1100 - A Multi-Model Machine Learning Framework Identifies Resistance-Associated Biomarkers in Axicabtagene Ciloleucel CAR T-Cell Therapy for Large B-Cell Lymphoma

Purpose/Objective(s): Although CD19-directed CAR T-cell therapy has yielded transformative responses in relapsed/refractory large B-cell lymphoma (LBCL), nearly half of patients fail to achieve durable remission. To address this critical gap, we developed a multi-model machine learning pipeline to interrogate single-cell RNA-sequencing (scRNA-seq) data and identify transcriptional determinants of therapy resistance.

Materials/Methods: We retrospectively analyzed CAR T-cell products from 28 patients with large B-cell lymphoma (LBCL) treated with axicabtagene ciloleucel across two academic centers (Harvard: n=11/8; Stanford: n=4/5; responders/non-responders). Resistance-associated biomarkers were identified through a three-layer computational framework utilizing single-cell RNA sequencing data. Our methodology integrated: 1) Leiden clustering with logistic regression to isolate high-predictive cell populations discriminating clinical outcomes; 2) differential expression analysis coupled with Multi-Layer Perceptron and SHAP interpretation to prioritize candidate genes with optimal discriminatory potential; and 3) PERMANOVA/XGBoost ensemble modeling to derive minimally sufficient biomarker signatures achieving robust classification metrics. This systematic approach revealed treatment-response determinants based on both transcriptional abundance and cellular prevalence within functionally distinct T-cell subpopulations.

Results: Non-responders displayed a pronounced shift toward exhaustion-like phenotypes and dysregulated signaling pathways. In the CD4? compartment, JUND and FYN co-expression emerged as a key predictor of resistance, with JUND?FYN? cells comprising a significantly higher fraction of CAR T cells in non-responders, indicating a hyperactivated yet dysfunctional transcriptional state. By contrast, in CD8? T cells, CD81? cells predominated among non-responders, suggesting that altered tetraspanin-mediated signaling may drive reduced efficacy. Both resistance signatures were characterized by elevated inhibitory receptor expression (e.g., CTLA4, PDCD1) alongside the maintenance of effector transcription factors, indicative of a partially exhausted but still activated state.

Conclusion: Using a robust multi-model machine learning pipeline, we identified cell-subset–specific biomarkers associated with axi-cel resistance in LBCL. Co-expression of JUND/FYN in CD4? T cells and the predominance of CD81? CD8? T cells define distinct transcriptional states linked to therapeutic failure. These findings offer clinically actionable targets to optimize CAR T-cell design and potentially guide combination strategies, such as with radiation or checkpoint blockade, to overcome resistance and improve patient outcomes.