Abstract
Background
Relapsed B-cell Acute Lymphoblastic Leukemia (B-ALL) remains a leading cause of cancer-related death in pediatric patients. To prevent relapse and improve ALL outcomes, it is critical to identify cellular populations and features causing treatment failure. Detection of Minimal Residual Disease (MRD) following remission induction therapy carries a poor prognosis for relapse. Despite now using MRD to guide risk assignment and therapeutic decisions, what makes MRD cells distinct and treatment resistant at early phases of the induction therapy remains unknown.
Methods
We performed single-cell analyses on longitudinal samples of 48 pediatric B-ALL patients enrolled on the AIEOP-BFM ALL 2009 clinical trial (NCT01117441). A total of 146 samples collected at diagnosis (Dx), early timepoints of treatment (day 8, day 15) and relapse (Rx) were analyzed by CyTOF using a 42-parameter panel. ALL cells were aligned to their most similar B cell population across patients, and a Random Forest (RF) model was trained to predict treatment-resistant cells. Feature importance was assessed using SHAP (SHapley Additive exPlanations) analysis. We performed CITE-seq, measuring 15 proteins overlapping with our CyTOF panel, on a subset of the cohort (n=11 patients) collected at Dx, MRD timepoints and Rx for a total of 26 samples. Cell types were annotated using the BoneMarrowMap dataset (Iacobucci et al. Nat Cancer 2025) and molecular subtype classification using the ALL CatchR classifier. CyTOF and CITE-seq data were then integrated based on proteomic phenotype.
Results
Our cohort included a total of 20 standard risk, 10 intermediate risk, and 18 high risk patients and was enriched for patients who would relapse (25%). Leukemic cells were detectable at Dx and at early timepoints of treatment (Day 8 and Day 15) in 98% of the samples, (range 0.008% - 96.1%). Developmental classification of CyTOF data demonstrated that patients who would go on to relapse had an increased proportion of late pro-B cells (p=0.0335) at Dx and late pre-B cells at Day 15 (p=0.0073), compared to patients in CR. To further determine features associated with relapse, we used a Random Forest model, which successfully predicted relapse-associated cells with an average AUC of 0.79. The model assigned a significantly higher proportion of treatment-resistant cells and higher average probabilities of resistance to patients who eventually relapsed (p < 0.01), regardless of their clinical risk. Both the built-in RF feature importance and SHAP analysis highlighted a combination of phenotypic and functional markers for the prediction of treatment resistance. Specifically, the model is positively driven towards higher probability when cells express high levels of the phenotypic markers CD34, CD24, CD73 and CD304; while it is driven towards lower probability when cells express high levels of pSYK, CD10, pBTK, pS6. Compared to our B-cell developmental classifier's output, highly predicted cells were not enriched in a particular B cell population, suggesting that functional cell state, depicted by the RF model, in combination with cell type, might be associated with early resistance. To further corroborate these findings and to study transcriptional programs of resistant cells, we complemented CyTOF analysis with CITE-seq, revealing distinct classification patterns based on the molecular subtypes and outcome. Notably, even within the same molecular subtype (KMT2Ar), B-cell classification showed significant differences. Specifically, one patient exhibited subclonal populations with multi-lineage potential (HSC/MPP and MPP-MyLy) and experienced a lineage switch at second relapse. Integration of transcriptomic data demonstrates enrichment in several metabolic pathways in resistant cells, including nucleotide metabolism, glycolysis, and oxidative phosphorylation genes in MRD cells. Further details on these analyses will be presented.
ConclusionsThis study demonstrates the effectiveness of integrating high-throughput CyTOF data with CITE-seq to uncover the transcriptional profiles of functionally distinct phenotypes linked to early treatment resistance in pediatric B-ALL. While preliminary, these findings hold the potential to identify a functional signature of early resistance that will enhance our understanding of the mechanisms underlying treatment failure and will pave the way for future therapeutic strategies aimed at preventing relapse.
