.998), 30 strains | 0.920 (0.907 to 0.934), 33 strains |
By the registered rule, the lowest depth from which the mean stays within
0.05 of its full-depth value, recall settles at 20x for both assemblers
(D3; results/reporting_summary.md). At 5x the measure reports the
comparison's limit as much as the assembly's: an element counts as intact
only if it and 2 kb on each side lie on one contig, and contig N50 was 1 to 3
kb. Counting instead the long-read calls whose exact sequence is among the
calls on the read assembly, an exploratory measure, gives a mean of 0.80
(SPAdes) and 0.67 (MEGAHIT) at 5x and 0.96 and 0.95 at 10x
(exploratory_long_calls_matched_by_sequence in
results/arm3/assemblies.tsv). MEGAHIT's recall stayed below SPAdes' at
every depth, and its assemblies stayed less contiguous.
Among intact calls, the fraction with the long-read CDEII length (D2) was at
least 0.989 for both assemblers at every depth; by the registered rule it
settles at 5x, where only 24 SPAdes and 5 MEGAHIT strains had any intact call
(results/arm3/confirmatory.tsv). In the 7,280 comparisons of a long-read
centromere with a read assembly, a region was intact but uncalled 7 times
(results/arm3/centromere_status.tsv).
At 10x the three seeds gave mean recall of 0.845, 0.854 and 0.839 with SPAdes
and 0.661, 0.665 and 0.655 with MEGAHIT. Within a strain, recall moved between
seeds by a median of 2 of the strain's 16 long-read calls with SPAdes and 3
with MEGAHIT, and by at most 5 and 6; CDEII agreement moved by at most 0.07
and 0.10 (results/arm3/seed_range_10x.tsv; per strain, Figure S3 in
figures/supplementary/figS3_read_depth_per_strain.png).
Breaks (D4). At 5x, breaks fell at centromeres about as often as in the
AT-matched null windows, with a break-rate ratio of 0.96 (0.94 to 0.98) for
SPAdes and 1.00 (1.00 to 1.01) for MEGAHIT. From 10x on they fell there less
often: ratios of 0.45, 0.09, 0.04 and 0.07 at 10x, 20x, 40x and 80x with
SPAdes, and 0.69, 0.49, 0.45 and 0.42 with MEGAHIT. Every one of these tests
has a Holm-adjusted p of 0.0011 across the 14 registered tests, except
MEGAHIT at 5x (0.56; results/reporting_summary.md).
The breakpoints of the published short-read assemblies and of the full-depth arm 3 assemblies fall in AT-rich windows more often than uniform placement predicts, with maximum-likelihood gammas of 5.0 (arm 2), 5.3 (SPAdes) and 6.2 (MEGAHIT), all below the 7.7 of arm 1's AT-weighted model.
Outcome C5 locates every short-read contig end on the long-read assembly
(36,791 ends in 69 strains; 2,414 more could not be located;
results/arm2/breakpoint_summary.tsv) and fits the gamma of arm 1's model to
the 500 bp windows that hold them. The log-likelihood is higher at arm 1's
gamma than at zero and the bootstrap interval of the fitted gamma excludes
zero, so by the registered rule the breaks resemble the AT-weighted model
(results/arm2/confirmatory.tsv). In 63 of 69 strains the windows holding a
break are richer in AT than the genome as a whole. The fitted gamma is lower
than the one arm 1 assumed: in these assemblies AT content raises the chance
of a break about 4.5-fold per 30 percentage points, where arm 1 assumed
tenfold. The closest arm 1 model is the sensitivity run at gamma = ln(5) /
0.30, about 5.4, under which mean recall in the simulation was 0.930 at a
contig N50 of 10 kb and 0.960 at 20 kb
(results/sensitivity/arm1_at_weight_gamma.tsv).
Outcome D5 applies the same fit to the full-depth arm 3 assemblies: 9,600
contig ends in 30 strains with SPAdes and 18,783 in 33 strains with MEGAHIT.
The fitted gamma is 5.30 (4.03 to 6.56) for SPAdes and 6.22 (5.32 to 7.09)
for MEGAHIT; both intervals exclude zero and the log-likelihood is higher at
arm 1's gamma than at zero, so by the registered rule both resemble the
AT-weighted model (results/arm3/confirmatory.tsv).
Positive controls. S288C against SGD: 16 of 16 called, as described above. The
planted variants of arm 1b, where the right answer for every edit is known in
advance: PCAn's response is reported above. The liftover compared each
long-read assembly with itself: all 1,103 calls came out intact and called
(results/arm2/liftover_control.tsv). The arm 3 strains are 33 of these 69,
compared against the same long-read assemblies.
Negative controls. Five PCAn runs on S288C gave byte-identical call tables,
FIMO passes and candidate tables, 4.7 to 5.8 s each
(results/arm0/determinism.tsv). The AT-matched null windows are the
comparison for C4 and D4.
Technical replicates. The ten fragmentation replicates per genome, model and
level give the interquartile ranges in Figure 2; the per-genome curves are in
figures/supplementary/figS1_fragmentation_per_genome.png. In arm 3, the
three 10x seeds are described above.
Every threshold the design calls arbitrary was re-run at alternative values
(results/sensitivity/summary.tsv). For each alternative, a registered test
counts as changed if its Holm-adjusted significance at 0.05 or the direction
of its effect changes, and an estimate counts as changed if it falls outside
the primary 95 per cent interval; the rule was written before any sensitivity
result was read.
- No alternative changes the confirmatory test of C4. Its ratio stays below 1 and significant with flanks of 1 or 5 kb (0.66, 0.71), AT bins of 2.5 or 10 points (0.50, 0.57), placement identity of 0.90 or 0.99 (0.52, 0.65), and with heterozygous strains excluded (0.04).
- Estimates without a test move under some alternatives. Flanks of 5 kb lower
the fraction intact and called to 0.793, because a longer flank is more
likely to cross a break; identity of 0.99 raises the split fraction from
0.056 to 0.102; flanks of 1 kb shrink the C4 difference to -0.060.
Excluding heterozygous strains raises C1 to 0.868, C2 to 0.992 and C3 to
0.993, each outside the primary interval
(
results/sensitivity/arm2_variants.tsv). - In arm 1, AT weighting of ln(5) / 0.30 and ln(20) / 0.30 moves mean recall
outside the primary interval at 10 kb and at 20 kb respectively
(
results/sensitivity/arm1_at_weight_gamma.tsv). In one replicate at the steeper weighting (Ashbya aceri, 10 kb), the length anchor removed two calls whose windows were intact: the only losses of that kind in 1,900 replicates. - In arm 3, three alternatives change a registered test, all at 5x. With
1 kb flanks the MEGAHIT break-rate ratio at 5x becomes 0.97 (0.94 to 1.00,
raw p 0.012), significant after Holm's correction where the primary 1.00
was not; with 5 kb flanks the SPAdes ratio at 5x becomes 1.00 (0.99 to
1.00, raw p 0.45), no longer significant, and the MEGAHIT ratio at 5x moves
from just above 1 to just below it, which the rule counts as a change of
direction although neither value is significant. No D4 test at 10x or
above changes, and neither D5 nor the settling depths are among the
alternatives (
results/sensitivity/arm3_variants.tsv). - Arm 3 recall depends on the flank length, as the flank method implies. Flanks of 5 kb lower it at every depth, to 0.906 (SPAdes) and 0.803 (MEGAHIT) at 80x and 0.854 and 0.769 at 20x; flanks of 1 kb raise it at 5x and 10x, to 0.244 and 0.900 with SPAdes and 0.087 and 0.758 with MEGAHIT. The break proportion difference of D4 moves outside its primary interval under some flank lengths and AT bin widths, while keeping its sign at every depth from 10x on.
pcan_assembly_dependence/
config/ what was fixed in advance, and dated records of later decisions
analysis_plan.md arms 2 and 3, registered (commits cbf81a5, a005d21, dd4693d)
arm1_design.md genome choice, levels, models, seeds
species_arm0.tsv every species-table row, its role and reason
arm2_pairs.tsv the 69 strain pairs and the evidence for each
arm3_runs.tsv the 69 strains, the 33 eligible runs, and why the rest are not
project.conf.example the format of the machine-specific project.conf
env/ explicit conda lock files, scrubbed of local paths
scripts/ 00 to 13, one per stage, and lib/ with the shared code
tests/ known-answer tests, run by bash run_all.sh --test
patches/ the one change PCAn needs to run Arxiozyma, with the reason
results/ tables, sensitivity analyses, reporting summary (see results/README.md)
figures/ main figures, with supplementary figures in figures/supplementary/
data/ assemblies, reads in transit, environments and the PCAn checkout; not tracked, see data/README.md
logs/ time, memory and disk per stage and per arm 3 strain; downloads; versions
run_all.sh the orchestrator
LICENSE
README.md
The data directory peaked at 8.83 GB before arm 3 (logs/resources.tsv):
3.3 GB of conda environments, the 4.0 GB Peter et al. archive kept once with
its checksum, 0.78 GB of NCBI assemblies, and the rest PCAn outputs. Arm 3
adds, at any time, the downloaded read files of one strain (0.9 to 7.2 GB)
and the subsamples and assembler working directories of another, and took
the directory to a peak of 17.98 GB; its per-strain peaks are logged in
logs/arm3_strains.tsv. Commands, in order, with the
times measured here:
bash scripts/00_configure.sh --disk 35 --yes # writes project.conf; add --data-dir to move data/
bash scripts/01_install.sh # 622 s, 3.3 GB after conda clean
bash run_all.sh --test # 64 known-answer tests
bash scripts/02_fetch_tables.sh # under a minute
bash scripts/03_fetch_assemblies.sh --set arm0 # 906 s, 0.51 GB
python scripts/05_reproduce.py --jobs 8 # 948 s
python scripts/06_fragment.py --jobs 8 # 4,520 s for 1,900 replicates, in two runs
python scripts/07_plant_variants.py --jobs 8 # 840 s for 310 runs
python scripts/08_score_perturbations.py # 45 s
python scripts/09_build_pairs.py # 5,043 s, mostly the 4.0 GB archive
bash scripts/03_fetch_assemblies.sh --set arm2 # 622 s, 0.27 GB
python scripts/10_compare_pairs.py --arm 2 --jobs 4 # 1,982 s
bash scripts/11_assemble_reads.sh # 47.3 h of wall time for 33 strains
python scripts/10_compare_pairs.py --arm 3 --jobs 4 # 1,967 s
python scripts/12_analyse.py
python scripts/13_figures.pyThe Python stages run in the tools environment (conda activate <data>/env/tools); bash run_all.sh does the activation and runs everything
above in order, and bash run_all.sh --arm 0 or --from compare runs part of
it. Serial runs are the default (--jobs 1) so that timings stay comparable
between machines; memory is not the reason, since no PCAn run here peaked
above 108 MB (logs/arm0_runs.tsv, results/arm2/pcan_runs.tsv).
Arm 1 is a simulation. Its breaks are clean cuts placed by a model, where real assemblies break at repeats, collapse duplicated sequence and carry misjoins, and its edits follow one mutational mechanism, tandem duplication or deletion of adjacent CDEII sequence.
Self-consistency is not accuracy. Arm 1 measures agreement with PCAn's own call on the unaltered assembly, and arms 2 and 3 agreement between two assemblies of one strain. Only a minority of the species have experimentally validated centromeres; the S288C comparison is the only external truth here.
Arms 2 and 3 cover one species, S. cerevisiae, so they say nothing directly about other genera, and PCAn's motifs and thresholds are clade-specific, so a result in one clade does not transfer to another. The arm 2 short-read assemblies all come from one study and one assembler, ABySS, and the long-read assemblies from one panel. Arm 3 uses two assemblers, one Illumina run per strain and only homozygous, euploid strains, and its recall needs the element and 2 kb on each side on one contig, so at 5x it measures the comparison's limit as much as PCAn's.
The flank method cannot tell a doubly matched flank from a missing one under the registered rule; the exploratory flank check shows how often that matters in heterozygous strains. Supplementary Data 4 identifies chromosomes for only part of the calls, so most broken calls in heterozygous strains carry no allelic flag either way.
One FIMO version, 4.11.2, was used. This repository describes how PCAn behaves on the assemblies analysed here and makes no biological claim.
Everything in results/, figures/ and logs/ is regenerated by the
scripts from public sources: PCAn at commit a5fa46f (Zenodo
10.5281/zenodo.17293587), the species and centromere tables of the Figshare
collection 10.6084/m9.figshare.c.7630151, NCBI assemblies
by accession and version (logs/assemblies_arm0.tsv,
logs/assemblies_arm2.tsv), the Peter et al. 2018 assembly archive from the
1002 Yeast Genomes project, the SGD S288C release R64-5-1, and reads from
ENA study PRJEB13017 (config/arm3_runs.tsv). No assembly or read file is
redistributed. Tables in results/arm0/ and results/arm1/ include the
sequences of individual calls and edits, which come from NCBI assemblies.
Every results table is described column by column in results/README.md.
The analysed tool and its data:
- PCAn v1.0, Zenodo. doi:10.5281/zenodo.17293587 (all versions: doi:10.5281/zenodo.17293586)
- Species and centromere tables, Figshare. doi:10.6084/m9.figshare.c.7630151
Tools:
- Grant CE, Bailey TL, Noble WS. FIMO: scanning for occurrences of a given motif. Bioinformatics 27, 1017-1018 (2011). doi:10.1093/bioinformatics/btr064
- Bailey TL, Johnson J, Grant CE, Noble WS. The MEME Suite. Nucleic Acids Research 43, W39-W49 (2015). doi:10.1093/nar/gkv416
- Li H. Minimap2: pairwise alignment for nucleotide sequences. Bioinformatics 34, 3094-3100 (2018). doi:10.1093/bioinformatics/bty191
- Camacho C et al. BLAST+: architecture and applications. BMC Bioinformatics 10, 421 (2009). doi:10.1186/1471-2105-10-421
- Shen W, Le S, Li Y, Hu F. SeqKit: a cross-platform and ultrafast toolkit for FASTA/Q file manipulation. PLoS ONE 11, e0163962 (2016). doi:10.1371/journal.pone.0163962; Shen W, Sipos B, Zhao L. SeqKit2. iMeta 3, e191 (2024). doi:10.1002/imt2.191
- seqtk 1.5, github.com/lh3/seqtk (no paper)
- Chen S, Zhou Y, Chen Y, Gu J. fastp: an ultra-fast all-in-one FASTQ preprocessor. Bioinformatics 34, i884-i890 (2018). doi:10.1093/bioinformatics/bty560
- Bankevich A et al. SPAdes: a new genome assembly algorithm and its applications to single-cell sequencing. Journal of Computational Biology 19, 455-477 (2012). doi:10.1089/cmb.2012.0021; Prjibelski A et al. Using SPAdes de novo assembler. Current Protocols in Bioinformatics 70, e102 (2020). doi:10.1002/cpbi.102
- Li D, Liu C-M, Luo R, Sadakane K, Lam T-W. MEGAHIT: an ultra-fast single-node solution for large and complex metagenomics assembly via succinct de Bruijn graph. Bioinformatics 31, 1674-1676 (2015). doi:10.1093/bioinformatics/btv033
- O'Leary NA et al. Exploring and retrieving sequence and metadata for species across the tree of life with NCBI Datasets. Scientific Data 11, 732 (2024). doi:10.1038/s41597-024-03571-y
- Cock PJA et al. Biopython. Bioinformatics 25, 1422-1423 (2009). doi:10.1093/bioinformatics/btp163
- McKinney W. Data structures for statistical computing in Python. Proceedings of the 9th Python in Science Conference, 56-61 (2010). doi:10.25080/Majora-92bf1922-00a
- Harris CR et al. Array programming with NumPy. Nature 585, 357-362 (2020). doi:10.1038/s41586-020-2649-2
- Virtanen P et al. SciPy 1.0. Nature Methods 17, 261-272 (2020). doi:10.1038/s41592-019-0686-2
- Hunter JD. Matplotlib: a 2D graphics environment. Computing in Science and Engineering 9, 90-95 (2007). doi:10.1109/MCSE.2007.55
- Grüning B et al. Bioconda. Nature Methods 15, 475-476 (2018). doi:10.1038/s41592-018-0046-7
- pigz 2.8, ShellCheck 0.11.0, openpyxl 3.1.5 and xlrd 2.0.1 (no papers)
Assemblies, reads and annotation:
- Peter J et al. Genome evolution across 1,011 Saccharomyces cerevisiae isolates. Nature 556, 339-344 (2018). doi:10.1038/s41586-018-0030-5
- O'Donnell S et al. Telomere-to-telomere assemblies of 142 strains characterize the genome structural landscape in Saccharomyces cerevisiae. Nature Genetics 55, 1390-1399 (2023). doi:10.1038/s41588-023-01459-y; data, Zenodo doi:10.5281/zenodo.8210254
- Engel SR et al. The reference genome sequence of Saccharomyces cerevisiae: then and now. G3 4, 389-398 (2014). doi:10.1534/g3.113.008995
- Cherry JM et al. Saccharomyces Genome Database: the genomics resource of budding yeast. Nucleic Acids Research 40, D700-D705 (2012). doi:10.1093/nar/gkr1029
- The NCBI assemblies listed in
logs/assemblies_arm0.tsvandlogs/assemblies_arm2.tsv, each cited by its accession and version, and ENA study PRJEB13017.
The code in this repository is under the MIT License (LICENSE). Licences of
the material it uses, each checked on 2026-10-09:
- PCAn is under the MIT License (its repository and the Zenodo record).
- The Figshare collection that holds the species and centromere tables is under CC BY 4.0.
- NCBI places no restrictions on the use or distribution of the assemblies it holds, though submitters may claim rights in their data (NCBI website policies). The ScRAP data are also on Zenodo under CC BY 4.0.
- The Peter et al. 2018 article is under CC BY 4.0. The 1002 Yeast Genomes website, which distributes the assembly archive, states no licence; the archive is downloaded by the scripts and not redistributed.
- Reads in ENA fall under the INSDC policy of free and unrestricted access; EMBL-EBI places no additional restrictions on their use.
- SGD data are under CC BY 4.0.





