CAR-T cell therapy has become one of the most powerful tools in hematologic oncology, achieving durable remissions in patients with relapsed or refractory B-cell malignancies including diffuse large B-cell lymphoma. However, its real-world benefit is far from universal. In non-Hodgkin lymphoma, up to 40% of patients who receive CAR-T infusions are later identified as non-responders, meaning they endure the full toxicity burden of a highly resource-intensive treatment without therapeutic gain. The ability to predict, before infusion, which patients are likely to respond has been an unmet need since the therapy's introduction.
Neurotoxicity as a compounding challenge: Layered on top of the response problem is immune effector cell-associated neurotoxicity syndrome (ICANS), a serious adverse event characterized by neurological symptoms ranging from confusion and tremor to life-threatening cerebral edema. High-grade ICANS (grade 3-4) occurs in approximately 25-32% of CAR-T recipients, and a meaningful subset of patients experiences both treatment failure and severe neurotoxicity simultaneously. Interventions for ICANS, including corticosteroids and IL-6 blockade with tocilizumab, are most effective when initiated early, making upfront risk stratification clinically valuable.
Why existing approaches fall short: Prior studies identified associations between specific T cell subsets in the CAR-T infusion product and clinical outcomes. Memory-like CD8+ T cells and central memory CD4+ T cells have both been linked to better responses, while T cell activation and exhaustion markers have been associated with ICANS. But these associations explain only a fraction of outcome variance. The fundamental problem is that infusion products are immunologically heterogeneous, and treating individual cell types or gene expression markers as independent predictors ignores the combinatorial, context-dependent dynamics that actually determine what happens in each patient.
Single-cell RNA sequencing (scRNA-seq) captures gene expression across thousands of individual cells simultaneously, offering unprecedented resolution into this heterogeneity. But converting that high-dimensional cellular landscape into actionable patient-level predictions is a major computational and conceptual challenge. CART-GPT was designed to bridge exactly that gap.
Before training any predictive model, the team needed a data foundation far larger and more comprehensively annotated than anything previously available. They assembled two complementary atlases. The first was a reference T cell atlas of approximately 216,000 T cells drawn from 17 human tissues including blood, bone marrow, kidney, spleen, and lymph nodes, sourced from publicly available scRNA-seq datasets hosted on the CELLxGENE database. This reference atlas captures the full range of natural T cell biology, from CD4+ helper subtypes to CD8+ memory cells to gamma-delta T cells and regulatory T cells.
The CAR-T-specific atlas: The clinically informative component was a manually curated CAR-T cell dataset comprising more than 1.12 million cells from 255 individuals, assembled from infusion product samples across multiple published studies. This is not peripheral blood drawn at some later time point. These cells come specifically from the residual material collected by washing the infusion bags after administration, preserving transcriptomic signatures that reflect the functional state of the actual cells delivered to each patient. Within this dataset, 782,558 cells from 199 patients were annotated with binary treatment response labels (responder vs. non-responder), and 343,061 cells from 56 patients were annotated with ICANS severity grades (0-4).
Quality control and batch correction: Assembling data from multiple independent studies inevitably introduces technical batch effects, where differences in laboratory protocols, sequencing platforms, and sample processing can distort biological signals. The authors applied scVI (single-cell variational inference), a deep generative model that learns latent representations of gene expression while simultaneously correcting for batch-specific variation. Standard quality control filters were also applied: genes expressed in fewer than three cells were removed, and cells expressing fewer than 100 genes were excluded. All datasets were processed into unified AnnData objects using the Scanpy framework, retaining layers for raw counts, normalized expression, and scVI-corrected embeddings.
The scale of this atlas, both the total cell count and the clinical outcome annotation depth, makes it the largest and most comprehensively labeled CAR-T scRNA-seq resource assembled to date. The authors also release the full atlas as a public community resource to accelerate future immunotherapy research.
CART-GPT is built on scGPT, a transformer-based foundation model pre-trained on large-scale single-cell RNA-seq data. The underlying transformer architecture treats gene expression profiles as sequences, applying multi-head self-attention to model gene-gene dependencies within each cell. Each gene's expression value is converted into an embedding vector of dimension d, and positional encodings are added to preserve gene ordering information. Twelve stacked transformer blocks process these embeddings, with attention weights computing for every gene pair within a cell. The resulting representations are passed to classification output layers, with the final prediction coming from a softmax function over the classification nodes.
Sequential fine-tuning strategy: Rather than training a single model from scratch, the authors applied a cascading transfer learning approach across three task-specific models. First, scGPT was fine-tuned on the reference T cell atlas (216,000 cells, 17 tissues) to produce TcellGPT, a general-purpose T cell subtype classifier. The gene embedding parameters were frozen during this step; only the transformer and classification layers were updated. TcellGPT was then further fine-tuned on the 782,558-cell response dataset to produce CART-GPT-response, with the output layer modified from 17 subtype classes to 2 nodes (responder vs. non-responder). Finally, CART-GPT-response was used as the starting point for fine-tuning on the ICANS dataset to produce CART-GPT-ICANS, with architectural modifications including a reduction in input embedding dimension from 512 to 64 and a four-layer transformer block with 64-dimensional hidden states.
TcellGPT performance: TcellGPT achieves overall T cell subtype annotation accuracy exceeding 90%, closely mirroring manual annotations in both rank order and relative abundance. For instance, Tnaive/CM_CD4 cells, correctly identified as the most abundant subtype at over 15%, and ILC3 cells, correctly identified as the least abundant at under 1%. Compared to CellTypist, a widely used general-purpose annotation tool, TcellGPT provides finer subtype resolution, distinguishing T follicular helper cells from broader helper T cell populations, while producing comparable distributions for shared populations such as Cycling T-NK, MAIT, and ILC cells.
This modular, sequential fine-tuning design is computationally efficient and biologically justified: each stage builds progressively more task-specific representations from a foundation that already captures general T cell biology, reducing the data requirements at each step.
One of the most technically novel contributions of the paper is the cell bagging strategy, which addresses a fundamental methodological challenge: CAR-T infusion products contain hundreds of thousands of individual cells, each of which the model classifies independently, but the clinician ultimately needs a single patient-level prediction. Naive aggregation, simply averaging cell-level predictions across all cells, performs poorly because not all cells in an infusion product are informationally equivalent for the task at hand. Some cells carry strong predictive signal for response or neurotoxicity; others are functionally irrelevant to the outcome being predicted.
Three-step aggregation: Cell bagging proceeds in three stages. In the first stage, cells with ambiguous classification layer outputs are excluded. Specifically, cells where the product of the two classification output node values is greater than or equal to zero (indicating that both outputs have the same sign, meaning the model is uncertain) are filtered out and do not contribute to the final score. This removes cells that the model cannot confidently assign to either the positive or negative class. In the second stage, each TcellGPT-predicted cell type cluster is assigned a contribution score by aggregating the softmax probabilities across all cells within that cluster. Cell type clusters are ranked by their average contribution to the positive classification node; only the top three ranking cell types with more than 100 cells per patient are retained for the final aggregation step.
Computing the bagging score: The final patient-level bagging score is defined as the difference between the aggregated softmax probabilities for the positive output node and the negative output node, averaged across the retained high-confidence cells from the top-ranked cell type clusters. A positive bagging score predicts the positive class (responder or high ICANS risk); a negative score predicts the negative class. This approach ensures that the patient-level prediction is driven by the most biologically informative, high-confidence subset of cells rather than by bulk averages diluted by uninformative populations.
The cell bagging strategy has a meaningful interpretability benefit as well. Because it selectively identifies which cell types are driving the prediction for each individual patient, the model reveals patient-specific cellular signatures rather than population-level averages. Two patients with the same clinical classification can arrive there through entirely different cellular mechanisms, and cell bagging captures that individual variation.
CART-GPT-response was evaluated using patient-level five-fold cross-validation on the 782,558-cell dataset from 199 patients. The model achieved a testing ROC-AUC of approximately 0.80, a strong result given the clinical complexity of CAR-T response prediction. On an independent dual-labeled validation cohort of 51 patients (where both response and ICANS annotations were available from the same individuals), the model correctly classified 46 out of 51 patients (90.2%), misclassifying only two responders and three non-responders. Statistically, bagging scores were significantly higher in responders than non-responders with a Mann-Whitney U test p-value below 10-6.
ICANS prediction performance: CART-GPT-ICANS was trained on 343,061 cells from 56 patients (46% high-grade, 54% low-grade) and evaluated using three-fold cross-validation. The model achieved a testing ROC-AUC of 0.781. On the dual-labeled validation cohort of 46 patients with ICANS annotations, CART-GPT-ICANS correctly stratified 39 out of 46 patients (84.8%), with three high-grade patients misclassified as low-risk and four low-grade patients misclassified as high-risk. The p-value for the bagging score difference between high- and low-risk groups was below 10-4.
Why individual cell types fail where CART-GPT succeeds: The authors systematically tested whether any single TcellGPT-predicted cell type could replicate CART-GPT's performance. For response prediction, single cell types produced AUC values ranging from 0.29 to 0.68 and adjusted Rand index (ARI) scores from -0.05 to 0.37. For ICANS prediction, single cell types produced AUC values of 0.29-0.70 with ARI values of -0.01 to 0.01. Unsupervised Leiden clustering similarly underperformed, yielding AUC of 0.31-0.67 for response (ARI -0.12-0.25) and 0.30-0.68 for ICANS. In both cases, individual cell type abundance and cluster composition were insufficient to discriminate outcome groups due to substantial intra-group heterogeneity.
Cell types identified as strong contributors to response prediction include Trm/em_CD8, Teffector/EM_CD4, and Th1/Th17 subsets. Notably, their overall abundance did not differ significantly between responders and non-responders across the dataset, confirming that predictive value comes from context-dependent functional behavior rather than population-level frequency.
Beyond prediction, the authors used CART-GPT's internal representations to identify gene programs associated with each clinical endpoint. Gene embeddings extracted from the fine-tuned model's output layers were used to construct a nearest-neighbor graph defining gene-gene relationships based on embedding similarity. From this graph, metagenes, clusters of co-regulated genes that collectively contribute to outcome prediction, were identified and grouped by their inferred contribution scores. These metagene programs were then visualized as gene regulatory networks (GRNs) using Cytoscape (v3.10.3).
Gene programs in responders vs. non-responders: The 12_SCORE metagene was prominently enriched in responders, featuring genes including GZMA, GZMK, GNLY, CCL5, and KLRD1. These genes encode granzymes, granulysin, and NK-cell activating receptors, collectively reflecting a cytotoxic and pro-inflammatory effector program consistent with effective anti-tumor immune activity. The 12_SCORE signal appeared most strongly in Tem/emra_CD8, Trm_Tgd, and Tem/em_CD8 subsets. In contrast, non-responders were enriched for the 31_SCORE metagene (GAS5, RPS4Y1, SNHG5, associated with stress-adaptive or exhausted T cell states), the 33_SCORE metagene (LEF1, TCF7, TRBC1, reflecting stem-like or memory-prone phenotypes), and the 62_SCORE metagene (ATP5F1A, ATP5F1B, ATP5MC3, pointing to altered mitochondrial metabolic activity).
Network topology differences: The GRN for responders was compact and modular, with FCRL6 (Fc Receptor-Like 6, typically expressed by cytotoxic lymphocytes) emerging as a central hub gene, suggesting a coordinated cytolytic regulatory program. The non-responder network was more diffuse and complex, with ICOS (Inducible T Cell Co-Stimulator) as the hub gene and prominent mitochondrial gene interactions, possibly reflecting a metabolically altered or functionally heterogeneous immune state.
Gene programs in high-grade vs. low-grade ICANS: For ICANS, high-grade patients showed enrichment in metagenes 31_SCORE, 65_SCORE (CCL5, IGHV7-81, linked to immune recruitment and B cell signaling), and 45_SCORE (CD69, NUCKS1, associated with T cell activation and chromatin remodeling). The high-grade ICANS GRN was densely connected, with GRINA (a calcium signaling and cellular stress gene implicated in neuronal survival) as the central hub. Low-grade ICANS was characterized by the 35_SCORE metagene (SRSF2, SELL, SPINT2, involved in mRNA splicing, leukocyte adhesion, and epithelial maintenance) and a sparser, less connected GRN with SRSF2 as the hub, suggesting a more transcriptionally homeostatic state.
CD19 CAR-T specificity: The most significant limitation is that the current model was trained almost entirely on CD19-targeting CAR-T datasets. CAR-T therapy is being developed against a growing list of targets, including CD22, BCMA, CD38, and FMC63, each involving different CAR construct designs, co-stimulatory domains, and immune activation patterns. It is not established that the T cell gene programs associated with response or ICANS in CD19 CAR-T therapy are generalizable to other targets. Expanding CART-GPT to incorporate multi-target CAR datasets will be necessary before the tool can be applied broadly across the evolving CAR-T treatment landscape.
Binary ICANS classification: CART-GPT-ICANS currently performs binary classification, separating grade 0-2 (low-risk) from grade 3-4 (high-risk). The full ICANS grading spectrum runs from grade 0 to grade 4, with clinically meaningful differences in severity and management across individual grades. The current binary approach aggregates a wide range of clinical presentations into two broad categories. As larger, more granularly annotated ICANS datasets become available, the authors anticipate extending the model toward multi-class stratification across the full severity spectrum.
Interpretability limits and the black-box problem: CART-GPT encodes complex, context-specific immune dynamics through its transformer attention layers, and the metagene analysis provides only a partial window into those representations. The model's full decision-making logic remains opaque. The gene programs identified through GRN analysis may reflect established immunological pathways, but they can also surface novel gene-gene associations with no current mechanistic explanation. The authors explicitly frame CART-GPT as a hypothesis-generating tool alongside its predictive function, acknowledging that many of its internal representations will require extensive experimental validation before biological conclusions can be drawn.
Dataset heterogeneity and validation scope: Despite the atlas's scale, the 56 patients in the ICANS training set represent a relatively small cohort for a three-fold cross-validation design. The response dataset (199 patients) is larger but still reflects data pooled from multiple published studies with varying clinical protocols and patient populations. Prospective, multicenter validation on independently collected cohorts will be essential before CART-GPT can be positioned as a clinical decision support tool.
Expanding to diverse CAR constructs and disease contexts: The most immediate priority identified by the authors is integrating scRNA-seq data from non-CD19 CAR-T therapies, including BCMA-targeting constructs used in multiple myeloma and CD22-targeting therapies in B-cell ALL. Each new CAR target introduces different co-stimulatory signaling (4-1BB vs. CD28 domains have distinct effects on T cell persistence and exhaustion), potentially reshaping which gene programs are predictive of response or toxicity. The framework's modular architecture and transfer learning design are well-suited for this expansion: each new target can be incorporated as an additional fine-tuning stage rather than requiring full retraining.
Refining ICANS grading resolution: Moving from binary to continuous or multi-class ICANS prediction is a stated future goal. As more institutional datasets with detailed grade-specific clinical annotations become available and are integrated into the atlas, the model could be retrained to discriminate among grades 0, 1, 2, 3, and 4 separately. This level of granularity would directly inform clinical decision thresholds, for instance, identifying which patients are at risk of grade 4 ICANS requiring ICU-level intervention versus grade 2 requiring outpatient corticosteroid management.
Generalization to broader immunotherapy contexts: The authors explicitly frame CART-GPT as a template for precision immunology beyond CAR-T therapy. The same framework, combining foundation model pre-training, sequential fine-tuning on outcome-labeled single-cell data, and cell bagging for patient-level aggregation, is directly applicable to predicting response to immune checkpoint inhibitors, bispecific T-cell engagers, and tumor-infiltrating lymphocyte therapies. It could also be adapted to characterize immune dysregulation in autoimmune diseases or infectious disease contexts where single-cell immune profiling generates large, phenotypically complex datasets requiring patient-level classification.
Integration with prospective clinical workflows: The ultimate translational goal is embedding CART-GPT into the pre-treatment evaluation process for CAR-T candidates. In practice, this would involve sequencing the leukapheresis product or infusion bag residuals, running the scRNA-seq data through CART-GPT before or at the time of infusion, and generating patient-specific response and ICANS risk scores to guide intensified monitoring, prophylactic interventions, or enrollment in alternative treatment protocols for predicted non-responders. Realizing this pathway requires prospective clinical trials where AI-guided decisions are compared against standard-of-care outcomes.