Mapping the proteogenomic landscape enables prediction of drug response in acute myeloid leukemia

Abstract

Acute myeloid leukemia is a poor-prognosis cancer commonly stratified by genetic aberrations, but these mutations are often heterogeneous and fail to consistently predict therapeutic response. Here, we combine transcriptomic, proteomic, and phosphoproteomic datasets with ex vivo drug sensitivity data to help understand the underlying pathophysiology of AML beyond mutations. We measure the proteome and phosphoproteome of 210 patients and combine them with genomic and transcriptomic measurements to identify four proteogenomic subtypes that complement existing genetic subtypes. We build a predictor to classify samples into subtypes and map them to a "landscape" that identifies specific drug response patterns. We then build a drug response prediction model to identify drugs that target distinct subtypes and validate our findings on cell lines representing various stages of quizartinib resistance. Our results show how multiomics data together with drug sensitivity data can inform therapy stratification and drug combinations in AML.

Keywords: acute myeloid leukemia; drug response; genomics; linear regression; multiomics; non-negative matrix factorization; proteomics; transcriptomics.

Graphical abstract

Figure 1

Multiomic clustering defines four biologically relevant AML subtypes (A) Circos plot of data collected as part of this study. Bar height correlates with number of drugs assayed for each sample. (B) Unsupervised multiomic clustering analysis combining mRNA, protein, and phosphosite measurements for a 159-patient cohort using NMF. Shading of squares indicates fraction of times patients were in the same cluster during randomization. Cluster annotations are located along the left-hand side. (C) Cluster enrichment via Fisher’s exact test, for gene-specific alterations and clinical variables, including “priorMDS,” which indicates a prior diagnosis of MDS; “Stage,” which indicates whether the sample was collected at initial diagnosis or at relapse; and “Post chemotherapy,” which indicates the patient had received treatment. The x marks p < 0.05. (D–F) Overrepresentation analysis using kinase-substrate enrichment analysis (KSEA) (D) and gene set enrichment analysis (GSEA) for RNA (E) and global proteomics (F), where ∗ indicates adjusted p < 0.05.

Figure 2

Subtype prediction expands classification to 210-patient cohort (A) Expression of the 147 features used to classify patients according to subtype (rows) and the expression for each patient (columns). Resulting classification is annotated across the top. Row color corresponds to source of feature, with the gray color showing features found to be predictive of multiple subtypes. (B) UMAP projections demonstrating how the features in (A) map patients into a two-dimensional landscape. (C) Kaplan-Meier plots of patients according to subtype classification. (D and E) Heatmap of enrichment scores for GSEA (D) and KSEA (E) results for each subtype using our full 210-patient proteomic or phosphoproteomic dataset. Color represents enrichment score; asterisk denotes an adjusted p < 0.05.

Figure 3

Mutational subtypes and proteomic subtypes stratify patient response of nonoverlapping sets of drugs (A) Association of mutation status or subtype with drug response via Welch’s paired t test. Shading represents t-statistic (legend inset). The x indicates adjusted p < 0.05. (B and C) SF3B1 mutation status effect on response to venetoclax (B) or panobinostat (C). (D and E) Three drug response profiles assessed by subtype (x axis) and FLT3-ITD status (color). Each ∗ indicates significance level (e.g., ∗p = 0.01, ∗∗p = 0.001).

Figure 4

Distribution of ex vivo drug responses by subtype (A–C) Distribution of AUC by subtype for (A) venetoclax, (B) panobinostat, and (C) sorafenib. The ∗ indicates FDR-corrected significance of differences between classes according to Welch’s t test p value; ∗ = 1e−2, ∗∗ = 1e−3. (D and E) Distribution of AUC values by subtype for the combination of (D) venetoclax and panobinostat and (E) sorafenib and panobinostat. (F and G) Signaling networks created using the shortest paths between drug targets, using the MAGINE subnetwork tool. (F) Subnetwork between the primary target of venetoclax (BCL2) and a primary target of panobinostat (HDAC1). (G) Subnetwork between a primary target of sorafenib (FLT3) and a primary target of panobinostat (HDAC1).

Figure 5

Feature analysis of drug prediction models (A) Venn diagram showing the number of features in each drug model. (B) Reactome terms overrepresented among the molecular features shown in (A). Colors represent the combined score from enrichR. + indicates significantly enriched terms based on adjusted p < 0.05. (C) Unsupervised clustering of the molecular features extracted from the drug response models. Molecular features from AUC regression model. (D) UMAP 2-dimensional projection of samples by the features extracted from the models, colored by subtype classification. (E) Same UMAP representation as (D) but colored by venetoclax AUC. (F) Same UMAP representation as (D) and (E) but colored by panobinostat AUC.

Figure 6

Predicting drug sensitivity following FLT3 inhibition induced landscape changes (A) Proteogenomic subtype projection of patient samples together with MOLM14 naive, early resistant, and late resistant cell lines using UMAP. (B) ElasticNet prediction of subtype of cell lines. Multiple measurements per cell stage allows a percentage projection of subtype. (C) AUC predictions of cell lines across the different MOLM14 cell lines, with median AUC of naive cells marked by dashed line. (D) Experimental results of venetoclax and panobinostat sensitivity on same cell lines; n = 4 for each condition.