Exercise biomarker claim rests on just 10 people

People respond very differently to the same training programme, and researchers have long hunted for a blood marker that could flag who is adapting. Circulating microRNAs are one candidate, because they are small regulatory molecules that change with physiological stress and are stable enough to measure in serum. Among them, miR-210 is of interest because it is hypoxia-inducible and linked to angiogenesis, mitochondrial control and cellular stress responses — all plausibly relevant to training adaptation. Existing exercise studies of miR-210 have produced mixed results, and whether it can classify an individual's training response had not been established.

Ten young adults completed standardised exercise training. Demographic, anthropometric and training characteristics were recorded, and serum miR-210 expression was quantified by qRT-PCR. The primary outcome was miR-210 upregulation, defined by a median split of expression values, so that those at or above the median counted as high and those below as low. Training frequency in sessions per week, and age, were used as the predictors. Three machine-learning models — ridge-penalised logistic regression, a random forest classifier and a support vector machine — were assessed using leave-one-out cross-validation, with accuracy, area under the curve, Brier score, sensitivity, specificity, precision and calibration reported.

Serum miR-210 values ranged from 0.677 to 1.220 with a median of 0.782, giving five subjects per outcome group. Ordinary logistic regression classified 80% correctly, while all three machine-learning models reached 90% accuracy: ridge-penalised logistic regression (AUC 0.88, sensitivity 1.00, specificity 0.80), random forest (AUC 0.92, specificity 1.00, sensitivity 0.80) and support vector machine (AUC 0.88, sensitivity 1.00, specificity 0.80). Calibration analysis suggested the predicted probabilities were reasonable, with the random forest showing the most consistent agreement. Notably, the direct correlations between miR-210 and training frequency (r = -0.256) and age (r = -0.117) were negative and non-significant.

This sits at the very earliest end of biomarker research: a proof-of-concept that a modelling approach can separate high from low miR-210 in a handful of people, rather than evidence that miR-210 tracks training adaptation. It does not resolve the mixed findings from earlier exercise studies, and the non-significant correlations with the two predictors used make the high classification accuracy hard to interpret. With five people per group, a single reclassified individual would move accuracy by a large margin. The authors themselves state that larger, adequately powered studies are needed before any claim of clinically meaningful predictive value.

Treat this as a signal that researchers are still searching for a blood marker of individual training response, not as evidence that one exists. Headline accuracy percentages from studies of ten people should not be read as reliability; they describe how a model behaved on one tiny dataset. There is no test here that a reader could ask for, and nothing that would change how anyone trains. The honest takeaway is that measuring how well you are adapting still depends on the things you can observe directly, such as performance over time.