Artificial intelligence predicts genetic network disruptions to overcome CAR-T resistance in Mantle Cell Lymphoma

Abstract Introduction Chimeric antigen receptor (CAR)-T cell therapy has significantly advanced the treatment of relapsed/refractory mantle cell lymphoma (MCL), yet therapeutic resistance remains a critical obstacle. Recent advances in artificial intelligence (AI), particularly transformer-based deep learning models, offer powerful tools to address this challenge. When trained on transcriptomic data, transformer models can learn intricate gene-gene dependencies and simulate the effects of perturbing specific genes on cellular states. This enables the prioritization of actionable targets to overcome resistance in cancer therapy. Methods We analyzed a single-cell RNA sequencing (scRNA-seq) dataset comprising 38 samples from 15 MCL patients treated with CD19-targeted CAR-T therapy, including 30 sensitive and 8 resistant tumors. Each tumor cell's transcriptome was converted into rank-ordered gene tokens and used to fine-tune a pretrained transformer model. This allowed the model to learn MCL-specific gene regulatory patterns while generalizing across patients and technical variation. To identify resistance-associated regulators, we conducted in silico single-gene perturbations. By computationally deleting each gene from resistant cell embeddings, we measured the resulting shift in transcriptional state. Genes whose removal moved resistant cells toward a sensitive-like profile were classified as resistance-reversing; conversely, genes whose deletion disrupted the sensitive state were termed sensitivity-promoting. Functional validation of top hits was performed in vitro using shRNA or CRISPR-mediated gene silencing in resistant MCL cell lines, followed by co-culture with CAR-T cells. Results The fine-tuned transformer model achieved high classification accuracy (0.97), enabling robust modeling of transcriptional programs across heterogeneous patient samples. Embedding space projections showed distinct clustering of resistant and sensitive cells, and in silico perturbations identified two distinct gene sets: resistance-reversing genes, whose loss reprogrammed resistant cells toward sensitivity, and sensitivity-promoting genes, whose loss disrupted the sensitive phenotype and induced resistance. Pathway enrichment of sensitivity-promoting genes revealed a strong overrepresentation of immune-related processes, including MHC protein complex assembly, antigen processing and presentation, interferon gamma signaling, B cell activation, and type II interferon production. These pathways are essential for immune recognition and CAR-T cell–mediated cytotoxicity. CD19, the direct target of CAR-T therapy, ranked top among the top sensitivity-promoting genes. Its in silico deletion shifted embeddings toward a resistant phenotype (adjusted p < 0.05), consistent with clinical observations linking CD19 loss to relapse. The identification of these canonical genes and pathways validates our perturbation-based approach. Resistance-reversing genes were enriched for translational and ribosomal biogenesis pathways, including cytoplasmic translation, translational initiation, and ribonucleoprotein complex formation. These programs may support the increased biosynthetic and metabolic demands of resistant cells. Furthermore, upregulation of epithelial–mesenchymal transition (EMT)-related programs, while atypical in hematologic cancers, suggests enhanced cellular plasticity that could facilitate immune evasion or dissemination. Notably, TCL1A, a known oncogene and AKT co-activator, enhances cell survival and impairs apoptosis. S100A4 and S100A6, genes involved in cellular migration and metastatic behavior, have been shown to promote in vitro chemotaxis of B cells and are co-expressed with stem-like markers in chronic lymphocytic leukemia. Functional validation of these targets is ongoing. Together, these results demonstrate that our model not only recapitulates known resistance mechanisms but also reveals novel regulatory programs that contribute to CAR-T failure. Conclusion This study demonstrates the power of transformer-based deep learning models to uncover mechanistic drivers of therapy resistance in hematologic malignancies. Through context-aware in silico perturbation, we identified transcriptional programs driving immune escape and tumor persistence, as well as novel therapeutic targets. Our findings offer a scalable framework for AI-guided precision oncology with broad implications for overcoming resistance across cancer types.