Five amino acids in the blood, twelve years before diabetes

Can a blood sample taken while a person is still healthy tell, years in advance, who is more likely to develop type 2 diabetes? The researchers looked beyond the usual tests (blood sugar, weight) at small molecules the body produces and consumes: metabolites.

The study drew on the Framingham Offspring Study, a group of American adults followed for decades. Among 2,422 people without diabetes examined between 1991 and 1995, 201 developed diabetes over the following 12 years. The researchers measured 61 metabolites in stored blood from 189 of these future cases and compared them with 189 people who did not develop diabetes, matched one to one for age, sex, body mass index and fasting blood sugar. The matching matters: it means the differences found cannot simply be explained by the usual risk factors. The findings were then tested in a second, independent group of Swedish adults (the Malmö Diet and Cancer study, 163 cases and 163 controls).

Five amino acids were higher, years earlier, in the people who later developed diabetes: isoleucine, leucine, valine, tyrosine and phenylalanine. Three of them (the "branched-chain" amino acids) and two "aromatic" ones. For each of them, being in the top quarter of the population meant two to three and a half times the odds of diabetes compared with the bottom quarter. A combination of three (isoleucine, phenylalanine, tyrosine) did better: five to seven times the risk for the top quarter. In the Swedish group the pattern held for four of the five amino acids, and the three-amino-acid combination gave about four times the risk. That is the observation. The interpretation the authors propose is that a disturbance of amino acid metabolism may be an early feature of the road to diabetes, perhaps linked to insulin resistance before any blood sugar test shows it.

The numbers are relative: "five times the risk" compares the top quarter with the bottom quarter inside a group that was deliberately rich in people at high risk, because controls were chosen to resemble the cases. When the researchers repeated the analysis in 400 people drawn at random from the whole cohort, the effect shrank to about twice the risk for the top quarter, and the improvement in the ability to tell who would get diabetes (the "c-statistic") was modest. In plain terms: the amino acids add information, but less than the headline figure suggests, and more in people already at risk than in the general population.

Type 2 diabetes is one of the few diseases for which prevention has been proven in randomised trials: changing diet and physical activity delays or prevents it in people at risk. A marker that flags risk earlier could, in principle, help choose who to offer such programmes to. This study does not test that: it shows the marker, not the benefit of acting on it.

Where this fits the Digital Twin: these amino acids are measured in a specialised laboratory, not by a watch or a phone. The Twin, today, does not read this kind of data. What the study supports is the idea behind the Twin: a signal measured once, while a person is well, can carry information about the years ahead, and that information is worth keeping next to the data that follows. To use this signature, a person would need a laboratory metabolomic test, and the model would need validation for that use.

It does not show that these amino acids cause diabetes, nor that lowering them prevents it. It does not provide a test you can order: no threshold is validated for individual use. The people studied were mostly white adults of European descent; the authors say the results need confirming in other groups. The measurement was done once, so the study says nothing about following the values over time. And in the fifteen years since, this signature has not become a routine clinical test, which itself suggests its added value over standard tests is limited.

Nested case-control within a prospective cohort (Framingham Offspring Study, exam 5, 1991–1995), with independent replication (Malmö Diet and Cancer study, Sweden) and an additional analysis in a random sub-cohort.

2,422 eligible adults aged ≥ 35 without diabetes or cardiovascular disease; 201 incident cases in 12 years; profiling in 189 cases and 189 propensity-matched controls (age, sex, BMI, fasting glucose). Replication: 163 cases, 163 controls (mean age 58, 55% women). Random sub-cohort: 400 controls.

Fasting plasma at baseline; targeted liquid chromatography–tandem mass spectrometry (LC-MS/MS), 61 metabolites (amino acids, biogenic amines, polar metabolites). Paired t-tests with a corrected threshold of p = 0.001; conditional logistic regression; c-statistic and likelihood-ratio statistics for model comparison; Cox models in the random sub-cohort.

Five amino acids with p ≤ 0.001 (isoleucine 0.0001, leucine 0.0005, valine 0.001, phenylalanine and tyrosine < 0.0001); odds of future diabetes +57% to +102% per SD of log marker, adjusted for age, sex, BMI and fasting glucose; top vs bottom quartile 2–3.5-fold for single amino acids; three-amino-acid score (isoleucine, phenylalanine, tyrosine) 5–7-fold (p for trend 0.007–0.0009). Replication: OR per SD 1.37–2.01 for four of five (isoleucine p = 0.09); three-amino-acid score 4-fold (p for trend 0.006). Random sub-cohort: adjusted OR 1.36 per SD (p = 0.008), ~2-fold for top quartile; changes in c-statistic described as modest. Associations unchanged after adjustment for HOMA-IR and HOMA-B. No AUC or confidence intervals are reported in the abstract.

Risk of bias (PROBAST, preliminary, now on full text): participants low; predictors low (standardised assay, blind to outcome not stated); outcome low (surveillance at serial exams; registry-based in Malmö, possible misclassification toward the null); analysis moderate (case-control enrichment, no calibration reported, sparse discrimination metrics). Generalisability: predominantly white European-descent cohorts; authors call for confirmation in other groups. Reporting quality: pre-TRIPOD study (2011); no reporting checklist.

Not demonstrated. No clinical use of this signature has been established in the years since; the authors themselves suggest the greatest value, if any, would be in high-risk individuals, and that confirmation is needed. No study of impact on outcomes.

B — potentially integrable, to be validated for the proposed use. Required data: targeted metabolomic assay (isoleucine, leucine, valine, tyrosine, phenylalanine) from a laboratory, plus age, sex, BMI, fasting glucose; a validated threshold does not exist. The signature is a candidate, not a tool.

PubMed record (metadata and abstract) and full text of the author manuscript in PubMed Central, read in full on 5 November 2026. Supplementary material not consulted. The author manuscript may differ in wording from the published version.

Among 2,422 normoglycaemic adults followed for 12 years, 201 developed diabetes. Amino acids, amines and other polar metabolites were profiled in baseline samples by liquid chromatography–tandem mass spectrometry; cases and controls were matched for age, body mass index and fasting glucose. Five branched-chain and aromatic amino acids (isoleucine, leucine, valine, tyrosine, phenylalanine) were strongly associated with future diabetes. A combination of three amino acids predicted future diabetes, with a more than fivefold higher risk for the top quartile. The results were replicated in an independent prospective cohort. The authors conclude that amino acid metabolism may play an early role in the development of diabetes and that amino acid profiles could help assess diabetes risk.

The authors' abstract cannot be reproduced here for rights reasons: read it at the link.