Expression levels of long non-coding RNAs are prognostic for AML outcome

First, we investigated to what extent individual lncRNAs were associated with overall survival in the Clinseq-AML cohort. Individual Cox proportional hazards regression models were fitted for each lncRNA using time-on-study as the time scale, adjusting for age, sex, ELN risk score, mutation status of CEBPA, NPM1, TP53, WT1, TET2, ASXL1, DNMT3A, RUNX1, IDH1, IDH2, and FLT3-ITD, and chromosomal abnormalities as covariates in the models. We found 33 prognostic (overall survival) lncRNAs (adjusted p value < 0.05, Fig. 2). These results suggest that there are individual lncRNAs that provide prognostic information beyond established risk classification scores (ELN risk score) and typical somatic aberrations in AML. We analyzed the association between lncRNA expression and overall survival in the TCGA-AML cohort (Additional file 1: Figure S1). However, none of the association have significant p value (< 0.05). A possible reason might be the small sample size of the TCGA-AML dataset.

Next, we investigated if subgroups of AML patients were present in the Clinseq-AML cohort that shares common multivariate lncRNA expression patterns. We applied an unsupervised consensus clustering approach (see the “Methods” section) to the lncRNA expression profiles and discovered four distinct lncRNA-based subtypes. Consensus clustering results indicated a high degree of co-clustering of subjects within these four groups (Fig. 3 and Additional file 1: Figures S2–S4). This indicates that AML patients in the Clinseq cohort could be stratified into four distinct subtypes based on their lncRNA expression abundances.

We assessed the prognostic information of the lncRNA-based subtypes in respect to overall survival (Fig. 4 and Additional file 1: Figure S5) and event-free survival (Additional file 1: Figure S6). The prognostic value of the four lncRNA-based subtypes was found to be significant (n = 274, p = 0.04 (log-rank test)). Among patients in cluster G1 (N = 65), the mean (±SE) overall survival at 60 months (5 years) was 61 ± 7%. Patients in clusters G2 and G3 had an intermediate rate of overall survival of 36 ± 8% and 26 ± 5% respectively. The cluster G4 has the worst survival outcome with an overall survival at 60 months of 18 ± 5%. The prognostic performance was also evaluated in the subset of cytogenetically normal patients (N = 130), which confirmed that the lncRNA-based subtypes were significantly associated with overall survival also in this subpopulation (Fig. 4b, p value = 0.02, log-rank test). Event-free survival for the lncRNA-based subtypes (Additional file 1: Figure S6) also provides a significant prognostic value (p = 0.015 (log-rank test)).

We validated the lncRNA expression-based subtypes in the independent TCGA-AML cohort (Fig. 4c, d). In the TCGA-AML cohort, the prognostic value of the lncRNA-based subtypes is significant (Fig. 4c, n = 172, p = 0.01, log-rank test). However, for cytogenetically normal patients in the TCGA-AML cohort, the prognostic performance is not significant (Fig. 4d, p value = 0.2, log rank) which potentially might have occurred due to the low sample size (n = 78). Details of validation using the TCGA-AML cohort are provided in the following section.

To ascertain if the subtypes were prognostic beyond established prognostic factors, we also fitted a multivariable Cox proportional hazards models, adjusting for established prognostic markers (Fig. 5), and this model was also found to provide a significant prognostic value (p value = 7.0 × 10⁻⁷). In particular, cluster G3 in this model was significantly different in overall survival compared with the reference group G1 (p value = 3.2 × 10⁻³, Fig. 5a).

To determine consistency of the subtype discovery, we implemented a nested cross-validation procedure that is analogous to repeatedly splitting our cohort into a training set (for model fitting, including parameter estimation) and an independent subset of patients for model evaluation in respect to prognostic value (test set). The misclassification rate of test set samples (nested cross-validation) was low, with overall classification accuracy in the nested cross-validation procedure of 85% (Additional file 1: Figure S7), using class labels assigned in the primary subtype discovery phase as reference. Cross-validation of lncRNA subtypes also revealed significant prognostic value (overall survival, Additional file 1: Figure S8) (p value = 0.012). These results indicate that the lncRNA subtypes, prediction model, and the prognostic value of the subtypes are robust.

Next, we assessed the reproducibility of newly discovered AML subtypes using independent TCGA-AML cohort. To handle intrinsic batch differences between the Clinseq and TCGA studies, we applied batch correction on Clinseq and TCGA lncRNA expression data [25]. We trained a random forest model [26] for subtype classification based on the Clinseq data and subsequently predicted subtypes in the TCGA-AML cohort. A list of lncRNAs selected using random forest models can be found in Additional file 2. Based on the predicted subtype labels in the TCGA cohort, we then assessed the prognostic information in respect to overall survival (Fig. 4c and Additional file 1: Figure S5B). In the TCGA-AML cohort, the prognostic value of the four lncRNA-based subtypes was found to be significant (n = 172, p = 0.01, log-rank test). In concordance with Clinseq-AML cohort, in TCGA-AML cohort, subtype G1 (n = 30) has the best survival outcome with mean (SE) overall survival at 60 months (5 years) which is 40 ± 12%. Similarly, subtypes G2 (n = 43) and G3 (n = 69) show intermediate survival with mean (SE) overall survival at 60 months which are 31 ± 8% and 22 ± 87% respectively. Similar to Cliniseq, subtype G4 (n = 30) has the worst survival outcome in TCGA cohort, where no patient survive at 60 months (Fig. 4c). We also fitted a multivariable Cox proportional hazards models, adjusting for age, sex, and established prognostic markers using TCGA clinical and mutation data (Fig. 5b). Prognostic value of this model was also found to be significant value (p value = 1.25 × 10⁻⁹). When compared with the reference group G1, in this model, subtype G4 was significantly different in overall survival (p value = 4.48 × 10⁻³). We also evaluated the prognostic performance in the subset of cytogenetically normal patients in the TCGA-AML cohort. In this subset of patients, the association to overall survival was not significant (Fig. 4d, p value = 0.2, log rank). However, this might be due to the low sample size (n = 78).

To determine if the lncRNA-based subtypes were associated with known cytogenetic or mutational aberrations, we applied association tests between subtypes and key genetic aberrations and clinical phenotypes (Table 2). Neither of the subtypes was found to be highly concordant with any of the conventional clinical or genetic factors (for details, see Additional file 1: Tables S1 to S20).

Patients belonging to group G1 are enriched for CEBPA mutations (2.56 and 1.46% single and double mutation respectively). CEBPA double mutations have been associated with favorable outcome in AML [27, 28]. Cluster G2 is enriched in NPM1 mutation (6.93%) but has low percentage of TP53 mutations (1.09%). The cluster G3 contains a substantial number of FLT3-ITD. This cluster is also enriched in CEBPA single and double mutations. Cluster G4 harbors a high percentage of TP53 mutations (4.75%). This cluster also contains the highest percentage (8.08%) of patients classified as high-risk category using ELN risk classification system.

We found that lncRNA expression-based subtypes were independent from the European LeukemiaNet (ELN) risk classification system [4] and the distribution of the ELN risk score is fairly even in all four groups (Fig. 1 and Additional file 1: Table S19). For each ELN risk type, we further stratified it using lncRNA subtypes (Additional file 1: Figure S9). These results indicate that for each ELN risk score, lncRNA subtypes can provide further stratification of patients. Although lncRNA-based subtypes were not found to be highly concordant with any specific mutations, cytogenetics, or clinical factors, we found that mutations in NPM1 and TP53 were associated with the lncRNA-based subtypes (Chi-square test p value is 1.09 × 10⁻⁵ and 2.99 × 10⁻³ respectively, see Additional file 1: Tables S1 to S20 for details).

LncRNAs have very limited functional assignments. In order to gain some overview of potential molecular mechanism related to the lncRNAs that define the lncRNA-based subtypes, we performed pathway analysis. First, we determined which mRNA transcripts were associated with the lncRNA-based subtypes, and subsequently, we utilized this set of mRNAs for pathway enrichment analysis (see the “Methods” section for details). This analysis revealed multiple significant pathways (Fig. 6) include “immune system” (adjusted p value = 0.01), “chromosome organization” (adjusted p value = 0.03), “mRNA processing” (adjusted p value = 0.01), and “transmembrane receptor protein tyrosine kinase signaling pathway” (adjusted p value = 0.02). List of all pathways and differentially expressed genes in four clusters can be found in Additional files 3 and 4 respectively.

Since lncRNA expression levels can in some cases be correlated to the expression levels of cis-located mRNAs [29] and potentially also be correlated with the global mRNA expression profile, we evaluated to what extent lncRNA-based subtypes were reflected in mRNA-based expression clusters. We applied an identical unsupervised consensus clustering methodology to determine mRNA-based clusters as for the lncRNA analysis (see Additional file 5 supplementary methods for details). Despite stratifying patients into groups that are substantially different (Fig. 7), a Chi-square test of dependence between mRNA and lncRNA subtype models did allow us to reject the null hypothesis of no relationship between the models (p value = 2.56 × 10⁻⁶⁸; Additional file 1: Table S21). For instance, mRNA subtype C2 is almost fully subsumed in lncRNA subtype G1, which might be a substantial contributor to the Chi-square statistic in this case. However, despite that mRNA and lncRNA models cannot be considered as independent, we note that mRNA and lncRNA expression profiling data stratify patients into markedly different groups (Fig. 7), suggesting that the information in mRNA and lncRNA expression profiles are different.

We used Clinseq-AML cohort, consist of 274 AML patients, treated according to the national guidelines in Sweden. The study was approved by the regional ethical review board in Stockholm, Sweden. All samples from the Clinseq-AML cohort were collected prior to the initiation of treatment. For detail characteristics of patients in Clinseq-AML cohort, see Table 1. In this study, we used data from 142 patients of the TCGA-AML study [1], who have received intensive induction treatment (chemotherapy) analogous to the Clinseq-AML cohort. Clinical and mutational data was retrieved from the data portal of TCGA (https://​gdc.​cancer.​gov) and TCGA-AML study publication [1]. Detailed characteristics of TCGA-AML cohort can be found in Additional file 7.

Transcriptomic RNA and somatic mutation panel of genes were sequenced using the Illumina HiSeq-2500 platform. Ribosomal RNA depletion was performed using the Ribo-Zero gold kit. HTSeq count version 0.6.1 [34] was used for gene expression estimation. RNAseq count data normalization was performed using the TMM method [35]. A total of 3030 lncRNAs were annotated using MiTranscriptome database [6].

Consensus clustering-based unsupervised learning was applied for subtype discovery [24]. Optimal number of cluster (k = 4) was determined using weighted silhouette index. For validation, first, we performed 10-fold cross-validation on Cliniseq-AML data. At each cross-validation round, data was randomly divided into train and test set. Unsupervised learning was performed on training set, and labels were used to train random forest model [26]. Labels for test dataset were predicted using this model.

For independent validation, common lncRNA in Clinseq and TCGA dataset were selected as features and batch correction was applied [25]. We trained random forest classifier [26] on batch-corrected Clinseq-AML data and subtype labels were predicted for TCGA-AML data.

For association analyses, Chi-square test was used. Overall survival was measured from the date of diagnosis to the date of death. Kaplan-Meier curve and non-parametric log-rank statistic were used for comparison. Uni-variable and multivariable Cox’s proportional hazards regression models were fitted to the survival data. In multivariate analysis, we adjusted for age, sex, etiology, ELN score, and mutational status of genes. Analysis was carried out using R (version 3.1.1).