Molecular precision medicine: Multi-omics-based stratification model for acute myeloid leukemia

Abstract

Acute myeloid leukemia (AML), as the most common malignancy of the hematopoietic system, poses challenges in treatment efficacy, relapse, and drug resistance. In this study, we have utilized 151 RNA sequencing datasets, 194 DNA methylation datasets, and 200 somatic mutation datasets from the AML cohort in the TCGA database to develop a multi-omics stratification model. This model enables comparison of prognosis, clinical features, gene mutations, immune microenvironment and drug sensitivity across subgroups. External validation datasets have been sourced from the GEO database, which includes 562 mRNA datasets and 136 miRNA datasets from 984 adult AML patients. Through multi-omics-based stratification model, we classified 126 AML patients into 4 clusters (CS). CS4 had the best prognosis, with the youngest age, highest M3 subtype proportion, fewest copy number alterations, and common mutations in WT1, FLT3, and KIT genes. It showed sensitivity to HDAC inhibitors and BCL-2 inhibitors. Both the M3 subtype and CS4 were identified as independent protective factors for survival. Conversely, CS3 had the worst prognosis due to older age, high copy number alterations, and frequent mutations in RUNX1, DNMT3A, and TP53 genes. Additionally, it showed higher proportions of cytotoxic cells and Tregs, suggesting potential sensitivity to mTOR inhibitors. CS1 had a better prognosis than CS2, with more copy number alterations, while CS2 had higher monocyte proportions. CS1 showed good sensitivity to cytarabine, while CS2 was sensitive to RXR agonists. Both CS1 and CS2, which predominantly featured mutations in FLT3, NPM1, and DNMT3A genes, benefited from FLT3 inhibitors. Using the Kappa test, our stratification model underwent robust validation in the miRNA and mRNA external validation datasets. With advancements in sequencing technology and machine learning algorithms, AML is poised to transition towards multi-omics precision medicine in the future. We aspire for our study to offer new perspectives on multi-drug combination clinical trials and multi-targeted precision medicine for AML.

1. Introduction

Acute myeloid leukemia (AML) manifests as a malignancy characterized by the abnormal proliferation of blast cells in the bone marrow and peripheral blood, stemming from the clonal expansion of immature myeloid cells [1]. Among adults, AML stands as the most prevalent and prognostically unfavorable form of acute leukemia, with an age-adjusted incidence rate of 4.3 per 100,000 person-years and a median age at diagnosis of 68 years in the United States [2]. The diverse prognoses observed in AML patients are influenced by the heterogeneous nature of disease characteristics, therapeutic strategies, and various other factors [3]. Concurrently, relapse and resistance pose persistent challenges in the realm of AML treatment [1]. Hence, enhancing the clinical outcomes for AML patients represents a pivotal objective.

Currently, in clinical practice, the common classification standards for AML include the French-American-British (FAB) classification and the World Health Organization (WHO) classification standards [4]. Based on the proportion of bone marrow blasts cells and bone marrow cell morphology, AML is classified into 8 subtypes (M0 to M7), according to the FAB classification system [5]. The FAB classification criteria are convenient in clinical practice, but fail to reflect the complex molecular heterogeneity of AML. Unlike the FAB classification of AML, the WHO classification (2022 updated) recognized 12 distinct subgroups of AML, defined by fusion genes or single-gene mutations, The WHO classification effectively reflects the status of gene rearrangements or mutations in AML [6], but does not incorporate information on non-coding RNA, epigenetics, and other biomarkers. The diagnosis and treatment of AML still require a more comprehensive and precise individualized classification standard.

The rapid development of multi-omics approaches, including genomics, transcriptomics, proteomics, and metabolomics, has opened up new perspectives for disease diagnosis and treatment [7,8]. Multi-omics-based classification of cancer integrates data on genetic mutations, non-coding RNA expression, epigenetic abnormalities, and other factors [9], and this also applies to studies on AML classification based on prognostic features and biomarkers. The classification of AML based on multi-omics aligns with the concept of precision medicine, aiming to provide the most effective treatment for each individual patient [10]. Processing vast quantities of medical data necessitates the application of computer technology. Notably, machine learning methods augment the efforts of clinicians in areas such as diagnosis, treatment, and prognosis [11]. Simultaneously, machine learning techniques offer innovative approaches to identify new biomarkers using multi-omics data, thereby holding the potential to contribute to precision medicine [12]. Multi-omics-based precision medicine holds promise in optimizing clinical decision-making and enhancing the diagnosis of AML patients.

To highlight the individualized and heterogeneous biomarker features of AML, develop AML stratification methods applicable to precision medicine using existing multi-omics data. In our study, we developed an AML stratification model by incorporating multi-omics data through a diverse array of clustering algorithms. In contrast to traditional AML clinical typing, this stratification model relies on biomarkers specific to AML subtypes, including mRNA, miRNA, lncRNA, DNA methylation, and somatic mutation data. This study presents a replicable and robust systematic approach to cancer stratification, providing a foundation for future research endeavors in AML precision medicine field.

2. Materials and methods

2.1. Data downloading and preparing

The multi-omics data of AML patients were chosen from the TCGA-LAML cohort in the Cancer Genome Atlas database (TCGA, https://portal.gdc.cancer.gov/). TCGA-LAML cohort contains RNAseq data (60660 probes × 151 samples, including mRNAseq, miRNAseq and lncRNAseq), Illumina DNA methylation 450 K data (486427 probes × 194 samples), copy number variation (CNV) data (n = 200), and clinical information (n = 200). "TCGAbiolinks" is an R package that provides integration, downloading, and analysis of TCGA data [13]. Transcriptome data, DNA methylation data, masked somatic mutation data, and clinical information of TCGA-LAML cohort were downloaded using the R package "TCGAbiolinks". The transcriptome data includes gene counts, transcripts per kilobase million (TPM), and fragments per kilobase million (FPKM) formats. The TPM format has been selected for annotating mRNA, lncRNA, and miRNA using the GENCODE v31 annotation [14]. Additionally, DNA methylation CpG sites were annotated by "TCGAbiolinks" package. M-values were used to measure DNA methylation levels.

Besides, GSE37642, which included 984 samples was obtained from the GEO (Gene Expression Omnibus) database [15]. 562 mRNA samples from GSE37642 were used for external validation, while 136 miRNA samples were used for external validation. The microarray platforms of mRNA sequencing used in GSE37642 were the Affymetrix Human Genome U133A, U133B Array (sample 1 to 422), as well as the Affymetrix Human Genome U133 Plus 2.0 Array (sample 423 to 562). The microarray platform of miRNA sequencing was Affymetrix Human Genome U133 Plus 2.0 Array.

2.2. Stratification of multi-omics data

“Multi-Omics integration and VIsualization in Cancer Subtyping” (MOVICS) is an R package for multi-omics integrative clustering algorithms and visualization in cancer subtyping [16]. The optimal number of AML subgroups was determined by identifying the clustering number at which the sum of the clustering prediction index (CPI) [17] and Gaps-statistics [18] reaches the maximum. Based on the selected cluster number, the "getMOIC" function, which includes 10 clustering algorithms such as SNF (Similarity Network Fusion), PINSPlus (Perturbation clustering for data INtegration and disease Subtyping), NEMO (Neighborhood based Multi-Omics clustering), COCA (Cluster-Of-Cluster-Assignments), LRAcluster (low-rank approximation cluster), ConsensusClustering, IntNMF (Integrative non-negative matrix factorization), CIMLR (Cancer Integration via Multikernel Learning), MoCluster, and iClusterBaye [19], calculated a consensus matrix to obtain a multi-omics stratification with a high robustness number of clusters. Simultaneously, the silhouette score was computed to quantify the similarity of subgroups [20].

2.3. Evaluation of clinical features and molecular function

The clinical baseline data, including the FAB classification, (Cancer and Leukemia Group B) CALGB categories, and characteristic cytogenetic abnormalities, were summarized in the identified model of AML. Additionally, the "MOVICS" package provided an approach to draw Kaplan-Meier curves and assess the overall survival (OS) of clinical prognosis among subgroups using the log-rank test [21]. Subsequently, univariable Cox regression analysis [22] was conducted to investigate the impact of single clinical features on survival risk. In order to further identify independent prognostic factors affecting survival risk and to avoid bias caused by statistically significant single clinical features affecting clusters in the model, we employed multivariate Cox regression analysis [23], hazard ratio (HR) > 1 indicates a risk factor affecting the outcome, whereas <1 indicates a protective factor affecting the outcome.

Moreover, subtype-pairwise upregulated and downregulated genes meeting the threshold of adjusted P-value <0.05 and |log2FC| ≥ 1 were filtered using the "limma" package in R [24]. To identify subtype-specific Gene Ontology (GO) pathways, gene set enrichment analysis (GSEA) [25] was performed on significant differentially expressed genes (DEGs) using the "limma" package. Additionally, the "clusterProfiler" package was utilized to enrich the gene functional units the gene set variation analysis (GSVA) approach [26].

2.4. Calculation of genetic alterations

The R package "MOVICS" [27,28]was utilized to obtain gene-level somatic mutations and calculate the tumor mutation burden (TMB) between two groups of AML patients within subgroups using the "compMut" function. The fraction of the genome altered (FGA) was employed as a measure to quantify the extent of genome affected by copy number gains or losses [29]. Furthermore, "MOVICS" also offered an interface to calculate FGA, as well as the fractions of the genome lost (FGL) or gained (FGG) within our stratification model [30].

2.5. Profiling of immune cells infiltration and drug sensitivity

CIBERSORT (Cell-type Identification By Estimating Relative Subsets Of RNA Transcripts, http://cibersort.stanford.edu/) is an algorithm based on deconvolution, which enables the characterization of cellular composition based on gene expression profiles [31]. In the context of AML subgroups, CIBERSORT can be applied to identify discrepancies in the immune cell microenvironment, with a set threshold of P-value <0.05.

Additionally, we assessed the potential responsiveness of AML subgroups to chemotherapy drugs by examining the differential half-maximal inhibitory concentration (IC50) values obtained from the Genomics of Drug Sensitivity in Cancer database (GDSC, https://www.cancerrxgene.org/) [32]. This analysis was performed using the ridge regression method implemented in the R package "MOVICS". Varied IC50 values reflect the sensitivity of different subgroups to specific therapeutic targets.

2.6. External validation of the stratification model

To assess the reproducibility and accuracy of the multi-omics-based AML classifier in an GEO validation cohort, we employed Nearest Template Prediction (NTP) and Partition Around Medoids (PAM) analyses to validate the concordance of overall survival (OS) between the TCGA-LAML model and the GEO validation cohort. NTP allows for multiclass predictions without the need for optimized parameters [33], while PAM identifies subsets of genes with the nearest shrunk centroids using Pearson correlation test [34]. The in-group proportion (IGP) was calculated to quantify the similarity of subtypes between the stratification model and the validation cohorts [35]. Furthermore, we utilized the Kappa-test to evaluate the consistency between the NTP and PAM models in the mRNA (n = 562) and miRNA (n = 136) validation cohort [15]. Additionally, we verified the similarity between the NTP model, PAM model, and the consensus omics model in the TCGA-LAML cohort. All of these analyses were performed using the "runNTP," "runPAM," and "runKappa" functions in the R package "MOVICS."

2.7. Statistical analyses

All analyses were conducted using R software version 4.1.2 (http://www.r-project.org). Continuous data were presented as mean ± SD. Shapiro-Wilk tests were used to assess the normality of the data. Non-normal data were compared using the Kruskal-Wallis rank sum test, while normal data were analyzed using Student's t-test.

For external validation, NTP analysis and PAM analysis were also performed to assess the similarity to the stratification model of the TCGA-LAML cohort. The consistency between the two classification models was evaluated using the Kappa test [36]. A P-value or an adjusted P-value less than 0.05 was considered statistically significant, the p-value adjustment method used is the Benjamini-Hochberg procedure.

3. Results

3.1. Overview of the stratification model

Multi-omics data including mRNA sequencing data (n = 151), miRNA sequencing data (n = 151), lncRNA sequencing data (n = 151), DNA methylation meta data (n = 194), masked somatic mutation data (n = 152), and clinical information data (n = 200) in TCGA-LAML cohort were downloaded by “TCGAbiolinks” package. Subsequently, using the "TCGAbiolinks" package to filter out samples with abnormal expression levels from the multi-omics data and match them with samples in the clinical data that have complete survival information, finally, a total of 126 AML samples (35514 mRNA data, 1331 miRNA data, 16172 lncRNA data, 35396 DNA methylation data, 955 somatic mutation data) were retained.

Next, utilizing multi-omics data from 126 AML samples, the optimal cluster number was determined to be 4 based on clinical considerations and the calculation of CPI and Gap-statistics using the 'getClustNum' function of the 'MOVICS' package. (Fig. 1A). Subsequently, we applied 10 algorithms to cluster the consensus ensembles, resulting in the division of the multi-omics data set of 126 samples into cluster 1 (CS1, 39 samples), cluster 2 (CS2, 37 samples), cluster 3 (CS3, 21 samples) and cluster 4 (CS4, 29 samples) with optimal robustness (Fig. 1B). The average silhouette width score was 0.42, indicating a clear distinction between the four clusters (Fig. 1C). To provide an overview of the five omics stratification model, including mRNA, LncRNA, miRNA, DNA methylation data, and mutational genes, we generated an overview map displayed in Fig. 1D. Furthermore, the top 10 OS-related biomarkers from each omic section were indicated on the right side of the map.

Fig. 1.

Stratification model construction and multi-omics analysis. (A) Identification of optimal cluster number by calculating cluster prediction index (CPI, blue line) and Gaps-statistics (red line) in TCGA-LAML cohort. (B) Consensus heatmap based on results from 10 multi-omics integrative clustering algorithms with cluster number of 2. Consensus Score: Dark blue (0) means the samples were never clustered together, while yellow (1) means the samples were always clustered together. (C) Quantification of sample similarity by silhoutte score based on consensus ensembles result. Silhoutte width quantify the similarity of 2 subgroups (D) Comprehensive heatmap of five omics data with classified and clinical information. mRNA_normalized: Normalized mRNA expression (after Z-score standardization), lncRNA_normalized: Normalized lncRNA expression (after Z-score standardization), miRNA_normalized: Normalized miRNA expression (after Z-score standardization), M-value: ,M-Value_normalized: Normalized M-Value (after Z-score standardization).

3.2. Clinical features and their survival risks in the stratification model

The survival analysis revealed a significantly poorer OS in patients belonging to the CS3 subgroup compared to other subgroups, with patients in the CS4 subgroup exhibiting the best overall survival. (p-value = 0.006, Fig. 2A).

Fig. 2.

(A) Kaplan-Meier survival analysis of the overall survival (OS) between 4 subgroups. (B) Univariable Cox regression analysis of clinical features of OS. (C) Multivariable Cox regression analysis of clinical features of OS. PB: peripheral blood; BM: bone marrow; HGB: hemoglobin; PLT: platelet; For "Cluster", the comparison baseline is set as "CS1"; for "Race", the comparison baseline is "White"; and for "FAB classification", the comparison baseline is "M1". Red bars: HR (Hazard Ratio) > 1, indicates the variable increases the risk of the event occurring. Green bars: HR (Hazard Ratio) < 1, indicates the variable decreases the risk of the event occurring.

The analysis of clinical characteristics among the four subgroups (Table 1) showed that the median age of patients in the CS4 subgroup was the youngest (50 [35.00, 59.00], p-value < 0.001). In terms of CALGB classification, the CS4 subgroup exhibited the highest proportion of favorable patients (65.5 %, p < 0.001). Moreover, the levels of cytogenetic abnormalities were lower in the CS1 and CS2 subgroup (p-value = 0.001). Interestingly, there was a significant difference in FAB classification among the four subgroups (P-value <0.001). Within the CS1 subgroup, the proportion of M1 (28.2 %) and M2 (28.2 %) subtypes were the highest. In the CS2 subgroup, the proportion of M4 (29.7 %) subtype was the highest. Within the CS3 subgroup, the proportion of M0 (28.6 %) subtype was the highest. In the CS4 subgroup, the proportion of M3 (37.9 %) subtype was the highest. Additionally, the platelet (PLT) count was lowest in CS4 (p-value = 0.021). There were no significant differences between four subgroups in terms of sex (p-value = 0.609), race (p-value = 0.646), prior malignancy (p-value = 0.173), prior treatment (p-value = 0.247), neoadjuvant chemotherapy (p-value = 0.247), peripheral blood blasts percentage (p-value = 0.304), bone marrow blasts percentage (p-value = 0.150), and hemoglobin (HGB, p-value = 0.727).

Table 1.

Summarization of clinical features in 4 subgroups.

	CS1 (n = 39)	CS2 (n = 37)	CS3 (n = 21)	CS4 (n = 29)	P.value
Age (year, median [IQR])	51 [40.00, 62.00]	57 [41.00, 67.00]	66 [60.00, 71.00]	50 [35.00, 59.00]	0.001
Sex (%)					0.609
Female	17 (43.6)	20 (54.1)	8 (38.1)	15 (51.7)
Male	22 (56.4)	17 (45.9)	13 (61.9)	14 (48.3)
Race (%)					0.646
White	35 (89.7)	34 (91.9)	20 (95.2)	25 (86.2)
Black or African American	4 (10.3)	2 (5.4)	1 (4.8)	3 (10.3)
Asian	0 (0.0)	1 (2.7)	0 (0.0)	0 (0.0)
NR	0 (0.0)	0 (0.0)	0 (0.0)	1 (3.4)
CALGB Category (%)					< 0.001
Favorable	3 (7.7)	5 (13.5)	0 (0.0)	19 (65.5)
Normal	25 (64.1)	27 (73.0)	11 (52.4)	7 (24.1)
Poor	11 (28.2)	4 (10.8)	10 (47.6)	2 (6.9)
NR	0 (0.0)	1 (2.7)	0 (0.0)	1 (3.4)
Cytogenetic abnormality (%)					0.001
One	8 (20.5)	7 (18.9)	4 (19.0)	17 (58.6)
Two	1 (2.6)	0 (0.0)	1 (4.8)	1 (3.4)
Complex	6 (15.4)	1 (2.7)	6 (28.6)	3 (10.3)
Normal	23 (59.0)	22 (59.5)	9 (42.9)	6 (20.7)
NR	1 (2.6)	7 (18.9)	1 (4.8)	2 (6.9)
FAB classification(%)					< 0.001
M0	5 (12.8)	1 (2.7)	6 (28.6)	0 (0.0)
M1	11 (28.2)	10 (27.0)	5 (23.8)	5 (17.2)
M2	11 (28.2)	6 (16.2)	4 (19.0)	10 (34.5)
M3	1 (2.6)	0 (0.0)	0 (0.0)	11 (37.9)
M4	8 (20.5)	11 (29.7)	3 (14.3)	3 (10.3)
M5	3 (7.7)	9 (24.3)	0 (0.0)	0 (0.0)
M6	0 (0.0)	0 (0.0)	2 (9.5)	0 (0.0)
M7	0 (0.0)	0 (0.0)	1 (4.8)	0 (0.0)
Prior malignancy (%)
No	36 (92.3)	36 (97.3)	17 (81.0)	27 (93.1)	0.173
Yes	3 (7.7)	1 (2.7)	4 (19.0)	2 (6.9)
Prior treatment (%)					0.247
No	27 (69.2)	26 (70.3)	19 (90.5)	23 (79.3)
Yes	12 (30.8)	11 (29.7)	2 (9.5)	6 (20.7)
Neoadjuvant chemotherapy(%)					0.247
No	27 (69.2)	26 (70.3)	19 (90.5)	23 (79.3)
Yes	12 (30.8)	11 (29.7)	2 (9.5)	6 (20.7)
PB blasts (%, mean ± SD)	67.85 ± 22.38	69.08 ± 21.72	57.10 ± 24.33	66.41 ± 22.31	0.304
BM blasts (%, mean ± SD)	48.31 ± 35.17	31.68 ± 28.72	33.38 ± 29.45	38.24 ± 30.08	0.150
HGB (g/L, mean ± SD)	9.69 ± 1.54	9.38 ± 1.50	9.38 ± 1.50	9.66 ± 1.11	0.727
PLT ( × 10^9/L, mean ± SD)	79.79 ± 61.30	67.89 ± 62.65	69.86 ± 46.22	44.34 ± 37.45	0.021

IQR: interquartile range, NR: not reported in current datasets; Cytogenetic abnormality, one: one chromosomal abnormal karyotypes; Cytogenetic abnormality, two: two kind of chromosomal abnormal karyotypes; Cytogenetic abnormality, complex: three or more chromosomal abnormal karyotypes; PB: peripheral blood; BM: bone marrow; HGB: hemoglobin; PLT: platelet.

Univariable Cox regression analysis (Fig. 2B) revealed that Age (HR = 1.033, 95 % CI [1.017, 1.050], P-value <0.001), CALGB Category (HR = 1.811, 95 % CI [1.273, 2.576], P-value <0.001), Prior treatment (HR = 0.559, 95 % CI [0.338, 0.927], P-value = 0.024), Neoadjuvant (HR = 0.559, 95 % CI [0.338, 0.927], P-value = 0.024), Multi-Omics Clusters (with CS1 as the reference) CS2 (HR = 1.311, 95 % CI [0.837, 1.731], P-value = 0.025), CS3 (HR = 1.856, 95 % CI [1.389, 2.203], P-value = 0.023), CS4 (HR = 0.289, 95 % CI [0.038, 0.346], P-value = 0.006), and the M3 subtype (HR = 0.216, 95 % CI [0.064, 0.731], P-value = 0.014) within the FAB classification (with the M1 subtype as the reference) were all factors influencing survival risk in TCGA-LAML cohort. Furthermore, the results of multivariable Cox regression analysis (Fig. 2C, P-value <0.05) indicated that Age, CALGB Category, Prior treatment, Neoadjuvant, Multi-Omics subgroups, and the M3 subtype within the FAB classification were all independent factors influencing survival risk in the TCGA-LAML cohort.

3.3. Molecular functions of four subgroups

A total of 887 up regulated and 410 down regulated DEGs were filtered by “limma” R package, using a threshold of adjusted P-value < 0.05 and |log2FC| ≥ 1 (Fig. 3A, Supplementary Table 1). We employed the GSEA approach to identify the GO (gene ontology) pathways, across the four subgroups, 32 GO pathways with significant differences in NES (normalized enrichment score) were identified. (adjusted P-value < 0.05, Fig. 3B–Supplementary Table 2). GSVA results demonstrated the expression profiles of 21 gene sets with significant differences between the four subgroups. (P-value < 0.05, Fig. 3C).

Fig. 3.

Gene expression status and enrichment analysis among 4 subgroups. (A) Heatmap of differential expression genes among 4 groups. mRNA_normalized: After Z-score standardization, the relative mRNA expression levels. (B) Gene Ontology (GO) signaling pathway of differential expression genes in 4 subgroups with gene set enrichment analysis (GSEA). NES: normalized enrichment score. (C) Differential functional units in subgroups by gene set variation analysis (GSVA).

3.4. Distribution of genetic alteration

Table 2 displays the genes with high mutant frequencies. Additionally, FLT3 (29 %), NPM1 (26 %), DNMT3A (25 %), RUNX1 (10 %), WT1 (7 %) and TP53 (6 %) were found to be significantly mutated genes between the 4 groups (Fig. 4A). Additionally, CS4 subgroup showed the fewest copy number alterations, with CS3 having the most (P-value < 0.05, Fig. 4B).

Table 2.

Top 11 mutated genes in TCGA-LMAL cohort.

Gene (Mutated)	TMB (%)	CS1	CS2	CS3	CS4	P-value
NPM1	33 (26 %)	12 (30.8 %)	19 (51.4 %)	0 (0.0 %)	2 (6.9 %)	< 0.001
RUNX1	12 (10 %)	5 (12.8 %)	0 (0.0 %)	6 (28.6 %)	1 (3.4 %)	0.0013
DNMT3A	31 (25 %)	11 (28.2 %)	14 (37.8 %)	5 (23.8 %)	1 (3.4 %)	0.0059
TP53	8 (6 %)	2 (5.1 %)	1 (2.7 %)	5 (23.8 %)	0 (0.0 %)	0.0067
WT1	9 (7 %)	4 (10.3 %)	0 (0.0 %)	0 (0.0 %)	5 (17.2 %)	0.0129
FLT3	36 (29 %)	13 (33.3 %)	16 (43.2 %)	2 (9.5 %)	5 (17.2 %)	0.0185
KIT	7 (6 %)	1 (2.6 %)	1 (2.7 %)	1 (4.8 %)	4 (13.8 %)	0.2530
TET2	11 (9 %)	6 (15.4 %)	2 (5.4 %)	2 (9.5 %)	1 (3.4 %)	0.3420
IDH2	15 (12 %)	4 (10.3 %)	4 (10.8 %)	5 (23.8 %)	2 (6.9 %)	0.3490
IDH1	14 (11 %)	6 (15.4 %)	5 (13.5 %)	2 (9.5 %)	1 (3.4 %)	0.4530
CEBPA	12 (10 %)	5 (12.8 %)	2 (5.4 %)	1 (4.8 %)	4 (13.8 %)	0.5550

TMB: tumor mutation burden.

Fig. 4.

Genetic alteration in subgroups. (A) The most significantly mutated genes between 4 subgroups. TMB: tumor mutation burden. (B) Fraction of genome altered (FGA) and fraction of genome lost or gained (FGL or FGG) in subgroups.

3.5. Immune cells infiltration and drug responsiveness

The results from the CIBERSORT algorithm indicated significant differences (P-value <0.05, Fig. 5A) in the composition of immune cells in the immune microenvironment among the four subgroups. These disparities were evident in “B cells naive”, “T cells CD8”, “T cells CD4 memory resting”, “T cells regulatory (Tregs)”, “NK cells resting”, “Monocytes”, and “Mast cells resting”.

Fig. 5.

Immune cells infiltration and drug responsiveness analysis. (A) Distribution of immunocytes content among four clusters. (B) The CS1 subgroup is potentially sensitive to Cytarabine. (C) The CS2 subgroup is potentially sensitive to Bexarotene. (D) The CS3 subgroup is potentially sensitive to Rapamycin. (E) The CS4 subgroup is potentially sensitive to Vorinostat. (F) Both CS1 and CS2 subgroups exhibit high sensitivity to Sorafenib. (G) Both CS1 and CS4 subgroups exhibit high sensitivity to Navitoclax. TME: tumor microenvironment, IC₅₀: half maximal inhibitory concentration.

Moreover, the responsiveness of AML to chemo drugs was assessed using data downloaded from GDSC (Supplementary Table 3). The results suggested that the CS1 subgroup might exhibit higher sensitivity to Cytarabine (Fig. 5B, P-value <0.001), while the CS2 subgroup could potentially be more sensitive to Bexarotene (Fig. 5C, P-value <0.001), the CS3 subgroup to Rapamycin (Fig. 5D, P-value <0.001), and the CS4 subgroup to Vorinostat (Fig. 5E, P-value <0.001). Additionally, the CS1 and CS2 groups might simultaneously benefit more from Sorafenib (Fig. 5F, P-value <0.001), while the CS1 and CS4 groups might simultaneously benefit more from Navitoclax (Fig. 5G, P-value <0.01).

3.6. Validation in external cohorts

The external validation cohort comprised 562 mRNA samples and 136 miRNA samples. According to the multi-omics AML stratification model, applying the NTP approach distinguished the predicted four subgroups in the mRNA cohort (Fig. 6A). Similarly, the predicted four subgroups were also clustered in the miRNA cohort (Fig. 6B). Additionally, the overall survival (OS) among the four subgroups in both the mRNA (Fig. 6C, P-value <0.001) and the miRNA (Fig. 6D, P-value = 0.002) validation cohorts were consistent with the multi-omics AML stratification model. Importantly, the IGP values of the PAM analysis for the four subgroups in the mRNA validation cohort were 0.737, 0.817, 0.710, and 0.691 (Fig. 6A), while the IGP values in the miRNA validation cohort were 0.695, 0.818, 0.731, and 0.667 respectively (Fig. 6B). Additionally, the mRNA validation cohort using the NTP method closely corresponded with results from the PAM method (Fig. 6E, Kappa = 0.677, P < 0.001). Similarly, the miRNA validation cohort using the NTP method also demonstrated consistency with results from the PAM method (Fig. 6F, Kappa = 0.682, P < 0.001). Furthermore, for the TCGA-LAML cohort, the Kappa test demonstrated high consistency between the multi-omics AML stratification model and the NTP approach (Fig. 6G, Kappa = 0.872, P < 0.001), as well as between the multi-omics AML stratification model and the PAM approach (Fig. 6H, Kappa = 0.766, P < 0.001).

Fig. 6.

External cohort validation of the multi-omics stratification model. (A) Predicted subgroups in mRNA validation cohort by Nearest Template Prediction (NTP) method with the subtype-specific TCGA-LAML biomarkers. (B) Predicted subgroups in miRNA validation cohort by Nearest Template Prediction (NTP) method with the subtype-specific TCGA-LAML biomarkers. (C) Kaplan-Meier curves of OS among 4 predicted subgroups in the mRNA validation cohort. (D) Kaplan-Meier curves of OS among 4 predicted subgroups in the miRNA validation cohort. (E) Consistency between NTP and Partition Around Medoids (PAM) in mRNA validation cohort. (F) Consistency between NTP and Partition Around Medoids (PAM) in miRNA validation cohort. (G) Consistency heatmap for TCGA-LAML between multi-omics and NTP. (H) Consistency heatmap for TCGA-LAML between multi-omics and NTP.

4. Discussion

In this study, we developed an AML stratification model using the R package "MOVICS," based on multi-omics data encompassing mRNA, lncRNA, miRNA, DNA methylation, and somatic mutations. Initially, building upon the CPI and Gaps-statistics approaches, we utilized ten clustering algorithms and classified a total of 126 AML patients into 4 distinct subgroups: CS1 (n = 39), CS2 (n = 37), CS3 (n = 21), and CS4 (n = 29). Survival analysis revealed that the prognosis is most favorable for the CS4 subgroup, whereas it is poorest for the CS3 subgroup. Similarly, the CS4 subgroup exhibited the fewest copy number alterations, while the CS3 subgroup displayed the most. FLT3, NPM1, DNMT3A, RUNX1, WT1 and TP53 were found to be significantly mutated genes between the 4 groups. Furthermore, greater age, poorer CALGB Category, CS2 and CS3 subgroups were all independent risk factors affecting prognosis, on the contrary, receiving neoadjuvant therapy, undergoing prior treatment, CS4 subgroups, and M3 subtype (FAB classification) were all identified as independent protective factors for prognosis. Moreover, The CS2 and CS3 subgroups exhibited higher expression profiles of GSVA gene sets, and there were significant differences among the four subgroups in terms of immune microenvironment cell composition. Additionally, we predicted the potentially sensitive chemical drugs for each subgroup. Finally, the validation of our stratification model was achieved through survival analysis and consistency testing using mRNA and miRNA cohorts obtained from the GEO database.

For the CS4 subgroup, which demonstrated the best prognosis, the median age was the lowest, CALGB Category was the most favorable, cytogenetic abnormalities were fewer, and the proportion of M3 subtype (FAB classification) was the highest. Retrospective analysis [37] has demonstrated that older patients with AML typically present with poorer performance status and a higher proportion of unfavorable cytogenetic abnormalities. AML with the M3 subtype, also known as acute promyelocytic leukemia (APL), is treated primarily with all-trans retinoic acid and arsenic trioxide. This treatment approach yields estimated 10-year survival rates of 80 % or higher, which is higher than other FAB subtypes of AML [38]. Despite age serving as a prognostic risk factor, there were significant differences among the four subgroups, with an HR of 1.043. This indicates that bias induced by age does not influence clustering in the model. However, the CS3 subgroup was the most hazardous factor (HR = 1.979). Despite the lack of significant differences in neoadjuvant and prior treatment among the four subgroups, they continued to function as independent protective factors for prognosis. Interestingly, we observed the lowest platelet (PLT) counts in the CS4 subtype, which correlated with a more favorable prognosis. Zhang et al. reported that low PLT counts were associated with improved outcomes in intermediate-risk AML patients at diagnosis [39], while Zhao et al. [40] identified platelet to white blood cell ratio (PWR) as an independent prognostic predictor in AML, showing a negative correlation between PWR and OS time and bone marrow blast percentage.

High-frequency mutant genes in AML serve as classification criteria and were distinctly identified in our study. These mutations not only influence disease phenotype but also impact therapeutic response and clinical outcomes [1]. In this study, the top four mutated genes are FLT3, NPM1, DNMT3A, and RUNX1. FLT3 (FMS-like tyrosine kinase 3), a member of the class III receptor tyrosine kinase family, is commonly found as the most prevalent genetic abnormality in AML, FLT3 represents a highly promising target for leukemia therapy [41,42]. Similarly, FLT3 has the highest mutation rate in the entire model. The nucleophosmin (NPM1) gene encodes a multifunctional chaperone protein that localizes primarily to the nucleolus. NPM1 mutations occur in about 30 % of AML, and frequently associate with FLT3-ITD, DNMT3A, TET2, and IDH1/25 mutations [43]. The overall model, as well as the gene mutation patterns observed in the CS1 and CS2 subgroups, also exhibit this trend. DNMT3A, a member of the DNA methyltransferase enzyme family, plays a crucial role in DNA methylation [44]. DNMT3A mutations have been associated with adverse overall survival and event-free survival in AML patients [45]. RUNX1, also known as Acute Myeloid Leukemia 1 protein (AML1), Core-Binding Factor Subunit Alpha-2 (CBFA2), or Polyoma Virus Enhancer-Binding Protein 2αB (PEBP2αB), is a master-regulator transcription factor involved in hematopoiesis [46]. Recent studies have highlighted the poor prognostic implications of RUNX1 mutations in AML and suggested it as a potential therapeutic target [47]. In the CS4 subgroup, which exhibits the best prognosis, WT1, FLT3 and KIT mutations have the highest frequency. WT1 mutations, as secondary events in AML, exhibit high instability and have a relatively minor impact on prognosis compared to mutations in DNMT3A, TET2, and IDH 1/2, which are mutually exclusive [48]. KIT mutations do not adversely affect complete remission (CR), relapse rates (RR), or OS in Caucasians. However, in non-Caucasian populations, the opposite effect is observed [49]. This is consistent with the predominantly White population in this study. In the CS1 and CS2 subgroups, which have the second and third best prognosis, FLT3, DNMT3A, and NPM1 mutations are most frequent. Conversely, in the CS3 subgroup, which has the poorest prognosis, RUNX1, DNMT3A, and TP53 mutations are highly prevalent, the high mutation rates of RUNX1 and DNMT3A in the CS3 subgroup may be associated with its relatively higher median age [50,51]. TP53 mutations are not only associated with increasing age but also serve as prognostic markers for poor chemotherapy response [52].

The proportion of mast cells is highest in the CS4 subgroup. Mast cells serve as immune sentinels and release chemokines, cytokines, and other factors [53]. However, mast cells exhibit complex immunomodulatory effects and can have both pro- and anti-tumorigenic properties [54]. AML often results in dysfunction of cytotoxic cells, such as CD8⁺ T cells or NK cells [55]. But in the CS3 subgroup, there is a high composition of naive B cells, CD8 T cells, CD4 memory resting T cells, regulatory T cells (Tregs), and resting natural killer (NK) cells. Furthermore, Tregs are a subpopulation of CD4⁺ T cells, the accumulation of suppressive Tregs can impair the activation and proliferation of cytotoxic T-lymphocytes in AML [56]. Williams et al. [57] reported significantly higher frequencies of Tregs in relapsed AML patients compared to newly diagnosed AML patients.

Our findings suggest that the CS1 subgroup may respond well to cytarabine, which is a basic chemotherapy drug for AML. Combining anthracycline drugs with cytarabine represents the longest-standing and most commonly used induction therapy for AML chemotherapy in history [58]. The CS1 subgroup is highly likely to benefit from conventional AML chemotherapy. Bexarotene, a specific agonist of RXRs, is a potential therapeutic agent that may benefit the CS2 subgroup, clinical trials have found that bexarotene induces differentiation responses in non-APL leukemias [59]. Experimental findings in different AML cell lines have shown that the combination of bexarotene with all-trans retinoic acid (ATRA) enhances the differentiation of certain AML cell lines [60]. Rapamycin's IC50 may be lowest in the CS3 subgroup. The mammalian target of rapamycin (mTOR), plays a vital role in regulating cell growth, metabolism and proliferation [61]. As an mTOR inhibitor, rapamycin has demonstrated efficacy in suppressing both immune responses and tumor growth [62]. Activation of the PI3K-Akt-mTOR signaling pathways contributes to blast cell proliferation [63], and targeted inhibition of mTOR holds promise for AML treatment [64]. A phase II clinical study found that decitabine in combination with rapamycin demonstrated a higher composite complete response rate compared to decitabine alone [65]. This finding may offer insights into potential treatment strategies for the CS3 subgroup. Histone deacetylases (HDACs) promotes extensive histone acetylation and leads to a significant increase in reactive oxygen species and DNA damage in malignant hematologic tumor cells [66]. Although it has been previously reported that HDAC inhibitors can enhance the cytotoxicity of anthracycline drugs and cytarabine against acute myeloid leukemia (AML) cells [67]. However, a recent randomized phase III study with 738 patients found that in young AML patients undergoing high-dose cytarabine induction therapy, the prognosis was not improved with or without vorinostat [68]. Further research is still needed to explore the potential of HDAC inhibitors as a therapeutic agent against hematologic malignancies. Sorafenib, a multi-kinase inhibitor, primarily targets FLT3 in AML treatment [69]. FLT3 mutations are present in nearly 30 % of AML cases and are associated with a poor prognosis [70]. Sun et al. [71] reported that sorafenib maintenance therapy improves outcomes after hematopoietic cell transplant (HCT) for patients with FMS-like tyrosine kinase 3 - internal tandem duplication (FLT3-ITD) AML. Additionally, a clinical study has shown that adding sorafenib to standard chemotherapy improves survival rates for newly diagnosed AML patients [72]. The high frequency of FLT3 mutations observed in both the CS1 and CS3 subgroups, where sorafenib IC50 is predicted to be lowest, aligns well with the results of some clinical trials. Members of the B-cell lymphoma 2 (BCL-2) protein family serve as core regulators of apoptosis, and resistance to apoptosis is a hallmark of cancer [73]. Venetoclax, a representative inhibitor targeting BCL-2, has been approved by the U.S. Food and Drug Administration (FDA) for the treatment of chronic lymphocytic leukemia (CLL) and AML, achieving favorable clinical outcomes [74]. Navitoclax and venetoclax have differences in their mechanisms of action and chemical structures. Navitoclax simultaneously inhibits BCL-2 and B-cell lymphoma-extra large (BCL-XL), while venetoclax specifically targets BCL-2 [75]. Although both the CS1 and CS4 subgroups may benefit from treatment with BCL-2 inhibitors, the prospect of using navitoclax is still under exploration, considering its potential to cause dose-dependent thrombocytopenia [76]. The sensitivity analysis of drug IC50 among the four subgroups may offer novel insights into exploratory clinical trial strategies involving the combination of multiple targeted therapies. However, the results of this study still require further clinical valid.

5. Conclusions

This study aimed to construct a multi-omics-based classification model for TCGA-LAML cohort, comprising four subgroups, through the integration of consensus ensembles derived from various advanced clustering algorithms. The conclusions are summarized as follows: Age, CALGB Category, CS2, and CS3 subgroups are all independent risk factors affecting the survival of AML patients, whereas prior treatment, neoadjuvant, CS4 subgroup, and the M3 subtype in FAB classification are independent protective factors influencing the survival of AML patients. For the CS4 subgroup, which has the longest survival time, although the high-frequency mutated gene WT1 is unfavorable, it exhibits high instability and limited prognostic impact. And another high-frequency mutated gene, KIT, does not adversely affect the prognosis of the main study population, Caucasian individuals. CS1 and CS2 subgroups are mainly characterized by typical FLT3, NPM1, and DNMT3A gene mutations. For the CS3 subgroup with the poorest prognosis, apart from factors such as advanced age and complex cytogenetic abnormalities, TP53 mutations are likely to be prognostic markers for poor chemotherapy response in CS3 patients. Of course, the M3 subtype, with the highest cure rate in AML FAB classification, will have an overall longer survival time due to its high proportion in the CS4 subgroup. The immune microenvironment of AML is highly complex. The role of a high proportion of mast cells in the CS4 subgroup remains to be further investigated. Although the CS3 subgroup contains a higher number of cytotoxic cells such as CD8⁺ T cells or NK cells, a high proportion of Tregs may actively inhibit or downregulate the anti-tumor response of immune cells in the tumor microenvironment. Predicting drug sensitivity among different subgroups is of great significance for conducting experimental clinical studies and advancing precision medicine treatment strategies. The CS1 and CS2 subgroups may benefit from FLT3 inhibitors, which align with the genetic mutation characteristics of both groups. Additionally, the CS1 subgroup is sensitive to the traditional chemotherapy drug cytarabine, while the CS2 subgroup may benefit from RXR agonists, with their mechanism inducing non-APL leukemia differentiation being worthy of investigation. In the CS3 subgroup, mTOR inhibitors have the lowest IC50, indicating potential for treating AML relapse and refractory cases through the PI3K-Akt-mTOR pathway. The CS4 subgroup shows better responses to HDAC inhibitors and BCL-2 inhibitors. Although the clinical efficacy of HDAC inhibitors in treating AML remains controversial, investigating whether their combined use with other small molecule targeted drugs confers survival benefits to patients would be an interesting research topic.

Integrated multi-omics analysis is pivotal in the development of precision medicine. Nonetheless, the comprehensive development of AML multi-omics databases still requires substantial funding, time, and researchers, multi-omics data with pairwise sufficiently detailed clinical information needs to be set up in the future. To a certain extent, there are a few limitations to our study. Insufficient sample size may affect the number of clusters in the clustering analysis. Additionally, due to the challenge of obtaining high-quality multi-omics data with corresponding clinical samples from external cohorts, external validation is only applied based on mRNA and miRNA data. Furthermore, the majority of samples in the TCGA-LAML cohort are from Caucasian individuals, and this cohort does not provide WHO classification information for AML patients. There is no direct literature evidence supporting the relationship between high-frequency mutated genes among different subgroups and prognosis in this study. Additionally, the drug sensitivity analysis is purely predictive in nature, and its results require validation through clinical trials.

The rapid advancements in sequencing technologies have ushered in a new era of molecular-based precision medicine, significantly advancing our understanding of the molecular pathogenesis of acute myeloid leukemia (AML) over the past decade. The successful clinical applications of FLT3, IDH1/IDH2, and BCL-2 inhibitors underscore the progress made in precision medicine for AML treatment [1]. However, given the intricate interplay of multiplex signaling pathways, epigenetic variations, and non-coding RNA networks, reliance on single-agent inhibitor treatments may lead to disease relapse. Therefore, there is a pressing need for multi-target treatments that combine novel agents to reduce the rates of AML relapse and refractory cases compared to monotherapy with single-agent inhibitors [77]. Undoubtedly, precision medicine heralds a transformative epoch in the domain of leukemia treatment, ushering in a new era characterized by personalized therapeutic interventions tailored to the unique attributes of individual patients.

Ethics approval and consent to participate

The research data were sourced from public databases. There is no applicable requirement for ethics approval and consent to participate.

Consent for publication

Not applicable.

Data availability statement

All the data in this article are sourced from publicly available databases. TCGA data set can be obtained from the following website: https://portal.gdc.cancer.gov/, GEO data set can be obtained from the following website: https://www.ncbi.nlm.nih.gov/geo.

CRediT authorship contribution statement

Teng Wang: Writing – original draft, Visualization, Validation, Conceptualization. Siyuan Cui: Writing – review & editing, Methodology. Chunyi Lyu: Writing – review & editing, Software. Zhenzhen Wang: Validation, Formal analysis. Zonghong Li: Formal analysis, Conceptualization. Chen Han: Validation, Data curation. Weilin Liu: Data curation. Yan Wang: Writing – review & editing, Supervision, Funding acquisition, Conceptualization. Ruirong Xu: Supervision, Funding acquisition, Conceptualization.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary data

The following are the Supplementary data to this article: