A novel immune prognostic model of non-M3 acute myeloid leukemia

Abstract

Acute myeloid leukemia (AML) is a common hematological malignancy in adults. AML patients exhibit clinical heterogeneity with complications of molecular basis. The leukemogenesis of AML involves immune escape, and the immunosuppression status of the patient might have great impact on AML treatment outcome. In this study, we established an immune prognostic model of AML using bioinformatics tools. With the data in the TCGA and GTEx datasets, we analyzed differentially expressed genes (DEGs) in non-M3 AML and identified 420 immune-related DEGs. Among which, 49 genes’ expression was found to be related to AML prognosis based on univariate Cox regression analysis. Next, we established a prognostic model with these 49 genes in AML by LASSO regression and multivariate Cox regression analyses. In our model, the expressions of 5 immune genes, MIF, DEF6, OSM, MPO, AVPR1B, were used to stratify non-M3 AML patients’ treatment outcome. A patient’s risk score could be calculated as Risk Score=0.40081 × MIF (MIF expression) - 0.15201 × MPO + 0.78073 × DEF6 - 0.45192 × AVPR1B + 0.25912 × OSM. The area under the curve of the risk score signature was 0.8, 0.8, and 0.96 at 1 year, 3 years, and 5 years, respectively. The prognostic model was then validated internally by TCGA data and externally by GEO data. At last, the result of single-sample gene-set enrichment analysis demonstrated that compared with healthy samples, the abundance of non-turmeric immune cells was significantly repressed in AML. To summarize, we presented an immune-related 5-gene signature prognostic model in AML.

Introduction

Acute myeloid leukemia (AML) is a common myeloid carcinoma with high heterogeneity [1,2]. Based on the French-American-British (FAB) classification system, AML is generally classified into 8 subtypes according to the leukemia cells’ morphology and cytogenetic characteristics of leukemic cells in hemogram and myelogram [2]. Different subtypes of AML have different prognosis. After the use of all-trans retinoic acid clinically, acute promyelocytic leukemia (APL) becomes the most curable subtype of AML with a cure rate of 80% [3,4]. However, for patients with other subtypes of AML, the 5-year overall survival (OS) is approximately 45% in young patients (≤ 60 years old), whereas it is less than 10% in old patients (> 60 years old) [5]. Thus, AML patients’ stratification and accurate treatment regimen selection remain important topics for researchers and clinical specialists.

It has long been recognized that immune features of AML patients affect the prognosis [6,7]. A previous study demonstrated that adult AML patients with high level of IgA2 B cells were related to poor OS [7]. In addition, TIGIT⁺ natural killer (NK) cells were related to poor outcome in AML patients [8]. On the other hand, immunotherapy, such as CD33 monoclonal antibody, has been used in AML treatment and showed promising in some patients. Other immunotherapies, including chimeric antigen receptor (CAR) T cell and bispecific T cell engagers (BiTEs), are widely tested in different clinical trials. European Leukemia Net (ELN) recommendations, which are based on cytogenetic abnormalities and genetic mutations, are commonly used hazard stratification in AML. To our knowledge, there is no recommended AML prognostic model that was based on immune genes expression. In this study, we established an immune model to predict AML treatment outcome. The model itself, as well as identified key genes expression, might help to improve AML immunotherapy.

Materials and methods

Data acquirement

RNA sequencing data of AML were acquired from the TCGA database (https://portal.gdc.cancer.gov/repository) and the GEO database (https://www.ncbi.nlm.nih.gov/geo/). The GTEx database (https://www.gtexportal.org/home/) was conducted to download gene expression data for 337 normal whole blood cell samples. The training data of 116 non-M3 AML patient samples were collected from TCGA-LAML, after excluding patient with incomplete clinical information. An external verification was performed using GSE37642 data from GEO. The immune-related genes (IRGs) were obtained from the ImmPort database (https://www.immport.org/).

Screening differentially expressed genes

Data processing and analysis were based on R (Version 4.1.0). The count matrices of the TCGA and GTEx datasets were converted to raw counts, and then the “cpm” function of “edgeR” package removed the bias of the library sequencing depth. After intersection, 26792 genes were obtained between AML patient samples and normal whole blood control samples. Gene difference analysis was conducted with the “limma”, “DSeq2” and “edgeR” packages, and genes with a Log₂ |Fold Change| of > 2 and P value ≤ 0.05 were identified as the differentially expressed genes (DEGs). Visualize DEGs obtained in 3 packages with “VennDiagram” package, and we used the “union” function to merge the genes. Immune-related DEGs in AML were obtained by intersecting DEGs with IRGs.

Biological functions and enrichment analysis

Enrichment analyses, including Gene Ontology (GO) functional enrichment analysis, Kyoto Encyclopedia of Genes and Genomes (KEGG) analysis and Gene Set Enrichment Analysis (GSEA), of DEGs and AML-immune-related DEGs were conducted with the “cluster Profiler” [9], “enrichplot” and “org.db.hus” packages of R. “Immunologic signature gene sets” and “hallmark gene sets” (http://www.gsea-msigdb.org/gsea) from GSEA were used.

Establishment and validating an immune prognostic model in AML

The correlation between target gene expression and OS of AML patients was analyzed using univariate Cox regression with R packages “survival”, “glmnet”, “survminer”, “My.stepwise”. Then we used LASSO regression analysis to identify candidate genes for the risk score signature. At last, a multivariate Cox regression analysis was performed to investigate candidate genes and establish the prediction model.

To validate the prognostic model with training dataset from TCGA as described earlier, non-M3 AML patients (n=116) were randomly separated into two parts according to a proportion of 1:1, and one group was defined as an internal validation dataset. For external validation, GSE37642 dataset from GEO was used. R packages used in survival analysis were “survival”, “glmnet”, “survminer”, “My.stepwise”, etc.

Mutation profiles analysis

Somatic mutation profiles of 51 none-M3 AML patients were obtained from TCGA. In addition, the waterfall plot showing the mutations were generated and analyzed by “maftools” R package.

Evaluation of tumor microenvironment immune cell infiltration

The abundance of non-turmeric cells in PBMCs of AML patients was analyzed as previous described [10]. A genomic set of immune cells was established according to Pornpimol Charoentong [11]. The single-sample gene-set enrichment analysis (ssGSEA) algorithm [12], which standardized the rank of genes expression in a specified sample and calculated the enrichment score with the use of the empirical cumulative distribution function, was used to evaluate the difference of enriched immune cells in tumor microenvironment (TME) between normal and AML samples. The ssGSEA was also applied for non-M3 AML patients’ data to perform immunocytroverance assessment to compare Cox prognostic risk score in high-risk and low-risk groups [13,14]. The principal component analysis (PCA) was used to reduce the dimension of the data, mainly through linear transformation to transform the original data into a group of linearly independent data, followed by extraction of the main feature components, so as to roughly judging the quality of the data. “GSVA”, “estimate” and “pheatmap” R packages were used.

Statistical analysis

All statistical analyses were conducted with R software (version 4.1.0). Based on the median hazard score, the patients were separated into high-risk group and low-risk group. DEGs were identified using students’ t-test. The Kaplan-Meier (KM) curve was conducted to show the prognosis of two groups of patients. Analyzing KM survival curves required the log-rank test. And the area under the curve (AUC) of the receiver operating characteristic (ROC) curve was conducted to assess the quality of the model by “timeROC” packages. We calculated correlations with Spearman’s correlation coefficients. A P value < 0.05 is considered statistically significant in all analyses.

Results

Identification of differentially expressed immune-related genes in AML

The research flow chart was shown in Figure 1. The DEGs between non-M3 AML patients and healthy individuals were screened as described in Method and Materials section. A summary of the basal clinical information of non-M3 patients from the TCGA-LAML dataset was presented in Table 1. The quality of original RNA sequencing data used in our study was analyzed by PCA (Figure 2A). AML and healthy individuals exhibited significantly different patterns of gene expression. As exemplified in a heatmap in Figure 2B, there were 14179 DEGs identified in AML, of which 11877 were upregulated and 2302 were downregulated. Next, we further extracted IRGs in DEGs. The list of IRGs, which comprised a total of 2483 IRGs and covered 17 immune categories, was obtained from ImmPort as described earlier [15]. After intersection of DEGs and IRGs, we found 420 differentially expressed IRGs in AML (Figure 2C). Detailed information of AML immune-related DEGs was presented in Table S1.

Figure 1.

The flow chart of establishing and validating an immune prognostic model in AML.

Table 1.

Clinical characteristics of non-M3 AML patients in the TCGA-LAML dataset

	Number	Percentage (%)
FAB category
M0	12	10.3%
M1	31	26.7%
M2	31	26.7%
M4	27	23.3%
M5	12	10.3%
M6	2	1.7%
M7	1	0.9%
Gender
male	64	55.2%
female	52	44.8%
Survival status
dead	75	64.7%
alive	41	35.3%
Age (years)
> 57	55	47.4%
≤ 57	61	52.6%
Median (range)	57 (21-88)
Overall survival days (median)	365

Figure 2.

Identification of differentially expressed immune-related genes in AML. A. Principal component analysis plot of normal and AML samples; B. Heatmap of differentially expression genes (DEGs) between normal and AML samples; C. Venn diagrams of immune-related genes (IRGs) and DEGs.

Next, we conducted GO and KEGG enrichment analyses with those 420 immune-related DEGs. In GO enrichment, the enriched pathways were categorized into molecular functions (MFs), cellular components (CCs) and biological processes (BPs). The top three enriched pathways in BPs were humoral immune response, cell chemotaxis and leukocyte chemotaxis. The top three enriched pathways in CCs were T-cell receptor complex, plasma membrane signaling receptor complex and immunoglobulin complex. And the top three enriched pathways in MFs were signaling receptor activator activity, receptor ligand activity, and cytokine activity (Figure 3A, 3B). In addition to annotating the function of genes, KEGG enrichment also showed relationship between genes and signaling pathways. The KEGG analysis results of immune-related DEGs revealed a high correlation with cytokine-cytokine receptor interaction, viral protein interaction with cytokine and cytokine receptor, chemokine signaling pathway (Figure 3C, 3D). In addition, GSEA enrichment provided alternative gene and pathway classification. Therefore, we performed GSEA pathway enrichment with 420 immune-related genes. Those 420 differentially expressed IRGs in AML were enriched in KRAS signaling, L6-JAK-STAT signaling and allograft rejection pathways in GSEA HALLMARK gene sets (Figure 3E). Specifically, we used immunologic signature gene sets in GSEA to analyze immunological pathway enrichment. A total of 36 immunological pathways were significantly enriched, with top 3 enrichments were NAÏVE BCELL vs NEUTROPHIL, HEALTHY vs RSV INF PBMC, CYANOBACTERIUM LPSLIKE vs LPS AND CYANOBACTERIUM LPSLIKE STIM DC 3H pathways (Figure 3F). More information of GSEA pathway enrichment was presented in Table S2. To summarize, with published gene expression data, we identified DEGs in AML. We selected immune-related genes in DEGs and analyzed those genes’ enrichment in pathways.

Figure 3.

Enrichment analysis of AML-immune-related DEGs. A. Barplot of Gene Ontology (GO) biological functions analysis; B. Circular cnetplot of GO pathways; C. Emapplot of Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways; D. Dotpplot of KEGG enrichment analysis; E. Enrichment plot of Hallmark pathways; F. Enrichment plot of Immunologic pathways.

The results of differential gene enrichment analysis revealed diverse signaling pathways related to cytokines. Of notice, many identified differentially regulated IRGs, such as IL-10, TGF-β, and IFN-gamma, had been shown to function as immunosuppressive factors in AML or other human cancers. Tumor cells could secrete chemokines to stimulate immune cell migration, thus modulated the infiltration of immune cells in TME. AML could recruit immunosuppressive cells, such as myeloid-derived suppressive cells (MDSCs), regulatory T cells (Tregs) and tumor-associated macrophages (TAMs), all of which were able to release cytokines that suppress the immune system [16-20]. There were evidences that TGF-β and IL-10 could suppress T cell activation and proliferation [21,22]. In addition, by binding to membrane-bound TGF-β, Tregs suppressed the functions of dendritic cells and the cytotoxicity of NK cells [23-26]. When stimulated by Galectin-9, NK cells produced IFN-gamma, which was an enzyme responsible for the breakdown of tryptophan, then tryptophan consumption and its metabolites would lead to T-cell apoptosis [27,28]. PD-L1 could also be induced by IFN-gamma, allowing AML cells to escape immune surveillance [29-32].

Establishing and validating an immune prognostic model in AML

We started with 420 immune-related DEGs in AML and examined the correlation between each gene’s expression and AML OS. The result of univariate Cox regression analysis suggested that 49 gene’s expression each correlated with AML OS (P < 0.05, Figure S1A). Next, we generated a prognostic model using LASSO regression analysis with of these 49 progno-stic genes. A total of 10 genes were identified as potential candidate genes according to the optimal value of λ (Figure 4A, 4B). The AUC of LASSO regression prognostic model was 0.757 (Figure S1B). Next, we conducted multivariate Cox regression analysis of those 10 genes, and identified a 5-gene signature in AML prognosis prediction with a C-index of 0.72 (Figure 4C). Those genes were MIF, MPO, DEF6, AVPR1B, and OSM. In order to calculate risk scores, the following formula was used: Risk Score=0.40081 × MIF (MIF expression) - 0.15201 × MPO + 0.78073 × DEF6 - 0.45192 × AVPR1B + 0.25912 × OSM. Time-ROC curve demonstrated that the AUC of the model was 0.8, 0.8, and 0.96 at 1 year, 3 years, 5 years, respectively (Figure 4D).

Figure 4.

Establishing an immune-related prognostic model in AML. A, B. Lasso regression of 49 AML-immune-related DEGs; C. Forest plot of multivariate Cox regression analysis; D. Time ROC curve of multivariate Cox regression; E. Kaplan-Meier (KM) curve between the high-risk group and low-risk group; F. The relationship between the risk score, overall survival, living status and the expression of 5 immune signatures; G. The expression level of 5 IRGs between the high-risk and low-risk group (**P < 0.01, ***P < 0.001, ****P < 0.0001).

To examine the risk-stratification efficiency of our model, we calculated risk scores for each of 116 AML patients from TCGA, and divided the patients into high-risk and low-risk groups with the median value. According to the KM curve, high-risk patients had a significantly shorter OS than low-risk patients (P < 0.0001, Figure 4E). As exemplified in Figure 4F, 4G, MIF, DEF6, OSM were upregulated in high-risk group, while MPO and AVPR1B were downregulated. In addition, the result of nomogram also indicated that the 5-gene signature could predict the survival of non-M3 AML patients with a C-index of 0.72 (Figure 5A, 5B). The prognostic model showed effective in different stage, gender, age, cytogenetic background groups of patients (Figure 5C). More information about above univariate Cox regression analysis was presented in Table S3.

Figure 5.

Results of multivariate Cox regression analysis. A. Nomogram plot; B. The calibration curve of nomogram plot; C. Survival analysis of 8 clinical subgroups according to risk scores.

Next, we evaluated our prognostic model with an internal validation and an external validation. For internal validation, we randomly divided non-M3 AML patients (n=116) from TCGA into two groups, the training set and test set, with the proportion of 1:1. The AUC of risk scores at 1 year, 3 years, and 5 years were 0.81, 0.78, and 0.92, respectively (Figure 6A). For external validation, we used GSE37642 dataset. The clinical information of patients in GSE37642 was presented in Table 2. KM curve demonstrated that high-risk patients had a noticeably shorter OS than low-risk patients (Figure 6B). And the AUCs were 0.66, 0.7, and 0.74 at 1 year, 3 years, and 5 years, respectively (Figure 6C). Overall, our 5-gene model provided good performance in AML prognosis prediction.

Figure 6.

Validation of the immune-related AML prognostic model. A. Time ROC curve of the internal validation; B. KM curve of GSE37642; C. Time ROC curve of external validation.

Table 2.

Clinical characteristics of non-M3 AML patients in the GSE37642 dataset

	Number	Percentage (%)
FAB category
M0	22	4.3%
M1	113	22.1%
M2	164	32.1%
M4	121	23.7%
M5	66	12.9%
M6	22	4.3%
M7	3	0.6%
Survival status
dead	382	74.8%
alive	129	25.2%
Age (years)
> 57	251	49.1%
≤ 57	260	50.9%
Median (range)	58 (18-85)
Overall survival days (median)	342

Mutation profiles in non-M3 AML

We also examined mutation profiles of non-M3 AML with published dataset in TCGA. As exemplified in Figure 7A, missense mutations were the most common genetic aberration in AML. Single nucleotide polymorphism was also common in non-M3 AML, with C > T was identified as the most prevalent single nucleotide variants (Figure 7B, 7C). The number of mutations in each patient was summarized in boxplot in Figure 7D. Top 10 mutated genes in non-M3 AML were exhibited in Figure 7E. Those genes were DNMT3A, RUNX1, NPM1, FLT3, IDH2, KIT, TP53, TTN, MUC16, and NRAS.

Figure 7.

Mutation profiles in non-M3 AML patients. A-C. Various classifications of mutation based on different groups; D. Burden of tumor mutation in selected samples; E. Waterfall plot of the mutation details.

Evaluation of immune cell infiltration in AML tumor microenvironment

According to ssGSEA analysis, most non-malignant cells were differentially infiltrated in AML samples versus healthy samples, except for immature B cell (Figure 8A). We investigated the immune activity and stromal activity in AML TME by the ESTIMATE algorithm [45]. The result demonstrated that immune and stromal activities were noticeably repressed in AML (Figure 8B). We also conducted correlation analyses to determine whether there was a relationship between risk scores and the type of TME cells in APL patients. As exemplified in Figure 8C, a positive correlation between risk scores and TME cell infiltration was observed in central memory CD8 T cell, T follicular helper cell, active CD4 T cell, active B cell, type 1 T helper cell, MDSC, neutrophil, immature B cell, Tregs, monocyte, NK T cell, CD56 bright NK cell, effector memory CD4 T cell and macrophage. For the purpose of exploring the relationship between target genes and TME infiltrating cells, we conducted spearman correlation analyses and found that DEF6 was positively correlated with the majority of the TME infiltrating immune cells, while AVPR1B was negatively correlated (Figure 8C). In summary, non-tumor immune cells were remarkably reduced in AML samples compared to healthy controls. This finding deserves further investigation.

Figure 8.

Evaluation of immune cell infiltration in AML tumor microenvironment. A. Boxplot of immune cell abundance between normal and AML samples; B. Using estimation algorithm to calculate immunity score and stromal score between normal and AML samples; C. The correlation between risk score and immune cells infiltration in non-M3 AML patients; D. The correlation between immune cells and 5 IRGs (*P < 0.05, **P < 0.01, ***P < 0.001).

Discussion

Immunotherapy has achieved good results in many human cancers [46-48]. In AML, allogeneic hematopoietic stem cell transplantation (alloHSCT) is the earliest and the most effective immunotherapy [49]. Emerging immunotherapy, such as CD33 monoclonal antibody has been shown to reduce recurrence risks of AML and improve outcomes among some patients [50,51]. At present, anti-CD33 monoclonal antibody is mainly used in newly diagnosed CD33+ AML patients, including those with favorable or intermediate cytogenetic risk profiles, or elderly patients (> 60 years) without actionable mutations who are not suitable for induced remission therapy. CD33 monoclonal antibody can also be used to treat CD33 positive relapsed/refractory AML patients [52,53]. In addition, many AML immunotherapy agents including immune checkpoint inhibitors [54], target antigen antibodies [55], and CAR-T cells have been developed and validated in pre-clinical and clinical stages [56]. However, because of the low mutation load, strong immunosuppression and low immunogenicity of tumor cells in AML patients, scientists have difficulties to identify antigen-specific cytotoxic immune cells for AML treatment [57]. In addition, the heterogeneity nature of AML also caused different responses to immunotherapy in patients. At present, the optimal single therapeutic target of AML has yet been identified. Most AML immunotherapies are used in conjunction with conventional therapies including chemotherapy and HSCT [53].

According to patients’ cytogenetic and molecular characteristics, ELN stratifies AML patients into three risk layers: poor, intermediate and favorable. Such stratification is also used to guide treatment and to predict prognosis [58]. Modifications of ELN stratification with additional clinical information, such as age and minimal residual disease status, result in better prediction of AML outcome [59,60]. Of notice, immune microenvironment is important for the development and treatment of AML [57], in particular in the era of AML immunotherapy. It might be interesting to develop immune-related prognostic model in AML and evaluate its significance. In this study, we first screened AML-specific IRGs, and then identified OS correlated IRGs. At last, we conducted LASSO regression and univariate Cox regression to generate a 5-gene panel to predict non-M3 AML prognosis. The AUC of our model was 0.8, 0.8, and 0.96 at 1 year, 3 years, 5 years, respectively in training cohort, with a C-index of 0.72. In the past decade, many models have been proposed to predict AML prognosis [10,61,62]. Recently, Zhu et al. identified six genes associated with immune response to predict AML prognosis, and the AUC was 0.6643 [63]. In 2021, Li et al. proposed an immune-17 signature prognostic model in AML with an AUC of 0.823 [64].

The 5-gene model, as well as related pathway enrichment, also provided molecular basis of prognosis-related immune aberrations in AML. Within the 5-gene panel, MIF and MPO have been investigated in AML. MIF is a pro-inflammatory cytokine. MIF binds to CXC-family chemokine receptors 2 and 4, and initiates cell signaling transduction in target cells [65,66]. A previous study showed that overexpression of MIF was associated with inferior AML prognosis [67]. Mechanistically, AML-derived MIF enhanced IL-8 expression, and promoted AML tumor cell proliferation and survival [68,69]. MPO (myeloperoxidase) is a pivotal lineage marker used for AML diagnosis [70]. A previous study showed that MPO enhanced ROS production in AML cells, and affected AML chemosensitivity [71]. High MPO expression was associated with favorable prognosis of AML [72,73]. Within the 5-gene model, the function of AVPR1B, DEF6 and OSM in AML has yet been recognized. DEF6 is a guanine nucleotide exchange factor for RAC1 and CDC42 [74,75]. The high level of DEF6, which contributes to the regulation of cell cycle [76], is associated with the unfavorable prognosis of various cancers, including osteosarcoma [77], clear cell renal cell carcinoma [78], ovarian carcinoma [76]. OSM is an active IL-6 family cytokine which inhibits tumor cell proliferation through the JAK/STAT3 pathway [79-81]. In addition, OSM presents essential to immune homeostasis and haematopoiesis, inflammation, so that the change of OSM activity promotes cancer [82]. A high expression level of OSM is related to poor outcomes in breast cancer [83] and the promotion of breast cancer metastasis [84,85]. AVPR1B is one of the receptors of Arginine vasopressin and positively correlated with the elevation of plasma adrenocorticotropin in corticotropinomas [86], the specific AVPR1B agonists could be used for rapid correction of anemia after bleeding, drug toxicity, or chemotherapy [87]. It is important to elucidate the functions of AVPR1B, DEF6 and OSM in AML in future studies. Since MIF and DEF6 are overexpressed in AML and positively correlated with AML prognosis, those two molecular might be used as diagnostic markers and therapeutic targets.

The success of 5-gene model in prediction of AML prognosis indicated the functions of those 5 genes in AML progression and/or treatment response. In a view to immunosuppression functions of those 5 genes, MIF promoted the polarization of M2-TAMs, which were critical immunosuppressive cells [88,89]. Moreover, previous studies demonstrated that MIF bond to macrophage receptor CD74 and activated the ERK pathway within cells, a pathway that promoted tumor cell proliferation [90,91]. Additionally, MIF was able to inhibit NK cell activity by downregulating NKG2D receptors and promoted ovarian cancer immune escape [92,93]. MPO is a marker for distinguishing myeloid or lymphoid leukemia, but its role in immune escape is not clear [94]. The immunosuppressive functions of OSM, DEF6 and AVPR1B have yet been recognized. Previous studies demonstrated that the immunosuppressive microenvironment of AML was related to the increase of Tregs and MDSCs in the peripheral blood and bone marrow (BM) [57]. Tregs promoted the expansion and recurrence of AML blasts by secreting immunosuppressive cytokines and increasing ATP hydrolysis [21,95]. Moreover, by expressing ectonucleotides CD73 and CD39, leukemia cells could produce adenosine, thereby suppressing the function of effector T cells and promoting Tregs [96]. The number of Tregs was negatively correlated with the prognosis of patients [57]. In MDSCs research, previous studies shown that AML tumor cells mimicked MDSCs in suppressing anti-AML T cell responses [97,98]. Specifically, AML tumor cells could induce T cells and NK cells apoptosis via ROS [97,99]. Alternatively, AML tumor cells also regulated arginine metabolism in T cells and inhibit cytotoxic lymphocytes via tryptophan metabolism [100-102]. At last, direct contact between NK cells and tumor cells could also lead to dysfunction of NK cells [103,104]. Overall, our findings from ssGSEA analysis agreed with those bench data.

As a conclusion, we developed an immune gene expression model to predict the outcome of AML patients. There were several limitations of this study. First, the model was only validated by clinical informatics data. A multicenter clinical study might provide further validation of the prediction model. Second, the infiltration of immune cells in AML TME was not validated. AML patient BM aspiration samples could be used for immune cells counting to validate our result from bioinformatics analysis. Third, mechanisms underline 5-gene prediction model, as well as characteristic immune cells infiltration in AML TME, were not addressed in this study. In particular, the immunosuppression functions of identified 5 genes in AML might provide great insight to AML therapy.

Disclosure of conflict of interest

None.