CD28 signaling complexes are correlated with patient outcomes in anti-CD19 41BB-costimulation CAR T cell therapy

Abstract

Background

Chimeric antigen receptor (CAR) T cells targeting CD19 achieve remarkable remissions in refractory B cell malignancies, yet deleterious side effects such as cytokine release syndrome (CRS) limit their broader application. Current preclinical assays on manufactured cell products do not predict human clinical function. We hypothesized that variability in the CAR proximal protein interaction networks that mediate CAR signal transduction may correlate with patient-to-patient differences in toxicity.

Methods

Using banked, preinfusion 41BB–CD3ζ CAR T cell products with known clinical outcomes, we applied quantitative multiplex co-immunoprecipitation (QMI) to profile ∼200 binary interactions among 21 key signaling proteins following CD19 stimulation. Bioinformatic analysis clustered interactions into functional modules, and correlated protein interaction patterns with clinical outcomes.

Results

Correlation network analysis, which clusters interactions into coregulated modules, identified a stimulation-responsive module with similar behavior in all products, and a second module that correlated with the presence of CRS. The CRS module was enriched for interactions among CD28, FYB, and the SRC family kinases LCK and FYN. In a head-to-head validation cohort, a similar CD28–FYB–kinase module again correlated with the presence of CRS. Using a combined dataset, a machine learning classifier trained on top QMI features retrospectively identified CRS samples with high accuracy.

Conclusions

These data indicate that subtle, batch-to-batch differences in CAR signalosome assembly may correlate with CRS, and they support the further development of a preinfusion proteomic assay to forecast CRS risk in CAR T cell products.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• Cytokine release syndrome (CRS) is a major complication of CD19-directed chimeric antigen receptor (CAR) T cell therapy, but its molecular basis remains unclear. Preclinical and in vitro models fail to predict CRS in humans, so studies using primary CAR T products are needed to understand patient-specific signaling mechanisms.

WHAT THIS STUDY ADDS

• Using quantitative multiplex immunoprecipitation on patient-derived CAR T cells from the PLAT-02 trial, Draper et al. identified a reproducible CD28-centered signaling module that correlates with CRS across independent cohorts.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• These results define a human signaling signature of CRS and point to endogenous CD28 signaling as a key contributor to toxicity risk even in 4-1BB–based CARs. The proteomic approach provides a framework for developing biomarkers and for refining CAR designs to reduce CRS while maintaining efficacy.

Introduction

Chimeric antigen receptor (CAR) T cells are the first commercially successful engineered cell therapy, representing a fundamentally new class of living, bioengineered drug.¹ CAR T cells targeted to CD19 positive leukemias and lymphomas have produced durable remissions in patients with poor prognoses, and are now an established part of the cancer pharmacopeia.2,6 Despite this success, CD19 CAR T cells still frequently produce undesired side effects including cytokine release syndrome (CRS) or neurotoxicity (NT).⁷ Moreover, CAR T cell therapy fails for some patients, producing no response (NR). Clearly, some CAR construct designs work better than others,^{8 9} and some patients’ batches of CAR T cell product produce better outcomes than others.¹⁰ Ultimately, a goal of the field is to produce more consistent patient outcomes through improved engineering. But first, we must understand how our engineering choices translate into differences in performance, and how patient-to-patient variability affects the translation.

41BB-CD3ζ (BBζ) CARs signal by recruiting protein–protein interactions. At a basic conceptual level, the CD3ζ domain recruits the kinase ZAP70 to initiate a LAT-SLP76 signalosome mimicking T-cell receptor (TCR) activation,^{11 12} while the 41BB costimulatory domain recruits TRAF-associated signaling proteins to add a costimulatory component mediated by TAK1 and NFΚB signaling.^{13 14} In addition, the CAR recruits more than 140 proteins on ligand recognition, spanning broad functions including TCR signaling, TRAF signaling, actin remodeling, endosomal engulfment, and nuclear signaling.¹⁵ Each of these recruited proteins can influence the overall signaling outcome of the CAR, resulting in products with varying levels of killing potential, cytokine release, exhaustion, and other qualities. Global research is actively focused on how different costimulatory domains,^{16 17} different CD3-derived activation domains,¹⁸ or different hinge¹⁹ transmembrane,²⁰ or svFv domains⁸ each contribute to determining the overall efficacy of a CAR product. Clearly, design choices influence the final milieu of recruited proteins, which in turn affects the clinical performance of the CAR.

Yet, even with well-designed, Food and Drug Administration-approved CAR T cell products, protein recruitment to the CAR and signal transduction may vary by individual. Product manufacturing involves creating a unique batch of bioengineered cells for each patient. While the manufacturing conditions are standardized using Good Manufacturing Practice certified laboratories, uncontrollable differences in initial T cell viability, infection history, differentiation state, or proliferation may affect the final product,^{21 22} as do differences in the in vivo environment into which the cells are injected.²³ In addition, single nucleotide polymorphisms (SNPs) have been identified in several proteins important for CAR signaling, and these SNPs have been correlated with autoimmune disease rates, suggesting an effect on immune function. For example, a SNP near the TRAF1 gene, rs3761847, is prevalent in the human population (∼25% AA, 50% AG and 25% GG); the GG variant correlates with an increased risk of autoimmunity, while the AA variant is protective.^{24 25} Stimulated T cells from individuals with the GG allele express less TRAF1 and produce less inflammatory cytokines when stimulated, compared with AA cells.²⁴ This SNP, and others like it, may provide a substrate for predictable and quantifiable intra-individual variability.

Might it be possible to predict, using the protein interactions recruited by each unique CAR product, which is likely to be successful, and which is likely to cause adverse side effects? Our group previously developed a quantitative multiplex co-immunoprecipitation (QMI) assay,²⁶ which measures several hundred protein–protein interactions among a network of CAR-signaling and TCR-signaling proteins that we identified in an unbiased mass spectrometry screen.¹⁵ We previously reported that CAR stimulation by CD19 causes widespread rearrangement of T cell protein interaction networks that is distinct from that produced by TCR engagement. Moreover, we identified small but potentially important differences in the signal transduction “biosignatures” produced by different variants of BBζ-CARs, providing a substrate of variability required for a predictive assay.¹⁵ Here, we use banked preinfusion CAR T cell products from a completed clinical trial,^{4 27} each with known clinical outcomes, to correlate protein interaction network “biosignatures” with clinical outcome. We report that CAR products that caused high-grade CRS show upregulation of a module of interactions involving the CAR, the costimulatory receptor CD28, the adapter protein FYB, and the SRC family kinases FYN and LCK. This “CRS-associated module” may directly contribute to individual variability in side effect profiles, and may pave the way for a predictive test to quantify the risk of CRS before the CAR product is infused.

Results

To compare CAR interactome signatures in CAR T cells that produced different clinical outcomes, we obtained viable, frozen aliquots of clinical autologous CAR T cell products (both CD4 and CD8 CAR T cells) that were infused to patients enrolled in the PLAT-02 study (NTC02028455).^{4 27 28} Optimal response (OR) cells produced a minimal residual disease negative complete response (MRDnegCR) in the absence of CRS or NT, defined as CRS and NT grades of 2 or less. CRS cells resulted in a MRDnegCR with CRS grade of 3 or 4, but NT grade of 2 or less. Neurotoxic (NT) cells yielded a MRDnegCR with CRS grade less than 2 but an NT score of 3. (Note this is not typical of NT patients in general, who often have co-occurring CRS²⁹; such patients were not included in this study). Non-responder (NR) cells did not produce a MRDnegCR regardless of toxicity. Patient and cell product characteristics are listed in table 1 and online supplemental Table S1; while CRS patients trended younger, there were no statistically significant differences in age, cell count or EGFR transduction marker expression.

Table 1. Patient characteristics.

	OR	CRS	NT	NR
CRS grade	1.1 (0.83)	3.4 (0.55)	1.8 (0.44)	0.6 (0.89)
Neurotox grade	1.1 (0.83)	0.8 (084)	3 (0)	1 (1)
Age (at enrollment)	14.75 (6.54)	8.0 (6.12)	15.8 (2.39)	10.4 (8.5)
Total N	8	5	5	5
CD4	5	4	3	5
CD8	8	5	5	5
Male/female (total)	9/5	6/3	8/0	8/2
CD4	4/2	3/1	3/0	4/1
CD8	5/3	3/2	5/0	4/1
Cell count (M)	49.05 (38.25)	117.2 (101.8)	30.53 (51.07)	61.26 (50.02)
CD4	24.84 (26.99)	63.70 (34.36)	9.47 (1.26)	63.62 (61.22)
CD8	62.5 (37.99)	160 (121.1)	43.16 (63.49)	58.90 (43.22)
% EGFR expression	65.32 (11.5)	56.82 (19.4)	59.13 (15.05)	67.44 (21.25)
CD4	64.72 (9.8)	66.41 (8.54)	51.87 (7.91)	72.56 (23.14)
CD8	65.65 (12.91)	49.16 (23.08)	63.48 (17.37)	62.31 (20.36)

Bolded numbers are significantly different compared with OR by Kruschall-Wallace test followed by Dunn’s multiple comparison test or by χ² test for age.

CRS, cytokine release syndrome; NR, non-responder; NT, neurotoxicity; OR, optimal response.

Cells were thawed, rested for 2 hours, and stimulated for 5 min with fixed K562 cells expressing CD19 (to stimulate the CAR), or anti-CD3 antibody OKT3 (to stimulate the TCR), or parental K562 controls (figure 1A). K562 cells were chosen as stimulators because they can be easily genetically manipulated and expanded for consistent and controlled stimulation across samples. We confirmed that K562-CD19 cells produce a qualitatively similar stimulation profile by QMI compared with primary lymphoblasts (online supplemental figure S1). A 5 min time point was chosen because the overall intensity of CAR signaling peaked at 5 min (online supplemental figure S2), which is consistent with peak QMI measurements during TCR stimulation.²⁶ CAR T cell stimulation by CD19-K562 cells was confirmed by a custom cytokine Meso Scale Discovery (MSD) assay following overnight co-culture; interferon gamma (IFNγ), interleukin-2 (IL2), tumor necrosis alpha (TNFα) and interleukin-10 (IL10) significantly increased with stimulation. However, no significant differences were observed in cytokine release between outcome groups, even when CD4 and CD8 cells were analyzed separately (online supplemental figure S3). In vitro cytokine release was therefore not a useful predictor of in vivo outcome.

Figure 1. Correlation network analysis identifies modules correlated with stimulation and outcome. A) Experimental design. (B) PCA of QMI dataset clusters data by stimulation type, either no stimulation (blue), CD3 stimulation (green), or CD19 stimulation (red). (C) PCA of the unstimulated CAR T cells labeled by outcome. OR (red), CRS (green), NT (blue) and NR (purple). (D) Topological overlap matrix clusters interactions (tips of dendrogram, labels shown in online supplemental figure S2) into modules (colored boxed) based on correlated behavior across all experiments. Note the yellow module nested within the turquoise. (E) A module-trait table reports CC (top number) and p value (lower number) for the correlation between each color-coded module’s eigenvector and each experimental variable listed below. Colored boxes indicate higher CC as shown in the bar at right. CAR, chimeric antigen receptor; CC, correlation coefficient; co-IP, co-immunoprecipitation; PCA, principal component analysis; QMI, quantitative multiplex co-immunoprecipitation.

QMI was performed using immunoprecipitation (IP) and probe antibody panels targeting 21 proteins critical to CAR T cell signal transduction (online supplemental table S2).¹⁵ The dataset contained stimulated and unstimulated, CD4 and CD8 cells from 8 OR, 5 CRS, 5 NT and 5 NR patients (inclusive of 2 OR, 1 CRS, and 2 NT patients who only had CD8 cells available), for a total of 108 samples. Each of 200 binary interactions that passed quality control cut-offs was measured by approximately 60 individual bead reads, in duplicate, for each sample and control, resulting in a dataset of approximately 3.5×10⁶ bead reads. For each sample, the median fluorescent intensity of each interaction was derived from the bead distribution of the duplicate samples. Following ComBAT normalization to minimize batch effects,³⁰ principal component analysis showed a major effect of stimulation (figure 1B). CD19-stimulated samples clearly clustered separately from unstimulated and CD3-stimulated samples, which partially separated from each other. To further explore if samples would cluster by outcome type, we plotted samples by outcome type within each stimulation group; there was no clear separation of outcome clusters (figure 1C).

Data were then input into a modified³¹ correlation network analysis³² (CNA) script in R, which clusters individual interactions into modules that share similar patterns of change across the entire dataset. This technique, borrowed from the transcriptomics field, identifies modules of coregulated interactions that correlate with experimental variables, and whose combined behavior may be more robust than measurements of any individual interaction.^{33 34} A topological overlap matrix (TOM plot) clustered the interactions into five modules, arbitrarily color-coded for identification (figure 1D). The plot represents the coregulation of protein interactions over 108 separate measurements, giving it high statistical power, and highlights both expected (eg, TCR_TCR and TCR_CD3) and unexpected (eg, TRAF1_TRAF2 and CD28_LCK) correlations among interactions (online supplemental figure S4). Modules were then correlated with experimental variables using a module-trait correlation plot (figure 1E), which identified three modules of interest: “turquoise” correlated with CD19 stimulation (correlation coefficient (CC)=0.86, p=7 × 10⁻⁴⁴); “blue” correlated with CD3 stimulation (CC=0.57, p=1 × 10⁻¹³); and a “yellow module” correlated with clinical outcome, specifically the presence of CRS (CC=0.37, p=4 × 10⁻⁶). Notably, no modules were correlated with the other outcome groups (OR, NR and NT), nor were they correlated with protein concentration or cell viability (figure 1E). These data were our first indication that the protein interaction signatures of a CAR product may correlate with clinical outcome, and prompted us to look more closely at these modules of interest.

We first focused on the “turquoise” module, which was correlated with CD19 stimulation. 54 interactions were significant members (p<0.005) of the turquoise module, of which 21 changed by at least Log2FC>0.2 comparing CD19 stimulated versus unstimulated for at least one outcome group. A heatmap (figure 2A) revealed two classes of interactions, those that increased with CD19 stimulation and those that decreased. As in our previous study,¹⁵ “up” interactions included an increase in the amount of coassociated CAR_ZAP70 and the formation of a downstream signaling complex containing SLP76, GRAP2, LAT, and PI3K. “Down” interactions largely involved the CAR itself, apparently reducing coassociations with a variety of partners following stimulation. It is important to note that, because QMI detects antibody-accessible epitopes within dynamic, multiprotein assemblies, apparent decreases may result from epitope masking as additional proteins are recruited into a complex, as well as from physical dissociation of complexes.¹⁵ Regardless, a decrease in the measured abundance of CAR-protein interactions including CAR_TRAF1, CAR_TRAF2, and CAR_FYN, as well as a reduction of TRAF1_BIRC3 and ZAP70_TAK1 indicates regulation of each interaction by CD19 stimulation.

Figure 2. CD19 stimulation-associated module is similar across outcome types. (A) A heatmap of all interactions that reached CNA significance and fold-change thresholds. Colored boxes represent z-scaled single interactions measured in individual CAR T cell samples, organized into rows (specific interactions) and columns (individual samples), and ordered by outcome type (labels at top) and stimulation (labels at bottom). Blue=low relative abundance, yellow=high relative abundance. (B, C) Averaged scaled value of interactions that increased (B) or decreased (C) with stimulation. *Indicates p<0.05 compared with the no stimulation condition within outcome type; # indicates p<0.05 compared with the OR outcome type within stimulation condition, by two-way ANOVA followed by Sidak post hoc testing. (D–G) Average intensity of select interactions that decreased (D, CAR_CD3, E, CAR_FYN) or increased (F, CAR_ZAP70, G, PI3K_LAT) following stimulation. * as in B. (H, I) Node-edge diagrams of interactions that changed on CD19 stimulation in OR and CRS cells. Nodes represent proteins, edges represent interactions that changed with stimulation, edge color and thickness indicate the magnitude and direction of change, respectively (red=up, blue=down). Note the similarity of the diagrams between conditions. ANOVA, analysis of variance; CAR, chimeric antigen receptor; CRS, cytokine release syndrome; ICANS, immune effector cell-associated neurotoxicity syndrome; MFI, median fluorescent intensity; NR, non-responder; NT, neurotoxicity; OR, optimal response; CNA, correlation network analysis.

We compared the averaged scaled value of module interactions, which approximates the overall activation state of the turquoise/stimulation module, among outcome groups. Upregulated interactions were significantly increased by both CD3 and CAR stimulation, with no between-group differences (figure 2B). Downregulated interactions showed no change with CD3 stimulation, and a decrease in abundance following CD19 stimulation (figure 2C). Comparing CRS to OR, the unstimulated condition was significantly higher, and the stimulated condition significantly lower, in CRS, which led to a larger overall fold change in response to CD19 stimulation in the CRS condition.

The interaction most correlated with the stimulation module was CAR_CD3 (figure 2D), which likely represents both CAR abundance and availability of the CD3z probe antibody epitope,³⁵ which lies between the first and second ITAM and overlaps a basic-residue-rich sequence that may interact with the kinase LCK.³⁶ Importantly, CAR abundance was not different between outcome groups, implying differences in CAR expression are not driving signaling differences. The interaction that showed the largest difference between OR and CRS was CAR_FYN (figure 2E); the initial abundance of the interaction was similar, but the stimulated abundance was 50% lower in CRS cells, resulting in a greater fold change. Among upregulated interactions, CAR_ZAP70 was significantly upregulated by CD19 stimulation regardless of outcome type (figure 2F), while PI3K_LAT was upregulated by both CD3 and CD19 stimulation (figure 2G). Plots of these interactions by individual show that trends were not driven by individual outliers (online supplemental figure S5). Overall, these data demonstrate that regardless of outcome type, all CARs engage previously described¹⁵ signaling networks on CD19 stimulation, involving ZAP70, TRAF and other downstream mediators. CRS-producing CARs may have slightly increased dynamic range among a subset of interactions (figure 2C), but the overall behavior of the stimulation module did not correlate with outcome.

The blue module correlated with CD3 stimulation (CC=0.57, p=1 × 10⁻¹³) and preliminary analysis indicated that CRS-producing CARs showed blunted CD3-induced signaling as indicated by interactions including TCR_CD3, LCK_TAK1 or CD28_TAK1 (online supplemental figure S6 A–F). However, internal controls revealed a significant technical artifact (online supplemental figure S6 G and H), suggesting that module clustering may have been driven by lot-to-lot variation in the TAK1 probe antibody rather than biological variation. This module was therefore excluded from further analysis.

The yellow module, nested within the turquoise/stimulation module (figure 1D), correlated with CRS (CC=0.37, p=4 × 10⁻⁶), and contained 49 interactions at p<0.005. We filtered interactions to 14 which changed more than Log2FC>0.1 between the OR and CRS conditions, revealing a module centered on CD28 and TRAF signalosomes (figure 3A,B), which were unexpectedly correlated (online supplemental figure S4). In OR cells, module interactions increased with CD3 and decreased with CD19 stimulation (figure 3C). Interactions were significantly elevated in both unstimulated and CD19-stimulated CRS cells compared with OR cells (figure 3C). Moreover, in OR cells, CD3 stimulation led to a significant increase in module interactions, but in CRS cells, interactions did not increase over their initial values, suggesting tonic activation in the unstimulated condition. Examining each interaction individually, CD28_CD3z was significantly more abundant in the unstimulated CRS cells compared with the unstimulated OR cells, and the interaction decreased with CD19 stimulation in both groups (figure 3D). CD28_FYN also trended towards greater abundance in unstimulated CRS cells and was significantly increased for OR cells with CD3 stimulation but not for CRS cells (figure 3E), a trend also seen in CD28_LCK (figure 3F). TRAF1_TRAF2, representing the ratio of the two major TRAF species in complex, was elevated in CRS, and showed a decrease with CD19 stimulation only in the CRS condition (figure 3G). Plots of these interactions by individual show that trends were not driven by individual outliers (online supplemental figure S7). Overall, these data suggest that, while all manufactured cell products were able to effectively respond to CD19 stimulation, increased activation of CD28-mediated and TRAF-mediated signaling, particularly in unstimulated CAR T cells, was associated with CRS in our exploratory sample.

Figure 3. CRS-associated module is increased in CRS-producing CAR T cells. (A) A heatmap of all interactions that reached CNA significance and fold-change thresholds. (B) Node edge diagram of interactions in the yellow/CRS module highlighting CD28 and TRAF signaling nodes. Edge thickness indicates the relative difference between OR and CRS in unstimulated cells. (C) Averaged scaled value of module interactions. *Indicates p<0.05 compared with the no stimulation condition within outcome type; # indicates p<0.05 compared with the OR outcome type within stimulation condition, by two-way ANOVA followed by Sidak post hoc testing. (D–G) Average intensity of select interactions in the CRS module, highlighting CD28 and TRAF signaling. Statistics as in C. ANOVA, analysis of variance; CAR, chimeric antigen receptor; CRS, cytokine release syndrome; NR, non-responder; NT, neurotoxicity; OR, optimal response.

Both CD4 and CD8 cells responded to CD3 and CD19 stimulation with similar changes to their protein networks, and no CNA modules correlated with a “CD4 vs CD8” variable (figure 1E). However, we did notice that, when we manually separated CD4 versus CD8 cells in comparisons of module behavior, CD4 cells tended to show an increased magnitude of change in the yellow module, and increased differences between OR and CRS samples (online supplemental figure S8), consistent with literature suggesting CD4 cells are the predominant driver of CRS responses.^{37 38} Moreover, in a de novo analysis using only CD4 or CD8 cells (online supplemental figure S9 and S10), while there were slight differences in the specific interactions that reached statistical significance, the overall trends were similar: a CRS-correlated module that contained CD28-mediated and TRAF-mediated interactions was identified in both cell types.

Validation cohort

The exploratory experiment described above was conducted over many weeks, running one patient sample at a time with internal controls to minimize batch effects, because it was not possible to run 108 samples from four experimental categories simultaneously. However, QMI comparisons are optimal if matched conditions are run simultaneously on the same 96-well plate to internally control for batch effects and other noise artifacts. To validate differences between OR and CRS-associated CAR T cells, we performed a validation experiment on five additional pairs of CRS and OR samples. We focused exclusively on CD4 cells due to the larger apparent magnitude of differences in the yellow module (online supplemental Figure S8). To maximize assay sensitivity, we both normalized and maximized protein concentration among rare clinical samples by limiting comparisons to unstimulated and CD19-stimulated cells.

Principal component analysis again revealed a major effect of CD19 stimulation, but no clear effect of OR versus CRS (figure 4A). CNA again identified a module that was correlated with CD19 stimulation (CC=−0.95, p=3 × 10⁻¹⁰), and a second nested module that correlated with CRS (CC=0.46, p=0.04) and with CD19 stimulation (CC=−0.49, p=0.03) (figure 4B). Importantly, modular organization was similar to the initial dataset, confirming the robustness of CNA analysis of protein complexes (online supplemental figure S11). The “stimulation”/turquoise module showed strong changes with CD19 exposure that were consistent with those in the exploratory dataset (figure 4C,D). In addition, this panel included an additional anti-CAR antibody targeted to the external FMC63 scFv domain (exCAR), as well as SHP2, CBLB and Sharpin probe antibodies not included in the prior experiment (online supplemental table S2). Note that both CAR antibodies identified the same dynamic interactions, demonstrating antibody independence. Differences between OR and CRS conditions in the Stimulation module were few and minor (figure 4E,F), confirming that both outcome types responded similarly to CD19 stimulation.

Figure 4. Validation cohort confirms elevation of CD28 interactions in CRS. (A) PCA shows clustering of QMI matrices by stimulation. (B) Module-trait table shows the turquoise and green CNA modules correlated with experimental variables. CC (top number) and p value (lower number) for each correlation are shown. Colored boxes indicate higher CC as shown in the bar at right. (C) A heatmap of all turquoise/stimulation module interactions that reached CNA significance and fold-change thresholds. (D) Averaged scaled value of turquoise/stimulation module interactions. *Indicates p<0.05 compared with the no stimulation condition within outcome type by two-way ANOVA followed by Sidak post hoc testing. (E, F) Node-edge diagrams of interactions that changed on CD19 stimulation in OR and CRS cells. Nodes represent proteins, edges represent interactions that changed with stimulation, edge color and thickness indicate the magnitude and direction of change, respectively (red=up, blue=down). Note the similarity of the diagrams between conditions. (G) A heatmap of all green/CRS module interactions that reached CNA significance and fold-change thresholds. (H) Node edge diagram of interactions in the green/CRS module overlaid on the yellow module from figure 3 to show overlap. Black edges were identified in both experiments. (I) Averaged scaled value of green/CRS module interactions. *Indicates p<0.05 compared with the no stimulation condition within outcome type; # indicates p<0.05 compared with the OR outcome type within stimulation condition, by two-way ANOVA followed by Sidak post hoc testing. ANOVA, analysis of variance; CAR, chimeric antigen receptor; CC, correlation coefficient; CRS, cytokine release syndrome; OR, optimal response; PCA, principal component analysis; QMI, quantitative multiplex co-immunoprecipitation.

The green/CRS module contained 58 interactions at p<0.005, but when we filtered interactions that changed Log2FC>0.1 between the CRS and OR conditions, we were left with only 5 (figure 4G). Of these five, three were in the “yellow-CRS” module shown in figure 3, and one (TCR_SHP2) was not included in the earlier analysis (figure 4H). The green module was significantly reduced by CD19 stimulation, but was significantly higher in CRS both with and without stimulation (figure 4I). Note, however, that the involvement of TRAFs did not replicate. This replication of the green module in a head-to-head, optimized QMI comparison increases confidence that a network of interactions involving CD28, FYB, PKCtheta and the SRK family kinases FYN and LCK play a role determining CRS.

Outcome prediction by machine learning

We next asked whether QMI biosignatures could predict clinical outcomes. We combined the exploratory and validation QMI datasets described above for maximum statistical power, and developed a novel normalization approach, Pair Vector Centralization (PVC), to compensate for variations in protein abundance and to normalize samples across experimental conditions (figure 5A). PVC builds on prior work using mean vector centering and alignment techniques to reduce inter-sample variability in high-dimensional datasets.^{39 40} While these methods typically normalize entire groups of samples to a shared centroid, we extend this concept further by performing normalization on a per-pair basis, aligning samples in a common reference frame while preserving their relative biological variance. Fold-difference scaling was incorporated with PVC (PVC-Fold) because it consistently produced the clearest separation between groups. Following PVC-Fold, samples that clustered by stimulation type in PCA space (figures1B 4A) now clustered by outcome (figure 5B). We compared PVC-Fold to several standard normalization methods by computing performance metrics across multiple clustering and classification models, and found that PVC-Fold consistently outperformed standard models (figure 5C). We then applied a leave-one-out bootstrapping approach, training a model on 29 out of 30 samples and classifying the 30th. The model correctly classified the left-out sample in all 30 of 30 iterations, demonstrating robust clustering of CRS and OR groups after PVC-Fold normalization.

Figure 5. Machine learning model performance following PVC-fold. (A) PVC computational pipeline. (B) Principal component analysis of OR versus CRS samples following PVC-fold normalization. Lines between points indicate sample pairing. (C) Comparison of clustering performance following different normalization methods. Rows represent normalization methods, columns represent computational approaches, numbers indicate performance metrics, colors indicate good (green) or poor (red) performance. CRS, cytokine release syndrome; OR, optimal response.

We next asked, could we achieve similar performance by monitoring a smaller subset of critical interactions? To simulate a laboratory test, we paired an “unknown sample” with a “banked” OR sample at a similar protein concentration, and built a predictive model using a random forest approach. Features (individual interactions) were ranked based on their contribution to the model’s predictive performance (online supplemental table S3). Of note, CD28_FYN was ranked highly for all models. After calculating the rankings, we rebuilt the models using just the top 10 interactions for each, which improved performance metrics (F1_{complete model} = 0.808, F1_Top10 = 0.883). We repeated model generation using only CD4 or CD8 samples, and found that F1 scores increased with fewer interactions for CD4 (F1_{complete model} = 0.778, F1_Top10 = 0.944), but not CD8 cells (F1_{complete model} = 0.788, F1_Top10 = 0.788). Overall, PVC-Fold normalization and machine learning built models that could target a small number of critical interactions to retrospectively predict the clinical outcomes of our CAR T cell dataset.

Discussion

It is well established that each functional domain of the CAR affects the proteins recruited to the CAR signalosome.816,20 However, batch-to-batch differences in protein complex recruitment across individuals have not, to our knowledge, been described previously. In theory, autoimmunity-linked SNPs in key CAR interactors like TRAF1,²⁴ LCK⁴¹ or FYB⁴² could affect expression levels²⁴ or binding kinetics of individual interactors, acting as a driver of signaling variability. Our finding that batch-to-batch differences in CAR-CD28-Kinase interactions may predispose to CRS offers a CAR-proximal molecular correlate of CRS risk that could potentially be corrected by improved bioengineering.

CD28 signaling in BBζ CARs

CD28 acts as “signal 2” in TCR stimulation; it modifies and strengthens a TCR-major histocompatibility complex recognition signal in the presence of CD80/86.⁴³ In humans, an ill-fated clinical trial that administered an activating antibody to CD28 resulted in severe CRS in 6/6 patients, requiring extended hospitalizations.⁴⁴ Additionally, CD28 costimulatory CAR T cells produce more toxicities in patients than 41BB costimulatory T cells,⁴⁵ supporting the idea that CD28 signaling may contribute to CRS.

Structurally, CD28 signaling occurs on ligand recognition, but only when Ca2+ elevations due to local TCR activation release the basic motifs in the CD28 intracellular tail from sequestration in the plasma membrane.^{46 47} This allows the SRC family kinases LCK and FYN to bind at one of the formerly sequestered basic motifs, and phosphorylate YMNM and PYAP motifs in the CD28 tail. These phosphorylations both stabilize the CD28-Kinase interaction, and increase binding by many interactors including PKCtheta, PI3K, Slp76 and GRAP2.47,49 The identified CRS-associated modules included many of these interactions, suggesting that tonic CD28 activation may increase in CRS-producing BBζ CAR T cells.

But how might CD28 interact with a BBζ CAR? Endogenous CD28 dimerizes through two important sequences in its transmembrane domain.^{50 51} The PLAT02 CAR contains a CD28 transmembrane domain, complete with both dimerization motifs. Importantly, Muller et al⁵² demonstrated that the CD28TM domain of BBζ CARs dimerizes with endogenous CD28, resulting in a hybrid receptor that does not respond to CD80, but that lowers the overall amount of CD28 on the host cell surface. Moreover, stimulation through CD28 ligands alone is sufficient to drive CAR T cell proliferation in CD28-TM-containing BBζ CARs,⁵² providing a plausible mechanism through which CD28 signaling may influence outcomes. Indeed, QMI data shows the CAR in shared complex with CD28, detected reciprocally by IP CAR probe CD28 and IP CD28 probe CD3ζ (both interactions are members of the yellow/CRS module). The SRC-family kinases LCK and FYN were both detected in complex with CD28 and the CAR, and their abundance was tightly correlated, suggesting a shared binding mechanism. In natural killer cells, FYB interacts with both FYN and LCK in a complex that recruits PKCtheta and activates TAK1 to mediate cytokine production.⁵³ CNA suggests a similar complex coassociating with CAR-containing and CD28-containing complexes, which would be a parsimonious explanation for the CRS-associated modules observed here. Increased coassociation of this complex could increase TAK1 signaling to NFKB, increasing the abundance of cytokine transcripts produced in CRS-causing cells. Mechanistic follow-up studies to test this hypothesis are warranted.

A limitation of this study is that we are unable to directly test this hypothesis using an appropriate outcome measure; CRS is a uniquely human toxicity, and current animal or in vitro models fail to reproduce the relevant biology.54,56 Indeed, cytokine release assays did not differ between outcome groups in our hands (online supplemental figure S3). Future studies comparing the in-human performance of CARs with different TM domains, or CAR products otherwise engineered to limit exogenous CD28 signaling, would be needed to confirm or refute this hypothesis.

Potential of QMI to screen CAR products for negative side effects

Certainly, multiple factors besides the CAR-proximal signalosome contribute to CRS. Myeloid cells, particularly macrophages, contribute to CRS expression, and blockade of granulocyte-macrophage colony-stimulating factor and IFNγ during CAR-T cell treatment reduces macrophage activation and cytokine release in mouse studies.^{57 58} Variations in patient immunophenotype due to previous antigen experience can change the transcriptomic and epigenetic profiles of CAR-T cells, affecting performance.⁵⁹ Indeed, several groups have applied immunophenotyping or transcriptional biosignatures to predict clinical response, but have met with only limited success.^{10 60} Our machine learning algorithm shows early promise, correctly categorizing CRS and OR cells based on CAR signalosome state, despite our limited sample size. We speculate that the CAR signalosome integrates genetic, transcriptomic and epigenetic variation into a single system measurement, and that the CAR signalosome is so central to CAR function and CRS generation that the model yields acceptable predictive power. The next step will be to develop a quantitative risk score, which will require a larger cohort of patient samples with known outcomes, ideally using different CAR designs to establish generalizability.

Such a model would be desirable at two phases of CAR production. First, during the development stage, a detailed analysis of protein interactions engaged by preclinical CARs could inform CAR designers of which protein complexes their CAR construct is recruiting. Second, postmanufacture, but preinfusion, a CRS score could predict which batches of CAR T cells may be likely to cause CRS, and allow for increased monitoring, prophylaxis, or even remanufacture of potentially dangerous products. Perhaps most importantly, our finding that CRS-producing CARs differentially activate signaling modules in a CAR T cell-autonomous fashion implies that with improved bioengineering of the receptor itself, we could limit undesirable side effects and advance CAR T cells to a first-line cancer therapy.

Materials and methods

Antigen presenting cell culture

Parental human K562 cells (ATCC #CCL-243) and K562 cells expressing either CD19 or OKT3¹⁵ were cultured in Roswell Park Memorial Institute (RPMI) medium containing 10% heat-inactivated fetal bovine serum, 1% Hepes, 1% GlutaMAX supplement (Thermo Fisher catalog no. 35050061), and 1% penicillin-streptomycin.

Patient sample preparation

The primary criterion for sample selection was clinical outcome (OR, NR, NT, CRS), defined as: OR=MRDnegCR in the absence of CRS or NT, defined as CRS and NT grades of 2 or less. CRS=MRDnegCR with CRS grade of 3 or 4, but an NT grade of 2 or less. NT=MRDnegCR with CRS grade less than two but an NT score of 3. NR=did not produce a best response of MRDnegCR regardless of toxicity. In the event that multiple cell vials were available for a given outcome type, we prioritized samples based on overall sample availability (greater number of vials remaining) and absolute cell count (higher numbers of cells per vial). Cryopreserved preinfusion product was stored in liquid nitrogen until preparation for QMI in cryopreservation tubes or CellSeal vials. The cells were thawed at 37C, centrifuged at 300×g for 10 min and resuspended in RPMI medium at a concentration between 2 and 5 million cells per milliliter of media. The cells were then rested for 2 hours in an incubator before further handling.

Antigen-presenting cell fixation

K562 cells were washed twice in ice-cold phosphate-buffered saline (PBS), and then fixed for 30 s in 4 mL of 0.1% glutaraldehyde in PBS. The reaction was immediately quenched by 16 mLs of 200 mM glycine. Cells were washed twice in PBS and resuspended at an appropriate concentration for downstream use (10–20 million/mL).

Flow cytometry

CD19 and OKT3 presence on K562s were quantified via flow cytometry prior to CAR stimulation. CD19-expressing K562s and parental K562s were stained with 1:200 anti-CD19 antibody (BioLegend clone no. HIB19). OKT3-expressing K562s and parental K562s were stained with 1:200 anti-mouse IgG (Thermo Fisher catalog no. A-21235). Both stains were incubated for 30 min on ice in the dark. After incubation, cells were washed twice in Fly-P buffer (100 mM NaCl, 50 mM tris, 1% bovine serum albumin, and 0.01 sodium azide (pH 7.4)). Cells were then analyzed for antigen presence on a flow cytometer (Novocyte).

Patient CAR-T cells were stained for transduction efficiency before stimulation. After the 2 hour post-thaw rest, a small sample from each patient cell culture was resuspended in Fly-P buffer and incubated with 1:200 anti-EGFR (BioLegend clone no. AY13) for 30 min on ice in the dark. Cells were then washed twice in Fly-P buffer before EGFR fluorescence was analyzed on the Novocyte.

CAR-T cell stimulation

After a 2-hour rest and flow confirmation of EGFR expression, thawed patient CAR-T cells were spun down at 300×g for 10 min, resuspended at 20–40 M/mL in ice-cold PBS and divided evenly into 1.5 mL Safe-Lock Eppendorf tubes. Fixed APCs were mixed with the CAR-T cells at a 2:1 effector:target (E:T) ratio, on ice. We used a 2:1 E:T ratio to maximize effector access to APCs to provide rapid simultaneous stimulation while minimizing APC protein content in the lysate. The cell mixtures were centrifuged at 300×g for 15 min at 4C. The PBS supernatant was removed, and the cells were agitated by wash-boarding against an uneven surface. The cell pellets were then warmed to 37C for 5 min to allow for stimulation. Stimulation was terminated by freezing in liquid nitrogen, and the frozen cell pellets were stored at −80C for downstream use.

Internal control preparation

A cryopreserved sample of healthy donor CAR-T cells from Seattle Children’s Therapeutics was thawed and rested for 2 hours. After rest, the cells were divided into 50 pellets, stimulated with fixed CD19-K562s for 5 min, and snap frozen in liquid nitrogen. These pellets were stored at –80C for downstream use. An internal control was included in each QMI run.

Meso Scale Discovery cytokine assay

CAR-T cells and fixed APCs were combined at an E:T ratio of 5:3 at a concentration of 1M effectors/mL and allowed to incubate in cytokine-free media for 18 hours at 37C, 5% CO2 in 96-well plates. Cells were then centrifuged at 300×g for 5 min. The media supernatant was removed and stored at −80C.

Cell culture supernatants were titrated for cytokine concentrations using a custom Human Biomarker V-PLEX for IFN-γ, IL-2, IL-10, and TNF-a from MSD. The samples were diluted at a 1:10 ratio for a total volume of 50 μL in the provided assay diluent and were plated in triplicate. Standards were resuspended and six serial dilutions at a 1:4 ratio were performed and were plated in duplicate per manufacturer recommendation. Standards and samples were allowed to incubate for 2 hours with orbital shaking at room temperature, then were washed three times with ELISA wash buffer (1 × PBS+0.05% Tween-20). Detection antibody solution was prepared in 2760 μL of assay diluent and 25 μL was distributed into each well. After a 2-hour incubation at room temperature with orbital shaking, the plate was washed three times with ELISA wash buffer. To read the plate, 150 μL of provided read buffer was distributed into each well, and plates were immediately read.

CAR-T cell lysis

For each experiment, the patient CAR-T cells and an internal control sample were moved from long-term storage at –80C onto dry ice, then allowed to thaw for 30 s each on wet ice. The cell pellet was lysed with lysis buffer (1% digitonin, 1× phosphatase inhibitor cocktail, 1× protease inhibitor cocktail, 1 mM sodium orthovanadate, and 1 mM sodium fluoride in 150 mM NaCl and 50 mM Tris, (pH 7.4)). Lysis was allowed to proceed for at least 15 min before the samples were centrifuged at 19,000×g for 15 min at 4C. The supernatants were then moved to another tube for IP and protein quantification.

BCA assay

BCA assay was performed according to the manufacturer’s instructions (Thermo Fisher, catalog no. A55865). Standards were prepared by creating a 20 mg/mL stock of BSA. Six 1:1 serial dilutions were performed to create the other standards, which were stored at –20C with a “blank” water sample. For each experiment, 3 μL samples of each standard, the blank, and the lysates from each cell pellet were pipetted in triplicate in a flat bottom 96-well plate. BCA assay reagent A and reagent B were mixed at a 1:50 ratio according to kit guidelines (Thermo Fisher, catalog no. A55865) and incubated at 37C for 30 min. After incubation, absorbance at 562 nm was read on a SpectraMax ID3 spectrophotometer and sample concentrations were calculated using a four-parameter logistic curve. For head-to-head comparisons between patients (the “validation” experiments), protein concentration was normalized based on the lowest concentration lysate using the lysis buffer as a diluent.

Quantitative multiplex co-immunoprecipitation

After lysates were obtained as described, a master mix of antibody-bound Luminex beads was prepared on ice. The master mix was distributed evenly into cell lysates and IP was allowed to proceed overnight with rotation at 4C. The next day, the samples were placed on a magnetic rack to isolate the immunoprecipitated protein complexes attached to the beads, which were then washed twice in Fly-P buffer. The beads were then distributed across a 96-well plate in duplicates for each probe antibody. Probes were mixed with the beads in the 96-well plate such that the final concentration was 2.5 ug/mL. The probe antibodies were allowed to incubate for 1 hour on an orbital shaker at 4C, then the plate was washed three times with Fly-P buffer using the Bio-Plex Pro II magnetic plate washer. Biotin-conjugated probes were then incubated with 1 ug/mL PE-conjugated streptavidin in Fly-P for 30 min with shaking at 4C. Another three washes were performed with the plate washer, after which the beads were resuspended in 125 μL of Fly-P buffer. The plate was shaken for 1 min before loading into a Bio-Plex 200 system customized to keep the samples at 4C. Data were recorded via the Bio-plex Manager Software, which reported both the class of bead (corresponding to the identity of the IP antibody) and PE fluorescence (corresponding to the intensity of the probe, the identity of which was defined by 96-well-plate address). Data in the form of IP(X)_Probe(Y) were exported using XML formatting for downstream analysis.

CNA analysis

Weighted correlation network analysis³² was performed with modifications, as previously described.^{15 31} Data were ComBat-normalized and samples for which multiple independent replicates were performed to validate assay consistency (three sets of samples) were merged to generate a single averaged data matrix. CNA was run to cluster interactions using standard settings. Soft power was selected using the lowest number that approximates scale-free network topology. Module membership was limited to interactions with a p<0.005 for experiment one and p<0.05 for experiment two, and interactions were filtered to those that changed by log2FC<0.1 for between-group comparisons, and log2FC<0.2 for within-group comparisons. Averaged scaled value of the module was computed using the z-scaled value for each interaction in the module across all samples. Statistical significance was evaluated by two-way analysis of variance followed by Sidak post hoc testing.

PVC-fold normalization

PVC is based on mean vector centering, a statistical technique where vectors are normalized relative to their mean or centroid to reduce intersample variability. Samples were paired based on matching for both stimulation type (K562, CD3 or CD19) and protein concentration. Difference vectors were calculated for each pair, and pairs were recentered at the vector midpoint, effectively translating samples into a common reference frame and eliminating differences introduced by protein levels. PVC equations are provided in supplementary material, and the code is available on GitHub (link provided prior to publication).

Predictive modeling

To evaluate whether PVC enhanced downstream classification, we initially applied the model in a known-outcome setting by pairing CRS and OR samples, to verify that PVC improved separability between conditions. However, clustering known classes is a retrospective test; it shows PVC can work, but not whether it helps predict. To test generalizability, we next simulated a realistic clinical use case: removing outcome labels and asking whether a model trained on PVC-transformed data could accurately classify unseen samples. Thus, the second model shifts from descriptive separation to predictive inference, and lets the model decide if a sample is OR or CRS without prior labels. This two-stage approach, first validating PVC on known outcomes, then testing its utility in blinded prediction, was essential to distinguish signal enhancement from artifact of the transformation.

To simulate a clinical prediction setting where the patient’s future CRS status is unknown, we created two types of sample pairs: (1) OR paired with CRS (representing known outcomes during training), and (2) OR paired with another OR (representing patients who never developed CRS). After normalization, we removed the “control” OR sample from each pair, leaving a dataset of samples with unknown outcomes. This simulates an “unknown” preinfusion CAR T sample coming into a testing lab. To identify consistently informative protein interactions, we trained the model using all available features and ran 34 iterations, leaving each sample out once.

To integrate these rankings across runs, we used the Borda count method—a rank-based aggregation approach that assigns scores to each feature based on its position in each ranking. Features that consistently ranked highly across multiple runs received higher cumulative scores, yielding a final, consensus ranking of feature importance.

We evaluated model performance using the F1 score, which is more informative than accuracy for imbalanced classification tasks because the F1 score balances false positives and false negatives, while accuracy could be misleading if the model just guessed the majority class right without learning anything useful. This approach better reflects clinical reality and emphasizes the model’s ability to detect CRS risk from subtle patterns.

Data availability statement

Data are available upon reasonable request.