Overview of AS events in AML
The profile of AS events for 178 patients with AML were analyzed from TCGA cohort. The clinical and molecular characteristics of these cases were summarized in Additional file 1: Table S1. The median age at diagnosis was 55 years (range, 18 to 88), and the median follow-up duration was 16.4 months (range, 1 to 118).
After preprocessing procedure, integrated RNA-Seq data with AS events in seven splicing types were included in the present study. We detected a total of 27,833 AS events of 8,337 genes, comprising 2,136 AA in 1641 genes, 1724 AD in 1367 genes, 5588 AP in 2760 genes, 6317 AT in 3173 genes, 9985 ES in 4762 genes, 135 ME in 133 genes, 1948 RI in 1335 genes, as illustrated in Fig. 1a. A schematic diagram of AS events are shown in Fig. 1b. It should be noted that several mRNA splicing events may be detected in a single gene, and up to 6 types of AS events were observable for one gene (Fig. 1c). In addition, most genes have more than one type of AS events. ES was the predominant type of AS events in AML (36.12%), and these events may provide perhaps a critical process for enriching transcriptome diversity.
AS events associated with prognosis of AML
We performed univariate Cox analysis to evaluate the impact of AS type on OS in AML patients. A total of 3016 survival-related AS events within 1992 genes were identified. Hazard ratios (HRs) greater than or less than 1 accounted for 2.74% and 5.88% of the total AS events, respectively (Additional file 2: Table S2). Meanwhile, several survival-associated AS events can also be detected in one single gene, and different spliceosomes of a few parent genes (such as NBPF11, LRRC23, etc.) exhibited opposite prognostic effects in AML (Additional file 3: Figure S1). Top 20 most significant survival-associated AS events of each type were presented in Fig. 2.
Subsequently, independent prognostic AS events for seven types were used to construct prognostic predictors models. LASSO analysis was used to screen the optimal combination (Additional file 9: Figure S3). The Kaplan–Meier analysis indicated that these molecular signatures of AS events can be used to differentiate patients with distinct prognosis. To compare the efficiency of eight predictive models, ROC curves were applied into each model. The AUC of AA, AD, ES, AT, AP, RI, ME and marge-AS models was 0.912, 0.875, 0.834, 0.829, 0.820, 0.819, 0.738, and 0.867, respectively. AA was related with better predictive performance in AML, showed in Additional file 4: Figure S2. Detailed data of models based on each type of AS signature is listed in Additional file 5: Table S3.
Clinical characteristics of bone marrow immune microenvironment
To evaluate the relationship between AS and bone marrow microenvironment, we analyzed immune scores of AML patients by using a bioinformatics tool. Immune scores ranged from 1645.34 to 4145.87, and stromal scores form −1809.22 to 333.04. To assess the correlation of OS and DFS with scores, we divided patients into low- and high-score groups using the median of immune/stromal scores as cutoff. As shown in Fig. 3, the survival distribution curves showed that high immune scores correlated with worse OS (P value = 0.0215) and DFS (P-value = 0.0058) than low immune scores (Fig. 3a, b). However, patients with lower stromal scores only had numerically longer OS (P-value = 0.6896) and DFS (P-value = 0.1138) (Figure c, d). Since previous studies have indicated that immune conditions are associated with clinical characteristics, we investigated the relationships of gender, age and cytogenetic risk classification with immune and stromal scores. Immune scores were greater in the elderly group (P-value = 0.0013), while stromal scores were independent of age (P-value = 0.1196) and gender (P-value = 0.5156) (Fig. 3e-h). Intriguingly, the average immune scores of the favorable cytogenetics subtypes ranked the lowest in risk classification, which is consistent with findings above (Fig. 3i, k). To reveal the biological basis involved in different bone marrow microenvironments, we performed GSEA analysis in the high and low immune scores groups. Multiple pathways (Fig. 3J), including RNA degradation, nucleotide excision repair, hematopoietic cell lineage, and several metabolism related pathways were identified.
Network of survival-associated AS events and splicing factors
Splicing factors (SFs) are the key players of AS events and promote differential splicing patterns under stress conditions [22]. To understand the regulatory pattern of SFs in AML, we constructed the interaction networks of OS-related AS events and SFs (Fig. 4). A total of 117 AS events including 56 AS events with inferior prognosis (red dots) and 61 AS event with protective activity (green dots) were identified and correlated significantly with the levels of SF expression (purple triangle). Figure 5a shows several examples of the relationship between SF expression levels and PSI values for AS events, with more details provided in Additional file 6: Table S4. We observed that many SFs were correlated with different type of AS events and may have opposite effects, indicating that SFs may negatively (green line) or positively (red line) regulates OS-related AS events. Similarly, partial AS events could be regulated by multiple SFs simultaneously, this phenomenon partly explains the multiple AS events that can occur in the same transcript.
Since abnormal expression of SFs in cancer is a potential mechanism for regulating AS events, we analyzed the relationship between the observed changes in SFs expression and promoter methylation. Among the 89 SFs, the promoters of 72 SFs were highly hypermethylated and negatively correlated with their corresponding mRNA expression in AML patients. Figure 5b shows some highly relevant examples including FAM50B, SRSF8, HNRNPF, CPSF6, etc., with more details provided in Additional file 7: Table S5. In addition, we observed that 19.46% of cases in the AML cohort had at least one SF copy number variation. Circos plot shows the chromosome position information of partial SFs (Fig. 5c). This result revealed that SFs could be regulated by epigenetics, and further increase the diversity of AS events.
AS-based clustering was associated with clinical characteristics and immune features
Our research observed that the PSI values of each AS event varied among AML individuals. To better understand the molecular heterogeneity of AML, we explored patterns of AS based on survival-associated events by performing consensus unsupervised analysis of all samples. As a result, three clusters of samples were determined (Fig. 6a, b): C1 (n = 60, 33.7%), C2 (n = 29, 16.3%), C3 (n = 89, 50.0%). To discern the characteristics of different subtypes, we first investigated the relationships between AS clusters and immune microenvironment. The results showed that C1 was associated with lower immune scores compared with C2 and C3 (C1 vs. C2, P-value = 0.0188; C1 vs. C3, P-value = 0.0463), see Fig. 6c. However, no significant difference was found between clusters and stromal scores (Fig. 6d). Additionally, we compared clinical characteristics (gender, age, cytogenetics risk category, FAB subtype) between clusters. As showed in Fig. 6e, favorable-risk cytogenetics was rarely distributed in C2 (3.57%), but much more common in C1 (30.0%). To explore the association between AS clusters and immune features, the landscape of 22 immune cells abundance within the bone marrow microenvironment of each AML case were plotted. Furthermore, Kaplan–Meier analysis was performed to assess the relationship between clusters and survival status (OS and DFS). The results suggested that AS clusters could be used to identify patient subgroups with different survival outcomes (Fig. 6f, g), among which C1 was both associated with good outcome in OS (C1 vs. C2, P-value < 0.0001; C1 vs. C3, P-value < 0.0001) and DFS (C1 vs. C2, P-value = 0.0159; C1 vs. C3, P-value = 0.0338) analysis, followed by C3 and C2. The overall median survival for clusters (C1–C3) was 31.5 months, 7.16 months, and 14.13 months, respectively.
To further reveal the molecular characteristics of samples with AML in TCGA cohort, we performed a comprehensive molecular analysis of the mutation pattern SNPs, mutation information of each gene was exhibited according to different classified categories. Among all variant classification, missense mutations accounted for the largest proportion, followed by frame shift mutation and nonsense mutation (Fig. 7a); insertion or deletion occurred less frequently than single nucleotide polymorphism (Fig. 7b). The C > T transversions was the predominant of single nucleotide variants (SNVs) in AML (Fig. 7c). The most frequently mutated genes were exhibited in Fig. 7d, e, and the top ten includes DNMT3A, NPM1, TP53, KIT, RUNX1, FLT3, IDH2, WT1, TTN, and IDH1. Based on the AS events clustering, we found that the distribution of common mutated genes was inconsistent in different clusters (Additional file 8), the mutation frequency of DNMT3A in cluster C1-C3 was 10%, 13.8% and 11.2%, respectively (Fig. 7f); and NPM1 was 5%, 13.4% and 10.1%, respectively (Fig. 7g). Notably, no TP53 mutations were detected in C1 (Fig. 7h), which might be related to the good prognosis of C1 or the insufficient number of TP53 mutations samples. These results indicate that distinct patterns of AS are associated with different molecular characteristics.
Immune checkpoint inhibitors are considered a promising treatment strategy in AML [23]. We investigated the relationship between the expression of immune checkpoints and AS clusters. As shown in Fig. 8a, CD279, CD276, CD27 were significant different in expression distribution among the three clusters of samples. To explore the immune characteristics in bone marrow microenvironment, we generated a bar chart to illustrate the distribution of 22 immune cells in each sample by CIBERSORT algorithm (Fig. 8b). The results revealed that the most representative cell composition in bone marrow microenvironment of AML patients were monocytes, T cells CD4 memory resting, mast cells resting, B cells naive, and eosinophils. We observed that the composition of mast cells resting and B cells naive were slightly higher in C2 than C1 or C3. Collectively, these findings suggested that AML displayed distinct patterns of survival-associated AS events, and splicing events are ubiquitous and influence clinical outcome. Our findings provide new insight into molecular targeted therapy and immunotherapy strategy for AML.