Precision Lineage Deconvolution in Mixed Phenotype Acute Leukemia Using Cite-Seq Derived Hematopoietic Stages Identifies Lineage Dynamics Associated with Treatment Response

Introduction:

Mixed phenotype Acute Leukemia (MPAL) is challenging due to ambiguous biology and lack of formal therapeutic guidelines. Although MPAL itself is rare, mixed phenotype lineage expression in secondary AML (sAML-MP) occurs with greater frequency and is associated with poor clinical outcomes. Multimodal single cell sequencing with cellular indexing of transcriptomes and epitopes (CITE-seq) offers more precise lineage characterization and assessment of lineage dynamics. Here we demonstrate a novel, unbiased approach to quantitative multistage hematopoietic lineage assessment in MPAL and sAML-MP using CITE-seq data. In addition, we demonstrate that this approach can be translated from single cell to bulk analysis.

Methods:

Sample preparation: Four MPAL and 6 sAML-MP samples displaying T-cell and myeloid subsets were identified and flow sorted into T and myeloid subgroups. RNA-seq was performed on each subset and analyzed using DESeq2. ComBat was used for batch correction when integrating with disparate data sources.

Single cell lineage deconvolution: CITE-seq samples were processed using Seurat. A single cell deconvolution library was constructed using previously published hematopoietic stage clustering and labeling, and applied to CITE-seq data from five MPAL samples in the same dataset. Deconvolution algorithm parameters were optimized through a bootstrapping approach on in-silico sample mixtures.

Bulk and translated lineage deconvolution: Bulk lineage deconvolution was performed using a published library of 13 stages of hematopoiesis derived from healthy donors. This was applied to bulk RNA-seq data from a cohort of published MPAL samples as well as the 6 sAML-MP and 4 MPAL samples described above. A pseudobulk was created for each CITE-seq lineage stage. Bulk lineage deconvolution was then re-run on the MPAL and sAML-MP samples using this library.

Clinical Outcomes: A Wilcoxon signed-rank test was used to analyze differences among deconvolution-derived patient clusters, including response to induction therapy, specific gene mutations, and likelihood of transplant.

Results:

Lineage evolution with treatment: Deconvolution was applied on 17,848 MPAL cells and 35,038 control PBMC/BMMC cells. Antibody Derived Tag (ADT) features were consistent with lineage deconvolution stage assignments. CD38 was negatively correlated with HSC (-0.34, p < 0.001) and LMPP (-0.18, p < 0.001) while CD34 was positively correlated with both HSC (0.11, p < 0.001) and LMPP (0.27, p < 0.001) stages. MPAL 5 in particular was assessed at diagnosis and at two later time points. Unsupervised hierarchical clustering of all timepoints resulted in 10 clusters (Figure A). The diagnostic time point was largely contained in clusters 4 (LMPP character, N = 1545 cells, 37.13%) and 9 (HSC character, N = 756 cells, 18.16%). The majority of both relapse timepoints were within cluster 4, with 76.77% (N=357 cells) of T1 and 73.92% (N=1100 cells) of T2 located in it.

Bulk Unsupervised Analysis: Unsupervised analysis of sorted MPAL (N = 8), sorted AML-MP (N = 12), and unsorted MPAL (N = 24) bulk samples resulted in six distinct clusters, each with a distinct lineage signature (Figure B). Some were enriched for more differentiated stages such as monocyte or GMP (clusters 6, 5), while others were enriched earlier stages such as LMPP (cluster 2). Secondary AML-MP samples were notably enriched in cluster 4. We found significant associations between lineage clusters and mutations in RUNX1 (p < 0.01), FLT3 (p = 0.004).

Clinical Outcomes: Bulk RNA deconvolution cluster was significantly associated with complete remission (p = 0.0009). On average, individuals with an incomplete/no response to induction chemotherapy had a higher NK signature (p = 0.001), while individuals assigned to transplant had decreased LMPP character (p = 0.0495). Among the single-cell stages, decreased CLP 2 levels were associated with a poor response (p = 0.0224) while CMP / LMPP (p = 0.040) and CD4 N2 (p = 0.022) levels were decreased in patients who were able to undergo HSCT.

Conclusions: Precise identification of lineage signatures in mixed phenotype leukemias shows promise in identifying clinically meaningful biological subsets of these diseases. Prospective analysis of lineage-derived biomarkers should be performed to undertake identification of formal risk stratification and treatment schemas.