Acute myeloid leukemia stratifies as 2 clinically relevant sphingolipidomic subtypes

TO THE EDITOR:

Acute myeloid leukemia (AML) is an aggressive, heterogeneous disease with genomic subtypes that are increasingly treated differently. There is a growing interest in going beyond mutations and cytogenetics to stratify AML. Recent work has combined proteomics,¹ signaling,^2,3 or immunophenotypes⁴ with integrated genomic-transcriptomic measurements to improve AML risk classifications.⁵ Although invaluable as resources, such approaches can neither extend retroactively to existing repositories nor prospectively to new AML cases lacking these data types. We sought to develop a more extensible approach involving sphingolipids (Figure 1A), a family of bioactive molecules implicated in AML pathogenesis and therapeutic resistance^6,7 that differentially regulate cell proliferation,⁸ differentiation,⁹ autophagy,¹⁰ apoptosis,¹¹ and immune cell activation.¹² Sphingolipid abundances in AML vary heterogeneously and ratiometrically,¹³ prompting us to ask whether systematic sphingolipidomic profiling could meaningfully stratify patients with AML and common AML cell lines.

We quantified 33 sphingolipid metabolites by mass spectrometry in 213 primary AML samples, 30 human AML cell lines, and 6 normal CD34⁺ bone marrow samples after carefully controlling for cell purity and viability. Normalized sphingolipid profiles in AML cell lines and primary AML cases were highly dispersed (while remaining intermixed) and separable from normal samples (Figure 1B), motivating a pan-AML stratification. Consensus clustering of the normalized lipidomics data identified 2 sphingolipidomic clusters that were statistically robust (supplemental Figure 1A-C). The 2 clusters were equally populated with cell lines and primary samples, and neither was differentially enriched for common AML mutations (supplemental Figure 1D; supplemental Tables 1-3). In contrast, the clusters were divergent in their abundance of hexosylceramide (Hex) and sphingomyelin (SM) species (Figure 1C). Cluster 1 exhibited proportionally less Hex and more SM (Hex^lowSM^high), whereas cluster 2 exhibited more Hex and less SM (Hex^highSM^low; Figure 1D-E). Additionally, the Hex^highSM^low cluster was elevated in long-chain, C14 to C20 carbon chains relative to the Hex^lowSM^high cluster (supplemental Figure 1E). There were no detectable differences in ceramide (Cer), other long-chain or sphingoid bases (LCB, comprising sphingosine and its derivatives), or very long chain, C22 to C26 carbon sphingolipids (supplemental Figure 1F-H). The analyses supported 2 sphingolipidomic subtypes with biochemical states that were uncoupled from AML derivation or mutation status.

We next examined whether the subtypes differed in clinical outcomes. Complete data were available for 70 Hex^lowSM^high and 72 Hex^highSM^low cases sampled at diagnosis before intensive induction chemotherapy (supplemental Tables 2-3). Based on European LeukemiaNet (ELN) 2022 criteria,⁵ patients with Hex^lowSM^high had twice the failure rate of Hex^highSM^low (supplemental Figure 2A). The Hex^lowSM^high subtype also trended toward shorter event-free survival (EFS) and overall survival (OS), although the difference was not statistically significant (Figure 2A-B). To eliminate confounding caused by events that occurred before the first round of induction chemotherapy was complete, we excluded patients with EFS <20 days and observed significant differences in EFS and OS (supplemental Figure 2B-C). These results were robust to EFS threshold (supplemental Figure 2D) and suggested prognostic value.

To associate broader transcriptional differences with the subtypes, we collected RNA-sequencing (RNASeq) data for 33 primary AML samples and 30 AML cell lines with sphingolipidomic profiles. We appended transcriptomes from additional AML cell lines available through the Cancer Cell Line Encyclopedia,¹⁵ which were batch-corrected and merged with 2 clinical RNASeq repositories for AML: TCGA-AML¹⁶ and BeatAML¹⁷ (supplemental Figure 3A-B; supplemental Tables 4-6). For the cell lines and primary samples with sphingolipidomics, we identified 734 increased transcripts in Hex^lowSM^high and 1125 increased transcripts in Hex^highSM^low (Figure 2C; supplemental Table 7), including 6 enzymes with complex roles in sphingolipid metabolism (UGCG, FUT4, NAGA, SMPD3, GANC, and B3GALT1 are discussed in supplemental Table 8). The clinically favorable Hex^highSM^low subtype was enriched for hallmark gene sets related to immune activation (supplemental Figure 3C-D). In contrast, the clinically unfavorable Hex^lowSM^high subtype was enriched for leukemic stem cell (LSC) signatures^14,18-20 (Figure 2D and supplemental Figure 3E-H). We unmixed bulk transcriptomes computationally by CIBERSORTx²¹ using customized signatures for AML differentiation states^22,23 and repopulating activities.²³ Hex^lowSM^high was highly enriched for leukemia stem-progenitor cell (LSPC) states with enriched repopulating activity, whereas Hex^highSM^low was enriched for differentiated myeloid lineages (Figure 2E-F, supplemental Figure 3I-J). Next, we identified sphingolipidomics-profiled AML cases that had not been sequenced to be analyzed using flow cytometry. Statistical power was limited because comparatively few samples had enough remaining viable material for recovery from thaw and staining (supplemental Figure 3K), but we identified multiple Hex^lowSM^high (n = 11) and Hex^highSM^low (n = 12) samples to evaluate general trends. Standard flow paradigms for LSCs involved surprisingly few genes from the LSC gene signatures themselves, but CD34 was increased (log₂ fold-change = 1.56, P = .02). Mean CD34 immunoreactivity was unchanged between sphingolipid subtypes (supplemental Figure 3L) and consistently lower rather than higher in CD34⁺ subsets of the Hex^lowSM^high subtype (supplemental Figure 3M), arguing against overall changes in gene expression. In contrast, upon gating with standard LSC markers, we noted suggestive and specific increases in CD34⁺CD38⁻CD96^+/− cells for the Hex^lowSM^high subtype (Figure 2G-H and supplemental Figure 3N). Although more cases are needed, the available samples favored increased proportions of LSC-like cells in the Hex^lowSM^high subtype. We concluded that the 2 sphingolipid subtypes were more coupled to transcriptomic states and differentiation status than AML driver mutations (supplemental Figure 1D), suggesting different disease mechanisms.

For inferring sphingolipidomic subtypes from transcriptional states alone, we developed a support vector-machine classifier of Hex-SM status using the 284 most variable and differential genes between subtypes (supplemental Table 9). When trained on 60% of the samples with paired transcriptomics and sphingolipidomics (including both primary cases and AML cell lines), the classifier showed excellent predictive performance on the remaining 40% of samples (Figure 2I). We then used the classifier with the batch-corrected transcriptomic data from TCGA-AML and BeatAML to infer sphingolipidomic subtypes. For both repositories, the classifier predicted a balanced proportion of Hex^lowSM^high and Hex^highSM^low cases, supporting that neither subtype is rare (supplemental Tables 4-5). Consistent with our independent cohort (supplemental Figure 2B-C), patients inferred to be Hex^lowSM^high had significantly worse outcomes than those predicted to be Hex^highSM^low (Figure 2J-K). Importantly, sphingolipid subtype remained an independent predictor of outcomes in a multivariable Cox regression analysis that included age, sex, white blood count, mutation statuses of NPM1 and FLT3, and study as covariates (supplemental Table 10). The Hex^highSM^low subtype was enriched for patients in the molecularly–defined favorable or good group, the Hex^lowSM^high subtype was enriched for the adverse or poor group, and the intermediate-risk group was not detectably skewed by subtype (supplemental Figure 4A-B). Thus, we stratified cases by risk group and examined whether the sphingolipid subtypes differed in their clinical outcomes. OS of Hex^lowSM^high was similar to that of Hex^highSM^low in the favorable or good risk group, slightly worse in the adverse or poor risk group, and significantly worse in the intermediate-risk group (Figure 2L and supplemental Figure 4C-D). This extension to public AML data sets strengthens the conclusion that Hex^lowSM^high is a high-risk subtype with poor outcomes, especially for patients whose molecular risk classification is intermediate.

By examining AML from the perspective of sphingolipid metabolism, our work uncovers a stratification that eluded prior gene-based classifications.^14,17,24 Sphingolipidomic subtypes are embedded in a fraction of the transcriptome that is not prominent when clustering is performed in an unsupervised manner. Indeed, the sphingolipid profiles of AML cell lines are much more concordant with primary samples (Figure 1C), unlike when their whole transcriptomes are coclustered (supplemental Figure 5A-C). The LSC programs in the high-risk Hex^lowSM^high subtype are consistent with the importance of sphingolipid homeostasis in LSC maintenance.²⁵ Our sphingolipid-guided transcript classifier could be useful for identifying patients who are least likely to benefit from intensive induction chemotherapy or stem-cell transplant and thus most eligible for experimental therapeutics. Stratification of the intermediate-risk subgroup provides an additional tool to suggest therapy in this highly heterogeneous population. Many patients with intermediate risk are offered transplants in first remission; thus, prognostic biomarkers are of great interest. Future studies should confirm these findings and investigate the pharmacologic vulnerabilities of the 2 subtypes, which presumably arise from complex differences in sphingolipid metabolism (supplemental Table 8). Given the promise of targeting sphingolipid biology in AML^7,25 we envision that sphingolipidomic subtyping could contribute to tailored treatment selections for patients with AML who otherwise lack targetable alternatives.

Acknowledgments: The authors thank Samuel Haddox for help with RNA preparation for RNA sequencing (RNASeq), Emily Sullins for processing patient samples and cell lines, Galina Diakova for help with RNASeq data from the Oncology Research Information Exchange Network, UVA Biorepository and Tissue Research Facility (Research Resource Identifiers [RRID]: SCR_022971), Memorial Sloan Kettering Tissue Bank for assistance with sample processing, and the Flow Cytometry Core facility at the Penn State University College of Medicine (RRID: SCR_021134). The authors appreciate comments from Cameron Griffiths, Russell Hawes, and Kenley Ellis on the manuscript. This work was financially supported by the National Institutes of Health (NIH) under the National Cancer Institute (NCI) award number P01 CA171983 (to T.P.L. Jr, M.K., and M.C.), NIH/NCI Cancer Center support grant P30 CA044579 (to T.P.L. Jr), NCI R35 CA197594 (to R.L.L.), NIH/NCI Cancer Center support grant P30 CA008748 (to R.L.L.), NIH/NCI F31 CA271809 and UVA Robert R. Wagner Fellowship (to J.U.), National Science Foundation Graduate Research Fellowship NSF grant #1842490 (to W.A.F.), NCI K08 CA215317 and Edward P. Evans Foundation (to A.D.V.), NCI R03 CA252825 (to B.M.B.).

Contribution: B.B.P., S.-F.T., T.E.F., U.G., M.C., F.E.G.-B., M.K., D.C., D.J.F., K.A.J., and T.P.L. designed the research; B.B.P., S.-F.T., T.E.F., W.D., I.L., U.G., A.S., A.D.V., and D.C. performed the research; T.E.F. and F.E.G.-B. contributed in providing vital new reagents or analytical tools; S.-F.T., T.E.F., J.U., U.G., W.D., I.L., S.P., A.S., A.D.V., F.E.G.-B., D.C., and D.J.F. collected data; B.B.P., S.-F.T., T.E.F., J.U., U.G., J.J.P.S., W.A.F., B.M.B., M.S.T., M.C., F.E.G.-B., R.L.L., D.C., D.J.F., K.A.J., and T.P.L. analyzed and interpreted data; B.B.P., F.E.G.-B., and K.A.J. performed statistical analysis; B.B.P., S.-F.T., T.E.F., J.U., F.E.G.-B., D.J.F., K.A.J., and T.P.L. wrote the manuscript; all authors read and approved the manuscript.

Conflict-of-interest disclosure: A.D.V. is a scientific adviser to Arima Genomics. B.M.B. is the owner and founder of Tahosa Bio, LLC (Rapid City, SD). M.S.T. has received research funding from AbbVie, Orsenix, BioSight, Glycomimetics, Rafael Pharmaceuticals, and Amgen; is on the advisory boards for AbbVie, Daiichi-Sankyo, Orsenix, KAHR Medical, Oncolyze, Jazz Pharma, Roche, BioSight, Novartis, Innate Pharma, Kura, Syros Pharmaceuticals, Ipsen Biopharmaceuticals, Cellularity; has received royalties from UpToDate (for writing); is chair for the data and safety monitoring board for HOVON 156; is chair of the adjudication committee for Foghorn Therapeutics; has received honoraria from Northwell Health, Japan Society of Hematology, MetroHealth Cleveland, Ohio State University, American Society of Hematology; and is on the board for the American Society of Hematology. R.L.L. is on the supervisory board of Qiagen; is a scientific adviser to Imago, Mission Bio, Zentalis, Ajax, Auron, Prelude, C4 Therapeutics, and Isoplexis; receives research support from Ajax and Zentalis; has consulted for Incyte, Janssen, AstraZeneca, and Novartis; and has received honoraria from AstraZeneca, Roche, Lilly, and Amgen for invited lectures and from Gilead for grant reviews. D.J.F. has received research funding, honoraria, and/or stock options from AstraZeneca, Dren Bio, Recludix Pharma, and Kymera Therapeutics. K.A.J. serves on the scientific advisory board of BridgeBio. T.P.L. has received scientific advisory board membership, consultancy fees, honoraria, and/or stock options from Keystone Nano, Flagship Labs 86, Dren Bio, Recludix Pharma, Kymera Therapeutics, and Prime Genomics. The remaining authors declare no competing financial interests.

Mark Kester died on 20 July 2022.

Correspondence: Thomas P. Loughran, Jr., Department of Medicine, Division of Hematology & Oncology, The University of Virginia School of Medicine, Charlottesville, VA, 22908; email: TL7CS@virginia.edu; and Kevin A. Janes, Department of Biomedical Engineering, Biochemistry & Molecular Genetics, The University of Virginia, Charlottesville, VA, 22908; email: kjanes@virginia.edu.