Transcriptional Profiling Directs the Classification of Mixed Phenotype Acute Leukemias into AML, B-ALL or T-ALL

Introduction

Mixed phenotype acute leukemia (MPAL) is a rare, poor-prognosis subtype of acute leukemia, defined by immunophenotypic features associated with multiple hematopoietic lineages. It is unclear whether this mixed phenotype is a consequence of the misexpression of a few markers in committed cells (lineage infidelity) or the transformation of a multipotent cell of origin (lineage promiscuity). Although MPAL patients are typically treated with acute myeloid leukemia (AML) or acute lymphoblastic leukemia (ALL) regimens, optimal treatment choice is hindered by their lineage ambiguity. Therefore, we investigated the added value of mutational and transcriptional data to improve lineage assignment, currently based mainly on surface markers.

Methods

Twenty-three adult MPAL cases were retrospectively studied with an in-house computational pipeline to identify genetic lesions from RNA-seq data: point mutations, indels, copy number amplifications (CNA), fusion genes and structural variants. We validated the results of this analysis with a battery of standard diagnostic techniques, including karyotyping (n=14), multiplex ligation-dependent probe amplification (MLPA, n=21) and targeted DNA sequencing (DNA-seq, n=17). For classification purposes, we used gene expression data derived from RNA-seq to train a multinomial logistic regression model with lasso regularization, followed by a 20-fold cross-validation.

Results

We comprehensively profiled the mutational landscape of adult MPAL (n=23, median age = 54) using RNA-seq data. Our pipeline detected genetic lesions with a sensitivity >90%, as confirmed by karyotyping, MLPA and targeted DNA-seq. The median of mutations per patient was 4, including recurrent lesions in signaling pathways (83% of cases), transcription factors (74%), epigenetic modifiers (70%) and cell cycle regulators (65%). The most frequently mutated genes were CDKN2A (35%), DNMT3A (30%), IKZF1 (26%) and TP53 (26%). Almost 40% of the mutated genes were targets of existing drugs (either approved or in clinical trials), and every patient in our cohort carried at least one targetable mutation.

Next, we compared the gene expression profiles (GEPs) of the 23 MPALs with representative cases of AML (n=145), B-ALL (n=223) and T-ALL (n=85). In a principal component analysis, MPALs were not distinguished as a separate group, but clustered among cases of AML, B-ALL or T-ALL. This is consistent with previous studies showing that MPALs do not constitute a single clinical entity (Alexander et al. 2018). Accordingly, a multinomial logistic regression classifier trained with GEPs of acute leukemias segregated 19/23 MPALs into myeloid-, B- or T-lymphoid leukemia. These 19 patients harbored genetic abnormalities known to be associated with the classifier-assigned leukemia. Furthermore, MPAL GEPs were deconvoluted with single-cell transcriptional profiles of normal hematopoietic cells using CIBERSORTx, revealing enrichment for signatures of lineages corresponding to the leukemic type predicted by our algorithm: myeloid for AML, B-lymphoid for B-ALL and T-lymphoid for T-ALL.

Finally, the classifier was validated on an external MPAL cohort (n=24) with both RNA-seq and methylation array data available (Takahashi et al. 2018), accurately assigning 87.5% of the patients to a lineage matching their immunophenotypic and methylation profiles.

Discussion

In conclusion, we have developed a computational pipeline that employs RNA-seq data to accurately detect the majority of genetic lesions present in MPAL, while simultaneously improving their lineage assignment. Our study demonstrates that RNA-seq can serve as a comprehensive diagnostic tool, potentially replacing other molecular biology techniques currently in use. Moreover, we classified most MPAL patients into lineage-restricted leukemias, suggesting that their phenotype results from lineage infidelity in committed cells. Conversely, the few cases that could not be clearly assigned to a specific lineage are likely to derive from early multipotent progenitors. This improved subclassification could have implications for diagnosis and guide therapeutic decisions, but additional studies are needed to evaluate whether MPAL patients classified as AML, B-ALL or T-ALL respond better to their corresponding treatment regimens.