Machine learning based characterization of high risk carriers of HTLV-1-associated myelopathy (HAM)

Abstract

HTLV-1-associated myelopathy (HAM) develops in a part of HTLV-1-infected individuals while most of the individuals remain asymptomatic. This complicates the identification of HTLV-1 carriers at elevated risk. In this study, we integrated HTLV-1 proviral load and antibody titers against Tax, Env, Gag p15, p19, and p24 proteins in a machine learning (ML) framework to identify and characterize high-risk individuals likely to develop HAM. We stratified asymptomatic carrier samples employing an anomaly detection model. We further developed and validated classifier models capable of distinguishing three clinical subgroups, carrier, ATL, and HAM for assessing the anomaly carrier samples as unseen test data. With most anomaly carrier samples (~ 76.47%) predicted as HAM, further statistical and interpretative analysis revealed the 'HAM-like' characteristics of the anomaly carrier samples indicating elevated risk. Additionally, significant heterogeneity in immune response was observed among other asymptomatic carriers. As an exploratory, hypothesis-generating study, our findings are preliminary and aim to propose potential biomarkers and computational strategies that warrant validation in future longitudinal investigations. Our machine learning-based approach offers a novel and insightful tool for identifying and evaluating high-risk characteristics for HAM, providing a holistic view of the complex immune dynamics of asymptomatic carriers of HTLV-1.

Fig. 1

Illustration of the prediction results of anomaly carrier samples by the random forest classifier model. The left bar shows the training and test data for the classifier model. The model was trained on three sample groups and predicted the anomaly carrier samples as unseen test data. The anomaly carrier samples (n = 17) were classified into three prediction groups: Around 76.47% of the anomaly carrier data were predicted as HAM, whereas only 17.64% and 5.88% of the samples were predicted as carrier (n = 3) and ATL (n = 1) respectively (shown on the right bar).

Fig. 2

PLS-based visualization of sample distribution across clinical groups. This PLS plane shows the distribution of all 369 samples from non-anomaly carriers (green), ATL patients (blue), HAM patients (yellow), anomaly carriers (red), and the CDH (dark red as edge color and white-centered). Each dot on the plot represents an individual sample. This plot depicts the clustering of the sample groups based on the analyzed variables, where anomaly carrier samples were positioned near the HAM cluster.

Fig. 3

The boxplots collectively illustrate the distribution of PVL and Antibody titers to HTLV-1 antigens Tax, Env, Gag p15, Gag p19, and Gag p24 across different clinical subgroups: non-anomaly carrier (green), ATL (blue), HAM (yellow), and anomaly carrier (red). The individual data points overlaid on the boxplots show the actual distribution and density of the data.

Fig. 4

SHAP feature analysis of HAM and anomaly carriers. Each bar shows an absolute median value of 300 iterations for different random seeds and the black line represents standard deviation. SHAP analysis for non-anomaly carriers and ATL is in Supplementary Fig. S7.