Single Cell Multimodal Framework Identifies CD84 As a Leukaemic Stem Cell Target in Paediatric AML

Outcomes with CAR T-cells for acute myeloid leukaemia (AML) have been disappointing to date. Targeting of leukaemic stem cells (LSC) is crucial to prevent relapse, but their heterogeneous antigen expression has precluded the discovery of an ideal CAR target. Moreover, existing targets (e.g. CD33, CD123, CLL1) are also expressed on normal haematopoietic stem and progenitor cells (HSPC) –requiring consolidative transplant– and paediatric AML (pAML) LSC knowledge is primarily derived from adults. We generated a single cell pAML dataset, unique in that it is highly enriched for LSC and characterises both their transcriptome and surface proteome (CITEseq). With this high-resolution dataset, we develop a framework for LSC characterisation and identification of targets specifically tailored to pAML LSC.

The dataset comprises >210 000 high-quality single cells from 15 pAML patients (15 diagnostic, 2 serial post-induction and 7 serial relapse samples) and 5 healthy controls (3 bone marrow and 2 cord blood). Median age was 4 years (range 0.2-18) and cytogenetic heterogeneity was well represented: 4 KMT2Ar, 2 inv(16), 2 NUP98::NSD1, 1 t(8;21), 1 CBFA2::GLIS2, 1 t(6;9) and 1 t(7;12)+19 (cytogenetic information was unavailable for 3 patients). Cells were stained with a custom panel of 81 CITEseq antibodies against potential LSC markers, sorted and enriched for CD34 (median 70%). Transcriptome and surface proteome libraries were generated with 10X Genomics.

We first differentiated malignant from healthy cells using an occupancy score, presence of copy number variations and a 7-gene pAML blast signature (PMID: 37798266) and validated this annotation with a machine learning classifier (scANVI). Cell-types/states were obtained by mapping our single cell transcriptomes to published datasets. Our integrated analysis identified a cell cluster highly enriched in functionally derived LSC gene expression signatures, with contributions from all AML samples (median 4.7% LSC, IQR 1.3-6.7). These cells express a unique gene program (e.g., higher MSI2, RUNX1 and IGF2BP2 expression) that is distinct from adult LSC and adult healthy HSPC signatures, but is recapitulated in cord blood, potentially reflecting their cell of origin. This may allow preferential targeting of pAML LSC over mature HSPC.

By analysing the surface proteome (custom CITEseq panel) and leveraging transcriptomes to discover surface-protein-coding genes, we identified targets overexpressed in LSC vs healthy HSPC, at the transcript (181) and protein (37) levels. Among these is CD84, a recently proposed pAML CAR target (Pigazzi et al., ASH 2022), restricted to haematopoietic tissue. High expression in LSC (56% across all patients, IQR 37-77) and blasts (83%, IQR 56-92) and low levels in healthy HSPC (21%, IQR 18-23, p<2.2^-16) and lymphocytes are in contrast with CD33, CD123 and CLL1, which -while expressed in 51-65% of LSC-­ are expressed by 47-82% HSPC (p<0.05). These data suggest that CD84 ­­may be ideal to target AML LSC whilst preserving normal HSPC, overriding the need for consolidative transplant. Further, other LSC targets are enriched in distinct pAML groups driven by cytogenetics and cellular hierarchies (PMID: 35618837), paving the way for rationally designing combinatorial strategies that enhance efficacy and limit toxicity.

Here we present a unique dataset to profile vulnerabilities in pAML LSC and provide a framework to identify pAML LSC-specific targets and tailor CAR target combinations, with CD84 emerging as a potential near-universal LSC marker. We are currently validating and expanding our findings in silico by applying our paediatric LSC signature to other AML datasets and using predictive algorithms to infer protein levels from single cell transcriptomes lacking protein information. A systematic analysis to elicit optimal target combinations, as well as functional validation of candidate and previously undescribed paediatric LSC markers are underway.

*MOE and SK contributed equally