Metabolomics profiling and machine learning prediction of efficacy and prolonged cytopenia in patients with large B-cell lymphoma undergoing CAR-T therapy

Abstract NTRODUCTION Achieving complete response (CR) within 3 months after CAR-T therapy is a key early indicator for predicting the long-term prognosis of patients with relapsed or refractory large B-cell lymphoma (R/R LBCL). The ability to predict the likelihood of treatment failure before deciding on CAR-T therapy could help clinicians and patients select more appropriate treatment options. Furthermore, early intervention for patients less likely to achieve CR after CAR-T therapy may improve overall survival. However, there is currently a lack of a unified risk model in clinical practice to assess early treatment response. METHODS Serum samples collected prospectively from 123 R/R LBCL patients before and after CD19 CAR-T therapy underwent untargeted metabolomic sequencing. Patients were randomly allocated into training (n=85) and validation (n=38) cohorts. Recursive feature elimination (RFE) combined with LASSO regression was used to screen for differential metabolites. The performance of algorithms including Random Forest (RF) and Logistic Regression was compared, with the area under the curve (AUC) serving as the primary evaluation metric for predicting early efficacy after CAR-T therapy and the risk of prolonged cytopenia. RESULTS Patients were categorized into CR and non-CR groups based on efficacy. The non-CR group exhibited a higher proportion of patients with advanced stage, elevated LDH, ECOG score ≥2, IPI ≥3, TP53 mutation, primary refractoriness, and progressive disease (PD) or stable disease (SD) status prior to lymphodepletion, indicating a higher tumor burden before CAR-T therapy. Using machine learning algorithms on the training cohort, models built with the most significantly differential metabolites effectively predicted whether patients achieved CR within 3 months (RFE: AUC=0.99; Logistic Regression: AUC=0.82). This prognostic model was further validated in the independent validation cohort (RFE: AUC=1.00; Logistic Regression: AUC=0.84). Additionally, pretreatment serum metabolite markers also predicted the risk of neutropenia (AUC=0.694) and thrombocytopenia (AUC=0.754) persisting beyond 1 month after CAR-T therapy. CONCLUSIONS Using serum metabolomics and machine learning, we developed a model that efficiently predicts early response to CD19 CAR-T therapy and the risk of prolonged cytopenia in R/R LBCL patients. This model can facilitate early clinical intervention to improve patient outcomes.