Multi-omic single cell sequencing reveals discordant MRD and mechanisms of relapse in pediatric AML Free

Introduction: Pediatric acute myeloid leukemia (AML) is a biologically heterogeneous disease. Prognosis is determined by risk of relapse, and the most powerful prognostic marker is measurable residual disease (MRD) at the end of induction therapy. Current clinical standard is to aggressively treat MRD with the assumption that the major clone at time of MRD assessment will drive relapse. However, there is no clear evidence confirming that the major MRD clone predicts the driver of relapse. Further, the biologic mechanisms underlying relapse after chemotherapy in pediatric AML remain unknown. We utilized multiomic single cell (SC) sequencing of DNA, RNA, and cell-surface proteins to assess MRD, changes in gene expression, and clonal composition driving relapsed disease.

Methods: We analyzed serial samples from 25 patients treated on COG AAML1031. 23/25 patients had samples from diagnosis, end of induction (EOI), and relapse; 2 patients lacked EOI samples. We performed high-throughput SC DNA sequencing with simultaneous cell-surface immunophenotyping (DAb-Seq, Mission Bio Tapestri) and SC RNA sequencing with simultaneous cell surface immunophenotyping (CITEseq, 10X).

Results: Using CITEseq, we analyzed 128,558 cells (7,142 cells/patient); using DAb-seq, we analyzed 33,813 cells (3,381cells/patient)

CITEseq data was annotated with a novel two-step cell classifier system to identify leukemic blasts and allow for subtyping of leukemia by fusion gene expression. First, a k-nearest neighbors model trained on the Da Marra et al (2024) pediatric AML dataset was used to identify malignant cells. Next, we used viewmasterR trained with the Granja et al (2019) dataset to subclassify leukemia subtypes based on fusion expression. For each fusion subtype, we assessed changes in gene expression over time via SC gene set enrichment analysis (GSEA). We identified clear changes in gene expression. For example, in KMT2A-rearranged patients, at relapse, we identified enrichment for gene sets associated with MYC targets and decreased enrichment for gene sets associated with heme metabolism and inflammatory signaling (TNFA signaling, inflammatory response, complement and TGFB signaling). Recent work from our group and others demonstrated selection for monocytic differentiation and RAS signaling activation at relapse on targeted therapy such as FLT3 and BCL2 inhibitors. In contrast, in this chemotherapy-treated cohort, our analysis demonstrated decreased enrichment for gene sets associated with activated RAS signaling and monocytic differentiation at relapse.

Using DAb-Seq, we reconstructed the genetic clonal architecture present at each time point. To do this, we developed a novel method for unbiased identification of genetic clones by clustering co-occurring mutations across timepoints. This method applies k-means clustering to time-series variant data and ranks variants by their importance, enabling the identification of biologically meaningful events. Using this method, we identified mutation-based MRD in 40% of patients, compared to 9% of patients with MRD by conventional flow-cytometry, suggesting DAb-seq is a more sensitive method of MRD detection. Importantly, the dominant genetic clone at the time of MRD assessment was not always identified as the dominant clone at the time of relapse. For example, one leukemia at diagnosis had a majority NPM1/ GATA2/ FLT3 co-mutant clone. At EOI, though clinically called MRD negative, SC DAb-seq identified a dominant leukemic MRD clone with SH2B3, EZH2, and ETV6 co-mutations. At relapse, however, there was an outgrowth of a WT1 mutant clone. In all patients, a pre-existing clone expanded at relapse, indicating that relapse is driven by selection for chemotherapy resistant clones present since diagnosis.

Conclusions: Using multiomic SC sequencing of a well-annotated cohort of newly diagnosed pediatric AML patients treated on the COG AAML1031 trial, we observed distinct changes in gene expression programs of AML blasts between diagnosis and relapse. We noted that relapse is driven by outgrowth of genetic clones pre-existing chemotherapy. Intriguingly, while presence of MRD may predict future relapse, the dominant clone at the time of MRD was not always the dominant driver of relapse. These data suggest that factors leading to clonal dominance at relapse are more complex than simple stoichiometry and that treatment of strategies purely targeted at MRD clones will be insufficient to prevent relapse.