# Urine Metabolomics and Machine Learning Identify Metabolic Features and Potential Biomarkers of HTLV-1-Associated Myelopathy (HAM)

## TL;DR

This study uses urine metabolomics and machine learning to find noninvasive biomarkers for a neuroinflammatory disease caused by HTLV-1.

## Contribution

The novel use of untargeted urine metabolomics and machine learning to identify potential biomarkers for HTLV-1-associated myelopathy.

## Key findings

- 175 metabolites were identified, with 86 showing significant differences across groups.
- Machine learning models achieved robust separation of HAM from other groups using specific metabolites.
- Key metabolites like histidine and 4-hydroxyphenylacetic acid showed reduced levels in HAM with high diagnostic accuracy.

## Abstract

Human T-cell lymphotropic virus type 1 (HTLV-1) can cause HTLV-1-associated myelopathy/tropical spastic paraparesis (HAM/TSP), a progressive neuroinflammatory disease that lacks noninvasive biomarkers. We used untargeted urine metabolomics with machine learning to profile 113 participants (39 with HAM, 17 with intermediate syndrome, 33 asymptomatic carriers, and 24 healthy controls). Gas chromatography–mass spectrometry identified 175 metabolites, 86 of which showed significant differences (fold change > 2, FDR p < 0.05). Multivariate analyses revealed distinct but partially overlapping metabolic profiles: sPLS-DA captured a reproducible yet moderately discriminative signal, while nonlinear machine learning models (Random Forest and SVM) achieved robust group separation, with HAM displaying a distinct metabolic signature. Key discriminators included Unknown_151, Unknown_127, histidine, alanine, and 4-hydroxyphenylacetic acid, which showed marked reductions in HAM and yielded ROC AUCs of 0.855–0.871. Pathway and disease enrichment analyses highlighted disturbances in amino acid metabolism, particularly beta-alanine and aromatic amino acids, along with disease signatures related to inherited amino acid handling disorders such as hyperlysinemia. These results demonstrate that urinary metabolomics combined with machine learning can identify potential noninvasive biomarkers for HAM and provide novel insights into HTLV-1-associated pathophysiology.