Reverse engineering of TLX oncogenic transcriptional networks identifies RUNX1 as tumor suppressor in T-ALL.

The TLX1 and TLX3 transcription factor oncogenes have a key role in the pathogenesis of T cell acute lymphoblastic leukemia (T-ALL). Here we used reverse engineering of global transcriptional networks to decipher the oncogenic regulatory circuit controlled by TLX1 and TLX3. This systems biology analysis defined T cell leukemia homeobox 1 (TLX1) and TLX3 as master regulators of an oncogenic transcriptional circuit governing T-ALL. Notably, a network structure analysis of this hierarchical network identified RUNX1 as a key mediator of the T-ALL induced by TLX1 and TLX3 and predicted a tumor-suppressor role for RUNX1 in T cell transformation. Consistent with these results, we identified recurrent somatic loss-of-function mutations in RUNX1 in human T-ALL. Overall, these results place TLX1 and TLX3 at the top of an oncogenic transcriptional network controlling leukemia development, show the power of network analyses to identify key elements in the regulatory circuits governing human cancer and identify RUNX1 as a tumor-suppressor gene in T-ALL.

TLX1 and TLX3 encode highly related homeobox transcription factor oncogenes frequently activated by chromosomal translocations in T-ALL [3-5].

To interrogate the transcriptional programs associated with aberrant expression of TLX1 and TLX3, we analyzed gene expression data from 82 human T-ALLs [6]. This analysis revealed that TLX1 and TLX3 tumors share a common expression signature including 319 up-regulated and 450 down-regulated gene transcripts respectively (Fold change >2, P < 0.005) (Fig. 1a; Supplementary Table 1). Moreover, non negative matrix factorization (NMF) and Principal Component Analysis showed that TLX1 and TLX3 leukemias are highly related and clustered together separate from the rest of T-ALL samples in our series (Supplementary Figure 1). These results support a broadly overlapping role of TLX1 and TLX3 in the induction of T-ALL, however, TLX1 and TLX3 leukemias have been associated with different prognosis in some series [1,7], suggesting important biological differences between these two groups. Consistently, comparative marker analysis identified a broad gene expression signature in TLX1 T-ALLs compared with TLX3 tumors (Supplementary Figure 2).

Figure 1

TLX1- and TLX3-expressing T-ALLs are associated with a distinct gene expression signature highly enriched in downregulated TLX1- and TLX3-ChIP-chip direct target genes

(a) Heat map representation of the top 25 differentially expressed genes (P < 0.005; Fold change > 2) in TLX1- and TLX3-expressing T-ALLs. (b) Venn diagram representation showing the overlap between TLX1 and TLX3 ChIP-chip direct target genes (c) GSEA analysis of TLX1 and TLX3 ChIP-chip direct target genes in TLX1- and TLX3-expressing T-ALLs. Enrichment plots (left) and heat map representations of the 25 top ranking genes in the leading edge (right) are shown. Genes in heat maps are shown in rows, each individual sample is shown in one column. The scale bar shows color coded differential expression from the mean in standard deviation units with red indicating higher levels and blue lower levels of expression.

Next, we analyzed TLX1 ChIP-chip data from ALL-SIL, a T-ALL cell line expressing high levels of TLX1 as result of the t(10;14)(q24;q11) translocation[3] and performed ChIP-chip analysis for TLX3 in HPB-ALL, a t(5;14)(q35;q32) TLX3-activating translocation positive line[8]. These analyses identified 2,236 promoters bound by TLX1 and 3,148 promoters occupied by TLX3 with a significance cutoff of P < 10−9 (Supplementary Table 2). Strikingly, 75% of TLX1 direct targets were also bound by TLX3 (Chi-square P < 0.001) (Fig. 1b). Finally, Gene Set Enrichment Analysis (GSEA) demonstrated a highly significant enrichment of genes whose promoter was bound by TLX1 and TLX3 in the expression signature associated with TLX1 and TLX3 leukemias (P < 0.001) (Fig. 1c) (Supplementary Table 3). Most notably, genes bound by TLX1 and TLX3 were characteristically downregulated in this group (Fig. 1c), strongly suggesting that TLX1 and TLX3 primarily function as transcriptional repressors in the pathogenesis of T-ALL.

We then used the ARACNe reverse-engineering algorithm[9,10] to generate a genome-wide T-ALL transcriptional network or T-ALL interactome (T-ALLi) using gene expression data from 228 T-ALLs. This analysis yielded a T-ALLi including 19,689 genes (nodes) connected via 471,824 interactions (edges) (Supplementary Figure 3). Notably, MYC target genes inferred in the T-ALLi were markedly enriched in MYC ChIP-chip direct target genes (74/252, Chi-square P = 2.5×10−5) supporting the soundness of this approach (Supplementary Figure 4). Analysis of TLX1 and TLX3 connected genes in this setting identified 325 candidate TLX target genes (Fig. 2a), including 70 TLX1- and TLX3- highly significant (P < 0.0001) ChIP-chip target genes (Chi-square P = 0.02) (Fig. 2b) and 117 genes differentially expressed (P < 0.0001) in TLX1- and TLX3- T-ALLs (Chi-square P < 0.001) (Fig. 2c).

Figure 2

An ARACNe transcriptional network identifies TLX1- and TLX3-direct targets and TLX1- and TLX3- differentially expressed genes

(a) Graphic representation of node first neighbors in the T-ALLi ARACNe transcriptional network connected to a TLX1 and TLX3 metagene (TLX). (b) TLX first neighbor genes in the ARACNe transcriptional network identified as TLX1- and TLX3-ChIP-chip direct target genes (ChIP-chip P < 0.0001). (c) TLX first neighbor genes in the ARACNe transcriptional network differentially expressed in TLX1- and TLX3-expressing T-ALLs (Differential expression P < 0.0001). Relative expression in TLX1- and TLX3-expressing T-ALLs is color coded with nodes in red indicating upregulated and nodes blue indicating downregulated genes.

Next, we defined the TLX-subnetwork (TLXi) as the space of the T-ALLi encompassing the 445 TLX1- and TLX3- direct target genes (ChIP-chip P < 0.0001) that are also differentially expressed in TLX1- and TLX3-expressing T-ALLs (P < 0.0001) and their most direct interconnections (Fig. 3a). The TLXi subnetwork retains the topological features of the TALLi. Thus, 411/445 (92%) of the genes were involved in at least one interaction, but only 8/445 (< 2%) showed 50 or more direct interactions (Fig. 3b)(Supplementary Table 4). Notably, and consistent with the role of TLX1 as transcriptional repressor TLXi genes transcripts were also characteristically downregulated by GSEA in a transgenic mouse model of TLX1-induced T-ALL[11] (Supplementary Figure 5). Moreover, GSEA analysis of the expression signatures induced by shRNA knockdown of TLX1 in ALL-SIL cells and of TLX3 in the HPB-ALL cell line demonstrated a high level of enrichment of genes in the TLXi among the transcripts upregulated upon inactivation of TLX1 and TLX3 respectively (Supplementary Figures 6 and 7).

Figure 3

Reverse engineering and structure analysis of the TLXi subnetwork

(a) Graphic representation of the TLXi subnetwork. Each node represents a TLX1- and TLX3- ChIP-chip direct target (P < 0.0001) gene differentially expressed (P < 0.0001) in TLX1- and TLX3-expressing T-ALLs. (b) Connectivity plot representing the connections of each gene within the TLXi. Transcription factors are colored in red and non transcription factor encoding genes are indicated in blue (c) Sun diagram showing the RUNX1 first neighbor genes in the TLXi. Relative expression in TLX1- and TLX3-expressing T-ALLs is color coded in the network representation with nodes in red indicating upregulated and nodes blue indicating downregulated genes. (d) GSEA analysis of RUNX1 ChIP-chip direct target genes in TLX1- and TLX3-expressing T-ALLs. Enrichment plots (left) and heat map representations of the 25 top ranking genes in the leading edge (right) are shown. Genes in heat maps are shown in rows, each individual sample is shown in one column. The scale bar shows color coded differential expression from the mean in standard deviation units with red indicating higher levels and blue lower levels of expression.

Based on these results we proposed that the hierarchical regulatory structure of the TLXi subnetwork could reflect, at least in part, the functional hierarchy of TLX1- and TLX3- target genes involved in T-cell transformation. In this context, RUNX1, a critical transcription factor in hematopietic development[12] frequently mutated in acute myeloid leukemias[13-15] stood up as the single most highly interconnected hub in the TLXi (Fig.3b,c). ChIP analysis of TLX1 and TLX3 confirmed the binding of these transcription factors binding to the RUNX1 promoter (Supplementary Fig. 8). In addition, RUNX1 was significantly more interconnected in the TLXi-subnetwork than in the T-ALLi as a whole (Chi-square P = 2.14×10−133) and stood up as one of the most prominent TLXi genes downregulated in mouse TLX1-induced T-ALLs (Supplementary Figure 5). Consistently, Master Regulator Analysis [16,17] identified RUNX1 as one of the top most prominent master regulators of the transcriptional program associated with human TLX1 and TLX3 induced leukemias (Supplementary Table 5). The model that emerges from this analysis is a regulatory feedforward loop in which downregulation of RUNX1 by TLX1 and TLX3 would subsequently affect the expression of numerous other TLX target genes (Supplementary Fig. 9). To test this possibility we performed ChIP-chip analysis of RUNX1 direct targets in HPB-ALL cells. In this analysis we identified 308 high confidence RUNX1 target genes (P < 0.0001) (Supplementary Table 6). Strikingly, and in concordance with our network analysis, 50% of RUNX1 occupied promoters were also bound by TLX1 and TLX3 (Chi-square P < 10−15). Moreover, GSEA analysis of RUNX1 direct target genes showed a high level of enrichment of RUNX1 targets among the top transcripts downregulated in T-ALL cells expressing high levels of TLX1 or TLX3 (P = 0.05) (Fig. 3d).

These results suggest that RUNX1 could mediate, at least in part, some of the oncogenic effects of TLX1 and TLX3 overexpression. Consistent with this hypothesis, retroviral expression of RUNX1 in TLX1-positive (ALL-SIL) and TLX3 positive (HPB-ALL) cells resulted in impaired cell growth (Supplementary Figure 10) indicating a possible tumor suppressor role for RUNX1 in T-ALL. Mutation analysis of RUNX1 in T-ALL revealed the presence of RUNX1 mutations in 4/12 (33.3%) T-ALL cell lines and 5/114 (4.4%) T-ALL primary samples (Fig. 4a, Supplementary Tables 7 and 8). Interestingly, all ALLs identified in kindreds with FPDMM (platelet disorder, familiar, with associated myeloid malignancy, MIM ID #601399), a leukemia predisposition syndrome caused by mutations in RUNX1, happen to be T-ALLs [18-20].

Figure 4

RUNX1 mutations in T-ALL

(a) Schematic representation of RUNX1 mutations identified in T-ALL. Runt: RUNX1 DNA binding domain; AD: activation domain; ID: inhibitory domain (b) Representative DNA sequencing chromatograms of paired diagnostic and remission genomic DNA T-ALL samples showing somatically acquired mutations in the RUNX1 gene. (c) Molecular surface rendition depicting the interaction between the RUNX1 runt domain (grey), DNA and CBFB (green) complex. RUNX1 mutations present in T-ALL and AML are indicated in orange. RUNX1 mutations found in AML are depicted in yellow. (d) Effects of RUNX1 T-ALL mutant alleles in the activity of a CSF promoter reporter construct. The size of the nodes in the star diagrams is proportional to the significance P value as indicated in the scale at the bottom.

RUNX1 mutations found in T-ALL were heterozygous frameshift truncating mutations (3/10) and missense single nucleotide changes (6/10) (Fig. 4a,b). Notably, DNA sequence analysis of samples obtained at the time of clinical remission demonstrated the somatic origin of RUNX1 mutations in each of the 2 cases with available material (Fig. 4b). Moreover, five of these RUNX1 mutant alleles (pL29S, pH58N, pH78Y, pS114fs and pG138fs) have been previously described as oncogenic mutations in myeloid tumors [21-25]. Interestingly, all four RUNX1-mutated samples with available immunophenotype data showed a CD4 and CD8 double negative immunophenotype indicative of a very early arrest in T-cell maturation (Supplementary Table 9). Mapping of T-ALL RUNX1 mutations on the structure of the RUNX1 runt domain (PDB 1H9D) showed clustering of these amino acid substitutions in the DNA recognition interface of RUNX1 (Fig 4.c). Most strikingly, the RUNX1 H78 residue resides within a highly structurally conserved 16.9 Å diameter cavity frequently targeted by RUNX1 AML mutant alleles, which is adjacent to the DNA binding interface and is predicted to be disrupted in the RUNX1 H78Y T-ALL mutant (Fig 4.c). Next we tested the functional significance of the RUNX1 mutants predicted to be most structurally disruptive in luciferase reporter assays. In these experiments RUNX1 H78Y, RUNX1 S114fs and RUNX1 G138fs showed marked (5 fold) reductions in their capacity to activate a RUNX1-responsive CSF promoter reporter construct compared with wild type RUNX1(Fig. 4d).

Next we analyzed the transcriptional programs and disease kinetics of leukemias occurring in Lck-TLX1 transgenic Runx1 wild type mice and in Lck-TLX1 Runx1 heterozygous knockout animals. This analysis revealed that TLX1 Runx1 +/+ and TLX1 Runx1+/− share a common gene expression program consisting of 215 commonly differentially expressed genes (fold change > 2, P <0.001). However, and consistent with the presence of 50% non overlapping target genes between RUNX1 and TLX1, loss of one copy of Runx1 partially changes the transcriptional signature of TLX1-induced leukemias resulting in 540 differentially expressed transcripts between TLX1 Runx1+/+ and TLX1 Runx1+/− tumors (fold change > 2, P <0.001) (Supplementary Figure 11). Notably, and despite these transcriptional differences, Lck-TLX1 transgenic Runx1 wild type and Lck-TLX1 Runx1 haploinsuficient mice developed T-ALL with identical kinetics (Supplementary Fig. 12), suggesting that, in agreement with the prediction of our network analysis, the oncogenic effects of TLX1 are overlapping with the tumor suppressor activity of Runx1.

Overall, the integrative analyses presented here (Supplementary Figure 13) show a high level of functional overlap between TLX1 and TLX3 in T-cell transformation and identify RUNX1 as a tumor suppressor gene in T-ALL. Notably, this work highlights the power network analysis to decipher the structure of complex oncogenic circuitries and to identify critical genes and pathways involved in the pathogenesis of human cancer. Moreover, reverse engineering of signaling and transcriptional networks controlling phenotypes associated with distinct gene expression signatures such as cell transformation, metastatic potential or drug resistance could be exploited to identify new therapeutic targets.

METHODS

Clinical samples

Leukemic DNA and cryopreserved lymphoblast samples were provided by collaborating institutions in the US [Eastern Oncology Group (ECOG) and Pediatric Oncology Group (POG)]. All samples were collected under the supervision of local IRB committees. Informed consent was obtained from all patients at trial entry according to the declaration of Helsinki.

Master Regulator Analysis

Master Regulators (MRs) Analysis was carried out as previously described[16]. Briefly, each set of transcription factor targets (regulon) was partitioned in positive and negative based on the correlation of the transcription factor and target. Positive and negative regulons were tested for enrichment in the TLX1 and TLX3 signature. Redundancy in inferred master regulators that have a large number of common targets was corrected for by removing “shadowed transcription factors”, identified as those master regulators whose enrichment is significantly reduced when the common targets are disregarded.

ChIP and ChIP-chip analysis

ChIP-chip analysis of TLX3 and RUNX1 target genes was performed in the HPB-ALL cell line. Briefly, 1×108 cells were used for chromatin immunoprecipitation using the A-17 goat polyclonal (sc-23397) and the H-55 rabbit polyclonal (sc-30185) antibodies recognizing TLX3 (Santa Cruz Biotechnology) or two rabbit polyclonal antibodies against RUNX1 (Ab980 from Abcam and 4336S from Cell Signaling Technologies). ChIP-chip was performed following standard protocols provided by Agilent Technologies using Agilent Human Proximal Promoter Microarrays (244K features/array) as previously described [26]. This platform analyzes ~17,000 of the best-defined human genes sourced from UCSC hg18 (NCBI Build 36.1, March 2006) and covers regions expanding from −5.5 kb upstream to +2.5 kb downstream of their transcriptional start sites. We scanned the arrays with an Agilent scanner and extracted the data using the Feature Extraction 8 software. TLX3 and RUNX1 direct target genes were identified using ChIP-chip Significance Analysis (CSA) as described before[26]. MYC and TLX1 ChIP-chip analysis in T-ALL have been previously reported[11,26].

Relative real-time PCR quantitation of RUNX1 promoter sequences was normalized to ACTB gene levels in chromatin immunoprecipitates performed with an antibodies against TLX1 (C-18 rabbit polyclonal antibody (sc-880), Santa Cruz Biotechnology) and TLX3 (A-17 goat polyclonal (sc-23397), Santa Cruz Biotechnology) . Primer sequences are listed in Supplementary Table 9.

Reverse engineering of the T-ALL transcriptional networks

To generate a T-ALL transcriptional network we processed Human U133 Plus2.0 Affymetrix microarray gene expression data from a series of 228 T-ALL primary samples using GC-RMA normalization ARACNe algorithm as described before[10] and named the resulting global T-ALL transcriptional network the T-ALL interactome (T-ALLi). Given the high level of overlap between TLX1 and TLX3 regulated direct target genes, and to avoid that the connections between genes showing high levels of mutual information with both TLX1 and TLX3 are eliminated by ARACNe during the Data Processing Inequality step aimed to filter out indirect connections, the expression of these two transcription factors was analyzed as a single node by assigning the same gene label (TLX) to TLX1 or TLX3 probes.

In a separate analysis, we defined the genes experimentally identified as TLX1- and TLX3-direct targets by ChIP-chip (P < 0.0001) and differentially expressed in TLX1- and TLX3-expressing tumors (differential expression P < 0.0001) as the core of the oncogenic program controlled by TLX1 and TLX3 in T-ALL. We then defined the TLX subnetwork (TLXi) as the subspace within the T-ALLi containing all these TLX1- and TLX3- differentially expressed direct target genes and their shortest path interconnections. The significance of the TLXi was tested performing in silico simulations of 10,000 random networks characterized by the same TLXi features (48 transcription factors and 1,655 connections). The significance of TLXi versus the random generated networks was obtained calculating a non parametric P value.

RUNX1 mutation analysis

All RUNX1 exon sequences were amplified from genomic DNA by PCR and analyzed by direct dideoxynucleotide sequencing. PCR and sequencing primer sequences are listed in Supplementary Table 10.

Structural depiction and analysis

Structural coverage of the RUNX1 protein was identified through use of the PSI-Blast and SKAN algorithms; viable structures were subsequently mapped to all RUNX1 isoforms, and analyzed with the MarkUS web annotation server [27]. Protein database (PDB) structures 1EAN, 1EAO, 1EAQ, 1H9D, 1IO4, 1HJB, 1HJC, and 2J6W were structurally aligned along the RUNX1 Runt domain-DNA interface, and the resulting composite structure was subsequently analyzed to assess conformational flexibilities[28]. Potential effects for the RUNX1 T-ALL mutations were investigated with SCREEN and VASP for cavity prediction and volumetric rendering, ConSurf for analysis of structural conservation, PredUS for protein-protein interface prediction, and DelPhi for highlighting potential alterations in electrostatic potential [27]. Probalistic classification of mutations through physical and evolutionary comparative considerations was conducted through use of the PolyPhen-2 batch servers and algorithms[29]. RUNX1 AML mutations were extracted from the COSMIC database, filtered, and mapped to RUNT domain structures[28]. All structural images were created using UCSF Chimera[28]

Statistical analysis

Significant overlapping between different groups of genes was calculated with the Chi-square test.