AUTHORS
Sun Ting, PhD, Liu Aixue2, Mingyu Jin3, Anqi Zhang4, Chen Yunfei, MD5, Yang Renchi6, Lei Zhang, MD7
AFFILIATIONS
1 Institute of Hematology and Blood Diseases Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Tianjin, China
2 Tianjin Medical University, Tianjing, China
3 Tianjin Medical University, Tianjing, China
4 Institute of Hematology and Blood Diseases Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Tianjin, China
5 Institute of Hematology and Blood Diseases Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Tianjin, China
6 Institute of Hematology and Blood Diseases Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Tianjin, China
7 Institute of Hematology and Blood Diseases Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Tianjin, China
ABSTRACT
Primary immune thrombocytopenia (ITP) is an acquired autoimmune bleeding disorder characterized by heterogeneous treatment responses. Rituximab serves as a second-line therapeutic option with an initial response rate of 50% - 60%; approximately 30% - 40% of patients exhibit poor treatment response or early relapse. However, reliable predictive tools are lacking to identify the population with response advantages. This study aimed to identify protein biomarkers associated with response to rituximab by integrating bone marrow proteomics with machine learning algorithms, and to construct a predictive model to inform individualized treatment decisions for ITP. This study prospectively enrolled 97 patients with ITP receiving rituximab as second-line therapy at the Institute of Hematology & Blood Diseases Hospital, Chinese Academy of Medical Sciences, between January 2023 and December 2024. Based on the therapeutic efficacy at 12 weeks after treatment, the patients were divided into a responder group (n = 65) and a non-responder group (n = 32). Protein expression profiles of pretreatment bone marrow samples were quantitatively analyzed using 4D-DIA proteomics technology. The expression values of all proteins were normalized using Z-score transformation. Differentially expressed proteins (DEPs) were identified based on the criteria of P < 0.05 and |fold change (FC)| > 1.2. Subsequently, Gene Ontology (GO) enrichment analysis was performed, with FDR < 0.05 set as the threshold for statistically significant enrichment. Protein features were selected using least absolute shrinkage and selection operator regression with cross-validation (LassoCV), which enables adaptive selection of the regularization parameter through 10-fold cross-validation. To compare the predictive performance of linear and nonlinear models, two predictive models were constructed based on the proteins identified by LassoCV: a logistic regression protein model and a random forest protein model. The superior algorithm was adopted. Subsequently, clinical variables with P < 0.1 from univariate analysis were incorporated along with the feature proteins to develop a combined model. Internal validation was performed using 10-fold cross-validation and bootstrap resampling with 1000 iterations. Receiver operating characteristic (ROC) curves were plotted based on the predicted outcomes from each validation method, and the area under the curve (AUC), sensitivity, specificity, and accuracy were calculated to comprehensively assess model performance. Independent predictors were identified through multivariate logistic regression analysis, and SHAP analysis was employed to interpret the decision-making logic of the model. A total of 5552 proteins were identified, with 60 DEPs selected. These DEPs were enriched in biological processes including adaptive immune response, monocyte activation, and immunoglobulin complex formation. Lasso regression identified four feature proteins: IGHG4, MMP8, CDKN1B, and CTTN. In 10-fold cross-validation, the logistic regression protein model achieved an AUC of 0.802, with accuracy of 76.1%, sensitivity of 87.9%, and specificity of 68.8%. The logistic regression combined model showed comparable performance (AUC = 0.808) but slightly reduced stability. The random forest model performed relatively poorly (AUC = 0.701). Multivariable analysis demonstrated that MMP8 (OR = 0.50, 95%CI: 0.28 - 0.90, P = 0.020) and CDKN1B (OR = 0.45, 95%CI: 0.23 - 0.85, P = 0.015) were independent negative predictors of treatment response. SHAP analysis confirmed CDKN1B and MMP8 as the primary contributing features. This study is the first to integrate bone marrow proteomics with machine learning for the prediction of response to rituximab in ITP. We identified bone marrow proteins with potential predictive value for rituximab efficacy, and the constructed prediction model provides a valuable tool for the early identification of patients likely to benefit from rituximab therapy in clinical practice. Additionally, our findings suggest the potential application of bone marrow proteomics in predicting rituximab efficacy in ITP patients, offering new directions for future exploration of the biological functions and mechanisms of action of relevant proteins.
DISCLOSURES
Sun Ting, PhD: No relevant conflicts of interest to declare.