Machine Learning-Based Prediction of Mononuclear Cell Collection Efficiency for CAR-T Manufacturing

Area of Science:

• Biotechnology
• Immunotherapy
• Data Science

Background:

• CAR-T cell therapy manufacturing relies on efficient leukapheresis.
• Predicting mononuclear cell (MNC) collection efficiency (CE) is crucial for optimizing CAR-T production.
• Current methods for predicting CE are limited.

Purpose of the Study:

• Develop and validate an interpretable machine learning model.
• Utilize routine precollection variables to predict MNC collection efficiency (CE).
• Support individualized leukapheresis planning for CAR-T manufacturing.

Main Methods:

• Retrospective study of 206 patients undergoing leukapheresis.
• Variable screening using correlation, Boruta, RFE, and variance filters.
• Six supervised algorithms trained with nested cross-validation and bootstrap resampling.
• SHAP analysis for model interpretability.

Main Results:

• Extreme gradient boosting (XGBoost) achieved the best prediction performance (R²=0.67).
• The model accurately discriminated high CE (AUC=0.84) with high sensitivity and specificity.
• Key predictors included precollection lymphocyte count, lymphocyte/WBC ratio, age, BSA, and monocyte count.

Conclusions:

• Routine precollection variables can effectively estimate MNC CE.
• The developed model demonstrates good internal validity.
• This approach may enhance individualized leukapheresis planning for CAR-T manufacturing.