henhaobio
Sign in

bioRxivBioengineeringPreprint

Dynamic, single-cell monitoring of CAR T cell identity and activation with Raman spectroscopy

Ariel StiberiD, Boi Bryant Quach, Babatunde Ogunlade, Antony Georgiadis, Kai Chang, Y. Li, Patrick J. QuinniD, Haoqing WangiD, Kristin CY Tsui, Charm Ang, Elena SotilloiD, David Bernard MiklosiD, Crystal Lee MackalliD, Zinaida GoodiD, Jennifer A. DionneiD

PDFDOIData

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (18 Jul 2026) · CC BY 4.0

Gold-nanorod surface-enhanced Raman spectroscopy with LightGBM separates CAR T from untransduced T cells, tracks antigen-specific activation dynamics, and discriminates sorted CAR+ from CAR- patient T cells.

Current CAR T monitoring trades depth for practicality: PCR lacks phenotype, flow cytometry needs labels and gives static snapshots, transcriptomics is slow and costly, and label-free morphology methods lack biochemical specificity. Prior Raman studies mostly compared static resting versus chemically activated T cells; this work detects CAR expression and antigen-specific activation dynamics in single live cells without targeted labels, including in patient blood.

Key results

  1. ShownSingle-cell SERS separates CD19-CAR T cells from donor-matched untransduced (Mock) T cells. CAR spectra show stronger protein/aromatic bands and weaker nucleic acid bands, and accuracy rises when CAR+ cells are FACS-purified away from the untransduced fraction.

    ~37,000 spectra, nine donors: mean donor-level 10-fold CV accuracy 75.0% (pooled cross-donor model 69.3%). FACS-purified CAR T cells (~10,200 spectra, four donors) raised mean donor-level accuracy from 70.3% to 80.6%, with every donor improving in accuracy and AUC. Nanorod-free Raman (~7,800 spectra, three donors) preserved the band trends. · Distinguishing CD19-CAR from Mock T cells; Machine learning classification and feature…

  2. SuggestedSemi-continuous SERS of CD19-CAR T cells co-cultured with CD19+ JeKo-1 cells distinguishes antigen-specific activation from unstimulated T cell controls, and an unsupervised pseudotime recovers the experimental time order without using time labels.

    Two donors, ~7,400 spectra collected 15-95 minutes into co-culture: 94.6% 10-fold CV accuracy. Pseudotime vs real time Spearman 1.00 and 0.94. Nucleic acid bands declined monotonically (Spearman −1.00); protein/aromatic bands rose in the first hour then partly fell (0.83 and 0.6). Flow CD3+CD19+ conjugate events declined only in CAR co-cultures. · Time-resolved SERS of antigen-specific…; Pseudotime analysis; Figure legends

  3. ShownThe method transfers to a tonic-signaling GD2-CAR with 4-1BB costimulation and separates it from Mock better than CD19-CAR. GD2-CAR spectra also show cytochrome c-associated mitochondrial bands, which orthogonal mitochondrial assays support.

    Five donors, ~11,800 spectra: mean donor-level accuracy 85%, average AUC 0.91 vs 0.78 for unsorted CD19-CAR (pooled cross-donor accuracy 81.4%). GD2-CAR-enriched bands at 755, 1,130, 1,310 and 1,584 cm^−1; nonyl acridine orange flow and mitochondrial-fraction protein assays showed higher mitochondrial content in GD2-CAR+ cells. · Generalizability to a tonic signaling CAR…; Machine learning classification and feature…; Flow cytometry / Fluorescence-Activated Cell…; Mitochondrial extraction

  4. SuggestedIn a four-patient axicabtagene ciloleucel cohort, patient-specific models separate PBMC spectra by pre-infusion, day 7 and day 21 timepoints. They also distinguish flow-sorted CAR+ T cells from CAR- T cells and from non-CAR background cells.

    ~8,200 mixed PBMC spectra: mean patient-level timepoint accuracy ~84%. In patient/timepoint groupings with at least 250 spectra, CD3+CAR+ vs CD3+CAR- T cells: mean accuracy 86%, AUC 0.93; vs pooled CAR- T plus Dump+ background: 88%, AUC 0.95. In patients 019, 022 and 024 (016 excluded), Peak-to-Late shifts in a CAR-associated Raman score matched the direction of flow-measured CAR+ frequency. · Longitudinal SERS profiling of CD19-CAR T…; Machine learning classification and feature…; CAR-associated Raman score analysis; Data size and handling

  5. ShownSingle-cell SERS discriminates four blood cell populations (T cells, primary B cells, JeKo-1 lymphoma B cells, red blood cells) using bands assigned to nucleic acids, proteins, cytochrome c, hemoproteins and lipids. Nanorods and laser exposure did not compromise viability.

    ~11,000 spectra; LightGBM 98.6% accuracy by 10-fold stratified CV; the ~725 cm^−1 adenine band had the largest classification contribution. Acridine orange/propidium iodide viability after 1-hour incubation with or without nanorods and 785 nm illumination (duplicates) supported biocompatibility. · Human blood cell classification; Nanorods enable rapid live-cell SERS; Cell viability

Methods, data and reproducibility

Live cells mixed with cationic gold nanorods were profiled by single-cell SERS (785 nm, 1.5-second acquisitions); preprocessed spectra were classified with per-donor or per-patient LightGBM models (Optuna tuning, 10-fold stratified cross-validation), with gain-based feature importance, UMAP and pseudotime used to interpret bands and activation trajectories.

  • Live single-cell SERS spectra of T cells, primary CD19+ CD20+ B cells, JeKo-1 mantle cell lymphoma cells and red blood cells in PBS · human · n = ~11,000 spectra, four cell types
  • CD19.28ζ CAR T cells vs donor-matched Mock T cells from healthy donors, unsorted and FACS-purified · human · n = ~37,000 spectra, nine donors; sorted ~10,200 spectra, four donors
  • Non-enhanced Raman spectra (no nanorods, 25-second acquisitions) of CD19-CAR vs Mock T cells · human · n = ~7,800 spectra, three donors
  • CD19-CAR and unstimulated T cells co-cultured with CD19+ JeKo-1 cells (3:1 effector-to-target) with parallel flow cytometry · human · n = ~7,400 spectra, two donors (DN7518, DN2058)
  • GD2.4-1BBζ CAR T cells vs donor-matched Mock T cells · human · n = ~11,800 spectra, five donors
  • Longitudinal PBMCs (pre-infusion, day 7, day 21) from axicabtagene ciloleucel-treated large B cell lymphoma patients, plus FACS-sorted CD3+CAR+, CD3+CAR- and Dump+ cells · human · n = ~8,200 PBMC spectra, four patients
  • Matched CyTOF data from the same patient cohort and timepoints (prior study, Good et al.) · human · n = four patients · Stanford Digital Repository qb215vz6111
  • scRNA-seq of CD19 CAR T infusion products from axi-cel-treated large B cell lymphoma patients (CAR+ vs CAR- differential expression, SCENIC regulons) · human · n = 102,359 cells (82,552 CAR+, 19,807 CAR-), seven patients
Methods & tools
HORIBA XploRA+ confocal Raman microscope (785 nm excitation), Gold nanorod SERS substrate (seed-mediated synthesis, CTAB-stabilized, washed), LightGBM, Optuna Bayesian hyperparameter optimization, scikit-learn (ROC/AUC, confusion matrices, PCA), scipy (Savitzky-Golay derivatives, Spearman correlation), airPLS baseline correction; skimage BayesShrink wavelet denoising, UMAP, k-NN graph minimum-spanning-tree pseudotime, Flow cytometry/FACS (FACSAria, Cytek Aurora, Novocyte Quanteon) with FlowJo, Seurat v5.0, SCENIC/AUCell, EnhancedVolcano, SEM, cryo-EM/ET (IMOD, UCSF ChimeraX), zeta-potential, CyTOF
Reagents & environment
Jupyter notebooks, Python 3.11.13; LightGBM, Optuna, scikit-learn, scipy and skimage (versions not stated); Seurat v5.0 for scRNA-seq.
Assay conditions & compute
Analyses executed on a CPU (no further specifications); SERS spectra acquired at 1.5 seconds each (non-enhanced 25 seconds), with 30.47 mW (20x objective) or 26.34 mW (50x objective) at the sample.
Reproducibility gaps
Raman source data and analysis code are promised only upon publication; the scRNA-seq GEO deposit is tied to a manuscript in submission and no accession is given. The workflow uses standard Python packages with no custom algorithm.
Limitations (6)
  • Variation between donors, uneven transduction and small cohorts limit generalization to unseen donors or patients. Inter-donor variability exceeded between-class differences, and pooled models reached 69.3% (CD19-CAR) and 81.4% (GD2-CAR), so per-donor models were used. · stated by the authors
  • The CAR-associated Raman score reflects spectral similarity, not calibrated CAR+ T cell frequency, and was compared with flow only for direction of change. Patient 016 lacked enough CAR+ spectra and was excluded from this analysis. · stated by the authors
  • Nanorod hotspot heterogeneity and SERS sampling bias add spectral variability. Nanorod-free confirmation covered three donors, and the authors propose metasurface substrates to reduce this variability. · stated by the authors
  • Co-culture spectra average over interacting and dying T and JeKo-1 cells at a 3:1 effector-to-target ratio above in vivo levels, and PBMC spectra average heterogeneous cell mixtures. The authors propose tracking single cell-cell interactions to address this. · stated by the authors
  • Within-donor cross-validation draws training and test spectra from the same sample acquisitions, so effects of the session or well could inflate accuracy. No evaluation on held-out donors or patients is reported. · noted in this summary
  • Patient timepoint classification cannot separate a CAR-specific signal from lymphodepletion and shifts in immune composition. No link to treatment response, cytokine release syndrome or neurotoxicity was tested in the four patients. · noted in this summary

Read this if you build label-free single-cell assays or monitor CAR T cell products or patients and want concrete accuracy numbers for Raman spectroscopy with machine learning.

Citations
1

Abstract

Chimeric antigen receptor (CAR) T cell therapies have reshaped treatment for cancers and immune-mediated diseases, yet their safety and efficacy depend on both the proliferation of engineered cells and their dynamic functional state -- features that remain challenging to monitor in real-time clinical settings. Current methods require targeted labels, extensive processing, and provide only static snapshots of cell identity and activation. Here, we introduce a surface-enhanced Raman spectroscopy (SERS) and machine learning (ML) approach that enables single-cell identification of engineered CAR T cells without molecularly-targeted labels and time-resolved, semi-continuous monitoring of their functional activation state through a single physical readout spanning donor-derived cells and patient blood. From intrinsic vibrational signatures of live cells, we detect spectral differences resulting from engineered receptor expression in donor-derived CD19- and GD2-targeted CAR T cells (nine and five donors, respectively) with 81-85% donor-level accuracy, and resolve dynamic antigen-specific activation trajectories with temporal precision. Applying this same SERS-ML method to longitudinal samples from a four-patient CD19-CAR T therapy cohort, we classify patient peripheral blood mononuclear cells across pre-and post-infusion timepoints with a mean patient-level accuracy of 84%, and show that isolated CAR-positive T cells are distinguishable from both CAR-negative T cells and background populations with average 86-88% accuracies. These capabilities stem from biochemical signatures consistent with processes such as receptor expression, tonic signalling, and immune synapse formation, demonstrating a single method that reports both cellular identity and activation state with biochemical specificity across cell engineering and clinical monitoring contexts. Our results extend CAR T cell monitoring beyond static phenotyping and support the potential of SERS-ML analysis for rapid, point-of-care assessment of engineered immune cells across the therapeutic lifecycle.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISXhangolli '18Friis '25Ran '23Morales '26Soemardy '25Eichholz '20
57% semantic matchBioengineering
I. Xhangolli, B. Dura, G. Lee+3 more·bioRxiv·Oct 4, 2018