AUTHORS
Xianjing Cheng, PhD, Yibo Wang2, Longfei Yang3, Huhai Yuan4, Haoqi Li5, Xinwen Zhang6, Xiaoli Liang7, Duo Li8, Rui Ji, PhD9, Ruihao Huang10, Renrui Liu11, Wei Fan12, Changzheng Li13, Xiaoqi Wang, MD, PhD14, Xi Zhang, MD15
AFFILIATIONS
1 Xinqiao Hospital Army Medical University, Chongqing, China
2 Xinqiao Hospital Army Medical University, Chongqing, China
3 Xinqiao Hospital Army Medical University, Chongqing, China
4 Xinqiao Hospital Army Medical University, Chongqing, China
5 Guangzhou Medical University, Guangzhou, China
6 Xinqiao Hospital Army Medical University, Chongqing, China
7 Xinqiao Hospital Army Medical University, Chongqing, China
8 Xinqiao Hospital Army Medical University, Chongqing, China
9 Xinqiao Hospital Army Medical University, Chongqing, China
10 Xinqiao Hospital Army Medical University, Chongqing, China
11 Xinqiao Hospital Army Medical University, Chongqing, China
12 Xinqiao Hospital Army Medical University, Chongqing, China
13 Xinqiao Hospital Army Medical University, Chongqing, China
14 Xinqiao Hospital Army Medical University, Chongqing, China
15 Xinqiao Hospital Army Medical University, Chongqing, China
ABSTRACT
Introduction
Ocular chronic graft-versus-host disease (ocGVHD) remains a sight-threatening complication following allogeneic hematopoietic stem cell transplantation (allo-HSCT). However, its clinical management is critically hampered by the lack of objective and quantitative assessment tools. Although tear cytokines and corneal confocal microscopy represent promising non-invasive biomarkers, their integration into a robust predictive framework—enabling both accurate diagnosis and dynamic severity grading—is urgently required. To address this, we integrated machine learning with multi-omics profiling to identify robust biomarkers, construct and compare diagnostic and severity stratification models, and elucidate the core regulatory network via single-cell transcriptomics, complemented by in vivo and in vitro validation.
Methods
We prospectively enrolled 132 eyes (62 with ocGVHD and 70 without) and randomly allocated them into training (70%) and validation (30%) sets(Registration Number: ChiCTR2500099386). A total of 39 parameters, including tear cytokines and corneal confocal microscopy indices, were assessed. Biomarker selection was performed using univariate analysis followed by LASSO logistic regression. Five algorithms—Random Forest (RF), XGBoost, Support Vector Machine (SVM), Multinomial Logistic Regression, and Generalized Linear Model (GLM)—were trained and evaluated using the AUC, accuracy, precision, and F1-score. SHAP (SHapley Additive Explanations) was applied for feature interpretation. Mechanistically, we performed single-cell RNA sequencing (scRNA-seq) on corneal tissues from a B10.D2→BALB/c cGVHD murine model, combined with in silico gene knockout analysis. Validation was conducted via siRNA-mediated knockdown of IL-6 in primary human corneal fibroblasts and anti-IL-6 neutralizing antibody therapy (topical vs. systemic) in the murine model.
Results
For the diagnostic model, 13 core features (including IL-6, IL-1ra, CCL17, ASEM grade, nerve density, immune cell density,etc.) were selected, with the GLM model achieving superior performance (AUC = 0.86). For severity stratification, the XGBoost-based multiclass classifier, incorporating ASEM grade, IL-6, IL-8, ICAM-1, CCL19, immune cell density,etc., performed best in the test cohort (AUC = 0.82). SHAP analysis revealed that inflammatory cytokines (particularly IL-6 and IL-1ra) exhibited the highest feature importance, outweighing clinical signs, indicating that inflammatory imbalance serves as the primary driver of disease progression. Mechanistically, scRNA-seq and virtual knockout perturbations identified IL-6 as a central inflammatory hub originating from a distinct corneal fibroblast subpopulation. In vitro, siRNA-mediated IL-6 knockdown markedly downregulated downstream effectors, including MMP-9 and IL-1β (all P < 0.05). Critically, in vivo anti-IL-6 neutralization significantly ameliorated corneal fluorescein staining, increased tear secretion, and reduced inflammatory infiltration in cGVHD mice(all P < 0.05). Notably, topical administration of anti-IL-6 was superior to systemic delivery.
Conclusions
Our study successfully established a robust machine learning-based framework integrating tear biomarkers and confocal microscopy for accurate ocGVHD diagnosis and severity stratification. We further identified and validated IL-6, specifically derived from corneal fibroblasts, as a non-invasive predictive biomarker and a central therapeutic target. Our findings advocate for the clinical translation of this objective scoring system and support topical anti-IL-6 therapy as a promising precision treatment for ocGVHD. Furthermore, this framework provides hematologists with a non-invasive, precise tool for diagnosis and prognosis assessment, holding significant clinical value.
DISCLOSURES
Xianjing Cheng, PhD: No relevant conflicts of interest to declare.