My recent snark-laden post about the microbiome was in part inspired by the coverage of this Science paper “Intestinal Inflammation Targets Cancer-Inducing Activity of the Microbiota.” Interestingly, the coverage of this paper didn’t make Jonathan Eisen happy either:
Some studies of course do an OK job of trying to test whether observations are correlations or have some causative connection. One seemingly well done case from the scientific publication point of view involves a recent paper on gut microbes and colon cancer: Gut Microbes Implicated in the Development of Colorectal Cancer. Alas it is not an open access paper so what most people out there have to go on is the press coverage of the work. The paper itself is quite interesting and the authors do a pretty good job of discussing how they went about testing the roles of specific microbes and even specific genes in the etiology of disease in a mouse model.
The press coverage has not been so clear alas. Some examples of the press coverage are below:
- Health News – Genes carried by E.Coli bacteria linked to colon cancer
- E. coli strain linked to cancer in mice
- Clear Links Found Between Inflammation, Bacterial Communities and Cancer
- Gut Bug Links Inflammation to Colorectal Cancer
- Inflammation, Bacteria and Colorectal Cancer
In many of the stories the key distinction between correlation vs. causation is nowhere to be found. Of course, I don’t always expect the press to cover such distinctions, but the more this is discussed in blogs, press stories the better off we all will be.
If you read the titles of the stories versus what’s reported in the article, there is quite a big gap. So I’m going to summarize the results as best I understand them. Here’s what they found (note: I’m trying to reduce the use of non-technical language and strip things down to their essentials for non-specialists; if you’re an expert, don’t be pedantic in the comments):
1) Germ-free mutant mice prone to forming tumors and developing colitis under the experimental conditions have a different gut microbiota than normal non-germ-free mice (germ-free are raised from birth in a sterile environment).
2) Germ-free mutant mice prone to forming tumors and developing colitis have a 1000-fold increase in the bacterium E. coli.
3) Germ-free mutant mice prone to forming tumors and developing colitis exposed to either Enterococcus faecalis or E. coli strain NC101 develop severe colitis, but only the mice exposed to E. coli NC101 routinely developed tumors.
4) A group of E. coli, known as the B2 phylogroup (a lineage of E. coli), to which E. coli NC101 belongs, produce colibactin (colibactin damages DNA, potentially causing tumors). More about these E. coli later.
5) 5/24 people (not mice) who did not have either IBD (inflammatory bowel disease) or colorectal cancer had colibactin-producing E. coli, while 14/35 IBD and 14/21 patients with colorectal cancer had colibactin-producing E. coli.
6) A variant of E. coli NC101 which lacks the ability to produce colibactin is unable to damage the DNA of a laboratory line of rat epithelial cells.
7) The absence of colibactin production in E. coli in germ-free mutant mice prone to forming tumors and developing colitis appears to reduce tumor formation, although colonic inflammation is unaltered. It is worth noting the normal mice seem perfectly fine: only mice prone to inflammation (the mutants) seem to run into trouble.
This is an impressive study. But is this really a link between E. coli and colon cancer? Come on. It’s important because the study indicates a phenomenon we need to explore further, ultimately in people. Yet there is a long way to go from these data to making causative statements about disease (as Eisen notes).
I also want to raise a couple points about the E. coli side of things too. B2 phylogroup E. coli are quite common, usually occurring in 20-30% of healthy people; they are also disproportionately associated with urinary tract infections and other ‘extraintestinal’ infection (i.e., not diarrhea). What’s interesting is that if you look for the colibactin production genes in commensal B2 E. coli (those not isolated from sick people–in some cases, not even from humans), almost all of them have these genes, and no other non-B2 commensals do (that I’ve examined so far; obviously, ‘no other’ is a very strong statement). That doesn’t invalidate their model, but there are a number of genes that are very common in and exclusive to B2 E. coli; those genes could also be responsible for these mechanisms. It should also be noted that other genotoxins are found in many E. coli, but, as the authors note, they are not found in E. coli NC101. What they observe in the study might not be a colibactin-specific or B2-specific phenomenon at all.
The other issue is that it’s not clear how this translates into healthy humans. I realize that’s often a generic criticism, but normal mice aren’t affected by E. coli NC101. Lots of healthy people carry B2 E. coli. People will also have a mixture of E. coli at various times, and E. coli are, in the long term, transient–this isn’t like Helicobacter pylori where strains will persist for years. It remains to be seen how long an association is required in a complex community, as opposed to “mono-associated mice” for these processes to occur (if they do at all). In short, this study highlights a potential mechanism underlying colorectal cancer, but we have no idea how potent this mechanism is in humans (doesn’t seem to have much effect in normal mice). As Eisen put it, “there is no evidence that the colibactin producing bacteria cause cancer.”
Finally, there’s one result that doesn’t make sense to me–the statistics don’t seem correct:
Although 5 of the 24 (20.8%) non-IBD/non-CRC controls [MB: ‘normal’ mice] harbored pks+ E. coli [MB: colibactin-producers], the genotoxic island was detected in 14 of 35 (40%, P < 0.05) IBD patients and in 14 of 21 (66.7%, P < 0.001) CRC patients (Fig. 3B and table S3).
I must be missing something obvious, because if you use any of the basic statistical tests (binomial sign, Fisher’s Exact, Chi-Square or Pearson’s tests), the IBD patients don’t have a significantly higher frequency of colibactin-producers (5 vs. 19 for normal; 14 vs. 21 for IBD). Anyone want to run the numbers and show me what I’m not getting?
For the non-scientists, it’s worth actually realizing what studies like these do well: illuminating potential mechanisms. This work is good science. The reason I went through this exercise is not to tear down this article, but because, like Eisen, I’m concerned by what I’ll call ‘microbiome creep’, where claims in the popular coverage keep escalating (hell, give it few months, and we will probably start to see people claim E. coli causes breast cancer). When we (and I include science communicators in “we”) act as if these kinds of studies concretely explain natural phenomena, we are stretching good studies too far. Inevitably, even among scientists who are skimming news reports, many readers will think, “Bacteria can cause cancer.” This then leads to unrealistic expectations (e.g., even if this mechanism does work, it could be very weak, so it’s not where you intervene medically). That’s not a good thing, especially for younger researchers just getting their feet wet in this field, and who could get hit with a nasty backlash when ‘we haven’t cured colorectal cancer.’
For the scientists, remember than even though science journalists are sympathetic to scientists (certainly relative to political reporters and politicians), their interests and research scientists’ interests do not always overlap. We don’t want to undersell things, but overplaying results won’t help us either.
Cited article: Arthur JC, Perez-Chanona E, Mühlbauer M, Tomkovich S, Uronis JM, Fan TJ, Campbell BJ, Abujamel T, Dogan B, Rogers AB, Rhodes JM, Stintzi A, Simpson KW, Hansen JJ, Keku TO, Fodor AA, Jobin C. 2012. Intestinal Inflammation Targets Cancer-Inducing Activity of the Microbiota. Science DOI: 10.1126/science.1224820This entry was posted in E. coli, Microbiology, Microbiome, News Media. Bookmark the permalink.
I don’t think you’re missing anything obvious, except that the legend for figure 3 says they used the binomial test. This is the wrong test to use if the question is whether there is a differing distribution of counts for control vs CRC or control vs IBD; as you know Fisher would be the obvious test there. Fisher(14,7,5,19) for CRC has P = 0.0027, while Fisher(14,21,5,19) has P = 0.16. Binomial(14, 21) and Binomial(14, 35) are not significantly different from chance anyway. It’s quite unclear to me what they meant to do here.
Thanks for doublechecking my math; I was also surprised they didn’t use Fisher’s.
It seems to me that this experiment doesn’t rise to “illuminating potential mechanisms” because then you would be able to point in the direction of the mechanism. It seems this paper – and it is a fascinating result – is more like a proximity detector for a mechanism. There is clearly at least one mechanism nearby, but where is it and how many of them are actually nearby……?