Validation of a High-Sensitivity Assay for Detection of Chimeric Antigen Receptor T-Cell Vectors Using Low-Partition Digital PCR Technology

Abstract

Although in vivo engraftment, expansion, and persistence of chimeric antigen receptor (CAR) T cells are pivotal components of treatment efficacy, quantitative monitoring has not been implemented in routine clinical practice. We describe the development and analytical validation of a digital PCR assay for ultrasensitive detection of CAR constructs after treatment, circumventing known technical limitations of low-partitioning platforms. Primers and probes, designed for detection of axicabtagene, brexucabtagene, and Memorial Sloan Kettering CAR constructs, were employed to validate testing on the Bio-Rad digital PCR low-partitioning platform; results were compared with Raindrop, a high-partitioning system, as reference method. Bio-Rad protocols were modified to enable testing of DNA inputs as high as 500 ng. Using dual-input reactions (20 and 500 ng) and a combined analysis approach, the assay demonstrated consistent target detection around 1 × 10^–5 (0.001%) with excellent specificity and reproducibility and 100% accuracy compared with the reference method. Dedicated analysis of 53 clinical samples received during validation/implementation phases showed the assay effectively enabled monitoring across multiple time points of early expansion (day 6 to 28) and long-term persistence (up to 479 days). CAR vectors were detected at levels ranging from 0.005% to 74% (vector versus reference gene copies). The highest levels observed in our cohort correlated strongly with the temporal diagnosis of grade 2 and 3 cytokine release syndrome diagnosis (P < 0.005). Only three patients with undetectable constructs had disease progression at the time of sampling.

Chimeric antigen receptor (CAR) T-cell therapy has emerged as a highly successful treatment modality in the field of cancer immunotherapy. Unprecedented efficacy has been observed in relapse and refractory hematologic malignant cancers of B-cell lineage (acute lymphoblastic leukemia, large B-cell lymphoma, follicular lymphoma, mantle cell lymphoma, and multiple myeloma),1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13 leading to the FDA approval of several CAR T-cell products for mainstream treatment in select settings. Research continues swiftly in this field, promising extended applications of existing and new products to other malignant cancers.

CAR T-cell therapy is a multistep treatment that involves the collection of the patient's own T cells from blood and his/her genetic modification to code and express CARs that can selectively target tumor surface antigens.^14,15 Once re-engineered and expanded ex vivo, the cells are transferred back to the patient's bloodstream, where they further multiply to eventually target and eliminate the antigen-expressing tumor cells. The in vivo engraftment, expansion, and persistence of these CAR T cells are pivotal components of treatment efficacy.^11,16,17 Peak expansion, particularly in the first 28 days,¹⁶ and subsequent long-term persistence have been correlated with overall response and durability in the setting of hematologic malignant cancers.^4,18,19 Studies also demonstrate that early loss of CAR T-cell detection correlates with disease reccurrence.^1,20, 21, 22, 23, 24 More important, the treatment and expansion of the cells in vivo have been associated with several severe adverse effects, such as cytokine release syndrome (CRS)25, 26, 27, 28, 29 and immune effector cell-associated neurotoxicity syndrome (ICANS).30, 31, 32, 33 Both responses and adverse effects may be highly variable and affected by a combination of factors, including the type of product, the tumor type, tumor burden, patient-specific immune status, IL-6 levels, lymphodepletion, the peak number of CAR T cells, and the time course of the proliferation, among many others.^2,3,5,11,33, 34, 35, 36, 37, 38 How these and other patient factors interact is an area of active investigation. Therefore, routinely monitoring the overall in vivo CAR T-cell status and kinetics by robust laboratory methods could prove highly valuable in correlating and elucidating both patient- and product-specific responses, ascribing risk of toxicity and potentially guiding immediate treatment decisions.

To date, there are no standardized or commercially available tests for monitoring of CAR T cells after infusion, and testing is not performed in routine clinical practice. Major efforts have focused on methods to assess manufactured products for release, ensuring genetically engineered cells meet a defined threshold of vector copy number per transduced cell. Quantification methods, such as droplet digital PCR (ddPCR)39, 40, 41 and real-time quantitative PCR,⁴² combined to vector transduction efficiency measured by flow cytometry are commonly utilized. Similar methods can be employed for longitudinal postinfusion monitoring, although only sparse published experience exists.⁴³

In contrast to initial product qualification, tests for monitoring CARs in vivo require high sensitivity and may pose challenges for implementation in routine clinical practice. Flow cytometric analysis, for example, is a method that relies on the identification of the CAR T cells based on immunophenotype and has been used in various studies.^44,45 Although it allows sensitive cell subset analysis, performance requires fresh samples with intact cells, and results may experience high technical variability related to individual operator handling, reagent/antibody issues, and data analysis. Alternatively, PCR-based methods targeting unique nucleotide sequences that code for the CAR may be more easily standardized and implemented across laboratories. Quantitative real-time PCR has been utilized in various studies to quantify administered CAR T cells and for post-treatment monitoring.^1,46 Limitations in this context arise from several introduced biases, including variations in DNA recovery, variable target amplification efficiencies affecting product measurement in the exponential phase, and the reliance on several independent reactions for the reference and calibration target DNA standards for quantitation. More recently, digital PCR (dPCR) has gained preference in clinical laboratories as a more robust and highly reproducible molecular method, capable of circumventing several real-time quantitative PCR limitations.

Digital PCR relies on the partitioning of the sample into thousands of independent PCRs, reducing target amplification biases, with products that are digitally detected at end point. dPCR systems allow absolute quantification of the intended targets and internal controls in the same reaction, without the need for independent reference or calibration curves. The accuracy and sensitivity of dPCR assays, however, depend on the system used, the level of partitioning, and the degree to which Poisson statistics can accurately correct for multiply occupied partitions during data analysis.

In previous years, our laboratory developed and validated several ultrasensitive dPCR assays for detection of genomic variants and CAR T constructs using a high-partitioning system (Raindrop dPCR system; formerly RainDance Technologies, Billerica, MA). Partitioning into millions of droplets (denoted as picodroplets) provided a high dynamic range that allowed accurate quantification of targets as low as 0.0001% in single reactions. The market termination of the Raindrop dPCR technology prompted the transition to the Bio-Rad dPCR system (Bio-Rad Laboratories, Hercules, CA), a broadly used platform across clinical laboratories. This low partitioning technology, commonly denoted as ddPCR, features a maximum of 20,000 droplets per reaction and is robust for DNA inputs up to of 20 ng per reaction. Theoretically, provided excellent DNA quality and perfect instrument performance, this input corresponds to approximately 3000 genome equivalents (6000 copies of a reference gene), enabling a maximal limit of detection of approximately 0.02% (1 in 6000). In practice, however, because of variations of DNA quality and system-related limitations (intact droplet formation, stability, and dead volumes in the microfluidics system), the attainable and consistently reproducible limit of detection is generally approximately 0.1%, which is insufficient for longitudinal monitoring of rare targets. Higher sensitivity may be obtained with higher input. Unfortunately, given the limited number of droplets (10,000 to 20,000) and the mode of quantitation, this leads to droplet overload, such that Poisson corrections no longer allow accurate quantification.

This manuscript describes the development and analytical validation of an ultrasensitive ddPCR assay for detection of CAR T constructs, designed to circumvent the limitations associated with low-partitioning digital PCR systems. The performance characteristics are compared with the previously validated clinical assay, and its application to clinical samples of patients treated with CAR T-cell therapy is preliminarily explored in the context of their treatment course.

Materials and Methods

Samples previously submitted for clinical CAR T-cell construct detection [axicabtagene ciloleucel (Yescarta), brexucabtagene autoleucel (Tecartus), and an institutional Memorial Sloan Kettering construct] were identified. Clinical testing was performed using a validated high-partitioning assay, which received approval through the New York State Department of Health for use in the authors’ Clinical Laboratory Improvement Amendments laboratory in 2020. Relevant clinicopathologic information was retrieved from electronic medical records as part of standard procedures for initial molecular diagnostic workup, including the tumor type, white blood cell count at the time of sampling, disease response based on the Lugano classification,⁴⁷ and information regarding the development of CRS or ICANS, with grading based on severity of symptoms on a scale of 1 to 4 (mild to severe), according to consensus recommendations established by the American Society for Transplantation and Cellular Therapy.⁴⁸

Genomic DNA was extracted through standard automated laboratory protocols using Chemagen magnetic bead technology (PerkinElmer, Waltham, MA) on a Hamilton ChemagicSTAR instrument (Hamilton, Manitowoc, WI); genomic DNA concentration and total yield were measured using Qubit dsDNA HS assay (Thermo Fisher Scientific Inc., Life Technologies, Carlsbad, CA). Clinical dPCR was performed with the Raindrop dPCR system using previously described methods.^49,50 The QX200 Droplet Digital PCR system (Bio-Rad Laboratories) was used as the new platform for assay validation. The performance characteristics of the assay were established, including accuracy, limit of blank, analytical sensitivity, analytical specificity, interassay reproducibility, and intra-assay reproducibility, and were compared with the reference method. Following validation and optimization, the longitudinal performance of the assay was assessed on additional patient samples.

Testing by Digital PCR Using the Bio-Rad System (Low Partitioning) and Comparison to RainDance Raindrop (High Partitioning)

Assay design: Proprietary primers and probes were designed to target the CAR T transfer mammalian expression vector unique to the axicabtagene, brexucabtagene, and CD19-directed CAR constructs developed internally at Memorial Sloan Kettering. Focal regions of an endogenous housekeeping gene (HKG) were also amplified to serve as internal quantitation control and to enable relative quantification of the incorporated construct. GUSB was used for the reference method HKG, whereas ATP2B4 was selected for the new platform because of cleaner performance.

Testing using the RainDance assay was performed using a single high input reaction of 500 ng of DNA. The Bio-Rad assay was performed in duplicate using two DNA inputs of 20 and 500 ng. The low-input testing (20 ng) followed previously published manufacturer protocols.⁵¹ Briefly, 20 ng of genomic DNA was loaded onto the automated droplet generator instrument (Bio-Rad Laboratories) to emulsify the samples in the presence of Droplet Generation Oil, ddPCR Supermix, 4-base cutter enzyme (MseI; New England BioLabs, Ipswich, MA), specific PCR primers flanking the targets, and fluorescently labeled, construct-specific probes that hybridize at the target site, followed by amplification. Testing of the 500-ng reaction followed similar protocols with additional adjustments, including a decrease of the PCR primer and probe concentration for the ATP2B4 gene in the PCR mixture from 900 and 250 nmol/L to 60 and 16.6 nmol/L, respectively. All reactions were performed using a maximal volume of 22 μL, including 2 μL of microfluidics system dead volume. Negative, low positive, high positive, and no template controls were tested as described in Supplemental Table S1.

Data Analysis

The data from cluster plots were analyzed using the Bio-Rad QX Manager data analysis software following provided manufacturer's protocols (QX Manager–Data Analysis Workflow Version 1.2 Standard Edition). Briefly, the QX Manager software uses the raw data from the QX200 Droplet Reader to produce data files containing analysis information based on the number of positive and negative droplets identified for each target in each sample. Poisson distribution statistics are applied to correct for multiple copies of the target in single droplets and determine the final number of copies in the reaction that are expressed in unit copies/μL. Final results for the target CAR T are expressed as a quantitative ratio, which is the percentage of CAR T copies relative to the internal control.

Data were analyzed using R version 4.1.3 (https://www.npackd.org/p/r/4.1.3); visualizations and graphs were generated using the ggplot2 package. The quantitative values between two groups were compared using U-test. The quantitative values between more than two groups were compared using Kruskal-Wallis test. Correlations between numeric variables were performed using linear and log-linear regression. The α value was set to 0.05.

Assessment of mVCN in Pre-Infused CAR T-Cell Products

The vector copy number (VCN) was determined by the manufacturer before infusion by amplifying the vector and the human albumin gene in a multiplex real-time quantitative PCR assay. The average number of vector copies per cell (transduced and non-transduced) was calculated by dividing the number of vector copies amplified by half the number of human albumin copies (two copies per cell). The average VCN per transduced cells was estimated by dividing the average VCN per cell by the percentage of transduced cells determined by flow cytometry (% CAR⁺).

By ddPCR, VCNs were assessed after analyzing both target and reference ATP2B4 results. Assuming two copies of the ATP2B4 gene per genome cell, mean VCN (mVCN) was calculated as the ratio of the number of target copies/half the number of reference gene copies, normalized by the reported transduction efficiency percentage.

For comparison, estimations of mVCN by ddPCR were also calculated without the provided transduction efficiency by applying Poisson statistics to the measured VCN for the total cell population (transduced and non-transduced). This assumes that all cells are equally susceptible to transduction.

Results

Determination of the Limit of Blank and Positivity Cutoffs

The limit of blank for the assay was determined in a two-step process. The first step encompassed the analysis of 52 replicate reactions of the commercial negative control sample (IVS-0000). Using this control, a single droplet was observed in one sample. A second step was to analyze data generated from 13 clinical samples from patients who had never received CAR T-cell treatment (Supplemental Table S2). In this set, a slightly higher level of noise could be observed as rare, dispersed droplets that fell in the region of the positive gate in the 500-ng reaction in both platforms. For the clinically validated RainDance assay, up to four highly dispersed target droplets in the 500-ng reaction were established as the noise of the system. For the Bio-Rad assay, up to two dispersed target droplets could be observed in the 500-ng reaction in 30.76% (4/13) of untreated samples (representative plots are shown in Supplemental Figure S1). Therefore, three or more droplets, distributed in a tight cluster within the positive gate, were established as the minimum acceptable number for an unequivocal positive call. For clinical reporting, samples with three positive droplets in the semiquantitative (SQ) reaction were repeated for confirmation with concordant results in all cases.

Accuracy Study (Comparison with the Raindrop dPCR Reference Method)

A total of 47 samples were tested in the accuracy study (34 positive and 13 negative), as outlined in Supplemental Table S3. This included 36 clinical samples (27 peripheral blood and 9 bone marrow aspirates), 6 samples from CAR T-cell products left over from the infusion bags (3 direct samples and 3 dilutions), 4 control samples generated from dilutions of a plasmid in tissue DNA at known levels of 1000, 500, 450, and 250 copies in the 500-ng reaction, and 1 commercially available negative control DNA. Among the positive samples, RainDance CAR T ratios (CAR T copies/internal control copies) ranged from 0.005% to 151.86%. Using the 20-ng reaction, only 65% (22/34) of the positive samples were detectable by the Bio-Rad assay and were confined to those with CAR T ratios >0.1%. Below this level, detection was only possible with the 500-ng reaction. The combined low- and high-input reactions delivered 100% concordance between the Bio-Rad and the reference method, as far as the binary determination of positive or negative for CAR T constructs.

Among the positive samples, there was excellent correlation between the calculated ratios (R² = 0.97; P < 0.0001) of the RainDance and the 20-ng reaction by Bio-Rad. However, accurate quantitation from the 500-ng reactions using the Bio-Rad software was precluded by droplet overload, leading to marked underquantification of the internal control copies.

Quantitation of Low-Level CAR T Events Detected by the High-Input (500-ng) Reaction

To circumvent the limitations of quantification in the high-input reaction, a combined quantification approach was developed. The first step was to determine the number of CAR T droplets from the 500-ng reaction. Because signals for the CAR T construct are absent in the 20-ng reaction, it is assumed that the CAR T target initially present in each droplet in the 500-ng reaction is low, mostly as single copies, such that the CAR T copy number is roughly equal to the number of positive droplets identified. Copies of the internal control are calculated in reference to the corresponding 20-ng reaction. A derived semiquantitative ratio of CAR T copies to internal control was calculated as follows:

$$\text{SQ Ratio}=\left[\mathrm{C}\mathrm{A}\mathrm{R}{\text{T}}_{500}{/\mathit{A}\mathit{T}\mathit{P}\mathit{2}\mathit{B}\mathit{4}}_{{500}^{\ast }}\right]\times 100$$ (1)

In this formula, CAR T₅₀₀ corresponds to the CAR T–positive droplets taken directly from the 500-ng reaction, whereas the number of ATP2B4 copies is calculated on the basis of the number from the corresponding 20-ng reaction using the following formula:

$${\mathit{A}\mathit{T}\mathit{P}2\mathit{B}\mathit{4}}_{{500}^{\ast }}=\mathit{A}\mathit{T}\mathit{P}2\mathit{B}\mathit{4}/\mathrm{n}\mathrm{g}\times 500\mathrm{n}\mathrm{g}$$ (2)

The $\mathit{A}\mathit{T}\mathit{P}\mathit{2}\mathit{B}\mathit{4}/\mathrm{n}\mathrm{g}=\mathrm{Q}\mathit{A}\mathit{T}\mathit{P}\mathit{2}\mathit{B}\mathit{4}$ copies/μL (given 20-ng DNA input in 20 μL reaction).

To further validate the relative quantitative accuracy of this approach, 23 previously characterized samples with known quantitative results between 1% and 0.1% were used to compare results of the quantitative (Q) ratio derived by the Bio-Rad instrument versus the SQ ratio (Supplemental Table S4). Results were compared by a linear correlation test showing good correlation (R² = 0.90). Running the linear regression through the origin to check for systematic error, and based on the slope in the equation, it was determined that a calibration coefficient was needed so that the SQ ratio would more closely correlate with the expected Q ratio. A linear regression equation was calculated using the best-fit line method to find the slope where the sum of squared differences of the points from the line was minimum. The regression equation is stated below:

$$\mathrm{C}\mathrm{o}\mathrm{m}\mathrm{p}\mathrm{a}\mathrm{r}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}\mathrm{m}\mathrm{e}\mathrm{t}\mathrm{h}\mathrm{o}\mathrm{d}\mathrm{r}\mathrm{e}\mathrm{s}\mathrm{u}\mathrm{l}\mathrm{t}=\mathrm{B}\times \mathrm{N}\mathrm{e}\mathrm{w}\mathrm{t}\mathrm{e}\mathrm{s}\mathrm{t}\mathrm{r}\mathrm{e}\mathrm{s}\mathrm{u}\mathrm{l}\mathrm{t}$$ (3)

$$\mathrm{A}\mathrm{d}\mathrm{j}\mathrm{u}\mathrm{s}\mathrm{t}\mathrm{e}\mathrm{d}\mathrm{S}\mathrm{Q}\mathrm{R}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}(\%)=1.84\times \mathrm{S}\mathrm{Q}\mathrm{R}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}(\%)$$ (4)

Comparison of the adjusted SQ ratios versus Q ratio are summarized in Figure 1A (detailed data in Supplemental Table S4). Further assessment by paired t-test showed no statistically significant differences between the two values (P = 0.81), and a difference plot showed no systematic bias after correction, supporting that the two values are equivalent.

Figure 1.

A: Comparison of adjusted semiquantitative (SQ) ratio versus quantitative (Q) ratio in positive samples with chimeric antigen receptor (CAR) T levels below 1% using the 500-ng reaction shows excellent correlation. The gray lines show the 95% CI of the fitted regression line. B: Comparison of all positive samples in the validation. Plotted Bio-Rad ratios reflect standard Q ratios for samples positive for a CAR T construct by both 20- and 500-ng reactions and adjusted SQ ratios for those only positive with the 500-ng reaction. C: Representative plots of a sample with Q ratio >0.1%. Top panel: Results of the 20-ng reaction display the fluorescence amplitudes for hexachlorofluorescein (HEX) and fluorescein amidite (FAM) on the x and y axes, respectively, for the targets in the assay. Four discrete clusters are identified: blue (top left) represents the CAR T–positive droplets; orange (top right) are dual-occupied droplets, positive for both CAR T construct and housekeeping gene (HKG); green (bottom right; high HEX) are HKG-positive droplets; gray (bottom left) are target-negative droplets, no HEX and FAM fluorescence. Bottom panel: The corresponding 500-ng reaction. Note the larger number of droplets comprising the blue cluster, because of greater DNA input. The dual-occupancy cluster is not seen, and the green cluster shifts to the left because of minimal corresponding HKG product. D: Low positive case with CAR T constructs only detectable with the 500-ng reaction. Top panel: Absence of droplets in the blue and orange cluster regions. Bottom panel: Corresponding 500-ng reaction shows several clustered blue droplets. C and D: The CAR T construct/HKG ratios are 1.3% and 0.003%, respectively. dPCR, digital PCR; ddPCR, droplet dPCR.

After establishing an alternate method of calculation, comparisons against the reference method were made for all positive samples, showing no statistically significant differences to those previously generated by the Raindrop assay (Figure 1B and Supplemental Figure S2). All negative samples (as established by Raindrop testing) were also negative on the Bio-Rad assay in both the low- and high-input reactions (Supplemental Table S3). Representative plots are included in Figure 1, C and D.

Reproducibility

Assessment of interassay and intra-assay reproducibility was performed using seven independent samples tested in triplicate using the 20- and 500-ng inputs: three negative samples from patients never treated with CAR T-cell therapy and four positive samples chosen to represent the high, intermediate, and low end of the dynamic range of the assay. All runs showed excellent interassay and intra-assay reproducibility for the Q and SQ ratios (Figure 2, A and B). Detailed results are included in Supplemental Tables S5 and S6.

Figure 2.

Reproducibility and sensitivity studies. Scatterplots showing chimeric antigen receptor (CAR) T interassay and intra-assay reproducibility studies. A and B: Three positive and three negative samples are run in triplicate in the same run and across independent runs for the quantitative (Q; A) and semiquantitative (SQ; B) studies. A: Detected CAR T levels (Q ratio; quantitative) based on standard analysis (20-ng reaction) using the Bio-Rad system. B: CAR T levels based on the adjusted SQ levels (combined 20- and 500-ng reaction). There is slightly higher variability of Q ratios compared with the adjusted SQ ratio values obtained for each sample in both interassay and intra-assay runs for levels around 1 and below. Corresponding calculated SD and CV for the positive cases for the quantitative and adjusted semiquantitative ratios. C and D: Results of the dilution series of the clinical sample (C) and plasmid control (D). Dilution points presented in the x axis and calculated CAR T ratio (log10 transformed) in the y axis. Study shows excellent linearity across all levels of detection. For the clinical sample, the CAR T construct was not detected at the 9th dilution (0.004%) or below; for the plasmid sample, the construct was detected down to the 10th dilution (0.001%). ddPCR, droplet digital PCR.

Sensitivity: Serial Dilution Study

The sensitivity study of the ddPCR was initially established using DNA derived from a plasmid control sample that was serially diluted with normal DNA starting with a Q ratio of 0.871%, in its undiluted state, down to 0.001% (Figure 2C). A similar dilution was also performed on the basis of a patient sample (bone marrow DNA) with a CAR T Q ratio of 1.217%, which was serially diluted with normal bone marrow DNA down to 0.001% (Figure 2D). Testing was performed using the 20- and 500-ng total DNA inputs at each level. Results are summarized in Figure 2, C and D, and detailed in Supplemental Table S7, showing detection of the construct down to 0.001% for the plasmid control. For the patient sample, the construct becomes undetectable at ≤0.004%. Regression curves showed excellent linearity in both plasmid and cell DNA dilutions (R² = 0.98 and 0.99, respectively).

Figure 3.

A–C: Time-series plots showing the estimated chimeric antigen receptor (CAR) T construct copies per microliter of blood for three representative patients with the highest CAR T levels. Plots display log-transformed quantitative results (y axis) versus days after infusion (x axis). The results of clinical evaluation for cytokine release syndrome (CRS) and immune effector cell–associated neurotoxicity syndrome, as well as treatment, are shown underneath the horizontal axis. A: Displays results for Patient 21 with diagnosis of diffuse large B-cell lymphoma; high levels of CAR T cells at day 10 (170 cells/μL; CAR T ratio of 7.73%) coincide with the diagnosis of CRS grade 2 (G2); clinically partial response is shown. Sequential decrease in levels is evident in subsequent testing points. Patient demonstrated disease progression at day 14 and beyond. B: Results for Patient 22 show similar sequential decreases and a small bump in CAR T levels in day 27. C: Patient 25 is a patient with acute B lymphoblastic leukemia/lymphoma. This patient maintained relatively high levels of CAR T beyond 28 days, documented in both the bone marrow and blood without evidence of disease involvement in these samples. However, there was soft tissue involvement based on biopsy of a breast mass. Next-generation sequencing of the marrow showed no somatic alterations corresponding to the B-cell acute lymphoblastic leukemia disease but identified a complex rearrangement t(1;2) (q24.2; q33.2) at low level, fusing CD28 exons 1 to 4 to CD247 exons 4 to 8. The fusion corresponds to a portion of the brexucabtagene autoleucel (Tecartus) construct: CD28 costimulatory domain and the T-cell receptor T3 ζ chain (or CD247), further documenting the presence of the construct at high level in the marrow (see Supplemental Figure S4 for Integrative Genomics Viewer visualization of the fusion). D and E: Box plots showing the estimated CAR T construct copies per microliter (D) and CAR T ratio (E) for cases with clinical evaluation for CRS within 6 days of the blood draw for CAR T digital droplet PCR. In this data set, high CAR T levels per μL of blood, as well as high ratios (CAR T signals/reference gene) correlated strongly with the history of CRS G2/G3. Toci, tocilizumab for treatment of CRS.

Assessment of mVCN in CAR T-Cell Products from Infusion Bags

ddPCR results were compared with qualification studies provided by the manufacturing facility at Memorial Sloan Kettering Cancer Center for two CAR T-cell medical products left over in the bag after infusion. For the two products, the calculated mean VCNs were 1.73% and 1.54%, which were in keeping with those provided by the manufacturing facility at 1.6% and 1.8%, respectively (Supplemental Table S8).

Estimations of VCN, without the provided transduction efficiency, were also calculated by applying Poisson statistics to the measured VCN (total cell population). Total VCNs of 0.36 and 0.59 corresponded to transduction rates of 30% and 40%, respectively, with corresponding estimated mean VCN for the transduced cells of 1.2 and 1.4.

Monitoring of Patients Treated with Approved CAR T-Cell Drugs

A retrospective review of all clinical samples tested during the validation and postimplementation phases was performed with characteristics and outcomes summarized in Supplemental Table S9. Among 53 samples tested (31 patients), 35 were from patients who received axicabtagene and 18 were from patients who received brexucabtagene. Monitoring samples were primarily blood (94%, 50/53); however, three bone marrow samples were also submitted for confirmation of findings.

Monitoring samples were submitted across several time points in the post-treatment period, covering early expansion (day 6 to 28) and long-term persistence (up to 479 days); 18 patients had only one sample tested, whereas 13 were monitored between two and five times each (Supplemental Figure S3). Across all patients treated, disease status at the time of last sample was as follows: 10 complete response, 11 partial response, 1 stable disease, and 9 progressions of disease. All patients had detectable CAR T constructs in all samples tested except for three; all three negative samples demonstrated progression of disease at the time of sampling. CAR T-cell ratios, compared with reference gene copies, ranged from 0% to 74%, with the highest levels observed between days 6 and 28. Among those patients with samples obtained beyond 28 days, only 2 of 10 (Patients 23 and 25) had ratios >1.5% (5.3% and 8.0%). The first patient had mantle cell lymphoma with widespread thoracic and abdominal lymphadenopathy, splenic involvement, and extensive extramedullary disease in the peritoneum. This patient had grade 2 CRS and grade 2 ICANS develop within 10 days of CAR T-cell infusion, with blood collected for dPCR on day 29. Similarly, the second patient had extensive thoracic and abdominal lymphadenopathy, breast and hepatosplenic involvement, and widespread extramedullary disease in the peritoneum. This patient had grade 2 CRS and grade 4 ICANS develop within 2 weeks of CAR T-cell infusion, with blood collected for dPCR on day 29.

To further assess the overall CAR T level, the number of CAR T cells per μL of blood was calculated for each patient. Calculation was based on the measured CAR T-cell ratio and the corresponding white blood cell count of the sample per μL. Because original product samples or inserts were not available for the patients to inform the mVCN per cell, an mVCN per cell of 2 was assumed, which corresponds to the low end of the range reported in the literature for Axi-cel (axicabtagene ciloleucel; 2 to 4).⁴⁰ Consequently, these numbers may overestimate the number of CAR T cells but still allow the authors to see general trends. Under this assumption, the established ratio of number of CAR T signals/the number of reference signals would be equivalent to the proportion of CAR T cells in a given volume. For Patient 4, for example, with a CAR T/reference gene ratio of 0.55% on day 13 and a white blood cell count of 6000/mL, the calculated number of CAR T cells in a μL of blood would be 33. Representative examples of patients monitored through several time points are presented in Figure 3, A–C. Remarkably, a high level of CAR T measured by ddPCR at day 29 in the bone marrow sample of Patient 25 (Figure 3C) could be detected as a fusion by next-generation sequencing (Supplemental Figure S4).

Across the samples obtained, the calculated CAR T cells/μL ranged from 0 to 1702. For samples assessed over the first 28 days, levels ranged from 0.21 to 1702, with a median of 5.5. No significant differences were seen between types of products [median, 5.3 versus 5.5 for Axi-cel versus Brexu-cel (brexucabtagene autoleucel)].

High CAR T levels per μL of blood, as well as high ratios (CAR T signals/reference gene) correlated strongly with the history of CRS grade 2/grade 3 (Figure 3, D and E) within 6 days of sampling. Patients with ICANS also showed high levels of CAR T; however, statistical analysis was precluded by the low number of samples obtained at the time of this diagnosis.

Discussion

CAR T cells are living drugs that directly interfere with the host immune system and rely on the harmonic interaction with a multiplicity of host-dependent factors for their expansion, disease targeting, and persistence. Monitoring both the engineered cells and the immune system after infusion is pivotal to further understanding such interactions and to enable therapeutic interventions in a more informed manner. On the basis of recent FDA recommendations (Guidance for Industry) for monitoring persistence and long-term follow-up after administration of CAR T cells and human gene therapy products,^52,53 PCR-based assays should attain high sensitivity, stating that persistence is indicated by detectable levels of gene therapy product sequences above the threshold of 50 copies/mg DNA. Translating this to dPCR signals, and assuming that each cell yields approximately 6.5 pg of DNA, 1 mg would be roughly equivalent to 150,000 cells or 300,000 copies of an HKG. In this context, the minimal required limit of detection should be at least approximately 0.01% (1 × 10^–4; 50/300,000).

Although reaching high sensitivity levels of 1 × 10^–4 and beyond is challenging with low-partitioning dPCR systems such as the Bio-Rad, previous studies have explored increasing individual reaction nucleic acid input above the recommended 20 ng. In a recent study by Haderbache et al,⁵⁴ the authors described their experience using 100 ng reporting detection of CAR T cells constructs as low as 1 × 10^–6 or 0.0001%. In our hands, this level of sensitivity is not reproducible, particularly considering that, even under perfect conditions of DNA quality and system performance, 100 ng of genomic DNA is equivalent to approximately 15,000 cells (30,000 gene copies of a chosen HKG) and hence insufficient to deliver this level of detection consistently. Alternatively, other groups have explored combining testing of several wells to increase the total template input. In the context of chronic myeloid leukemia, for example, researchers have reported variable detection of BCR-ABL fusion products between 1 × 10^–4 and 1 × 10^–6 (0.01% and 0.0001%, respectively).55, 56, 57, 58, 59 Although technically possible, the need for numerous low-input reactions has performance drawbacks. For instance, depending on the assay, background noise may interfere with the detection of low-level target sequences. In addition, with individual low-input wells, there is a high likelihood that a rare event will remain undetected or at the level of noise in several of the replicates, leading to both high false-negative rates as well as false-positive rates, depending on how aggressive the calling thresholds are established. At the end, a substantially high CV can result from the lack of precision near the limit of detection when applying this model.^57,59,60

In consideration of the above limitations, the approach presented here provides an alternate solution that consistently delivers detection in the range of 0.001% (1 × 10^–5) with excellent accuracy and reproducibility. Although the ultimate results are similar in comparison to a high-partitioning system, analysis relies on two reactions and should be considered semiquantitative rather than absolute quantification. This assay received New York State Department of Health approval as a replacement assay for the previous high-partitioning assay. Taking advantage of the high performance of the Bio-Rad platform for absolute quantification at low input, the 20-ng reaction for quantification of CAR T constructs down to the 0.1% level was used. For negative cases in the 20-ng reaction, further assessment is based on the combined analysis of the 20- and 500-ng reactions. Several optimization steps were necessary to use this approach. As previously shown, concentrations of genomic DNA >66 ng in a 20-μL reaction (the maximum input volume allowed for the platform) are accompanied with an increase in viscosity, which changes the droplet volume and affects the accuracy of quantitation.⁶¹ Overloading has also been associated with droplet deformation, leading to decline in the total droplet count and further interference with the quantitation step (ie, fewer target-containing droplets reaching the fluorescent intensity detection threshold at the end of thermocycling).⁶² By strategically decreasing the primer and probe mixes to reduce the hexachlorofluorescein (HEX) fluorophore intensity generated by ATP2B4 reference gene probes, it is possible to minimize the droplet deformation and quantitation interference to accurately detect and quantify rare CAR T-cell constructs without the generation of double-occupancy fluorescein amidite (FAM) and HEX droplets. Although quantification of the highly abundant HKG remains compromised in the 500-ng reaction, low-level amplification of ATP2B4 maintains the necessary presence of both HEX and FAM signals required by the platform for calibration and plotting and serves as an internal reaction control to discriminate between negative results and PCR failure. Corrected normalization using high input is therefore based on the ATP2B4 counts derived from the low-input reaction, which provides final CAR T quantitative results without significant differences or biases compared with the reference high-partitioning system. Furthermore, when the DNA input is increased (to calculate the SQ ratio), although the signal is amplified, it also leads to an increase in noise and potential double occupancy of the template per droplet. To overcome this challenge, a correction coefficient needs to be introduced. The correction coefficient acts as a compensatory factor that adjusts the semiquantified measurement to align it with the expected Q ratio. Not surprisingly, without the use of a high-input reaction, the false-negativity rate was >30% on clinical samples, solely confined to those events <0.1%, reflecting the limitations of an insufficient template (Supplemental Table S3). An important observation is that further increases in input are not possible as this generally leads to PCR inhibition and overall failure of testing. In extended experience with this platform, PCR inhibition may be occasionally seen in 500-ng reaction, often related to DNA impurities and easily resolved with further DNA clean-up steps.

In the review of all clinical samples submitted during the validation and implementation, the robustness and dynamic range of the assay to detect CAR T vectors in blood across early expansion phases and long-term persistence (up to 479 days) were demonstrated, with detection of vectors as low as 0.004%. The utility of this assay to detect the presence of the construct in target tissues, such as bone marrow, was similarly documented. Of particular interest, more recently, we have also identified vector sequences in DNA of formalin-fixed, paraffin-embedded and fresh tissue samples, as well as cell-free DNA from cerebrospinal fluid, raising the possibility of utilizing our assay for the study of ICANS and assessment of tissue and central nervous system infiltration by CAR T cells. As part of the analytical validation, we also confirmed the ability of the assay to determine vector copy number levels of CAR T-cell infusion products with levels similar to those provided by the manufacturing facility, which would allow a second quality confirmation of vended products. The clinical efficacy of CAR T-cell therapy is dependent on both the expansion and persistence of the cell product, and hence longitudinal quantification of cells in peripheral blood is invaluable to the study of response, resistance, and toxicities associated with treatment. Utilizing reliable detection assays in a prospective manner could aid in the following: i) delineating the mechanism of response and the establishment of predictive models, ii) informing on timing of resistance, allowing for potential redosing of CAR T cells or changes in therapy, and iii) alerting on risk of CRS and ICANS independent of pretreatment patient characteristics.

The assessment of overall CAR T-cell kinetics in this study was retrospective and limited by lack of baseline samples and irregular sampling at sequential time points in most patients. Nonetheless, we were able to confirm previously published observations, such as high levels of CAR T cells primarily confined to the expansion phase between 6 and 28 days and progression of disease in those patients without detectable CAR T constructs at the time of sampling.^11,17,63 Also, although a comprehensive assessment of magnitude and duration of CD19 CAR T-cell engraftment related to patient- and product-specific factors was also precluded and not the focus in this study, we observed the significant relationship of CRS grades 2 and 3 with high levels of the construct (both as a ratio and normalized to volume of blood). Previous studies as part of clinical trials, and more recently, in a study by Ayuk et al⁶³ assessing the in vivo expansion of CAR T cells with ddPCR in a real-world setting, have observed a trend for higher incidence of CRS and ICANS in patients with higher CAR T levels. On the basis of these findings, the role of close CAR T-level monitoring by molecular assays deserves further exploration as part of the comprehensive assessment of CRS severity and other adverse events to potentially provide timely interventions to minimizing the toxicity and maximizing the anti-tumor benefit of CAR T cells.

Finally, clinical guidelines on how testing should be performed, the analysis of data, normalization units, and interpretation of results are currently lacking in this field. In this study, we used the CAR T ratios to establish the validation of performance parameters of the assay. In a clinical setting, these values would be commonly normalized, but there is high heterogeneity on how this normalization is performed, precluding accurate correlations across studies. Future efforts to successfully implement monitoring assays for CAR T-cell therapy will require the development of consensus guidelines and recommendations through the concerted effort of multiple institutions to decide the specific technology or combination of technologies required.

Supplemental Data

Supplemental Figure S1.

Representative two-dimensional droplet digital PCR plots of cases from the clinical validation demonstrate the difference between background noise and true low-level target positivity. A: Two samples of patients never treated with chimeric antigen receptor (CAR) T-cell therapy display rare single or dispersed target droplets (two in the left panel, and one in the right panel) in the 500-ng reaction, circled in red. Similar findings may be observed in other tests for other targets and are interpreted as background noise. B: Two samples of patients treated with CAR T-cell therapy show low positivity; target droplets are observed clustered in the expected gate (circled in green). We have established three or more target droplets, distributed in a tight cluster within the positive gate, as the minimum acceptable number for an unequivocal positive call. HEX, hexachlorofluorescein; FAM, fluorescein amidite.

Supplemental Figure S2.

This figure shows the results of Raindrop digital PCR (dPCR) chimeric antigen receptor (CAR) T ratio (percentage) on the x axis and the results of the Bio-Rad droplet dPCR (ddPCR) CAR T ratio (percentage) on the y axis. The blue dots represent Bio-Rad ddPCR quantitative (Q) ratio measurements obtained with a 20-ng input, which shows excellent correlation with Raindrop ratio above 1% cutoff, with performance of Q ratio rapidly deteriorating below the 1% threshold. The orange and gray dots represent the adjusted and unadjusted semiquantitative (SQ) ratio from the Bio-Rad assay obtained with a 500-ng input and show excellent correlation below the 1% threshold and poor performance above it. Log10 scales are shown on both axes. To show the Q ratios that were 0 on a logarithmic scale, they were assigned a value of 0.0001.

Supplemental Figure S3.

Plot displaying the estimated chimeric antigen receptor (CAR) T-cell levels per μL of blood for all patient samples tested (y axis) and the corresponding day after infusion (x axis). Colored dots connected by a line correspond to the time points for individual patients tested multiple times. Gray dots correspond to all other patients tested at a single time point.

Supplemental Figure S4.

Integrated Genomics Viewer visualization of the CD28 (NM_006139) to CD247 (NM_198053) rearrangement: t(1;2) (q24.2; q33.2) [chromosome (chr) 1:g.167407952::chr2:g.2204599633], consistent with the fusion of CD28 costimulatory domain and the T-cell receptor T3 ζ chain (or CD247) of brexucabtagene autoleucel (Tecartus) construct.