A five-day fast produced a fall in body mass index that was still partly present twelve weeks later, and the mix of bacteria in a person’s gut beforehand helped predict how much of it they kept, according to a randomised trial published on 16 September in Genome Medicine. The study involved 38 healthy adults under clinical supervision. It is a mechanistic study of who responds, not evidence that anyone should attempt a prolonged fast, and it says nothing about doing so outside a supervised setting.
What the trial did
The LEANER study randomised healthy adults to a five-day fasting intervention or to a waitlist control, with follow-up at twelve weeks; 38 people completed it. Researchers at the Max Delbrück Center and Charité in Berlin measured body mass index and body composition, sequenced the gut microbiome, and profiled metabolites in both blood and stool.
Fasting reduced BMI acutely, and the loss came predominantly from fat mass rather than lean tissue. Part of that improvement persisted at twelve weeks, though not all of it. The microbiome shifted markedly, as did plasma and faecal metabolite profiles, and how much microbial diversity changed — both immediately and later — tracked with how diverse a person’s microbiome had been to begin with.
The prediction is what the authors treat as the finding. A machine-learning model built from baseline microbiome composition plus routine clinical variables predicted who would still be lighter at twelve weeks. The variables carrying most weight were an unclassified Faecalibacterium species, Oscillibacter sp. 50_27, LDL cholesterol and systolic blood pressure. The model was cross-validated and then tested in three independent cohorts that had undergone prolonged fasting — people with metabolic syndrome, people with multiple sclerosis undergoing repeated fasts, and healthy volunteers fasting for six to twelve days — and it held up.
Why the external validation is the interesting part
Prediction models built on microbiome data have a poor track record: they often perform well in the cohort that produced them and fail elsewhere, and the field carries a long list of signatures that never replicated. Testing this model in three separate groups, including patients with quite different underlying conditions, is what distinguishes it from that pattern, and it is the reason a 38-person study is worth reporting at all.
What the trial does not do is turn fasting into a prescription. It tested one supervised protocol in screened, monitored healthy adults. It did not compare fasting with any other approach to weight, did not follow anyone past twelve weeks, and did not establish that the people who responded well were better off a year later. Prolonged fasting also carries real contraindications and is not a neutral intervention — people taking glucose-lowering drugs, among others, cannot simply attempt it. Nothing in this paper changes that, and readers weighing it against other options can see how we cover the wider evidence base in our research news.
Thirty-eight completers is a small sample, and the trial was not powered to detect harms. The control arm was a waitlist rather than an active comparison, so the contrast is against doing nothing in particular. Partial persistence at twelve weeks is not durability; nobody was followed to a year. The model identifies associations, not causes — nothing here shows that altering a person’s Faecalibacterium would alter their response. And one author is employed by the commercial fasting-clinic group whose protocols the work is built around, a conflict disclosed in the paper and worth holding in mind when reading how the intervention is framed.
Sources: Kaufhold G, Bartolomaeus TUP, Schütte K, et al. “Machine learning identifies microbiome and clinical predictors of sustained weight loss following prolonged fasting.” Genome Medicine 2026;18(1):136, published 16 September 2026 (doi:10.1186/s13073-026-01765-0). Trial registration NCT04452916.
