Label-free chimeric antigen receptor T-cell expression analysis using neural networks and statistical distribution modeling

Chimeric antigen receptor T (CAR-T)-cell therapy has emerged as a promising treatment for hematologic malignancies. Accurate monitoring of CAR expression levels is essential for optimizing therapeutic efficacy and ensuring patient safety. Conventional assessments rely on antibody-based flow cytometry (FCM)-a labor-intensive process that may alter cellular properties. This study introduces a novel label-free method for predicting CAR expression rates using bright-field microscopy and deep learning. The proposed approach integrates convolutional neural networks with classification score distribution analysis by employing Gaussian fitting to estimate CAR expression rates. It achieved a prediction accuracy within 5 % of standard FCM measurements across four donors, reducing the maximum prediction error from 13 % to 4.4 %. This study represents the first successful label-free prediction of CAR expression rates in T cells, offering a practical solution for real-time monitoring and manufacturing of CAR-T cells. This method is readily applicable to manufacturing and research settings, as it relies solely on standard bright-field microscopy.