Blood is a gossip. Every organ leaks its chemistry into it, and the molecules that circulate carry traces of what the liver did with what the gut made, of which fats a cell membrane is building and which it is tearing down. For decades, Parkinson's disease has been read almost entirely from the outside: the slowed hand, the stiff arm, the tremor that arrives at rest. A team at Xiangya Hospital, Central South University, went looking for the same disease written in a language of small molecules, and published what they found in eBioMedicine on 14 July 2026.
The design matters more than the headline number. Using untargeted ultra-high performance liquid chromatography-tandem mass spectrometry, the researchers profiled plasma in two independent case-control datasets. The first compared 153 patients with drug-naive, de novo Parkinson's against 153 healthy controls, 306 people in total; the second added 270 more. Drug-naive is the crucial word. Antiparkinsonian medication leaves its own chemical shadow, and the paper shows it: treatment was significantly associated with shifts in tyrosine, tryptophan and polyamine pathways. Studying people before they start treatment is how you avoid mistaking the fingerprint of a drug for the fingerprint of a disease.
Across both datasets, 111 metabolites moved in the same direction in early Parkinson's. Three stood out: 12-hydroxyeicosatetraenoic acid, spermine and niacinamide. When the differences were sorted into pathways, sphingolipid metabolism came out strongest, with the lowest adjusted p-value (0.0002) and the highest pathway impact score (0.54), seven altered metabolites mapping onto it. Sphingolipids are structural and signalling fats, the scaffolding and switchboard of cell membranes, and they have surfaced in Parkinson's genetics before. A machine learning classifier built on six metabolites reached an area under the curve of 0.976, with the single best metabolite alone reaching 0.916. Those are impressive figures, and they were validated internally, inside the same population that produced them. The authors say plainly that external validation is still pending.
The part that speaks most directly to families is not the classifier at all. In 477 patients spanning different stages of disease, two metabolites made by gut bacteria and then processed by the liver, phenylacetylglutamine (PAGln) and p-cresol glucuronide (Pcg), were elevated and tracked symptom severity across both motor and non-motor domains, with the strongest links to autonomic and gastrointestinal features. Pcg rose with Hoehn and Yahr stage (p = 0.0128, adjusted p = 0.0383). Neither molecule was a main driver of the diagnostic model; their value was in describing how much disease, not whether. The researchers also tested the obvious objection, that these compounds simply reflect a sluggish bowel. Among patients without functional constipation, the correlations with UPDRS scores held (PAGln r = 0.23; Pcg r = 0.24; p < 0.0001), and among those with constipation they did not, which argues against gut dysmotility being the whole explanation.
This is not the first time these molecules have appeared. A 2023 study of 500 participants in Taiwan reported higher plasma PAGln and p-cresol metabolites in Parkinson's, correlated with motor severity and with specific gut bacterial species, and a separate 2023 metabolome-wide study in two California populations also flagged p-cresol glucuronide. Reproducibility across continents and platforms is the quiet achievement here: metabolomics has a long history of findings that do not survive the next laboratory. What is genuinely new is the combination of drug-naive early cases, deep clinical phenotyping and a severity signal that is chemically separable from the diagnostic signal.
What would this mean in a clinic room? Not yet a blood test. These were snapshots, taken once, in people who already had a clinical diagnosis, in a single Chinese population, compared with healthy volunteers rather than with the conditions that genuinely mimic Parkinson's in its first year. A model that distinguishes patients from healthy controls is not a model that distinguishes Parkinson's from essential tremor or from multiple system atrophy, and an AUC earned inside one cohort is a promise, not a performance. But the direction of travel is worth understanding: a peripheral, cheap, repeatable readout that might one day describe the kind of Parkinson's a person has, not merely that they have it.