Where all of this comes from.

Split by person, not by sample. Six of the fourteen cohorts sampled the same individual repeatedly. Splitting at the sample level would put one person's Monday and Thursday samples on opposite sides of the train/test boundary, and the model would score well by recognising the person. Every split here assigns whole individuals.

Missing is not zero. Different cohorts ran different metabolomics panels, so a blank is usually "this lab did not measure that compound," not "the level was zero." Each metabolite is therefore trained only on the samples where it was genuinely measured, tracked by an explicit mask.

Accuracy travels with every prediction. Each metabolite carries the correlation it achieved on held-out patients, shown next to the value, always, because 185 of 1098 metabolites are predicted poorly, and rendering those identically to the good ones would be the single most misleading thing this interface could do.

Batch confounding is severe. Each cohort studied one disease on one platform, so disease status and laboratory are nearly inseparable. The consequence is quantified on Inside the model: a classifier using only cohort identity nearly matches one using the actual biology.

Association is not causation, and group is not individual. A microbe reported reduced in a condition across a population says nothing about why, and nothing about any one person. Both directions of causality are plausible for most of these findings, and many are likely downstream of diet or medication.

Composition only, no function. The model sees which bacteria are present, not which genes they are expressing. Two communities with identical membership can behave differently. Functional profiling would likely improve accuracy and is the clearest next step.

Genus-level resolution. Only six cohorts have species-level data, so the common feature space stops at genus. Meaningful differences exist below that level; E. coli strains range from harmless to dangerous.

Cohorts are not representative. The source studies skew toward particular geographies, ages and diets. A profile from an under-represented population will be compared against a reference that does not describe it well.