CART-GPT

CAR-T cell therapy can drive durable remissions in hematologic malignancies, but two outcomes remain hard to anticipate before infusion: whether a given patient will respond, and whether they will develop immune effector cell-associated neurotoxicity syndrome (ICANS). Both depend on the composition of the infusion product — a heterogeneous mixture of T cell states whose clinical consequences are not captured by any single marker. CART-GPT approaches this as a representation-learning problem, adapting a pretrained single-cell transformer to read the transcriptional makeup of an infusion product and translate it into patient-level risk.

Developed by researchers at Atrium Health Wake Forest Baptist and Houston Methodist with collaborators at the National Research Council Canada, CART-GPT is a fine-tuned extension of scGPT. Rather than a single predictor, it is packaged as three linked modules sharing one transformer backbone: TcellGPT for fine-grained T cell subtype annotation, CART-GPT-response for therapy efficacy, and CART-GPT-ICANS for neurotoxicity risk. Each is trained by transfer learning from the module before it, so the clinical heads inherit a representation grounded in general T cell biology.

The work also contributes the underlying resource: a manually curated atlas of more than 1.12 million CAR-T cells from 255 individuals, assembled from published infusion product studies and annotated with clinical response and ICANS grades.

#Key Features

• Three-module toolbox on one backbone: A single fine-tuned transformer supports T cell annotation, response prediction, and ICANS risk stratification, with each downstream head initialized from the module above it.
• Cell bagging for patient-level calls: Instead of averaging over all cells, the models aggregate only high-confidence, informative subpopulations, bridging single-cell predictions to a patient-level metric while suppressing uninformative signal.
• Fine-grained T cell annotation: TcellGPT resolves 17 subtype classes and distinguishes closely related populations such as T follicular helper cells that coarser reference-based annotators tend to merge.
• Interpretable subset attribution: Classification-layer contribution scores identify which T cell states drive each prediction, showing that neither efficacy nor neurotoxicity traces to a single cell type.
• A released CAR-T single-cell atlas: The 1.12-million-cell, clinically annotated atlas is offered as a community resource independent of the models themselves.

#Technical Details

TcellGPT was produced by freezing scGPT's gene embedding layers and fine-tuning the remaining components on a reference atlas of roughly 216,000 T cells spanning 17 human tissues drawn from CELLxGENE; it uses 12 transformer layers, 512-dimensional cell embeddings, and a 17-way output head. CART-GPT-response preserves that architecture, replacing the classifier with two output nodes, and is trained on 782,558 response-labeled cells from 199 patients under patient-level five-fold cross-validation, reaching a test ROC-AUC of about 0.8 (0.796 at the patient level). CART-GPT-ICANS is initialized from the response model and compressed — input embedding dimension reduced from 512 to 64, four transformer blocks with 64-dimensional hidden states, and a three-layer classifier — then trained on 343,061 cells from 56 patients, reaching AUC 0.78 under three-fold cross-validation. Batch effects across source datasets were corrected with scVI before training. On an independent dual-labeled cohort, the fold-1 checkpoints classified 46 of 51 patients correctly for response (90.2%) and 39 of 46 for ICANS (84.8%) without retraining.

#Applications

The intended use is pre-infusion risk assessment: profiling a CAR-T infusion product by scRNA-seq, annotating its T cell composition, and producing patient-level estimates of response likelihood and neurotoxicity risk to inform monitoring intensity and prophylaxis. TcellGPT is also usable on its own as a T cell annotation tool for immunology datasets beyond CAR-T. For translational groups, the released atlas provides a clinically annotated substrate for studying which T cell states in a manufactured product track with outcome.

#Impact

CART-GPT demonstrates that a general-purpose single-cell foundation model can be specialized into a clinically framed prediction toolbox, and that response and toxicity can be modeled jointly from the same infused-product representation. Its interpretation analyses argue against single-marker explanations of CAR-T outcomes, favoring combined contributions from multiple T cell states. Important limitations remain: the work is a preprint that has not been peer reviewed, the dual-labeled evaluation cohort comes from the same publicly sourced atlas rather than a separate institution, all cells derive from axi-cel infusion products, and the repository named in the preprint's code availability statement is not publicly accessible, so neither code nor weights can currently be obtained.