Life · at the detection floor

The Microbiome in the Blank

In a 2016 study of six placentas, reads within 3% of one DNA sequence, a sequence identical to a stretch of C. difficile 16S, made up 63% to 78% of the reads from every placental biopsy extracted with one kit, and 66% to 71% of the reads from every extraction blank, a tube that held only the kit's reagents. With the study's other kit none of 358,808 reads came that close. This page recounts the deposited reads, then tries the two obvious ways to tell placenta from blank, subtracting the blanks and demanding agreement between kits, and shows what each can and cannot say.

Twenty-eight tubes

In 2016 Abigail Lauder and colleagues at the Perelman School of Medicine at the University of Pennsylvania looked for bacterial DNA inside placentas from six uncomplicated deliveries at term, beside control tubes that held no tissue. Biopsies from the fetal and the maternal side of each placenta were extracted with each of two DNA extraction kits, the MO BIO PowerSoil kit and the PSP Spin Stool DNA Plus kit (each biopsy went whole into one kit, so the two kits worked on different pieces of the same side), and every tube was sequenced for a stretch of the bacterial 16S rRNA gene.

Below are the 28 tubes from the MO BIO kit, unlabelled and shuffled: 12 placental biopsies and 16 controls. Each tube is filled with its sequencing reads, sorted into clusters of near-identical sequences; the largest clusters each have their own colour, the same in every tube. Mark the ones you think hold placenta, then reveal.

In 27 of the 28 tubes the largest layer is the same cluster, and in 26 it holds more than half of all the reads: 63.2% to 77.9% in every placental biopsy, and 65.8% to 71.1% in every extraction blank. The one tube where it is not the largest is unopened sterile swab 4, where it holds 6.5% and another cluster holds 40.6%. The other tube where it is under half is air swab 5, at 33.9%.

Switch to the PSP kit and that cluster is gone: none of the 358,808 reads from the 23 placental and control tubes extracted with that kit falls in it. A different pair of clusters leads instead: one of the two is the largest layer in all 11 PSP controls and in 10 of the 12 PSP placental tubes. The other 2, both maternal-side biopsies, are each led by another cluster, and both come back below. The paper names this kit's background: For samples purified using the PSP kit, all placental samples and background controls contained Sediminibacterium, Bradyrhizobiaceae, Methylobacterium, and Propionibacterium.

The printed count and the deposit

The paper gives the MO BIO result in two sentences:

For DNA purified from the MO BIO kit, the dominant lineage in the controls was an OTU aligning 100 % to Clostridium difficile. This OTU was predominant in 22 out of 23 samples, including fetal side and maternal side placenta, extraction blanks, unopened sterile swabs, and swabs exposed to air in the clinical laboratory.

Lauder and others, Microbiome, 2016, Results

The deposit holds more tubes than that sentence counts. The reads are in NCBI's Sequence Read Archive as project PRJNA309332, and its sample records list 28 MO BIO placental and control samples: 6 extraction blanks, 5 unopened sterile swabs, 5 air swabs, 6 fetal-side and 6 maternal-side biopsies. The paper's own Figure 2b draws the same 28 columns, and in it sterile swab 4 is the one column that is red in the row labelled Cyanobacteria o__Streptophyta.

That the deposit is the paper's data can be checked another way. Its run read counts average 25,044.6 per MO BIO run (851,516 over 34 runs) and 59,197.9 per PSP run (2,071,925 over 35 runs), and the paper prints an average of 59,198 (PSP extraction) and 25,044 (MO BIO extraction) 16S rRNA gene reads.

Recounted from those reads, the cluster is the largest in 27 of the 28 tubes and more than half of the reads in 26. The tube where it is not the largest, sterile swab 4, fits the last sentence of the same paragraph: Additionally, one sample was high in a lineage of the order Streptophyta, possibly a sequence derived from chloroplasts in pollen. This page does not guess why the printed count is 22 out of 23.

What the layer is, and is not

The cluster is founded by its most frequent read, a sequence 297 bases long. Reads identical to it are 25.3% to 32.3% of the reads from each MO BIO placental biopsy and 25.3% to 28.6% of those from each extraction blank; the rest of the cluster is near-copies, within 3% of it. All 297 bases match RefSeq record NR_113132.1, part of the 16S ribosomal RNA gene of Clostridioides difficile strain ATCC 9689, starting at its base 21, and record NR_112172.1 from the same strain, starting at its base 13. Of the 47 RefSeq 16S records that NCBI files under the family Peptostreptococcaceae, none but the two of C. difficile comes closer than 21 edits in 297 (92.9% identity); the nearest is Romboutsia ilealis, NR_125597.1.

That is a match of DNA sequence, not a finding of living bacteria: 297 bases of one gene. The authors were careful about where it came from:

We have previously amplified C. difficile 16S rRNA gene sequences in our laboratory, so we cannot rule out that the low-level contamination originated from this source. However, as C. difficile only appeared in negative controls worked up with the MO BIO kit, we tentatively associate this contamination with the kit.

Lauder and others, 2016

So the safe statement is the narrow one: a sequence identical to C. difficile 16S appeared only in samples extracted with that kit. In this deposit it is in all 28 MO BIO placental and control tubes, and in all 6 MO BIO saliva samples at 3 to 268 reads; in the 29 PSP runs recounted here it has no reads at all. Three more clusters hold shorter copies of the same stretch, 255 to 282 bases long; the 97% rule keeps them apart because their lengths differ by more than 3%. Adding them in changes neither the 27 nor the 26, and none of them occurs in a PSP run.

Just subtract the blanks

The obvious reply is that everyone knows kits carry DNA: sequence the blanks, remove whatever they contain, and what is left is the placenta. Try it on every tube, the controls included. For each tube, the instrument removes every cluster found in any other control from the same kit, and reports the share of reads that is left. A read whose sequence occurs only once in the whole study and joins no cluster counts in its tube's total but is never counted as left.

With the defaults, where a single read in another control is enough to remove a cluster, the MO BIO placentas keep 1.2% to 6.0% of their reads and its controls keep 1.6% to 42.1%: extraction blanks 1.6% to 5.8%, air swabs 1.7% to 14.1%, sterile swabs 1.9% to 42.1%. A blank with every other control subtracted is not empty, and the placentas' residues are the size of the blanks'. The authors saw the same thing: apparently enriched lineages could be found readily in comparisons between contamination data sets.

The PSP kit is less tidy, and the page shows it. Its controls keep 1.3% to 8.2%, but its placentas keep 3.0% to 26.4%, and three maternal-side biopsies keep 14.2%, 20.6% and 26.4%. That is the pattern a subtraction would call placental. The paper also reports maternal-side samples high in lineages of their own, looked for those in the same women's saliva and vaginal samples, and wrote: Thus, we were unable to trace the origin of these outlier lineages to specific body sites in our paired samples. A residue says that a tube holds something the other controls did not catch. It cannot say where that came from, and the next test can say more.

The test that can say yes

The authors' own test starts from a sound idea: We reasoned that authentic microbiota should be mostly reproducible regardless of the DNA extraction kit used and so focused on OTUs recovered in common using both extraction kits from placental or saliva samples (Fig. 4). They ran the same test on saliva, where bacteria certainly live; that is what lets an empty result for placenta mean something. Here a cluster passes at a threshold T when it reaches T in both kits' extractions of the same woman's tissue, or of her saliva, and stays below T in every one of the 27 controls of both kits. The two halves of the test are not quite alike: for saliva, each kit took a portion of the same sample, while for placenta each took a different biopsy, and among its limits the paper writes: If there is patchy microbial colonization, it might have gone undetected.

At the paper's threshold of 1%, saliva passes 25 clusters and placenta none. The paper reports the same shape: There were no placental OTUs that achieved at least 1 % abundance in both PSP and MOBIO in at least one sample and not in controls, In contrast, the same filtering procedure recovered 28 OTUs from salivary samples. At 0.5% saliva passes 31 and placenta 1; at 2%, 19 and 0. Lower the threshold and placental clusters appear: 2 at 0.1%, 11 at 0.01%, 18 at 0.001%. The lower chart shows why. Each bar is one placental cluster in one woman, drawn from its highest share in any control to the smaller of its two kit shares, and it passes at any threshold inside its bar. 29 of the 31 bars end below 0.2%.

The closest call at 1% is one cluster in one woman's maternal-side biopsy: 3.7% of the PSP reads and 0.62% of the MO BIO reads, while no control holds more than 0.37% of it. It passes at any threshold above 0.37% and up to 0.62%, so it counts at 0.5% and not at 1%. The page shows it rather than picking a threshold that hides it, and does not call it placental: one woman, one cluster, six times more abundant in one kit than in the other.

The largest share any other cluster reaches in a placental tube fails the other way. That cluster holds 66.0% of the PSP reads from one woman's maternal-side biopsy and 3.1% of the MO BIO reads from a replicate biopsy of the same side, so it clears 1% in both kits. It fails because a MO BIO extraction blank holds 3.6% of it. That share belongs to this recount's clusters: where a cluster's edges fall decides how much of it a control holds.

At the lowest threshold the paper's own counts disagree with each other. Its Results say 16 in the maternal side placental samples, and 5 in the fetal side placental samples. Two OTUs appeared in both fetal and maternal side samples, leaving 19 unique placental OTUs. Its Discussion says we found 18 OTUs reproducible between extractions for the maternal side sample and 5 for the fetal side sample, and its Table S3 lists 21 clusters, 18 maternal-side and 5 fetal-side with 2 in both. This recount finds 18 at 0.001%: 15 maternal-side and 5 fetal-side, 2 in both. The authors read their low-level clusters with care: Some of these extremely low-abundance lineages may represent sporadic, rare placental microbiota. However, at this low level of abundance, noise is abundant in the data—several unclassified OTUs apparently arose from mispriming on human DNA.

A statistical filter

In 2018 Nicole Davis and colleagues, who wrote the decontam software, ran it on this dataset. Its prevalence method rests on the observation that sequences from contaminating taxa are likely to have higher prevalence in control samples than in true samples, and they used amplicon sequence variants (ASVs), inferred by DADA2, rather than 97% clusters. They reported:

We found that six of the 810 ASVs present in at least five samples were identified as non-contaminants at an FDR threshold of 0.5 (Table 1). Five of those six ASVs matched Homo sapiens rather than any bacterial taxa.

Davis and others, Microbiome, 2018

The sixth was a Ruminococcaceae variant, a known member of human gut microbial communities, and they were unable to establish its ground truth. This page reports their result as they computed it; it does not rerun decontam.

One cluster in this recount does set the placentas apart: it is in 18 of the 24 placental extractions and in none of the 27 controls, at most 0.64% of a tube's reads. This page does not identify it. Of its own clusters found in more than one placenta, the Lauder paper wrote: The three OTUs found in multiple placentas could not be assigned to any bacterial taxa, and further analysis showed the sequences were 100 % identical to human chromosomal sequences.

Why the blank wins

In 2014 Susannah Salter and colleagues made the mechanism visible. They grew a pure culture of Salmonella bongori, a species picked because we have not observed it as a contaminant in any of our previous studies, diluted it tenfold five times, and had Imperial College London, the University of Birmingham and the Wellcome Trust Sanger Institute each extract every dilution and amplify it, with different production batches of one kind of kit. They estimate about 10⁸ cells went into the undiluted extraction and about 10³ into the fifth dilution.

Recounted here from their deposited reads, the share classified as S. bongori falls from 95.3%, 95.2% and 94.6% in the three undiluted cultures to 3.9%, 19.8% and 6.3% at the fifth dilution (Imperial, Birmingham, Sanger). The undiluted cultures fall short of 100% mostly because of the 8-edit limit: in the three of them, 4.0%, 4.0% and 4.8% of the reads have first 150 bases more than 8 edits from the S. bongori record. The rule also asks that a read be nearer to that record than to any of 122 other Enterobacteriaceae records, which sets aside another 0.6% to 0.7%. The fall is not smooth: at the Sanger Institute the second dilution sits below the third, 45.3% against 63.1%. Salter's own summary: S. bongori was the sole organism identified in the original undiluted culture but with subsequent dilutions a range of contaminating bacterial groups increased in relative abundance while the proportion of S. bongori reads concurrently decreased (Figure 1).

The curve is the simplest mixture, as the decontam paper sets it out: Let total sample DNA (T) be a mixture of two components (T = C + S): contaminating DNA (C) present in uniform concentration across samples and true sample DNA (S) present in varying concentration across samples. The sample's share of the reads is then S/(S + C), half when S equals C. The points follow it only roughly; no single C fits all three laboratories, or every dilution of one.

A placental biopsy sits at the low end of that axis. By quantitative PCR, the Lauder study found that placental and control samples showed low and indistinguishable numbers of 16S rRNA gene copies. Nor is this one brand's failing: the 2014 study tested DNA extraction kits from four different manufacturers, and found that Regardless of kit, contamination was always the predominant feature of the sequence data by the fourth serial dilution.

Is there a placental microbiome?

This page reports a lesson in measurement; it does not settle the biology. In 2014 Aagaard and colleagues, working from placental specimens of 320 subjects, wrote: We characterized a unique placental microbiome niche, composed of nonpathogenic commensal microbiota from the Firmicutes, Tenericutes, Proteobacteria, Bacteroidetes, and Fusobacteria phyla. The Lauder paper tested that idea, and its own limits paragraph says: Our data do not rule out the existence of placental microbiota—we can only conclude that we could not use our data to distinguish placental samples from contamination controls.

In 2019 de Goffau and colleagues concluded that the human placenta does not have a microbiome. Of the bacterial DNA signals they did detect, they wrote: Almost all signals were related either to the acquisition of bacteria during labour and delivery, or to contamination of laboratory reagents with bacterial DNA. The exception was Streptococcus agalactiae (group B Streptococcus), for which non-contaminant signals were detected in approximately 5% of samples collected before the onset of labour. In 2023 a review by Kennedy and colleagues began: Whether the human fetus and the prenatal intrauterine environment (amniotic fluid and placenta) are stably colonized by microbial communities in a healthy pregnancy remains a subject of debate. Its own reading: Our analysis indicates that the detected microbial signals are likely the result of contamination during the clinical procedures to obtain fetal samples or during DNA extraction and DNA sequencing.

What the reads above do establish is narrower. In this study, the cluster that fills a MO BIO tube that held nothing fills every MO BIO placental tube too, and the test that saliva passes at 1%, placenta does not.

curl -sO https://artwaste.land/checks/verify-the-microbiome-in-the-blank.mjs
node verify-the-microbiome-in-the-blank.mjs