The proteogenomic subtypes of acute myeloid leukemia

Abstract

Acute myeloid leukemia (AML) is an aggressive blood cancer with a poor prognosis. We report a comprehensive proteogenomic analysis of bone marrow biopsies from 252 uniformly treated AML patients to elucidate the molecular pathophysiology of AML in order to inform future diagnostic and therapeutic approaches. In addition to in-depth quantitative proteomics, our analysis includes cytogenetic profiling and DNA/RNA sequencing. We identify five proteomic AML subtypes, each reflecting specific biological features spanning genomic boundaries. Two of these proteomic subtypes correlate with patient outcome, but none is exclusively associated with specific genomic aberrations. Remarkably, one subtype (Mito-AML), which is captured only in the proteome, is characterized by high expression of mitochondrial proteins and confers poor outcome, with reduced remission rate and shorter overall survival on treatment with intensive induction chemotherapy. Functional analyses reveal that Mito-AML is metabolically wired toward stronger complex I-dependent respiration and is more responsive to treatment with the BCL2 inhibitor venetoclax.

Keywords: BCL-2 inhibitor; acute myeloid leukemia; chemotherapy; mitochondrial oxidative phosphorylation; multi-omics data integration; proteogenomics; proteomics; venetoclax.

Figure 1:. Proteomic AML subtypes

Mass-spectrometry-based label-free quantification of proteins in bone-marrow-derived AML blasts from 177 untreated AML patients. A Survival probability stratified by ELN classification of cytogenetic risk. Overall, relapse-free and event-free survival are shown for the 177 patients belonging to the discovery cohort. B Unsupervised, hierarchical clustering of patient–patient distances. Rows and columns correspond to patients, with each entry in the heatmap quantifying the distance between two patients. The clustering is used to define six distinct clusters (C1–C6, color-coded). Annotations at the top of the heatmap show patient risk category according to the ELN and FAB classification, mutations and cytogenetic aberrations. C Cluster-wise up- and downregulation of gene set variation scores (GSVA) based on differentially regulated Gene Ontology (GO)-pathways (left heatmap; shown as red and blue respectively). Cluster-wise Spearman correlation coefficient of the proteome and the transcriptome for the annotated pathways (right heatmap, dots mark significantly correlated pathways with p<0.05, Benjamini–Hochberg-corrected). D Cluster-wise up- and downregulation of transcription factors. The top annotation color-codes DNA-binding domains. E Cluster-wise up- and downregulation of human kinases. The top annotation color-codes kinase family. See also Figures S1, S2,Tables S1, S2, and Data S1–S3

Figure 2:. The proteogenomic characteristics of AML

Multi-omics data analysis integrating proteome, transcriptome, mutations and cytogenetics. A 28 factors drive variation in all data layers. Rows represent latent factors; columns represent data modalities. Each square color-codes whether a factor is active in a specific data layer (red marks active, blue marks inactive). B Patient-specific enrichment scores for LF1-NPM1/HOX, where each line corresponds to a patient and color encodes the pathway- and patient-specific normalized enrichment score. MHC explains variation in a coordinated fashion for both proteome and transcriptome (top). HOX signaling is enriched in NPM1 mutant patients in the transcriptome (bottom; annotated with NPM1 mutation status). C LF1-NPM1/HOX scores color-coded by NPM1 mutation status. Each symbol corresponds to a patient. D Pathway analysis for factor weights of LF6-Mito show that LF6-Mito is enriched for pathways related to mitochondrial processes. E LF6-Mito scores are shown for each proteomic cluster; LF6-Mito scores largely separate patients assigned to C-Mito and C5. F Scatter plot of the LF6-Mito score (y-axis) versus the first PC of the mitochondrial proteome (x-axis) for each patient (Spearman’s rho of 0.88). Proteomic cluster membership is color-coded. G Principal component analyses of the proteome and transcriptome for genes annotated to the mitochondrion only. C-Mito patients can be recovered from the proteome, but not the transcriptome when considering mitochondrial genes only. Each symbol corresponds to a patient; patients assigned to C-Mito are highlighted in red. See also Figures S3, S4 and Table S2

Figure 3:. Relationship between AML proteomic subtypes and survival after intensive induction therapy

Clinical phenotypes of proteomic clusters C-Mito – C5 A Kaplan–Meier model of overall survival for the Mito and non-Mito patients. B Forest plot based on hazard ratios (HRs) of a multivariate Cox regression analysis for the overall survival of patients in the discovery cohort. Covariates used in the Cox regression are shown in the leftmost column. The second column lists all modelled levels of the covariates, with the top-most level being the reference level for each covariate. HR estimates with 95% confidence intervals relative to these references are shown in the third column and visualized as boxes and horizontal lines respectively in the fourth column. The dotted vertical line marks a HR of 1. P-values indicating significance are shown on the right. C (Left panel) Group-wise frequency of cytogenetic and molecular genetic features in the Mito and non-Mito groups. (Right panel) Group-wise frequency of ELN cytogenetic risk categories in the Mito and non-Mito groups. D Kaplan–Meier model of overall survival for the C5 and non-C5 patients. E Forest plot based on hazard ratios (HRs) of a multivariate Cox regression analysis for the overall survival of patients in the discovery cohort. Covariates used in the Cox regression are shown in the leftmost column. The second column lists all modelled levels of the covariates, with the top-most level being the reference level for each covariate. HR estimates with 95% confidence intervals relative to these references are shown in the third column and visualized as boxes and horizontal lines respectively in the fourth column. The dotted vertical line marks a HR of 1. P-values indicating significance are shown on the right. F (Left panel) Group-wise frequency of cytogenetic and molecular-genetic features in the C5 and non-C5 groups. (Right panel) Group-wise frequency of ELN cytogenetic risk categories in the C5 and non-C5 groups. See also Figure S5 and Table S3

Figure 4:. Validation cohort and Machine learning-based C-Mito prediction

Super-SILAC-based quantification of proteins in bone-marrow-derived AML blasts from 75 untreated AML patients (validation cohort). A Unsupervised, hierarchical clustering of patient–patient distances. Rows and columns correspond to patients, with each entry in the heatmap quantifying the distance between two patients. The clustering is used to define three distinct clusters (C1–C3; color-coded). Annotations at the top of the heatmap show NPM1 mutation status, FLT3-ITD status, FLT3-ITD allelic ratio and NPM1/FLT3 status. Annotated at the bottom is a C-Mito class membership predicted by a classifier trained on the discovery cohort (predicted C-Mito in red). B Characterization of the three proteomic clusters using a pathway analysis of differentially expressed proteins between each cluster and the remaining patients. Mitochondrial terms were the distinct features characterizing cluster 1. C Precision versus recall and receiver operating characteristic (ROC) curve of the C-Mito classifier, which was trained on the discovery cohort and tested on the validation cohort. Metrics were computed based on unsupervised clustering of the validation cohort (C1 patients were considered as Mito). Areas under both curves close to 1 indicate a high concordance between the unsupervised analysis of the validation cohort only and classification supervised by the discovery cohort. D Bee swarm plot of the contributions of the 5 most relevant proteins to the classification of patients to C-Mito in the validation cohort. Each symbol represents a patient in the validation cohort and patients classified as C-Mito are shown in red, patients classified as non-Mito in grey. E Kaplan–Meier model of overall survival for the patients of the validation cohort predicted to be Mito and non-Mito by using the classifier trained on the discovery cohort. Patients classified as Mito have significantly shorter overall survival. F Kaplan–Meier model of overall survival for the Mito and non-Mito patients of the validation cohort based on hierarchical clustering; Mito patients again have a significantly shorter overall survival. G Forest plot based on hazard ratios (HRs) of a multivariate Cox regression analysis for the overall survival of patients in the validation cohort. Covariates used in the Cox regression are shown in the leftmost column (hierarchical cluster-based C-Mito). The second column lists all modelled levels of the covariates, with the top-most level being the reference level for each covariate. HR estimates with 95% confidence intervals relative to these references are shown in the third column and visualized as boxes and horizontal lines respectively in the fourth column. The dotted vertical line marks a HR of 1. P-values indicating significance are shown on the right. See also Figure S6, Table S4

Figure 5:. The Mito-AML-defining proteomic network

A StringDB-network of the top 100 differentially (FDR<0.01) upregulated proteins in C-Mito versus non-Mito patients. Each node represents a protein and edges indicate interaction type (purple = experimental evidence, light blue = database evidence, yellow = text-mining evidence, black = coexpression evidence). Three major protein hubs emerge, corresponding to mitochondria, chromatin regulators and nuclear envelope respectively. Node colors mark enriched pathway terms (red = mitochondrial protein complex and cellular respiration, yellow = chromatin remodeling, violet = nuclear envelope). B Expression of mitochondrial proteins (log2-intensity) stratified by proteomic cluster. Proteomic cluster C-Mito has significantly higher expression of mitochondrial proteins (GO:0005739) than Cluster 3 to Cluster 6 (Wilcoxon rank sum test, outliers not shown). C Expression of OXPHOS complex subunits in the proteome (log2-intensity) stratified by C-Mito. Patients in C-Mito have a significantly higher expression of OXPHOS complex subunits in the proteome than the remaining patients. D mRNA expression of mitochondrial genes (GO:0005739), stratified by proteomic cluster (Wilcoxon rank sum test, outliers not shown). E mRNA expression of OXPHOS complex subunits, stratified by C-Mito. All boxplots in the figure are defined as follows: middle line corresponds to the median; the lower and upper hinges correspond to first and third quartiles, respectively; the upper whisker extends from the hinge to the largest value no further than 1.5× the interquartile range (or the distance between the first and third quartiles) from the hinge and the lower whisker extends from the hinge to the smallest value at most 1.5× the interquartile range of the hinge. Data beyond the end of the whiskers are called ‘outlying’ points and are plotted individually if not stated otherwise. See also Figure S6.

Figure 6:. Metabolic wiring and therapeutically relevant vulnerabilities of Mito-AML

A Hierarchical clustering of acute leukemia cell lines based on expression of 45 OXPHOS related proteins. Columns correspond to cell lines, rows correspond to proteins. The clustering is used to split acute leukemia cell lines into two groups (Mito-high and Mito-low). B Volcano plot depicting the gene-set enrichment for the genome-wide differential CERES dependency score in Mito-high vs Mito-low cell lines available in the CRISPR (Avana) 21Q2 data. Significance: abs(NES) > 1.3 and, p <0.1. Complex I gene sets are highlighted red. C Scatter mean +/− SD dot plot depicting mean difference rank normalized CERES dependency scores in Mito-high vs. Mito-low cell lines across OXPHOS complexes I-V (One-sample t-test, CRISPR (Avana) 21Q2 dependency data). D The half-maximum inhibitory concentration (IC₅₀) of Mito-high AML cell lines (n=9) and Mito-low AML cell lines (n=10) is shown. Mito-high AML cell lines have increased sensitivity towards OXPHOS targeting drugs (Wilcoxon rank sum test). E Respiratory chain complex I -dependency is quantified in AML cell lines (n Mito-high = 8, n Mito-low = 9) using a Seahorse 96 extracellular flux analyzer. Mito-high cell lines exhibit significantly stronger complex I dependency (p<0.01, Wilcoxon rank sum test). F Oxygen consumption rate in cell lines measured after 24 h treatment with 2 μM venetoclax shows a marked decrease in Mito-high AML cell lines (n=8; p<0.001, Wilcoxon rank sum test) while no significant decrease is seen in Mito-low cell lines (n=8, p = 0.13, Wilcoxon rank sum test). The decrease is significantly greater in Mito-high compared to Mito-low cell lines (p<0.01, Wilcoxon rank sum test). G Complex I dependency is quantified in primary patient samples (n Mito = 9, n non-Mito = 11, all from the discovery cohort) using a Seahorse 96 extracellular flux analyzer. Mito samples exhibit significantly stronger complex I dependency (p<0.01, Wilcoxon rank sum test). H Flow cytometry-based quantification of Annexin V⁻/7AAD⁻ patient-derived AML blasts after incubation with 100 nM venetoclax ex-vivo reveals a significant difference between Mito and non-Mito patients (n Mito = 8, n non-Mito = 17, p<0.001, Wilcox rank sum test). All boxplots in the figure are defined as follows: middle line corresponds to the median; the lower and upper hinges correspond to first and third quartiles, respectively; the upper whisker extends from the hinge to the largest value no further than 1.5× the interquartile range (or the distance between the first and third quartiles) from the hinge and the lower whisker extends from the hinge to the smallest value at most 1.5× the interquartile range of the hinge. Data beyond the end of the whiskers are called ‘outlying’ points and are plotted individually if not stated otherwise. See also Figure S6