In February, Nature published a correction to a paper about coffee. The correction says that in the results section and in the tables, decaffeinated coffee was incorrectly referred to as caffeinated coffee. The paper's entire contribution is that different coffee preparations do different things.
I went looking for this one for a specific reason. It is the paper Nature's own recommendation engine puts next to the Cork gut microbiome study I wrote about earlier, and it turns out to land on the same bacterium by a completely different route. That is worth something. The correction is worth something too, though not what the authors intended.
The setup: a group at Peking University's School of Nursing ran a Mendelian randomization analysis asking whether coffee lowers glycated hemoglobin (HbA1c, the three month average of blood sugar) by way of the gut bacteria that make propionic acid. Mendelian randomization uses inherited gene variants as a stand in for a lifetime of exposure. Because your genotype is fixed at conception and shuffled at random, it sidesteps the healthy user problem that haunts every coffee cohort study. The American Heart Association leaned on this design in its statement last month, so it is already in the family.
What they actually did. No new participants. They combined two existing genome wide association datasets: the UK Biobank, for how people say they take their coffee, and the MiBioGen consortium, for the abundance of gut bacterial genera. They asked whether the genetic tendency toward a given coffee preference travels with the genetic tendency toward four propionic acid producing genera, and whether those in turn travel with HbA1c. Inverse variance weighting was the primary method, with the usual sensitivity analyses behind it.
Filtered coffee tracked with more Veillonella (beta 0.26, 95% CI 0.03 to 0.49, p = 0.03). More Veillonella tracked with lower HbA1c (beta minus 0.05, 95% CI minus 0.10 to 0.00). The indirect path came out at minus 0.013, and the direct effect of filtered coffee survived at an odds ratio of 0.97 (95% CI 0.943 to 0.99, p = 0.04). From those two paths they compute that Veillonella carries 43.33% of the total effect.
Where it thins outEvery number in that chain sits close to the edge. The first path's confidence interval nearly touches zero at the bottom. The second one touches zero at the top. Multiply two borderline estimates together and you get a third estimate that is more borderline than either, then report it to two decimal places. Microbiome Mendelian randomization also carries a structural weakness the paper does not raise: the genetic instruments for bacterial abundance are weak, because human genotype explains only a small fraction of who lives in your gut. Diet, drugs, and geography do most of the work, and none of them are randomized here.
The abstract offers unsweetened filter coffee as "a dietary intervention to improve glycaemic control" and closes on "practical guidance for diabetes management, emphasising the importance of coffee preparation methods."
What the design supportsMendelian randomization estimates the effect of a lifetime of genetically nudged exposure. It is a good tool for asking whether an association is likely to be causal. It is a poor tool for telling a 58 year old with an HbA1c of 7.4 what switching to a pour over will do for him, because nobody in this analysis switched to anything. The authors concede in the limitations that they built no dose response model at all, which is the first thing you would need before offering guidance. They also note the underlying data is almost entirely of European ancestry, and flag that this may hide metabolic patterns specific to other populations. That caveat matters more than usual here, since the outcome is diabetes.
The Cork feeding study, published four months after this one, put 62 people through two weeks without coffee and three weeks back on it. The genus that bloomed when coffee came back, on day 21, in both the caffeinated and the decaf arms, was Veillonella. Neither group of authors cites the other. One study is arithmetic on European biobanks, the other is stool samples from 31 people in Ireland, and they arrive at the same genus. In a field where most findings are a single cohort with a single questionnaire, two designs that share no data and no participants pointing at the same organism is the strongest signal in this whole corner of the newsletter.
Do not oversell itConvergence is not confirmation. Both studies are looking for bacteria that respond to coffee, and Veillonella is a lactate eater that turns up whenever fermentable substrate arrives, so it is a plausible finding and also a usual suspect. The Cork study found it rose in both coffee arms, which fits the polyphenol story rather than the caffeine one. This paper's claim is about filtered coffee specifically. Those are compatible but they are not the same claim, and no one has yet measured propionic acid and HbA1c in the same people.
Published 20 February 2026. Two items. In the results section and the tables, decaffeinated coffee was incorrectly referred to as caffeinated coffee. And the confidence intervals reported for both mediation paths were wrong. Both are now fixed in the online version, which is where the numbers above come from.
Why it is worth a card of its ownFor a decade the central question in coffee research has been which effects belong to caffeine and which belong to the several hundred other compounds in the cup. It is the question the Cork study was built around. It is the question the AHA statement answers organ by organ. It is the axis I sort every coffee paper on. And here is a paper in the Nature family, on precisely that axis, that had the two arms labelled backwards through peer review, production, and eight weeks of publication. The lesson is not that the paper is worthless. The lesson is that the distinction everyone in this field treats as obvious is easy to lose track of, even for the people whose results depend on it. Read the arm labels. Then read them again.
Three things, in descending order of confidence.
It puts a candidate mechanism under a finding I already believed. The lower diabetes risk in coffee drinkers is one of the most consistent results in the literature, and the American Heart Association's statement last month noted that decaf shows the same benefit, which means the credit belongs to something other than caffeine. A gut organism that ferments coffee's leftovers into a short chain fatty acid is exactly the sort of candidate that story needs. Candidate, not answer.
It reinforces the preparation advice from a new direction. The ledger already says filtered and black, for two unrelated reasons: a paper filter removes the cafestol that raises LDL, and the mortality signal in the cohort data belongs to coffee without much sugar or cream in it. This paper arrives at unsweetened and filtered from a third direction entirely. Three independent reasons for the same instruction is a good place to be.
And it is a reminder about attention. This paper has been out since December and has drawn almost no notice. The Cork paper, published four months later, was covered by 142 news outlets and sits in the 99th percentile of everything Nature Communications published that season. It has also, in the four and a half months since, been cited six times, by five review articles and one study of mice on a high fat diet. Nobody has tried to repeat it. The quiet Mendelian randomization paper that nobody wrote about is the one that independently corroborated its most interesting bacterium.
Two studies that share no data, no participants, and no authors both landed on Veillonella as the organism coffee acts through. That is the most encouraging thing to happen in this corner of the newsletter in a while. It arrived in a paper that had to be corrected for calling decaf regular, in a journal nobody covered, four months before the study everybody covered. Filter your coffee, skip the sugar, and hold the mechanism loosely.
Sources
Primary source: Cao Z, An Y, Du Y, Xu G, Wang J, Lu Y. Different coffee consumption patterns affect HbA1c via propionic acid-producing gut microbiota. npj Science of Food, published online 8 December 2025. doi.org/10.1038/s41538-025-00655-w (open access)
The correction: Author Correction, npj Science of Food, 2026. doi.org/10.1038/s41538-026-00834-3
The companion study: Boscaini S, Bastiaanssen TFS, Moloney GM, et al. Habitual coffee intake shapes the gut microbiome and modifies host physiology and cognition. Nature Communications 17:3439 (2026). doi.org/10.1038/s41467-026-71264-8. WAiR review: Fifteen People on Decaf. Citation and attention figures taken from that article's Nature metrics page and a Semantic Scholar citation query, both on 4 September 2026.
Underlying datasets: UK Biobank (coffee consumption phenotypes) and the MiBioGen consortium (gut microbiota genome wide association data).
Running tally: The Coffee Ledger