1126 Rational CAR T cell design via attention-based multiple instance learning of infusion product scRNA-seq data

Background Chimeric antigen receptor (CAR) T therapies have transformed the treatment options for patients with hematological malignancies, yet more than 50% of patients do not derive a durable benefit. The variability in clinical outcomes is partially driven by heterogeneity in the composition and state of infused cellular products.Methods Here, we introduce tcellMIL, a biologically informed machine learning framework designed to predict patient-level CAR T cell therapy outcomes from infusion product single-cell RNA-sequencing (scRNA-seq) data and identify genetic engineering strategies predicted to enhance patient outcomes. tcellMIL represents each infusion product as a ‘bag of cells’ using multiple instance learning (MIL) and incorporates regulatory priors via the SCENIC algorithm-inferred transcription factor (TF) regulon activity. The TF regulon activity scores are denoised via a self-supervised autoencoder and combined with explicit batch encoding to improve cross-cohort generalization. An attention-based MIL mechanism identifies the most outcome-relevant sub-populations, providing interpretability at cell and regulon levels.Results To train tcellMIL, we assembled a scRNA-seq dataset comprising axicabtagene ciloleucel (axi-cel; Yescarta) CAR T cell infusion products from 64 patients with relapse/refractory large B cell lymphoma (LBCL) treated across publicly available or internal clinical cohorts spanning 3 U.S. institutions. By comparing the adjusted Rand index based on batch labels, we found that data represented as TF regulon activity scores had significantly reduced batch effects compared to normalized scRNA-seq data representation (p = 2.87 × 10^ (-5); Wilcoxon signed-rank test). In addition, computing TF regulon activity scores distilled noisy transcriptomics data to a set of 154 biologically interpretable features. In leave-one-patient-out cross-validation, tcellMIL outperformed (i) multiple pseudobulk classifiers, (ii) foundation models, and (iii) other MIL baseline methods in predicting response at 3 months for unseen patients (accuracy: 0.72 vs. 0.53-0.69; F1 score: 0.74 vs. 0.58-0.73).To identify specific genetic edits predicted to improve patient outcomes following axi-cel for LBCL, we applied attention-based model interpretation methods and leveraged tcellMIL to perform in silico perturbation screens. This approach nominated TBX21 (T-bet) overexpression as a consistently beneficial edit, improving predicted response probabilities across all analyses. Notably, this finding is supported by prior experimental studies demonstrating enhanced in vitro and in vivo efficacy with T-bet overexpression in CAR T cells for lymphoma1 and CD19-low leukemia,2 thus validating our approach.Conclusions In summary, this study offers a generalizable framework to leverage primary scRNA-seq datasets for predictive modeling and rational engineering of CAR T cell therapies.References Gacerez AT, Sentman CL. T-bet promotes potent antitumor activity of CD4+ CAR T cells. Cancer Gene Ther. 2018 Jun;25(5-6):117–128. doi: 10.1038/s41417-018-0012-7. Epub 2018 Mar 7. PMID: 29515240; PMCID: PMC6021366.Cimons JM, DeGolier KR, Burciaga SD, Yarnell MC, Novak AJ, Rivera-Reyes AM, Kohler ME, Fry TJ. T-bet overexpression enhances CAR T cell effector functions and antigen sensitivity. J Immunother Cancer. 2025 Apr 17;13(4):e010962. doi: 10.1136/jitc-2024-010962. PMID: 40246581; PMCID: PMC12007057.