Research ArticleCOMPUTATIONAL BIOLOGY

TNF-insulin crosstalk at the transcription factor GATA6 is revealed by a model that links signaling and transcriptomic data tensors

See allHide authors and affiliations

Science Signaling  07 Jun 2016:
Vol. 9, Issue 431, pp. ra59
DOI: 10.1126/scisignal.aad3373
  • Fig. 1 A compendium of ligand-induced signals and transcriptional responses.

    (A) Overview of the experimental design. HT-29 cells were pretreated with IFNγ, stimulated with various combinations and concentrations of TNF, EGF, and insulin, and profiled for the indicated signaling receptors, adaptors, and effectors by kinase assay (KA), immunoblot (IB), or antibody array (AA) and for the associated transcriptomic signatures by microarray. The goal is to determine whether global ligand-induced mRNA regulatory states (Y) can be predicted from the upstream signaling network activation (X). (B) Hierarchical clustering of the signaling compendium for saturating (High) and subsaturating (Low) concentrations of TNF, EGF, and insulin (42, 43). Data are means of n = 3 to 6 independent biological replicates. (C) Hierarchical clustering of the dynamic transcriptomic responses resulting from the ligand combinations in (B). Data are means of n = 2 independent biological replicates.

  • Fig. 2 Structuring and modeling biological data sets as tensors.

    (A) Structured data sets are conventionally unfolded with time to create a concatenated data matrix of ns signals and nt time points. Using the unfolded matrix, data-driven modeling approaches (20) treat each time point of each signal as a separate predictor variable, yielding ns × nt regression (regr) coefficients that must be inferred. (B) Recasting stimulus-signal-time data sets as a third-order tensor. The tensor structure (X) considers each time point as a predictor variable for all signals and each signal as a predictor variable for all time points, resulting in ns + nt regression coefficients and thus a more parsimonious model. (C) A dependent third-order transcriptomic tensor (Y) structured by stimulus, nc gene clusters, and nt2 time points. (D) Decomposing third-order data tensors as sums of latent variables composed of triple products. The decomposed tensor for each latent variable is reconstructed as the triple product (purple) of a scores vector (t or u) and two weight vectors (wj and wk or ql and qm). Latent variables are iteratively calculated to capture the maximum covariance between X and Y that remains from the preceding latent variable. X and Y are connected by a linear inner relationship between t and u with slope = b. (E) Prediction with tensor models involves projecting a new stimulus onto the latent variables of X, predicting the dependent scores vector u from the linear inner relationship (u = bt), and then backprojecting onto the latent variables of Y.

  • Fig. 3 A tensor PLSR model linking ligand-induced signaling and changes in transcript abundance.

    (A) Time-unfolded measurements of transcriptional clusters (blue) compared to cross-validated predictions of the tensor PLSR model (brown). Standardized z scores of measured transcriptional clusters are means ± SD of n = 897 (#1), 841 (#2), 119 (#3), 106 (#4), 66 (#5), 49 (#6), 42 (#7), 33 (#8), and 26 (#9) probe sets (file S2). High (H) indicates saturating concentration of ligand, 0 indicates absence of ligand, and low (L) indicates subsaturating concentration of ligand. (B) Latent variable time weights for the signaling and transcriptomic tensors. LV#4 has a negative inner relationship (orange), indicating that LV#4 signaling is anticorrelated with LV#4 transcription. (C to E) Projections of the indicated stimulus conditions (C), signals (D), and transcriptional clusters (E) onto LV#3 and LV#4. For (D) and (E), the null projections of reshuffled data tensors are means (solid gray) ± SD (dashed gray) of n = 500 randomizations (79). In (D), the type of assay used to measure the signaling protein is indicated in parentheses (see Fig. 1A for details). ClvC8, cleaved caspase 8; ProC3, procaspase 3; ProC8, procaspase 8; lowercase p prefix represents phosphorylated protein; lowercase t prefix represents total protein; lowercase pt prefix represents the ratio of phosphorylated protein to total protein.

  • Fig. 4 Multipronged bioinformatics of TNF-insulin crosstalk suggests posttranslational regulation from GSK3 to GATA6.

    (A to D) qRT-PCR validation of selected cluster #9 transcripts upon pretreatment of HT-29 cells with IFNγ and stimulation with TNF with or without insulin for 2 hours (A and B) or 6 hours (C and D). Data are geometric means ± log-transformed SEM of n = 4 or 16 biological replicates. Full cluster #9 data are shown in fig. S5. (E) Promoter bioinformatics (86, 92, 93) suggests GATA and TCF4 as candidate regulators of TNF-insulin crosstalk. (F) Relative copy number estimates (96, 97) for the six GATA isoforms in HT-29 cells. Data are medians ± range of n = 3 biological replicates. n.d., not detected. (G to I) Transcriptional dynamics of GATA isoforms in response to TNF and insulin. Data are geometric means ± log-transformed SEM of n = 4 or 8 biological replicates. (J) Scansite (98) identification of candidate GSK3 phosphorylation sites (red). Each site’s percentile rank is averaged across the indicated sequences. (K) Mass spectrometry identifies 11 phosphorylation sites on GATA6L. Previously unreported sites (New) are shown below the primary sequence, those consistent with reports in the literature (104) (Reported) are shown above, and reported sites not detected in this study are gray. Start methionines (arrows) for the long and short forms are indicated along with the conserved GATA core and zinc finger (ZnF) domains.

  • Fig. 5 GATA6L abundance is not altered by prolonged rapamycin treatment or feedback phosphorylation of AKT.

    (A) HT-29 cells exhibit rapamycin-induced feedback phosphorylation of AKT, but do not stabilize GATA6. a.u., arbitrary units. (B) AC16 cells exhibit stabilization of GATA6S, but not GATA6L, without rapamycin-induced feedback phosphorylation of AKT. Vinculin, tubulin, and glyceraldehyde-3-phosphate dehydrogenase (GAPDH) used as loading controls (51). Quantitative immunoblot data are means ± SEM of n = 4 biological replicates across two separate experiments.

  • Fig. 6 Extensive phosphorylation of GATA6L is blocked by S37A mutation, reversed by TNF stimulation, and stabilized in HT-29 cells.

    (A) Electrophoretic mobility of FLAG-tagged GATA6L is downshifted upon S37A mutation in lipofected 293T cells. (B) Phos-tag electrophoresis (112) reveals that TNF stimulation for 1 hour causes the dephosphorylation of GATA6L. (C and D) Phos-tag electrophoresis (C) and quantification (D) of the upper and lower forms of GATA6L in response to IFNγ sensitization for 24 hours, pretreatment with 1 μM CT99021 for 1 hour, and stimulation with TNF or insulin for 1 hour. Data are median proportion ± range of n = 3 biological replicates. (E and F) Doxycycline (DOX)–inducible addback in HT-29 cells replaces endogenous GATA6S with epitope-tagged GATA6L. Cells were treated with doxycycline (1 μg/ml) for 48 hours. (G and H) The less phosphorylated form of wild-type (WT) GATA6L is unstable. Cells were treated with TNF (100 ng/ml) + 50 μM cycloheximide for the indicated times, and half-lives were estimated by nonlinear least-squares curve fitting. Quantitative immunoblot data are means ± SEM of n = 3 (F and H) or 5 to 6 (B) biological replicates.

  • Fig. 7 Phosphorylation and destabilization of endogenous GATA6L at 60 kD.

    (A) Knockdown (shGATA6) and FLAG-tagged addback of GATA6L at ~60 kD (red). Samples were immunoblotted for total (modified and unmodified) GATA6 (upper), FLAG (lower), and the indicated loading controls. (B) Destabilization of the 60-kD form of WT GATA6L compared to the S37A-mutant addback cells. (C and D) Endogenous phospho-GATA6L (Ser37) immunoreactivity is not detectably affected by stimulation with TNF for 1 hour, inhibition with 20 μM CT99021 for 6 hours, or both. (E and F) Phosphorylation and destabilization of the 60-kD form of GATA6L upon serum starvation. Specificity was confirmed by preincubation of cells with 20 μM CT99021 for 1 hour before serum starvation. (G and H) Phospho-GATA6L (Ser37) immunoprecipitation and total GATA6 immunoblot of HT-29 cells pretreated with IFNγ and stimulated with TNF ± insulin for 1 hour. The gamma of the immunoprecipitation image is set to 1.5 to minimize background from the immunoprecipitating antibody heavy chain. Input (0.5%) of each immunoprecipitate was immunoblotted for total GATA6 and the indicated loading controls. See file S5 for details. Data are means ± SEM of n = 3 (B, E, F, and H) or 6 (C and D) biological replicates.

  • Fig. 8 S37A mutation of GATA6L mimics and competes with the repression of transcript abundance in the TNF-insulin crosstalk cluster.

    (A) Doxycycline-inducible overexpression of WT GATA6L in HT-29 cells. Cells were treated with doxycycline (1 μg/ml) for 24 hours. (B) Ratio of TNF-induced transcript abundance for crosstalk cluster genes in the presence or absence of GATA6L overexpression. Data are mean ratios of n = 3 independent biological samples assessed by microarray profiling, with bias in the ratio assessed by two-sided binomial test. (C) S37A mutation of GATA6L mimics insulin stimulation (green) and antagonizes TNF-insulin crosstalk (purple). qRT-PCR data for cluster #9 genes in WT and S37A mutant (S37A) GATA6L addback cells pretreated with IFNγ and stimulated with TNF, insulin, or both for 2 or 4 hours. Data are row-standardized geometric means of n = 6 biological replicates across two separate experiments, with interactions between GATA6 status and TNF or insulin assessed by log-transformed five-way ANOVA with the following factors: GATA6, transcript, TNF, insulin, and time. (D) Three-state conceptual model for GATA6L regulation by TNF and insulin and its relation to the crosstalk cluster of transcripts. Ovals annotate the figure subpanels supporting the links depicted.

  • Fig. 9 Diversity of GATA6L forms across different cell lineages.

    (A to D) Arrows indicate the GATA6 forms confirmed earlier by knockdown or observed with multiple antibodies. Red asterisks indicate nonspecific bands. The gap represents a deleted lane in each blot. All samples were analyzed together on the same blots. Data are representative of n ≥ 3 independent experiments.

Supplementary Materials

  • www.sciencesignaling.org/cgi/content/full/9/431/ra59/DC1

    Text S1. Detailed description of tensor PLSR.

    Text S2. Detailed description of GATA6L mass spectrometry.

    Fig. S1. Tensor PLSR modeling predicts overall transcript abundance but cannot link changes in transcript abundance to cytokine-induced signaling.

    Fig. S2. Accuracy of tensor PLSR predictions.

    Fig. S3. Induction of cluster #9 probe sets and repression by insulin.

    Fig. S4. Disruption of NF-κB signaling does not widely affect the TNF-induced transcriptional response of cluster #9.

    Fig. S5. Widespread TNF-insulin crosstalk among genes in transcriptional cluster #9.

    Fig. S6. Immunolocalization of β-catenin is not altered by TNF stimulation or insulin costimulation in HT-29 cells.

    Fig. S7. HT-29 cells lack GATA1, GATA4, and GATA5.

    Fig. S8. Phylogeny of the human GATA family.

    Fig. S9. Prolonged rapamycin treatment alters the proportion of GATA6S to GATA6L independently of TNF or insulin treatment.

    Fig. S10. Insulin and CT99021 perturb GSK3 phosphorylation and activity.

    Fig. S11. Phospho-GATA6L (Ser37) antiserum is specific for mobility-shifted wild-type GATA6L but not the S37A GATA6L mutant.

    Fig. S12. GATA6L phosphorylation on Ser37 and GS phosphorylation on Ser641 are lost in a dose-dependent manner upon treatment with the GSK3 inhibitor CT99021.

    Fig. S13. GATA6L occupies GATA binding sites in the promoters of genes within the crosstalk cluster.

    Table S1. Top-scoring PEST sequences in the indicated proteins according to PEST-FIND.

    Table S2. qRT-PCR primer sequences.

    Table S3. ChIP primer sequences.

    File S1. Microarray probe sets differentially altered with time or by stimulation with TNF, EGF, or insulin.

    File S2. Transcriptional clusters identified by CLICK.

    File S3. Tensor PLSR model and associated files.

    File S4. Transcripts differentially altered by wild-type GATA6L overexpression with or without TNF stimulation.

    File S5. Raw files and densitometric analysis of phospho-GATA6L (Ser37) immunoprecipitation and total GATA6 immunoblots.

    References (137140)

  • Supplementary Materials for:

    TNF-insulin crosstalk at the transcription factor GATA6 is revealed by a model that links signaling and transcriptomic data tensors

    Zeinab Chitforoushzadeh, Zi Ye, Ziran Sheng, Silvia LaRue, Rebecca C. Fry, Douglas A. Lauffenburger, Kevin A. Janes*

    *Corresponding author. Email: kjanes{at}virginia.edu

    This PDF file includes:

    • Text S1. Detailed description of tensor PLSR.
    • Text S2. Detailed description of GATA6L mass spectrometry.
    • Fig. S1. Tensor PLSR modeling predicts overall transcript abundance but cannot link changes in transcript abundance to cytokine-induced signaling.
    • Fig. S2. Accuracy of tensor PLSR predictions.
    • Fig. S3. Induction of cluster #9 probe sets and repression by insulin.
    • Fig. S4. Disruption of NF-κB signaling does not widely affect the TNF-induced transcriptional response of cluster #9.
    • Fig. S5. Widespread TNF-insulin crosstalk among genes in transcriptional cluster #9.
    • Fig. S6. Immunolocalization of β-catenin is not altered by TNF stimulation or insulin costimulation in HT-29 cells.
    • Fig. S7. HT-29 cells lack GATA1, GATA4, and GATA5.
    • Fig. S8. Phylogeny of the human GATA family.
    • Fig. S9. Prolonged rapamycin treatment alters the proportion of GATA6S to GATA6L independently of TNF or insulin treatment.
    • Fig. S10. Insulin and CT99021 perturb GSK3 phosphorylation and activity.
    • Fig. S11. Phospho-GATA6L (Ser37) antiserum is specific for mobility-shifted wild-type GATA6L but not the S37A GATA6L mutant.
    • Fig. S12. GATA6L phosphorylation on Ser37 and GS phosphorylation on Ser641 are lost in a dose-dependent manner upon treatment with the GSK3 inhibitor CT99021.
    • Fig. S13. GATA6L occupies GATA binding sites in the promoters of genes within the crosstalk cluster.
    • Table S1. Top-scoring PEST sequences in the indicated proteins according to PEST-FIND.
    • Table S2. qRT-PCR primer sequences.
    • Table S3. ChIP primer sequences.
    • Legends for files S1 to S5
    • References (137140)

    [Download PDF]

    Technical Details

    Format: Adobe Acrobat PDF

    Size: 1.08 MB

    Other Supplementary Material for this manuscript includes the following:

    • File S1 (Microsoft Excel format). Microarray probe sets differentially altered with time or by stimulation with TNF, EGF, or insulin.
    • File S2 (Microsoft Excel format). Transcriptional clusters identified by CLICK.
    • File S3. Tensor PLSR model and associated files.
    • File S4 (Microsoft Excel format). Transcripts differentially altered by wild-type GATA6L overexpression with or without TNF stimulation.
    • File S5. Raw files and densitometric analysis of phospho-GATA6L (Ser37) immunoprecipitation and total GATA6 immunoblots.

    [Download File S1]

    [Download File S2]

    [Download File S3]

    [Download File S4]

    [Download File S5]


    Citation: Z. Chitforoushzadeh, Z. Ye, Z. Sheng, S. LaRue, R. C. Fry, D. A. Lauffenburger, K. A. Janes, TNF-insulin crosstalk at the transcription factor GATA6 is revealed by a model that links signaling and transcriptomic data tensors. Sci. Signal. 9, ra59 (2016).

    © 2016 American Association for the Advancement of Science

Stay Connected to Science Signaling

Navigate This Article