The contamination you cannot see in the flask

There is a category of laboratory problem I find more interesting than outright error, and it is the problem that produces clean-looking data. A failed experiment tells you it failed. A contaminated one hands you numbers, plots them nicely, and lets you write them up.

Mycoplasma is the canonical example. It is a genus of very small bacteria that colonize cell cultures, and the reason it matters is not that it is exotic. It is that essentially every property that makes a contamination noticeable is a property mycoplasma lacks.

Three reasons nobody sees it

Start with the visual check, which is what most bench work actually relies on. Anthony Olarerin-George and John Hogenesch, then at the University of Pennsylvania, put it directly: mycoplasmas "are able to reach high concentrations in the media of infected cells without noticeable turbidity." No cloudiness. The flask looks like every other flask on the shelf.

Second, the antibiotics. Cell culture medium routinely carries penicillin and streptomycin, and researchers reasonably treat that as background protection. It is not, for this organism:

mycoplasmas lack cell walls. This makes them impervious to cell culture antibiotics that inhibit cell wall synthesis, such as penicillin

Penicillin works by disrupting cell wall synthesis. An organism with no cell wall is not a hard target for it, it is simply not a target at all. The routine antibiotic in the medium is, against mycoplasma, an inert additive.

Third, filtration. The same paper notes that mycoplasma cells are "small (0.3–0.8 µM in diameter) and pleomorphic, allowing them to pass through standard filtration membranes." Pleomorphic means they have no fixed shape, which is what makes size alone a poor guide to what a membrane will stop.

So: not visible, not killed by the standing antibiotic, not reliably filtered out. Three independent safety nets, none of which is catching it. Uphoff and Drexler, writing for the DSMZ, the German national collection of microorganisms and cell cultures, name the consequence in one clause: mycoplasmas matter "because they do not conspicuously overgrow the human or animal cell cultures and can only be detected applying special assays."

"Only ... applying special assays" is the whole story. If you are not running a test specifically for this, you are not screening for it. You are assuming.

The number, and where it came from

Prevalence estimates for mycoplasma have circulated for decades, and they mostly came from sentinel testing: a cell bank or a testing service reports what fraction of the samples sent to it were positive. That is a real number, but it is a number about samples people chose to send.

The 2015 survey took a different route. Olarerin-George and Hogenesch analyzed sequence data from 9,395 rodent and primate samples across 884 series in NCBI's Sequence Read Archive, that is, RNA-seq data that hundreds of labs had generated for entirely unrelated purposes and deposited publicly. They then looked for reads that mapped to mycoplasma genomes. Their result:

We found 11% of these series were contaminated (defined as ≥100 reads/million mapping to mycoplasma in one or more samples).

Be precise about what that is. It is 11 percent of series, at a stated threshold, in a sample of rodent and primate RNA-seq deposited over roughly 2012 to 2013. It is not "11 percent of all cell culture," and the threshold is a choice the authors made and disclosed. What makes it valuable is that it is unchosen from the labs' side: nobody in those 884 series submitted their data to be screened for contamination. The evidence was simply already in what they had published.

For context, the same paper's account of the older sentinel record: by the early 1990s the FDA had tested over 20,000 cell cultures and found 15 percent contaminated, and a 2002 report found 28 percent of 440 cell lines tested were contaminated. The DSMZ article, drawing on its own testing, breaks it down by culture type, at roughly 1 percent for primary cultures, 5 percent for early passage cultures, and 15 to 35 percent for continuous cell lines. That gradient is itself informative: contamination rises with passage number, which means it is not coming from the donor tissue. It is being introduced in the lab, during propagation, and the most common source in the DSMZ accounting is other contaminated cultures.

What it actually does to your results

If mycoplasma simply sat in the medium doing nothing, none of this would matter much. It does not sit still.

The mechanism most worth understanding is nutrient competition. Mycoplasmas consume amino acids, nucleic acid precursors, vitamins and lipids from the medium, and the arginine-hydrolyzing species produce ammonia as they do it. The DSMZ authors note that ammonia is "a highly toxic agent inhibiting cell growth," and that arginine depletion has been linked to inhibited proliferation, induced apoptosis, and possibly to chromosomal aberrations, since arginine is a major component of histones.

That is the descriptive account. The controlled version arrived in 2019, when a group at the FDA's Center for Drug Evaluation and Research deliberately contaminated CHO cell bioreactors with Mycoplasma arginini and watched what happened. Arginine concentrations dropped by more than 90 percent as the mycoplasma population peaked. Ammonia accumulated to 25 to 30 mM, against a CHO toxicity threshold the authors give as around 4 mM. Cell-specific IgG1 productivity started near average and fell to about 10 percent of average by day seven after contamination.

And the detail that ties it back to the visibility problem: they observed no changes in CHO cell health or process conditions until at least two days after inoculation. There is a window in which the culture is contaminated, the metabolism has already turned, and every readout still looks normal.

Back on the gene expression side, the 2015 survey looked at a single-cell RNA-seq dataset and found 61 host genes significantly associated with mycoplasma read counts at P < 0.001. Sixty-one genes is not a catastrophe and it is not nothing. It is a systematic shift in the measured transcriptome that correlates with an organism the experimenters did not know was present, in the same direction across the affected samples, which is exactly the shape of a confound that survives statistical review.

The finding I keep coming back to

One result in the 2015 paper is a small methodological gem. Ninety percent of the mycoplasma-mapped reads aligned to ribosomal RNA sequences. That is unremarkable on its own, bacterial RNA is dominated by rRNA. What is remarkable is the next sentence: 37 percent of the contaminated series had used poly(A) selection to enrich for messenger RNA.

Poly(A) selection is supposed to pull down polyadenylated transcripts and leave bacterial rRNA behind. It did not, at least not completely. The contaminating signal came through a step whose stated purpose was to exclude it.

I work in microwave spectroscopy, so my instinct here is a spectroscopist's instinct, and it generalizes: every selection or cleanup step in a pipeline has a rejection ratio, not a rejection guarantee. We describe these steps by what they are for, and then we reason as though they achieved it perfectly. A filter that is 99 percent efficient against a signal that is 10,000 times larger than your analyte still leaves you with a contribution the size of your analyte. The poly(A) result is that arithmetic showing up in a sequencing library.

There is an upside to it, and the authors used it. Because the contamination survives into the data, the data can be screened retrospectively. Any lab with archived RNA-seq can go back and check cultures that were discarded years ago.

What actually detects it

Two assays are recognized in the European Pharmacopoeia, according to the DSMZ account: microbiological culture on mycoplasma-specific medium, where positives form colonies with a characteristic "fried eggs" appearance, and DNA fluorochrome staining with DAPI or Hoechst 33258. Both work. Both have caveats the authors state plainly, that some M. hyorhinis strains grow poorly or not at all on the culture media, and that fluorochrome staining results "are sometimes difficult to interprete and some experience is definitely necessary," with misinterpretation frequent when the culture is not in good condition.

PCR against 16S rRNA sequences is the modern workhorse, and the DSMZ authors describe it as easy, sensitive, specific, fast, reliable and cost effective, with the design caveat that primers must be broad enough to catch Acholeplasma as well as Mycoplasma while staying narrow enough to exclude other bacteria.

Worth pairing with the FDA bioreactor result: those authors note that conventional mycoplasma assays require 14 to 28 days, which is why real-time detection during a production run is not feasible with the standard methods. In a research setting the equivalent problem is that a quarterly test tells you about the quarter, not about the experiment you ran on a Tuesday.

The signal

Three things to carry out of this.

First, "the culture looked fine" is not a contamination test, and neither is the penicillin in your medium. Both are widely treated as informal evidence of cleanliness, and against this organism neither carries any information at all. If a paper's methods do not state that cell lines were tested for mycoplasma, the correct reading is that the question was not asked, not that the answer was no.

Second, when you are told that a preparation step removes something, ask for its efficiency rather than its purpose. Poly(A) selection is for excluding non-polyadenylated RNA, and it let bacterial rRNA through in 37 percent of the affected series. The same reasoning applies to washes, filters, gates, background subtraction and every other "we removed that" step in a protocol. Purpose is a design intent. Efficiency is a number, and if nobody measured it for your conditions, it is an assumption.

Third, notice the shape of how this was found. The strongest evidence about a widespread laboratory problem came from re-reading data that had been collected for other reasons and deposited in public. That is the same move as Report 050 on misidentified cell lines and Report 074 on what reproducibility actually means: the archive knows things the individual papers did not report. Deposited data is not just a courtesy to the next researcher. It is the audit trail, and occasionally it audits us.