Deep learning-based tumor microenvironment profiling predicts CAR-t efficacy and guides combinatorial immunotherapy in B-cell lymphoma

Abstract

Abstract Introduction CAR-T cell therapy efficacy in B-cell lymphoma remains limited by treatment resistance. To address this fundamental challenge and elucidate underlying mechanisms, we undertook a comprehensive single-cell analysis of the CAR-T-tumor microenvironment interface. By leveraging five public scRNA-seq datasets spanning 1.48 million cells from CD19-targeted CAR-T patients, this study establishes an advanced computational framework to decipher the cellular ecosystem governing therapeutic resistance. Methods Methodologically, we developed three integrated innovations: (1) A biologically informed hierarchical graph convolutional network (GCN) modeling patient-specific tumor microenvironments as cell-cell correlation graphs for response prediction; (2) A novel Grad-CAM extension enabling quantitative interpretation of GCN outputs at single-cell resolution to establish gene-outcome relationships and resistance biomarkers; (3) GeneTransformer—a hybrid transformer architecture predicting transcriptional responses to genetic perturbations while resolving cross-platform heterogeneity via stochastic triple-masking during finetuning. Results Analysis revealed system-wide transcriptional dysregulation and extensive cell-subset-specific alterations distinguishing non-responders. Functional integration identified four core resistance modules: pro-survival signaling (BCL2/XIAP), metabolic rewiring (AMPK exhaustion), immune dysfunction (PD-1/TIGIT), and homeostatic collapse (Hippo/YAP-mediated stemness). Crucially, we exposed mechanistic crosstalk revealing actionable combinatorial targets—notably PI3K inhibition synergizing with anti-PD-1 blockade. Cross-population validation confirmed model robustness, with 84.65% (p<0.05) of reverse-engineered resistance genes exhibiting differential expression. The GCN achieved superior predictive performance (AUC=0.900). Conclusion This study establishes a transformative computational framework that deciphers the single-cell architecture of CAR-T resistance, providing a mechanism-guided blueprint for designing combinatorial therapeutic strategies. The GeneTransformer platform further enables optimization of individualized therapies through perturbation response prediction. Our findings illuminate fundamental resistance pathways while delivering clinically actionable targets for overcoming treatment failure in B-cell malignancies.