Skip to content

cooccur on real data: a known CPR–host pair

cooccur had been checked on simulations and against propr. This page tests it on real data with a positive control. The best-known CPR–host pair is the Saccharibacteria TM7x (“Candidatus Nanosynbacter lyticus”), which grows attached to Schaalia odontolytica (formerly Actinomyces odontolyticus) strain XH001 from the human mouth (He et al. 2015). In culture, 13 of 27 Actinomyces strains tested, XH001 included, could host TM7x (Utter et al. 2020). The question: in many oral metagenomes, does cooccur put Saccharibacteria together with A. odontolyticus, or with Actinomyces in general?

Run on 2026-11-15 with cooccur from irit 0.6.0 (commit 07b274c). Numbers are measured here unless a source is given.

  • The positive control is not recovered. In 128 supragingival plaque and 136 tongue dorsum metagenomes from the Human Microbiome Project, neither Saccharibacteria profile has a positive association with A. odontolyticus or any other Actinomyces. In every run A. odontolyticus ranks in the lower half of the CPR genome’s partners (61st to 111th of 112–121), with rho between −0.11 and −0.35. On the full plaque table the negative association passes q ≤ 0.05 (rho −0.26 and −0.25 for the two Saccharibacteria profiles). At genus level, Saccharibacteria against Actinomyces gives rho −0.37 in plaque and −0.24 on the tongue, both passing.
  • The partners that do pass are anaerobes. The Saccharibacteria profiles’ top positive partners in plaque are Prevotella, Leptotrichia, Selenomonas, Centipeda and Treponema (rho 0.34–0.49). These organisms tend to be found in older, more anaerobic plaque, so the pattern looks like plaque ecology rather than host dependence (interpretation).
  • It’s not a computing error. An independent Python recomputation gave the same rho and Spearman values for the control pair to all six printed decimals. Without the log-ratio transform, the plain Spearman correlation of the percentages is also near zero (+0.04 to +0.05 in plaque). No measure tried gives a positive association.
  • The negative control is close to nominal. With each genome’s values shuffled across samples (section 4), 4.86–5.00% of null pairs had p ≤ 0.05 and 0.95–1.03% had p ≤ 0.01, and 50 shuffled tables per run gave 4 or 5 false pairs passing q ≤ 0.05 in total.
  • Real oral tables are densely associated. 28–44% of species pairs pass q ≤ 0.05 (50–56% at genus level), against 0.002% in the shuffles. With 90–136 samples, cooccur gives small q-values to weak associations too: the control pair passed at rho −0.26. On data like this, significance alone does not pick out a host. Ranking by rho, with --min-stat, is needed, and even that did not work here.

Source. curatedMetagenomicData (Pasolli et al. 2017), ExperimentHub resource EH2491, “20190423.HMP_2012.metaphlan_bugs_list.oralcavity”: MetaPhlAn2 (Truong et al. 2015) relative abundances of the 415 oral metagenomes of the Human Microbiome Project’s 2012 data set, with per-sample metadata (subject, body subsite, read count). URL: https://mghp.osn.xsede.org/bir190004-bucket01/ExperimentHub/curatedMetagenomicData/20190423/20190423.HMP_2012.metaphlan_bugs_list.oralcavity.rda (708,007 bytes, MD5 9e7575da931c5612d20143f108a712aa, which matched the server’s ETag; status “Public” in the ExperimentHub metadata of 2026-11-13). curatedMetagenomicData is under the Artistic-2.0 licence.

The file is an R ExpressionSet. R isn’t installed here, so the matrix (1,988 clades × 415 samples) and the sample table were read with the Python package rdata (eset2tsv.py, kept with the results). The 415 samples come from 91 people: 128 supragingival plaque, 136 tongue dorsum, 119 buccal mucosa and 32 from six other subsites.

Why this table. Three other public tables were checked, and none of them works:

Table Saccharibacteria? Note
HMP DACC HMSMCP/HMP.ab.txt (MetaPhlAn 1, 2012), on the HMP AWS bucket No Has A. odontolyticus, no TM7.
curatedMetagenomicData 3, EH5584 “2021-03-31.HMP_2012.relative_abundance” (MetaPhlAn 3, 748 samples, all sites) No No Saccharibacteria phylum among its 18 phyla.
Shaiber et al. 2020 Table S07 (paper, figshare, CC BY 4.0) Yes, 50+ TM7 genomes in 481 HMP oral samples Coverage of the TM7 genomes only. No host genomes in the same table.

The MetaPhlAn2 profiles have two Saccharibacteria species: the single-cell genomes TM7b and TM7c (Marcy et al. 2007). There is no TM7x entry. Reads of other oral Saccharibacteria, TM7x among them, may hit these two marker sets. So TM7b and TM7c here stand for “Saccharibacteria close to these genomes”, not for one strain. Presence (subset of the table above):

Subsite Samples TM7c present TM7c mean A. odontolyticus present its mean Actinomyces (genus) mean
supragingival plaque 128 107 1.03% 120 0.16% 9.4%
tongue dorsum 136 124 0.03% 136 0.54% 8.7%
buccal mucosa 119 48 0.10% 104 0.22% 2.8%

Plaque, where Saccharibacteria are most abundant, is the main test. Tongue is the second one.

Inputs (scripts in bioinfo-lab/cooccur-real/scripts/): make_inputs.py writes one species × sample table per subsite (MetaPhlAn rows with s__ and no t__, species absent from every sample dropped) and a two-column taxonomy (species, MetaPhlAn lineage with ; between ranks). With --one-per-subject it keeps each person’s first sample only. HMP sampled some people twice, and cooccur assumes exchangeable samples.

python3 make_inputs.py oral_mpa2.tsv oral_meta.tsv supragingival_plaque supragingival_plaque
cooccur supragingival_plaque.tsv -t supragingival_plaque.tax.tsv \
--cpr-taxon p__Candidatus_Saccharibacteria --method rho --min-prevalence 0.5 --fdr 0.05 \
--all-pairs --summary supragingival_plaque.rho.json -o supragingival_plaque.rho.all.tsv
# the same with --method spearman, for tongue_dorsum, for the one-per-subject tables,
# with --min-prevalence 0.8, and for genus tables (rows with g__ and no s__)
python3 rank.py supragingival_plaque.rho.all.tsv # partner ranks of each CPR genome

MetaPhlAn’s percentages go in as they are. cooccur’s statistic does not depend on per-sample scale, so the missing closure (MetaPhlAn sums to 100% over all ranks, not over the kept species) doesn’t matter. Each run took under a second for rho and a few seconds for Spearman.

The plaque summary (rho, all 128 samples):

cooccur: 128 samples, 114 of 389 genomes kept (prevalence >= 0.5), 2830 zeros replaced by 4.000e-5
rho two-sided, 6441 pairs tested, 2469 pass q <= 0.05 (1160 positive, 1309 negative)
labels: 2 CPR, 112 non-CPR, 0 unclassified
CPR–non-CPR pairs passing: 125 (50 positive, 75 negative)

A. odontolyticus as a partner of each Saccharibacteria profile: its rank among all that genome’s partners (1 = most positive), the statistic and q. For genus tables, the partner is the genus Actinomyces.

Table Samples Method Pairs passing TM7b – A. odontolyticus TM7c – A. odontolyticus
plaque 128 rho 2,469 of 6,441 (38%) 95 of 113, −0.26, q 0.009, passes 88 of 113, −0.25, q 0.02, passes
plaque 128 spearman 2,759 (43%) 75 of 113, −0.24, q 0.02, passes 73 of 113, −0.23, q 0.02, passes
plaque, one per person 88 rho 1,761 of 6,328 (28%) 85 of 112, −0.20, q 0.13 69 of 112, −0.16, q 0.20
plaque, one per person 88 spearman 2,060 (33%) 63 of 112, −0.16, q 0.26 61 of 112, −0.11, q 0.35
plaque, prevalence ≥ 0.8 128 rho 614 of 1,770 (35%) (TM7b dropped) 41 of 59, −0.19, q 0.07
plaque, 17 low-diversity samples dropped 111 rho 2,215 of 7,140 (31%) 93 of 119, −0.21, q 0.07 84 of 119, −0.17, q 0.12
tongue 136 rho 3,123 of 7,260 (43%) 106 of 120, −0.26, q 0.004, passes 100 of 120, −0.19, q 0.06
tongue 136 spearman 3,204 (44%) 111 of 120, −0.35, q 0.001, passes 111 of 120, −0.20, q 0.055
tongue, one per person 90 rho 2,226 of 7,381 (30%) 99 of 121, −0.21, q 0.09 96 of 121, −0.15, q 0.24
tongue, one per person 90 spearman 2,459 (33%) 105 of 121, −0.28, q 0.03, passes 82 of 121, −0.12, q 0.44
plaque, genus 128 rho 496 of 990 (50%) Saccharibacteria – Actinomyces: 40 of 44, −0.37, q 0.0004, passes
plaque, genus 128 spearman 558 (56%) 39 of 44, −0.45, q 0.0004, passes
tongue, genus 136 rho 705 of 1,378 (51%) 43 of 52, −0.24, q 0.006, passes
tongue, genus 136 spearman 759 (55%) 39 of 52, −0.22, q 0.03, passes

Every other Actinomyces species, and Propionibacterium propionicum, also ranks low against the Saccharibacteria profiles in plaque: ranks 63–107 of 113 for TM7b, all with rho ≤ −0.08. On the tongue, TM7c’s positive partners include A. johnsonii, A. naeslundii and A. oris (rho 0.21–0.28, ranks 5, 7 and 15 of 120, passing). TM7b’s are not (rho −0.05 to 0.14), and neither is A. odontolyticus. With the two profiles disagreeing and the signal only on the tongue, this is weak evidence at most.

TM7b and TM7c are each other’s strongest partner (rho 0.84 in plaque, 0.52 on the tongue). That is expected for two marker sets of related genomes, and it is the one pair here a reader could predict.

Top positive partners in plaque (rho, all 128 samples):

TM7b rho TM7c rho
Leptotrichia buccalis 0.43 Prevotella saccharolytica 0.49
Prevotella saccharolytica 0.42 Treponema socranskii 0.43
Prevotella maculosa 0.41 Prevotella maculosa 0.41
Centipeda periodontii 0.40 Selenomonas infelix 0.41
Selenomonas artemidis 0.38 Mitsuokella (unclassified) 0.41

Recomputed independently. check_pair.py (plain Python: prevalence filter, minimum non-zero pseudocount, CLR, rho, tie-averaged Spearman) gave rho −0.249734 and Spearman −0.230830 for TM7c – A. odontolyticus in plaque, and −0.258052 and −0.242961 for TM7b. cooccur printed the same values.

Each genome’s row was shuffled across samples independently (shuffle.py, seeds 1–50). That breaks every association but keeps each genome’s values and prevalence. cooccur then ran with the same settings (--seed = shuffle seed, --all-pairs):

Table Method Null pairs (50 tables) p ≤ 0.05 p ≤ 0.01 Pairs passing q ≤ 0.05 Tables with any
plaque rho 322,050 5.00% 1.03% 5 5 of 50
plaque spearman 322,050 4.86% 0.95% 4 2 of 50
tongue rho 363,000 4.90% 1.00% 5 4 of 50
tongue spearman 363,000 4.88% 0.97% 4 3 of 50

With every pair null, Benjamini-Hochberg at 0.05 lets through at least one false pair in at most about 5% of tables. Here it was 4–10% of tables (2 to 5 of 50; with 50 tables the binomial range is wide), with one or two false pairs when there were any. The real tables, at 28–56% passing, are nowhere near this null. The pass rate measures real structure in the data, not a broken test.

On real oral profiles, cooccur doesn’t find the cultured host. The control pair ranks near the bottom, and the genus-level association is significantly negative. Some reasons, none of them tested here:

  • The relationship is not a proportional one. TM7x is a parasite. In culture, permissive hosts grow, crash and recover after infection (Utter et al. 2020), and the host evolves lower susceptibility within weeks (Bor et al. 2018). Such dynamics need not leave a positive correlation in one snapshot per person.
  • Hosts are a few strains, and the profile is a species or genus. Of 27 Actinomyces strains, 13 hosted TM7x. A MetaPhlAn species sums permissive and non-permissive strains, and the Saccharibacteria profiles mix several Saccharibacteria. A host signal could be diluted by both.
  • Habitat dominates. The strongest axis in plaque runs from Actinomyces/Rothia-rich plaque to plaque rich in anaerobes. The Saccharibacteria profiles sit with the anaerobes, Actinomyces on the other side (interpretation of the partner lists). Co-occurrence measures shared habitat first (Faust & Raes 2012).
  • The CLR reference follows the zeros. In the plaque table, the per-sample mean log, which every CLR value is measured against, had a Spearman correlation of −0.95 with the number of zeros in the sample. Zero counts ranged from 2 to 113 of 114 species. One sample was 100% Rothia dentocariosa. The zeros track sequencing depth (Spearman −0.47 with read count) and community type. TM7c is absent mostly from low-diversity samples (Spearman −0.64 between its abundance and the sample’s zero count); A. odontolyticus is present almost everywhere (−0.10). That turns a near-zero association of the percentages (+0.05) into a negative CLR one (−0.25). Dropping the 17 samples with ≥ 40 zeros, or keeping only species in ≥ 80% of samples, weakened the control pair to rho −0.17 to −0.21, no longer passing. It was still not positive.

What the test doesn’t cover.

  • Association, not interaction, as the tool page says. A positive pair would not have proven a host either.
  • One body site at a time, one cohort (healthy US adults, 2009–2011), one or two samples per person. No time series, so lagged parasite–host dynamics are invisible.
  • Profiles, not MAGs. MetaPhlAn2 has no TM7x entry. Its two Saccharibacteria species are 2007 single-cell genomes, and its A. odontolyticus is a species, not the XH001 strain. A MAG-based table (all genomes mapped across the same samples, as in Shaiber et al. 2020 for TM7 only) would test strains, at the cost of mapping 400+ metagenomes.
  • MetaPhlAn relative abundances are already relative, and cooccur’s CLR handles that: rho is the same whether a sample sums to 100% or not. What the CLR cannot fix is zeros. A MetaPhlAn zero means “no marker reads”, and how often that happens depends on depth, so the pseudocount carries depth into every CLR value. n_both shows how many samples each pair rests on (TM7c – A. odontolyticus: 101 of 128).

No bug was found, and cooccur’s code was not changed. Two things came up:

  • q printed as 0.0500 for pairs that don’t pass. With --all-pairs, q has four significant digits, so a pair at q = 0.05003 shows as 0.0500 and is not in the pass count (e.g. 1,762 rows show q ≤ 0.0500 where the summary says 1,761 pass). A pass column, or more digits, would make the file unambiguous.
  • A zero diagnostic would help. On real profiles, the zeros dominated the CLR reference. A summary line with the per-sample zero range, and a warning when it is wide, would flag this. Also --min-prevalence 0.8 as a suggested setting for MetaPhlAn tables.

Kept in bioinfo-lab/cooccur-real/ (not in the repo): oral_mpa2.tsv.gz (the 415-sample MetaPhlAn2 table as TSV) and oral_meta.tsv (sample table), both from EH2491 (Artistic-2.0); the per-subsite inputs, taxonomies, --all-pairs edge lists, summaries and logs of every run above; and the scripts (eset2tsv.py, make_inputs.py, rank.py, summarise.py, check_pair.py, shuffle.py).