High-throughput bead-based cytokine profiling reveals dynamic immune signatures and identifies candidate exploratory biomarkers in CAR-T therapy Authors: Keyuan Pu, Wanying He, Jiayi Feng, Hongyan Li, Chun-Lung Chan, Thomas S. Y. Chan, Eric Tse, Hongzhe Sun Published: 2026 ABSTRACT Chimeric antigen receptor T-cell (CAR-T) therapy has revolutionized the treatment of refractory B-cell malignancies, yet severe cytokine release syndrome (CRS) and unpredictable long-term outcomes remain major clinical challenges. Here, we developed a rapid, scalable, 96-well plate-based strategy for the high-throughput preparation of lanthanide-encoded classifier beads. This approach substantially increases production throughput, from several types per day to 96 types within 8 hours, enabling sensitive multiplexed cytokine profiling by mass cytometry. By applying this platform to 136 longitudinal plasma samples from 32 patients with diffuse large B-cell lymphoma receiving CD19-directed CAR-T therapy, we achieved fg mL−1-level sensitivity and broad dynamic ranges across 14 key cytokines. Longitudinal profiling revealed distinct cytokine dynamics between patients with low- and high-grade CRS. Baseline TGF-β1 showed a significant univariate association with subsequent high-grade CRS (p = 0.001) and inferior event-free survival (p = 0.02). These findings demonstrate the utility of a scalable bead-based platform for high-resolution immune monitoring in CAR-T therapy and identify baseline TGF-β1 as a candidate exploratory biomarker associated with CRS severity and clinical outcome. This approach provides a practical tool for sensitive and high-throughput cytokine profiling and deeper understanding of immune responses in immunotherapy. Introduction The emergence of chimeric antigen receptor T-cell (CAR-T) therapy has revolutionized the treatment of relapsed and refractory B-cell malignancies, offering a potential cure when conventional options have failed. Optimizing treatment to maximize durable responses while minimizing life-threatening complications is now a central priority for clinical translation. Despite impressive efficacy, long-term outcomes remain unpredictable as some patients achieve sustained event-free survival (EFS) whereas others experience rapid relapses. Additionally, cytokine release syndrome (CRS), a systemic inflammatory response driven by massive cytokine release (notably IL-6, TNF-α, and IL-1β), can escalate to multi-organ failure and death, frequently necessitating prompt intervention and intensive care.1 Identifying reliable biomarkers distinguishing patients likely to obtain long-term benefit from those at high risk of severe toxicity or early relapse remains a major unmet need in the field. This challenge is compounded by the current reliance on post-infusion monitoring (e.g., peak cytokine levels and clinical grading scales), which limits opportunities for risk-adapted interventions such as selection of different CAR-T products and CRS prophylaxis. A scalable, high-sensitivity platform for early immune profiling could enable personalized risk stratification, improve patient safety, and guide the design of next-generation CAR-T strategies. Cytokines serve as critical signaling molecules that orchestrate immune responses, inflammation, tissue repair, and cellular differentiation.2–4 Recent advances in suspendable particle-based analytical systems have enabled a transition from conventional static assay formats toward highly parallel, compartmentalized functional screening platforms. Lab-on-a-Particle (LoP) technologies, in particular, have demonstrated the utility of hydrogel or polymeric microcompartments as modular assay units compatible with standard laboratory analytical instrumentation such as flow cytometry and mass cytometry. This approach enables scalable, high-throughput, single-cell and molecular profiling workflows.5,6 In the context of cellular immunotherapies such as CAR-T, such platforms hold significant promise for complex networks governing therapeutic efficacy and toxicity. However, existing cytokine detection platforms impose significant limitations for clinical translation. Enzyme-linked immunosorbent assay (ELISA) offers high sensitivity but has inherently low-throughput, typically detecting only one or few analytes per assay.7,8 Bead-based flow cytometry platforms (e.g., Luminex xMAP) enable multiplexing but are constrained by fluorophore spectral overlaps, limiting simultaneous detection to hundreds of targets.9 Liquid chromatography-tandem mass spectrometry (LC-MS/MS) provides broad proteomic coverage but requires extensive sample preparation and long chromatographic separation and specialized expertise, severely restricting throughput.10,11 The SomaScan platform achieves impressive multiplexing (up to 11 000 proteins) but requires 4–6 weeks of turnaround time and high cost, limiting its utility in time-sensitive clinical settings.12,13 These constraints highlight the urgent need for a rapid, scalable, high-dimensional cytokine detection method that can support longitudinal immune monitoring and biomarker discovery in CAR-T therapy and other immunotherapies. Mass cytometry (MC) represents a powerful alternative to conventional flow cytometry integrating flow cytometry principles with mass spectrometry detection. By labelling antibodies with stable heavy metal isotopes instead of fluorophores, MC eliminates spectral overlap, enabling simultaneous detection of over 80 biomarkers at single-cell resolution with high sensitivity and minimal background noise.14,15 Bead-based multiplexed immunoassays represent an important translational implementation of particle-based analytical systems for soluble biomarker detection. These classifier bead technologies share conceptual foundations with emerging LoP platforms, in which functionalized microparticles serve as suspension-compatible assay units enabling high-throughput biochemical analysis in standard laboratory workflows. MC-bead immunoassays utilize polystyrene (PS) microspheres encoded with diverse lanthanide compositions as classifier beads.16 The theoretical multiplexing capacity of this system exceeds 1021 unique bead populations, far surpassing the needs of even the most extensive biomarker panels.16 However, realizing this potential has been constrained by the lack of scalable and reproducible methods for generating large libraries of distinctly encoded beads. Existing metal-encoded bead technologies primarily rely on two strategies. The first employs dispersion polymerization to incorporate lanthanide complexes during bead synthesis.16–20 However, these approaches typically require multi-day reactions under deoxygenated and high-temperature conditions, resulting in batch-to-batch variability and limited scalability.16 The second utilizes nanoparticle-assembly (“Lego-style”) strategies, which expand barcode diversity through sequential surface functionalization via affinity interactions.21 While effective, this method involves multi-step processes, including nanoparticle synthesis, ligand exchange, and iterative assembly. Collectively, these limitations hinder throughput, reproducibility, and cost-efficiency of existing bead-encoding systems, particularly for large-scale clinical multiplexing applications. To address the limitations of the current technologies, we developed a scalable 96-well plate-based strategy that enables rapid, parallel generation of uniform lanthanide-coded bead libraries for mass cytometry. This approach overcomes key manufacturing bottlenecks of conventional methods. By applying this platform to 136 longitudinal plasma samples from 32 patients with diffuse large B-cell lymphoma (DLBCL) receiving CD19-directed CAR-T therapy, we performed multiplexed quantification of 14 key cytokines with fg mL−1 sensitivity and a broad dynamic range. The resulting high-resolution profiles uncovered elevated baseline TGF-β1 showed strong univariate association with high-grade CRS (p = 0.001) and inferior event-free survival (p = 0.02). These observations indicate the potential of baseline TGF-β1 as an exploratory biomarker for risk stratification and suggest that pre-treatment immunosuppressive status may influence CAR-T therapy and outcomes. Results A scalable 96-well platform for high-throughput lanthanide-encoded bead preparation To enable high-throughput preparation and multiplexed analysis, we miniaturized all key steps of bead encoding, functionalization, and immunoassay into a 96-well plate format (Fig. 1). Defined combinations and concentrations of lanthanide complexes [Ln(DBM)3Phen, Ln = Ce, Er, Lu, Yb, Gd, Sm, Y] were dispensed into individual wells of a 96-well plate and air-dried to form solid metal identifiers, which serve as distinct encoding reservoirs. Emulsified dichloromethane (DCM) swelling solutions were prepared, aliquoted, mixed with microbeads and the preloaded lanthanide complexes, and incubated on a magnetic stirrer to generate batches of uniquely metal-encoded beads in parallel. Subsequent silica coating, piranha activation, and epoxy functionalization were carried out in small volumes to yield a uniform and densely functionalized surface suitable for covalent antibody conjugation. Finally, distinct antibody-modified classifier beads were pooled and deployed in a bead-based sandwich immunoassay within a 96-well filter plate for multiplexed cytokine detection. Compared with conventional dispersion polymerization-based bead encoding, which typically requires over 24 hours per batch and yields limited batch-to-batch reproducibility (CV = 10–15%),22 the present plate-based swelling strategy reduces preparation time to 6 hours while improving uniformity (CV < 1%). In contrast to nanoparticle assembly-based encoding approaches that depend on iterative synthesis and surface conjugation steps, the current method eliminates multi-step chemical assembly by decoupling barcode generation from both polymerization and nanoparticle construction. As a result, the workflow enables parallelization, with 96 bead populations generated simultaneously in a single processing cycle. The workflow for multiplex bead encoding and functionalization. Defined combinations of lanthanide complexes [Ln(DBM)3Phen] were dispensed into individual wells of a 96-well plate and air-dried to form solid metal identifiers serving as distinct encoding reservoirs. Emulsified dichloromethane (DCM) swelling solutions were then added to the wells containing preformed polystyrene (PS) microbeads. The plate was placed on a magnetic stirrer for batch-mode mixing, allowing the lanthanide complexes to diffuse into the expanded polymer network during swelling. Subsequent silica coating, piranha activation, and epoxy functionalization provide a uniform surface for covalent antibody conjugation. Antibody-modified classifier beads from each well were then pooled and deployed in a bead-based sandwich immunoassay for multiplexed cytokine detection. This integrated workflow enabled rapid, parallel preparation of up to 96 distinct bead populations in approximately 6 hours, with high uniformity, reproducibility, and compatibility with downstream mass cytometry analysis. The swelling strategy exploits the reversible expansion of polystyrene beads in organic solvents (Fig. 2a). When exposed to good solvents, polystyrene chains expand, increasing free volume within the bead matrix and allowing lipophilic lanthanide complexes to diffuse into the polymer network. Subsequent solvent exchange with a non-solvent (ethanol) collapses the chains, physically entrapping the complexes without requiring covalent bonding. This process occurs under mild conditions compatible with downstream biomolecule stability. Optimization of DCM concentration for efficient lanthanide encoding in polystyrene (PS) beads. (a) Schematic illustration of the swelling-based metal encoding process. PS beads are swollen in DCM emulsions containing lanthanide complexes [Ln(DBM)3Phen], allowing diffusion into the expanded polymer network. Solvent exchange with ethanol collapses the matrix, physically entrapping the metals. (b) (i) Representative SEM images and (ii and iii) corresponding bead size distributions showing the effect of DCM concentration (0.8–15% v/v in 0.75% SDS aqueous solution) on bead morphology and uniformity after swelling–deswelling cycles. The 6.4% DCM condition yielded the lowest size variations (average diameter is 1.95 µm; CV < 2%), while higher concentrations increased bead size but reduced uniformity (n = 3 independent experiments). (c) High reproducibility of lanthanide-encoded beads prepared using 6.4% DCM across five independent batches, as evidenced by consistent ion intensity signals in mass cytometry. (d) Mass cytometry ion intensity profiles for (i) 140Ce- and (ii) 166Er-encoded beads demonstrating that the 6.4% DCM condition facilitates robust and uniform incorporation of lanthanide complexes into PS beads. (e) Long-term stability of lanthanide entrapment: six-element-encoded beads stored in ethanol at 4 °C showed no detectable signal loss over 160 days, confirming firm physical trapping of the complexes within the bead matrix. Polystyrene swelling is highly sensitive to solvent polarity. We screened a range of solvents spanning polarity indices from 2.1 (toluene) to 4.0 (tetrahydrofuran) (Fig. S2). Dichloromethane (DCM, polarity index is 3.1) emerged as optimal, maintaining spherical morphology and uniform bead size throughout swelling–deswelling cycles. DCM's intermediate Flory–Huggins interaction parameter (χ = 0.45) and diffusion coefficient (D = 2.5 × 10−10 m2 s−1) provide balanced polymer–solvent affinity, avoiding the incomplete swelling seen with toluene (χ = 0.35; D = 1.5 × 10−10 m2 s−1) and the excessive, non-uniform swelling observed with THF (χ = 0.30; D = 4.0 × 10−10 m2 s−1). We further optimized DCM concentration by preparing emulsions with 0.8–15% v/v DCM in 0.75% sodium dodecyl sulfate (SDS) aqueous solution (Fig. 2b(i)). Sonication in ice water produced stable DCM-in-water emulsions stabilized by SDS. The bead diameter increased proportionally with DCM content (Fig. 2b(ii)). At 6.4% v/v DCM, beads expanded to an average diameter of 1.95 µm with excellent size uniformity (coefficient of variation, CV < 2%). Higher DCM concentrations produced larger beads but increased size heterogeneity. The 6.4% DCM condition was therefore selected for all subsequent encoding. This protocol enabled highly reproducible preparation of lanthanide-encoded beads across multiple batches, with minimum batch-to-batch variation (CV < 1%, Fig. 2c). Mass cytometry analysis confirmed efficient incorporation of the loaded metals, with beads prepared using 6.4% DCM emulsion exhibiting strong, characteristic ion signals (Fig. 2d). Long-term stability was assessed by storing six-element-encoded beads in ethanol for extended periods; no detectable decrease in the lanthanide signal occurred after one month (Fig. 2e), demonstrating robust physical entrapment and ensuring consistent classification performance during immunoassays and long-term storage. Characterization of the classifier bead library To demonstrate multiplexing capability, we designed a proof-of-concept library comprising six distinct classifier beads encoded with unique lanthanide combinations (Fig. 3a). Each bead type incorporated 20 µg of [Er(DBM)3Phen] as a common identifier for gating during data analysis. Variable lanthanides (Lu, Yb, Gd, Sm and Y) were incorporated at either 10 µg or 60 µg, generating unique metal signatures through this dual-encoding strategy (fixed common element + variable elements), facilitating both positive identification and quantitative distinction. Mass cytometry analysis revealed that different loading amounts of metal complexes produced clear differences in ion intensity (Fig. 3b). The medium CV for ion intensity across six independently prepared batches was <15%, highlighting the high-reproducibility of the swelling-based metal-encoding strategy. Reproducibility and multiplexing performance of lanthanide-encoded classifier beads prepared by the swelling-based strategy. (a) Schematic overview of the six-classifier bead type, each encoded with unique combinations of lanthanide complexes [Ln(DBM)3Phen; Ln = Er, Lu, Yb, Gd, Sm, Y]. Dosages of variable metals were set at 10 µg, 20 µg and 60 µg, while 20 µg Er(DBM)3Phen served as a common identifier for gating. (b) Normalized ion intensity profiles from mass cytometry showing dose-dependent signal responses for each lanthanide across the six bead types. The intensity was normalized to the highest value for each element, revealing clear differentiation between dosage levels with low batch-to-batch variation. (c) Isotopic dot plots demonstrating unambiguous separation of the six-classifier bead. Initial gating on the common element (166Er and 167Er) identified all beads, followed by pairwise projection of variable elements (e.g., 167Er vs.175Lu and 167Er vs.89Y), confirming excellent population discrimination and high reproducibility across batches. (d) t-SNE dimensionality reduction map visualizing the clear separation of 40 representative classifier bead types prepared in a single batch. Each cluster corresponds to a unique lanthanide combination and dosage, with minimal overlap and robust differentiation based on metal composition. Owing to the visual complexity, the complete 96-bead library is presented as a heatmap in Fig. S4. Two-dimensional dot plots revealed clear separation of all six bead populations based on their distinct metal combinations and content (Fig. 3c). Initial gating of the dual isotopes of the common element (166Er and 167Er) identified all classifier beads, followed by projection onto dot-plot diagrams of variable elements (e.g., 167Er vs.175Lu and 167Er vs.89Y) for population discrimination. Overlapping signals from beads prepared across multiple batches further confirmed the reliability and consistency of the 96-well plate-based swelling method. To further validate scalability, we prepared 40 distinct classifier bead types in a single batch (Table S1). All 96 bead types could be reliably distinguished using an event-matching strategy based on pre-gated lanthanide intensity populations. Briefly, individual FCS files corresponding to each element–dosage combination (Y, Sm, Gd, Lu, and Yb; coded as −, +, or ++) were generated through manual gating (Fig. S3). Events from these pre-gated populations were then matched back to the balanced dataset using exact channel-value correspondence, enabling assignment of element–dosage profiles without threshold-based classification. These profiles were subsequently compared with a predefined bead design matrix to determine bead identities. Events with profiles not represented in the design matrix were retained as unmatched and were not reassigned, thereby preserving data integrity. The expanded 96-bead analysis demonstrated a low overall misclassification rate of 4.8% (Fig. S4), confirming the robustness of the encoding and classification strategy. The heatmap and t-SNE dimensionality reduction map revealed clear, well-resolved clusters corresponding to each unique element combination, with minimal overlap (Fig. 3d). These results collectively establish the 96-well plate-based swelling platform as a robust, reproducible, and scalable method for generating high-dimensional classifier bead libraries suitable for multiplexed cytokine detection in clinical samples. Surface functionalization of metal-encoded beads for covalent antibody conjugation To enable covalent antibody attachment, we functionalized the surface of metal-encoded beads with epoxy groups, which react efficiently with primary amines under physiological conditions without requiring prior activation (Fig. 4a). The metal-encoded beads were first coated with a ∼17 nm silica layer using the Stöber method. Transmission electron microscopy (TEM) confirmed the homogeneous silica coating on the beads (Fig. 4b). To maximize the density of reactive silanol groups for subsequent epoxy modification, the silica-coated beads were treated with piranha solution (Fig. 4a). To directly quantify the increase in reactive sites following piranha treatment, we labelled the beads with 5-amino tetramethylrhodamine, an amine-reactive fluorescent probe. Confocal microscopy analysis showed a 1.5-fold increase in surface fluorescence intensity after piranha washing (Fig. 4c), confirming the successful generation of high-density reactive groups. This enhanced epoxy density enabled greater antibody immobilization, resulting in 5-fold improvement in the detection signal when paired with biotinylated antibody and streptavidin-gold nanoparticle in mass cytometry assays (Fig. 4d). Silica coating and epoxy-based biofunctionalization of metal-encoded beads. (a) Schematic workflow for surface modification. PS beads were first coated with a uniform silica layer using the Stöber method, followed by washing with piranha solution to expose high-density surface silicon hydroxyl groups. Subsequent reaction with 3-glycidoxypropyltrimethoxysilane (GOPTS) introduces bio-reactive epoxy groups, enabling covalent immobilization of primary amine-containing antibodies under physiological conditions without additional activation. (b) Representative TEM images confirming ∼17 nm silica coating on the bead surface after the Stober process. (c) Confocal microscopy analysis of amine-reactive site density. Piranha solution treatment significantly enhances mean fluorescence intensity of 5-amino tetramethylrhodamine labelling on the bead surface (1.5-fold enhancement, p < 0.05), demonstrating enhanced availability of reactive silanol groups for epoxy modification. (d) Mass cytometry signal response validation. Piranha-activated and GOPTS-functionalized beads showed a nearly 5-fold increase in immunoassay signal intensity when probed with biotinylated secondary antibodies and streptavidin-conjugated gold nanoparticle tags, confirming improved antibody conjugation efficiency and detection sensitivity. Larger beads offer increased surface area for antibody conjugation and higher internal volume for lanthanide loading. We compared beads of different bead sizes and identified 2.57 µm as the optimal diameter, balancing signal enhancement with uniform size distribution (Fig. S5). The size distributions of three different bead populations after swelling are shown in Fig. 5a, with corresponding isotopic dot plots demonstrating clear separation (Fig. 5b). These 2.57 µm classifier beads produced enhanced immunoassay signals (Fig. 5c) and high detection reproducibility (CV < 12%, Fig. 5d). As epoxy-functionalized beads are ready for immediate antibody conjugation, we evaluated the stability of their reactivity during storage. When stored in ethanol at 4 °C, the detection sensitivity remained unchanged for up to 160 days, as assessed based on the response to physiological analyte concentrations (Fig. 5e). In contrast, storage in PBS or water led to a progressive decline during extended storage and immunoassay applications (Fig. S6). Performance advantages of 2.57 µm classifier beads in mass cytometry detection. (a) Size distribution histograms of classifier beads prepared at three different final diameters (e.g., ∼1.95 µm, ∼2.13 µm, and ∼2.57 µm) after swelling–deswelling cycles, demonstrating the controlled size tuning achievable with the 96-well plate-based method. (b) Isotopic dot-plots showing the ion intensity signals of loaded lanthanides for beads of the three sizes. The 2.57 µm beads exhibited the strongest and most clearly separated signals, reflecting increased internal volume for metal loading and improved detection efficiency. (c) Signal response of biotinylated secondary antibody in mass cytometry assays, with 2.57 µm beads producing the highest immunoassay intensity, confirming enhanced antibody conjugation capacity and detection sensitivity due to greater surface area. (d) High detection reproducibility (n = 8) of 2.57 µm classifier beads across eight independent preparations (n = 8), as evidenced by a low coefficient of variation (CV < 12%) in ion intensity and the immunoassay signal. (e) Long-term stability of bio-reactive epoxy groups on 2.57 µm beads stored in ethanol at 4 °C. No significant loss in assay sensitivity was observed over 160 days when detecting physiological analyte concentration (<100 ng mL−1), demonstrating robust storage performance for clinical and research applications. Development of a high-throughput multiplexed cytokine immunoassay We developed a bead-based sandwich immunoassay for simultaneous quantification of 14 cytokines central to CAR-T therapy, including IL-6, IFN-γ, IL-10, IL-12, IL-15, GM-CSF, MCP-1, IL-1β, TNF-α, CRP, TGF-β1, IL-18, IL-8, and IL-4 (Table S2). Cytokine-specific classifier beads were prepared by covalently immobilizing capture antibodies on epoxy-functionalized beads via overnight incubation, followed by blocking with 2% BSA. The 14 bead populations were then pooled into a master mix. The assay workflow involved incubating the bead mixture with a sample or standard for 2 hours to capture analytes, washing via vacuum manifold to remove the unbound material, adding a biotinylated detection antibody cocktail, washing again, and finally labelling with streptavidin-conjugated gold nanoparticle amplification tags. All steps were performed in 96-well filter plates, enabling parallel processing of up to 96 samples per run (Fig. 6a). Development and performance of the bead-based sandwich immunoassays for multiplex cytokine detection in biofluid samples. (a) Schematic workflow of bead-based sandwich immunoassay. Distinct classifier beads encoded with unique lanthanide compositions are pooled in equal proportions to form a multiplex master mix. The assay proceeds by incubating the bead mixture with a sample or standard to capture target cytokines, followed by washing, addition of a biotinylated detection antibody cocktail, additional washing, and labelling with streptavidin-conjugated gold nanoparticle amplification tags. The entire process is conducted in 96-well filter plates, enabling parallel analysis of up to 96 samples with high throughput. (b) Representative calibration curves for four cytokines (IFN-γ, IL-6, TGF-β1 and TNF-α) generated using serial dilutions from 45 ng mL−1 (or 0.45 ng mL−1) to 2 fg mL−1. The platform demonstrated fg mL−1-level sensitivity and broad dynamic ranges spanning 5–8 orders of magnitude. Calibration curves for the remaining 10 cytokines are provided in Fig. S7. (c) Validation of detection reliability for TNF-α, TGF-β1 and IL-6 in clinical plasma samples, showing strong concordance between the bead-based platform and conventional ELISA measurement. (d) Comparative performance on serial dilution (3- and 9-fold) of a patient plasma sample. The bead-based platform produced dilution-proportional, quantifiable concentrations that aligned closely with expected dilution factors for all targets. In contrast, ELISA often underestimated TNF-α and TGF-β1, with values near or below its detection limit, underscoring the superior sensitivity and dynamic range of the multiplexed platform for complex, clinically relevant samples. Calibration curves were generated for each cytokine using serial dilutions from 45 ng mL−1 (or 0.45 ng mL−1) to 2 fg mL−1. Representative calibration curves for IFN-γ, IL-6, TGF-β1 and TNF-α are shown in Fig. 6b, with curves for the remaining 10 cytokines provided in Fig. S7. To further evaluate analytical performance, we determined the limit of detection (LOD) for each cytokine using blank-subtracted calibration curves, with the LOD defined as the lowest concentration producing a signal-to-noise ratio of 3 : 1 relative to blank controls. The resulting LOD values are summarized in Table S2. The platform achieved fg mL−1-level sensitivity and broad dynamic ranges spanning 5–8 orders of magnitude. Importantly, it retained an IL-6 recovery rate of 68.9% in complex matrices, demonstrating robust assay performance and reliable sensitivity even in the presence of biological interferents. Application of this method to longitudinal plasma samples from a DLBCL patient before and after CAR-T infusion revealed consistent temporal trends with the ELISA results (Fig. 6c). Direct comparison using serial dilution (3- and 9-fold) of a patient plasma sample further validated performance (Fig. 6d). While both methods yielded comparable IL-6 concentrations, ELISA under-estimated TNF-α and TGF-β1, often falling near or below its detection limit. In contrast, the bead-based platform produced concentrations that remained consistent across dilutions upon back-calculation, indicating reliable quantification with minimal matrix interference. Our platform's enhanced sensitivity, wide dynamic range, and minimal sample volume requirement (for a 14-plex measurement) enable accurate profiling of complex, dynamic cytokine responses in CAR-T therapy, addressing key limitations of single-plex ELISA and offering a substantial reduction in sample consumption for multi-analyte detection. Longitudinal cytokine profiling and biomarker discovery in CAR-T therapy Study design and patient cohort We applied the MC-bead platform to profile 14 key cytokines in plasma samples from 32 patients with diffuse large B-cell lymphoma (DLBCL) receiving CD19-directed CAR-T therapy (Table S3). All procedures involving human participants were performed in accordance with the Declaration of Helsinki and were approved by the Institutional Review Board of the University of Hong Kong/Hospital Authority Hong Kong West Cluster (HKU/HA HKW IRB; approval number: UW 24-219). Written informed consent was obtained from all participants. Plasma samples were collected at the baseline (day 0) and multiple post-infusion time points, yielding a total of 136 samples (sample distribution is summarized in Fig. 7a). Patients were stratified according to CRS severity using American Society for Transplantation and Cellular Therapy (ASTCT) consensus criteria:23 low-grade CRS (grades 0–1, n = 22 patients, and 91 samples) and high-grade CRS (grade ≥ 2, n = 10 patients, and 45 samples). This longitudinal cohort design facilitated direct comparison of cytokine dynamics between risk groups and enabled identification of potential predictive biomarkers. A representative profile from a patient who developed high-grade CRS illustrates the temporal cytokine fluctuations following CAR-T infusion (Fig. 7b), showing the platform's ability to capture dynamic immune responses in real-world clinical samples. Longitudinal cytokine profiling in DLBCL patients receiving CD19-directed CAR-T therapy. (a) Overview of sample collection and patient-level sample availability across time points. A total of 136 plasma samples were collected from 32 patients, with each patient contributing no more than one sample at a given time point. Patient-level sample counts and missing samples at each time point are summarized in Table S4. (b) Representative longitudinal cytokine expression profile from a patient who developed high-grade CRS. (c) Boxplot showing significantly elevated IL-6, CRP and TGF-β1 on day 7 post infusion in patients with high-grade CRS compared to those with low-grade CRS. A box-plot for the remaining 10 cytokines is shown in Fig. S8. (d) Longitudinal cytokine trajectory patterns stratified by CRS severity. Cytokine concentrations were normalized to individual baseline (D0) values and expressed as fold changes relative to the baseline. Individual patient trajectories are shown as background lines, while bold lines represent the median trajectory of the most prevalent pattern within each CRS subgroup. Trajectories were classified into four categories (early peak, late peak, decreasing, and stable) based on peak timing and direction of change. Percentages indicate the proportion of patients assigned to the corresponding most prevalent pattern. Complete trajectory classification results are presented in Fig. S10. (e) Low baseline TGF-β1 levels were associated with a significantly lower incidence of high-grade CRS and longer EFS following CAR-T therapy. Longitudinal cytokine dynamics and patient-level trajectories Longitudinal cytokine profiling revealed distinct temporal patterns between patients with high-grade CRS and those with low-grade CRS. At the baseline (day 0), TGF-β1 levels were significantly higher in patients who subsequently developed high-grade CRS compared to the low-grade CRS group (p = 0.001; Fig. 7c). This difference was independently confirmed by ELISA in the same samples (Fig. S9). During the acute CRS phase (day 7 post-infusion), three cytokines showed significant elevation in the high-grade CRS group: IL-6 (39.4-fold change, p = 0.014), CRP (4.70-fold change, p = 0.033), and TGF-β1 (8.58-fold change, p = 0.045; Fig. 7c). These early peaks align with the known pathophysiology of CRS, in which CAR-T cell activation triggers rapid cytokine release, macrophage activation, and acute phase response.24,25 To capture individual heterogeneity beyond group-level comparisons, we classified longitudinal cytokine trajectories into four distinct patterns based on peak timing and direction of change: early peak, late peak, decreasing, and stable (Fig. 7d and S10). This patient-level analysis highlighted marked differences between high- and low-grade CRS groups. In high-grade CRS patients, decreasing trajectories were observed for TGF-β1, MCP-1, and IL-1β in 30%, 40%, and 50% of cases, respectively, representing the most prevalent trajectory category for these cytokines. This pattern after CAR-T infusion may reflect initially elevated baseline or early cytokine levels, which have been associated with high-grade CRS and adverse clinical outcomes.24–27 However, IL-6, CRP, and IFN-γ exhibited early peak trajectories in 50%, 60%, and 50% of high-grade CRS patients, respectively, consistent with the expected CRS pathophysiology timeline where rapid CAR-T-cell expansion triggers acute cytokine release during the first week following infusion.24–27 In contrast, low-grade CRS patients more frequently displayed early peak trajectories for TGF-β1, TNF-α, IL-1β, and CRP in 40%, 43%, 32%, and 45% of cases, respectively. These longitudinal trajectories provide a granular view of CRS evolution and severity, enabling dynamic tracking of immune responses and supporting the identification of predictive biomarkers for clinical stratification. Baseline TGF-β1 predicts high-grade CRS and poor outcomes in CAR-T therapy Elevated baseline TGF-β1 levels were strongly associated with subsequent development of high-grade CRS (Fig. 7e(i)). To evaluate its clinical predictive value, patients were dichotomized using a median-split cutoff: 6.79 pg mL−1. Among those with high baseline TGF-β1, 60.0% (6/10 patients) developed high-grade CRS compared with only 9.1% (1/11) in the low TGF-β1 group. This difference was statistically significant as determined by the chi-square test (χ2 = 4.033, p = 0.044, Fig. 7e(ii)). Furthermore, baseline TGF-β1 was associated with event-free survival (EFS). For survival analysis, the optimal cutoff value was determined independently using the maximally selected rank statistics (MaxStat) method, yielding a threshold of 9.93 pg mL−1. Patients with baseline TGF-β1 below this threshold exhibited significantly longer EFS (p = 0.02; Fig. 7e(iii)). Collectively, these findings indicate that lower baseline TGF-β1 levels are associated with both reduced high-grade CRS incidence and improved clinical outcomes following CAR-T therapy. Discussion Recent advances in Lab-on-a-Particle (LoP) systems and secretion-based single-cell profiling platforms have shifted biological measurements from fixed instrumentation toward highly scalable, suspension-based assay architectures. Within this evolving landscape, bead-based mass cytometry assays occupy a unique translational position, enabling high-dimensional soluble protein quantification using standard cytometry infrastructure while maintaining compatibility with particle-based assay principles. However, their broader adoption has been constrained by manufacturing limitations in generating large and reproducible encoded bead libraries. The present work addresses this bottleneck by introducing a scalable 96-well plate-based encoding strategy for lanthanide encoding, thereby aligning bead fabrication with the scalability principles demonstrated in emerging LoP technologies. By decoupling bead synthesis from metal encoding through a swelling-based strategy, we eliminated the need for harsh, multi-day dispersion polymerization, achieving parallel preparation of 96 distinct bead types in approximately 6 hours under mild conditions compatible with downstream biomolecule stability. The resulting beads exhibited uniform size distribution, high lanthanide incorporation efficiency, excellent batch-to-batch reproducibility, and long-term stability in ethanol. Subsequent silica coating, piranha activation, and epoxy functionalization generated high-density reactive sites on the bead surface, enabling robust antibody immobilization and significantly enhanced immunoassay performance. In contrast to existing bead-encoding platforms, which rely either on in situ polymerization or stepwise nanoparticle assembly, the present approach provides a unified post-synthesis encoding framework that achieves simultaneous improvements in throughput, reproducibility, and operational simplicity. Compared with conventional NHS/EDC-based antibody conjugation workflows, epoxy-based immobilization provides simplified chemistry and improved long-term stability without activation steps. The down to fg ml−1 detection sensitivity exceeds that of conventional ELISA-based and several bead-based multiplex immunoassays, which typically operate in the pg mL−1 range, thereby significantly expanding the detectable dynamic range for low-abundance cytokines in clinical plasma samples. When applied to 136 longitudinal plasma samples from 32 patients with diffuse large B-cell lymphoma receiving CD19-directed CAR-T therapy, the platform enabled simultaneous quantification of 14 clinically relevant cytokines with fg mL−1 sensitivity and dynamic ranges spanning 5–8 orders of magnitude. Longitudinal profiling revealed distinct cytokine patterns between patients who developed high- and low-grade CRS. Univariate analysis identified elevated baseline TGF-β1 as significantly associated with subsequent high-grade CRS (p = 0.0011) and inferior event-free survival (p = 0.02). Nevertheless, these findings underscore the value of high-dimensional cytokine profiling in capturing the complex immune dynamics following CAR-T infusion. Baseline TGF-β1 is a biomarker that potentially reflects a pre-existing immunosuppressive milieu that disrupts normal immunoregulatory networks, including regulatory T-cell function and negative feedback loops restraining cytokine production. In this context, the sudden and massive activation of CAR-T cells could overwhelm residual regulatory capacity, resulting in compensatory hyperactivation and an exaggerated inflammatory response characteristic of severe CRS. This mechanism complements the classical view of CRS driven by pro-inflammatory mediators such as IL-6 and CRP28,29 and reveals the importance of baseline immunosuppressive status in shaping clinical outcomes. These findings align with prior work linking elevated baseline TGF-β1 to impaired CAR-T cell function and poorer prognosis in B-cell malignancies30,31 suggesting that TGF-β1 may serve as a previously underappreciated biomarker for risk stratification prior to CAR-T infusion. The platform's minimal sample volume requirement, rapid throughput (calculated at 6 min per sample), and compatibility with clinical workflows represent a practical advance for longitudinal immune monitoring in CAR-T therapy and beyond. By enabling high-resolution, longitudinal cytokine profiling from limited plasma volumes, it offers a scalable tool to support biomarker discovery and deepen our understanding of immune responses in immunotherapy. Although the present CyTOF-based LoP platform is primarily designed for high-dimensional multiplexed analysis, future development may also adapt the underlying assay strategy toward simpler, more rapid, and user-friendly formats for focused cytokine monitoring in clinical settings.32 The broader implications of this work extend beyond CAR-T therapy. This scalable bead technology provides a versatile framework for high-dimensional protein profiling in other disease settings where dynamic cytokine responses drive outcomes, including checkpoint inhibitor therapy, autoimmune disorders, and infectious diseases. Future development could expand multiplexing capacity and integrate automated pipelines for large-scale precision medicine applications. Several limitations of the current study should be acknowledged. The modest sample size (n = 32) and single-centre design limit the generalizability of the clinical observations and necessitate validation in larger, independent cohorts and ideally multi-centre cohorts. Furthermore, only one CD19-directed CAR-T product, tisagenlecleucel (Kymriah®), was evaluated, and the findings may not be directly applicable to other CAR-T constructs. While baseline TGF-β1 showed significant univariate associations with CRS severity and event-free survival, these observations remain exploratory and require confirmation through multivariate analyses in adequately powered studies. Future investigations should focus on validating TGF-β1 as a predictive biomarker and clarifying its mechanistic contribution to CRS pathogenesis, potentially through its influence on Treg activity, immune regulation, or pre-existing inflammatory states. Such studies will be critical for determining its potential role in risk stratification and personalized management of patients receiving CAR-T therapy. Conclusions In conclusion, we establish a rapid and scalable 96-well plate-based platform for the preparation of lanthanide-encoded classifier beads, enabling high-throughput, sensitive, and reproducible multiplexed cytokine profiling by mass cytometry. By overcoming key limitations of conventional bead-encoding methods, this approach provides a practical framework for generating diverse bead libraries and performing high-dimensional protein detection from limited clinical samples. Application of this platform to longitudinal plasma samples from patients with diffuse large B-cell lymphoma receiving CD19-directed CAR-T therapy revealed distinct cytokine dynamics associated with CRS severity and identified baseline TGF-β1 as a potential exploratory biomarker linked to high-grade CRS and inferior event-free survival. Although further validation in larger, independent cohorts is required, these findings highlight the clinical value of high-resolution cytokine monitoring for understanding immune toxicity, improving risk stratification, and guiding personalized management in CAR-T therapy and broader immunotherapy settings. Author contributions K. P., H. L., and H. S. conceived and designed the study. K. P. performed the majority of the experiments, analyzed the data, and wrote the manuscript. W. H. and J. F. performed experiments and assisted with data analysis. H. L. assisted with manuscript organization and revision. C.-L. C. performed experiments. T. S. Y. C. assisted with the organization and analysis of clinical information and reviewed the manuscript. E. T. and H. S. reviewed and revised the manuscript and contributed to experimental design. Conflicts of interest The authors declare no competing interests. Supplementary Material