The establishment of a prognostic scoring model based on the new tumor immune microenvironment classification in acute myeloid leukemia

We aimed to use all databases that were accessible and included overall survival (OS) data for the patients. At the time of this study, there was a total of eight AML cohorts containing gene expression data and corresponding clinical information. Seven of the cohorts were from the Gene Expression Omnibus (GEO) database (https://​www.​ncbi.​nlm.​nih.​gov/​geo/​): GSE10358, GSE66525, GSE8970, GSE12417, GSE37642, GSE6891, and GSE71014. There were 1799 AML samples, of which 1229 samples had OS records. The Affymetrix microarray data sets of GSE10358, GSE12417, GSE37642, and GSE6891 were downloaded as CEL files and normalized by multi-chip averaging (R package affy, V1.60.0). The data of GSE71014, GSE66525, and GSE8970 were downloaded in the form of a normalized expression matrix. The eighth database came from the Cancer Genome Atlas (TCGA) database, which contained 173 AML patients with gene expression and prognosis data. Among these datasets, GSE12417 was only composed of cytogenetically normal (CN-) AML patients. The removed Batch Effect function of the R language limma package was used to remove batch effects [18], the expression matrices of the 7 GEO databases were merged, and then quantile normalization was performed through R package preprocessCore for model construction [19].

The XCell tool (https://​xcell.​ucsf.​edu) was used to analyze the expression matrix to infer the cellular components in the immune microenvironment of 1799 AML samples in the GEO datasets [20]. According to the median of each cell infiltration score, patients were divided into high and low groups, and the survival differences between the two groups were compared to evaluate the prognostic value of various cell infiltrations.

According to the infiltration scores of all cells with prognostic significance, all samples were hierarchically clustered based on Euclidean distance and Ward linkage to construct tumor immune microenvironment (TIME) classification.

The differential expression genes (DEGs) were analyzed by DEGseq. Those genes were considered as DEG by using thresholds for both the (adjusted) p-value and a fold-change (adj.P.Val < 0.05, FC > 1.5) [21]. Log-rank test and univariate COX regression analysis were used to screen differentially expressed genes with prognostic significance. DAVID was used (https://​david.​ncifcrf.​gov/​) for GO enrichment analysis, and the R software package was used for “GOPLOT” visualization [22]. Then the Cox-PH method based on Lasso was applied to establish the AML prognosis model. The risk index score of each patient was calculated, which was the sum of the expression of all genes in the model multiplied by their corresponding weighting coefficients. The median of the risk index scores in each group was identified as the cutoff value. Time-dependent receiver operating characteristic (ROC) curve analysis and Kaplan-Meier survival analysis were used to evaluate the predictive effect of risk index scores on the prognosis of AML patients (R package, survival ROC, v1.0.3).

To validate the predictive efficiency of the model, we firstly selected three GEO databases with a large number of patients: GSE10358, GSE37642, and GSE6891 for internal validation, and TCGA database for external validation; secondly, we chose GSE12417 to test the efficiency of our prognosis model in cytogenically normal AML (CN-AML) patients. At the same time, the GSE6891 and TCGA databases were selected to assess the prediction effects of the model in three different groups of AML patients, including good-risk, intermediate-risk (IR-), and poor-risk AML patients. All patients in the database were scored using the model, and the median score was used to divide the patients into high and low groups. Kaplan-Meier survival analysis was used to draw survival curves. The ELN system, LSC17, Wang and Yang models were collected, and the time/survival ROC curve and the concordance index (C-index) analysis were used to evaluate the prognostic performance.

The clinical information of AML patients from 7 GEO databases was shown in Table 1. Data information mainly came from 4 platforms, GPL10558 (n=104), GPL10532 (n=22), GPL570 (n=1055), and GPL96 (n=618). A total of 1799 patients were enrolled, of which 1299 patients had OS information. There were 479 patients in the group older than 60 years and 931 patients younger than 60 years. Regarding survival status, there were 449 patients alive, 884 patients died, and 457 patients with missing data.

To analyze the TIME of AML, we normalized the obtained GEO data after removing the batch effect and used the XCell tool to simulate and infer it by silico analysis. We found that there were 33 cell components in the AML immune microenvironment. Correlation matrix analysis showed that there were mainly B cell groups, T cell groups, and other cell groups (Fig. 1A). In order to evaluate the prognostic value of the infiltration level of these cells, all patients were divided into a high infiltration group and a low infiltration group according to the median cell infiltration score of each type, and the survival differences between the two groups were compared. The results of survival analysis showed that the high infiltration group of CD4⁺/CD8⁺ T cells, B cells, CD8⁺ central memory T cells, Class-switch memory B cells, eosinophils, fibroblasts, mast cells, and NKT cells were all conducive to survival (Fig. 1B–J; all P < 0.05); on the contrary, the group with high hematopoietic stem cell (HSC) infiltration had inferior OS (Fig. 1K; P < 0.0001).

Based on the obtained AML immune microenvironmental cell information, we used infiltration levels of 10 types of cells with prognostic significance for hierarchically clustered AML patients to establish a TIME classification. All the patients were divided into three groups based on the TIME classification. There was a significant difference in survival between the three groups. Cluster 1 had the shortest survival time and Cluster 3 had the longest survival time (Fig. 2A, B). The score characteristics showed that Cluster 1 had the lowest immune score and microenvironment score, and the stroma score of Cluster 1 was the highest; Cluster 3 had the highest immune score and microenvironment score, and the stroma score of Cluster 3 was the lowest; the scores of the Cluster 2 were in the median (Fig. 2C–E).

To establish a prognostic model, we analyzed the DEGs in Cluster 1 and Cluster 3 which demonstrated the largest differences. Compared with Cluster 3, Cluster 1 had 489 upregulated genes and 588 downregulated genes. Among these 1077 DEGs, 366 genes had prognostic significance (Fig. 3A, B). GO analysis showed that these 366 genes were mainly involved in the regulation of the immune system, immune response, defense response, leukocyte migration, inflammatory response, and so on (Fig. 3C). LASSO-Cox was used to identify the genes which were most relevant to prognosis among the 366 DEGs with prognostic significance. The coefficient of each gene was calculated and a proportional hazard model containing 121 genes was established (Fig. 3D, Additional file 1: Table S1).

After establishing the prognostic model, we verified the model with diversity AML cohorts. For 1229 AML patients, the calculated cutoff value was 0.0097, which equally divided the patients into the high and low groups (Fig. 4A). The higher the score, the shorter the patients’ survival time and the higher proportion of deaths (Fig. 4B). Subsequently, we used Kaplan-Meier survival analysis to compare the prognosis of two groups. Compared with the low score group, the survival status of the high score group was worse (Fig. 4C, P < 0.001). The clinical characteristics of patients in the two groups showed that the high score group had more old patients (age ≥ 60), fewer good- and intermediate-risk patients, and more poor-risk patients (Table 2, all P < 0.001). And the area under the curve (AUC) of 1, 2, 3, and 5 years were 0.77, 0.79, 0.81, and 0.77, respectively (Fig. 4D), indicating that our scoring model had high accuracy.

After analyzing the prediction efficiency across all the patients, we used independent GEO databases for validation. After processing the data in GSE34642 and GSE10358 with the same method, there was a significant difference in prognosis between the high score group and the low score group, and the high score was a poor prognostic factor (Fig. 4E, F). In a cohort with the same subtype of AML patients, such as GSE12417, which was all comprised of CN-AML patients, the model successfully divided the patients into high and low score groups with significant prognosis differences. The survival time of patients in the high score group was shorter, which was consistent with the results of other databases (Fig. 4G).

Moreover, GSE6891 and the TCGA database were selected to test the model’s predictive performance in a different stratification. Our model performed well in these two databases, which divided patients into two groups with significant differences in prognosis, and the OS of patients in the high score group was shorter (Fig. 5A, B; both P < 0.001). Similar results were also found in the intermediate risk AML (IR-AML) patients in two cohorts (Fig. 5C, D; both P < 0.001). For patients in the good-risk group, there was no significant difference in both cohorts (Additional file 1: Fig. S1A, B). Finally, the model divided the poor-risk patients of GSE6891 into two groups with different prognoses, and the group with higher scores had inferior OS (Additional file 1: Fig. S1C, P=0.0094). However, the same result was not found in the TCGA database, most likely due to the small number of poor-risk patients (Additional file 1: Fig. S1D).

In recent years, some new AML prognostic models have been proposed. We selected the latest three models to compare their prediction effects with our model: Wang’s model, LSC-17, and Yang’s model [5‐7]. Wang et al. established a model based on the gene expression profiling (RNA sequencing), which demonstrated the best predictive performance compared with previous studies. Therefore, we also compared the models published after Wang. LSC-17 was established on the basis of AML hematopoietic stem cells, and Yang et al. constructed a model using the gene expression profiling. In addition, we also compared our model with the classic ELN risk stratification system. Multivariate survival analysis found that in the GSE6891 and the TCGA database, our 121-gene prognostic model was the only independent prognostic factor for AML patients (Table 3; both P < 0.05). In GSE6891 and the TCGA database, our prognostic model demonstrated the largest AUC of the survival ROC curve and highest C-Index among the five prognostic models, indicating that our prognostic model was more reliable (Fig. 5E–H). At the same time, we also calculated the AUC of the time ROC curve and statistically analyzed the results of the four models. The final results showed that in GSE6891, our model had the highest AUC. The comparison of AUC between our model and LSC17 displayed the significant differences at all times (Additional file 1: Fig. S2A, all P < 0.05). Comparing our model with the Wang and Yang models, there were significant differences in AUC from the 600th day (Additional file 1: Fig. S2A, all P < 0.05). In the TCGA database, although our model demonstrated the largest AUC from the 500th day, there was only a statistical difference between our model and LSC17 on the 1000th day (Additional file 1: Fig. S2B, P=0.042). The possible reason for this result was the small number of patients in the TCGA database.