Study Identifies Key Predictors of Successful CD3+ Cell Apheresis for CAR T Manufacturing in DLBCL

A new machine-learning analysis is shedding light on one of the least understood but most consequential steps in chimeric antigen receptor (CAR) T-cell therapy: the collection of T cells through leukapheresis. According to the study, a small set of biological and physiological factors can reliably predict whether patients with diffuse large B-cell lymphoma (DLBCL) will achieve the T cell collection efficiency needed for smooth CAR T manufacturing.¹

The researchers argue that their findings, published in Scientific Reports, could help clinicians anticipate which patients are at risk for poor apheresis yield and intervene earlier in the treatment process by helping guide scheduling decisions, enabling early optimization of blood counts, or informing the need for repeat or delayed apheresis.

CAR T-cell therapy requires a successful extraction of CD3⁺ T cells before they can be engineered into cancer-fighting cells. When collection efficiency is low, manufacturing may be delayed or may fail entirely, putting clinical outcomes at risk.² Past research has attempted to anticipate apheresis success, but inconsistent study designs and heterogeneous patient populations have limited the reliability of available tools.

The new study aims to change that by focusing on a biologically consistent group of 98 patients with DLBCL treated at a single center, all undergoing mononuclear cell apheresis before CAR T production. Although the authors noted limitations of the trial, including the single-center design and relatively modest sample size, they emphasized that the homogeneity of the cohort was essential for isolating meaningful biological signals.

“The use of a highly homogeneous cohort of DLBCL patients in this study ensured a focused understanding of the CD3⁺ cell apheresis yield, minimizing variability due to patient heterogeneity,” described the researchers.

The team categorized patients based on whether they reached a CD3⁺ collection efficiency of at least 50%, a threshold that reflects the performance expectations in contemporary CAR T programs. Surprisingly, those who met or exceeded that mark did not have higher circulating T-cell numbers. Instead, many had lower absolute CD3⁺ counts, lower lymphocyte levels, and lower NK-cell proportions than their lower-yield counterparts.

Patients with higher blood volume were also more likely to achieve strong yield, identifying a new contributor. These patterns, explained the researchers, challenge the intuitive assumption that having more T cells in circulation automatically leads to better collection. Instead, the data suggest that overly high T-cell or NK-cell concentrations may interfere with the centrifugation and separation mechanics used during apheresis, reducing the proportion of T cells successfully captured.

The researchers also trained 3 machine-learning models—logistic regression, random forest, and XGBoost—to predict whether a patient would achieve the target apheresis efficiency. The models learned from pre-apheresis variables such as blood counts, cell proportions, and blood volume. On the test data, logistic regression consistently performed best, achieving strong accuracy, a high AUC (0.824), and stable results under cross-validation. While more complex models often promise superior performance, the simpler logistic regression framework outperformed its tree-based counterparts and provided clearer interpretability.

Using Shapley Additive Explanations, the researchers identified the features most responsible for model predictions, with absolute CD3⁺ count emerging as the single strongest driver, followed by NK-cell proportion, total blood volume, and CD3⁺ percentage. High absolute CD3⁺ and NK-cell levels tended to reduce predicted yield, while larger blood volume had the opposite effect.

References

1. Carbonell D, Rodríguez-Sosa A, Gómez-Serrano L, et al. Predictive biomarkers validation of CD3+ cell apheresis yield in CAR-T manufacturing for diffuse large B-cell lymphoma: a machine learning approach. Sci Rep. Published online November 28, 2025. doi:10.1038/s41598-025-27061-2

2. Baguet C, Larrghero J, Mebarrki M. Early predictive factors of failure in autologous CAR T-cell manufacturing and/or efficacy in hematologic malignancies. Blood Adv. 2024;8(2):337-342. doi 10.1182/bloodadvances.2023011992