Your Gut Is Keeping a Food Diary — Scientists Can Read It in Your Microbiome

For years, gut microbiome advice has sounded frustratingly general: eat more fibre, eat fermented foods, feed your good bacteria. Reasonable advice — but not very specific.

A 2026 study in Nature Medicine went much further. Researchers analyzed data from 10,068 people who logged their food intake using a smartphone app while their gut microbiome was characterized using metagenomic sequencing.

Then they asked a deceptively simple question:

If we know what people eat, how much can we predict about the microbes living in their gut?

The answer was: to a surprising extent, quite a lot.

But there is an important distinction. The study found statistically significant diet–microbiome relationships across a large population. It does not mean that a food diary can identify one person's microbiome with 92% accuracy.

Diet significantly predicted the relative abundance of 669 of 724 bacterial species and 313 of 320 microbial pathways. This does not mean the model identified an individual's microbiome with 92% accuracy.

What they found

1. Diet was linked to an unusually large part of the measurable microbiome.

Of the 724 bacterial species included in the analysis, diet significantly predicted the relative abundance of 669 species. It also significantly predicted 313 of 320 microbial metabolic pathways.

That is impressive, but the wording matters. The study did not report that the computer correctly “identified” 669 bacteria in each person. Rather, dietary information contained a statistically significant signal for the abundance of those species across the study population.

Prediction of overall microbial diversity was much more modest: the reported correlations were approximately r = 0.26 for richness and r = 0.24 for Shannon diversity. Diet explained about 6.7% of the variation in microbial diversity.

So diet clearly matters — but it is only one part of what shapes the gut microbiome.

2. Some individual foods showed surprisingly specific microbial signatures.

Several food–microbe associations stood out:

• Coffee was strongly associated with Lawsonibacter asaccharolyticus.
• Yogurt was associated with Streptococcus thermophilus.
• Milk was associated with several Bifidobacterium species.
• Nuts were associated with a bacterial species labeled UBA11774.

Some foods showed particularly strong microbial signatures. For coffee, the linked bacterium was not detected in the coffee itself, suggesting that the association is unlikely to be explained simply by swallowing the microbe with the drink.

The coffee finding was especially interesting. The researchers examined the food itself and did not detect Lawsonibacter asaccharolyticus in coffee.

That makes simple microbial transfer from the drink unlikely. Instead, coffee consumption may create conditions that favor the organism, or the association may arise through another indirect mechanism. This study cannot tell us exactly why.

The distinction is important. A food being associated with a bacterium does not automatically mean that the food “feeds” that organism directly.

3. Food processing emerged as an important microbiome signal.

Across the dietary measures examined, a higher proportion of calories from ultra-processed foods was strongly associated with lower microbial alpha diversity. Foods such as white bread and pastries were among the dietary features linked to lower diversity.

This does not mean that the usual plant-versus-animal distinction is irrelevant. Rather, the findings suggest that food processing may represent another important dimension of the diet–microbiome relationship, alongside food source, fibre content and overall dietary pattern.

4. The diet–microbiome relationship was not just a one-time snapshot.

The researchers also examined longitudinal data. Four years later, 82.5% of the species that were predictable from diet showed significant longitudinal tracking between predicted and observed abundance.

In simpler terms, when dietary patterns changed over time, the microbiome predictions and the measured microbial changes continued to move together for many species.

That strengthens the argument that the original associations were not simply artifacts of one stool sample collected on one particular day.

5. The models were also tested beyond the original dataset.

The investigators evaluated the diet–microbiome models using external datasets, including an Australian cohort and an Israeli nutritional intervention dataset. This kind of external testing is important because microbiome findings often fail to reproduce when populations, laboratory methods or dietary patterns change.

Then came the personalized nutrition experiment

This is where the study becomes more ambitious.

The researchers built a computer model that could simulate something like this:

The simulation was applied to 2,070 people. Interestingly, it often proposed only small dietary adjustments. The median recommendation involved changing approximately one food item.

Common simulated changes included reducing foods such as white bread and processed chicken products and increasing foods such as apples, avocado and nuts.

That sounds remarkably close to personalized nutrition. But this is exactly where the strongest caution is needed.

Now the part the headline cannot fit

Nobody ate the personalized diet. This is the most important limitation. The personalized recommendations were generated inside a computer. The participants did not follow those recommendations, and the researchers did not subsequently measure whether their microbiome, weight, blood tests or clinical outcomes improved.

The authors themselves describe this as an exploratory framework that requires empirical validation in controlled dietary intervention studies.

Prediction does not mean certainty. Diet contained a detectable signal for hundreds of bacterial species, but overall microbiome diversity was only modestly predictable. The associations become visible across thousands of people; they are not equivalent to an accurate microbiome readout for one individual.

This was largely observational research. People were not randomly assigned to drink coffee, eat yogurt or avoid ultra-processed foods. Coffee drinkers, for example, may differ from non-coffee drinkers in many other biological and behavioral ways.

Statistical adjustment can reduce confounding, but residual confounding from factors such as medication, physical activity, smoking, antibiotic exposure and supplement use can remain.

The detailed food logging covered a limited period. App-based food logging is substantially more precise than a traditional food-frequency questionnaire, but it still captures a relatively short window of eating compared with the many years during which a person's gut microbiome develops.

Greater microbial diversity is not automatically the same as better health. Higher diversity is often associated with healthier states, but “more diverse” should not itself be treated as a clinical endpoint. This study primarily measured microbial patterns and statistical health indices, not prevention of heart attacks, diabetes or other hard clinical outcomes.

Culture and cuisine matter. The main study population came from Israel, and dietary patterns differ substantially between countries. A Korean diet built around rice, kimchi, fermented soybean foods, soups and different patterns of meat and vegetable consumption was not directly tested here.

Whether the same food–microbe relationships look identical on a Korean table remains an open question.

This post is a plain-language summary of a published research paper and is not medical advice. Nothing here is a reason to start, stop or change a medication, supplement or diet. Discuss major dietary or medical changes with your own clinician.

Source
Segev T, Barak D, Zahavi L, et al. Diet–microbiome associations in 10,068 individuals from the Human Phenotype Project to guide personalized nutrition. Nature Medicine. 2026;32:1884–1894. doi:10.1038/s41591-026-04312-x

Key study results
Diet significantly predicted the relative abundance of 669 of 724 bacterial species (92.4%; FDR <0.05).
Diet significantly predicted 313 of 320 microbial pathways.
Alpha-diversity prediction: r = 0.26 for richness and r = 0.24 for Shannon diversity.
Diet explained approximately 6.7% of variation in microbial diversity.
Selected food–species associations: coffee–Lawsonibacter asaccharolyticus r = 0.43; yogurt–Streptococcus thermophilus r = 0.42; milk–Bifidobacterium associations approximately r = 0.31–0.36.
Among predictable species, 82.5% showed significant longitudinal tracking over four years.

Background
The Human Phenotype Project