Towards Improved CAR T Therapy Monitoring with Label-free Raman-Based Immune Cell Quantification

Chimeric antigen receptor (CAR) T cell therapy has emerged as a transformative immunotherapy for cancer, autoimmune diseases, and transplantation. In this treatment, a patient’s immune cells are isolated and engineered to express a synthetic CAR on their surface to direct T cell reactivity towards target cells. Patient CAR T cell quantity throughout the therapeutic process is a contributing indicator of the efficacy of immune response as well as an immediate sign of toxicity through complications such as cytokine release syndrome and immune effector cell-associated neurotoxicity syndrome. However, current methods for quantification are costly in both time and money, precluding point-of-care clinical decision support. There is a critical need for real-time, low-cost CAR T cell quantification. Here, we develop a label-free, and cost-effective technique for CAR T cell counting in patient blood using surface enhanced Raman spectroscopy (SERS) and machine learning (ML). SERS provides a platform for non-destructive, fast, label-free cellular identification with single-cell resolution, eschewing the need for complicated sample preparation. Integrating ML algorithms with the ability to recognize patterns in complex datasets provides precise differentiation of cellular phenotypes from these minute spectral differences. Here, we achieve single-cell SERS spectral collection and classification of engineered T cells from natural T-cells.<br/><br/>We synthesize gold nanorods with a localized surface plasmon resonance peak around 660-700 nm for Raman excitation at 785 nm. These nanorods will be used for spectral collection of live immune cells in liquid, therefore we select a resonance peak close enough to 785 nm to be excited by our laser but blue-shifted enough to minimize the competitive extinction component of SERS signal intensity. We overcome the low and inconsistent electrostatic interactions between immune cells and as-synthesized gold nanorods through thiol-gold mediated nanorod surface modification. With these modifications, we demonstrate increased nanorod-cell binding, decreased nanorod aggregation, and significantly decreased single cell spectral acquisition times. Using these nanorods we collect and classify SERS spectra of live engineered (CAR) and non-engineered T cells to create a dataset of 1000 single-cell spectra of each cell time over 10 donors. This is used to train a robust ensemble ML model that accounts for variations cell-to-cell, patient-to-patient, and in nanorod-cell binding, with which we demonstrate successful classification of CAR T cells with high accuracy (≥ 90%). Our ML algorithm allows us to identify the specific vibrational modes that distinguish CAR from non-engineered cells by calculating feature importance scores for each wavenumber or groups of biologically relevant peaks, which we use to enhance our understanding of the differences between CAR and non-engineered cells. We apply this method both to cultured cells in 10% EDTA, as well as the more clinically relevant application of cells spiked into donor blood, and validate our results with flow cytometry and molecular assays. We also describe work to digitize patient blood samples for rapid processing and CAR T cell quantification, using acoustic bioprinting to encapsulate single cells in droplets. By conducting this procedure on patient blood throughout CAR treatment, we describe how we can generate efficient cell quantification data that can be correlated to therapeutic impact with significant time and financial advantage over previous methods. Our results demonstrate the promise of this platform for rapid, cost-effective, label-free, real-time CAR T cell monitoring during cancer therapy.