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

Abstract

Signal -transduction networks coordinate transcriptional programs activated by diverse extracellular stimuli, such as growth factors and cytokines. Cells receive multiple stimuli simultaneously, and mapping how activation of the integrated signaling network affects gene expression is a challenge. We stimulated colon adenocarcinoma cells with various combinations of the cytokine tumor necrosis factor (TNF) and the growth factors insulin and epidermal growth factor (EGF) to investigate signal integration and transcriptional crosstalk. We quantitatively linked the proteomic and transcriptomic data sets by implementing a structured computational approach called tensor partial least squares regression. This statistical model accurately predicted transcriptional signatures from signaling arising from single and combined stimuli and also predicted time-dependent contributions of signaling events. Specifically, the model predicted that an early-phase, Akt-associated signal downstream of insulin repressed a set of transcripts induced by TNF. Through bioinformatics and cell-based experiments, we identified the Akt-repressed signal as glycogen synthase kinase 3 (GSK3)–catalyzed phosphorylation of Ser³⁷ on the long form of the transcription factor GATA6. Phosphorylation of GATA6 on Ser³⁷ promoted its degradation, thereby preventing GATA6 from repressing transcripts that are induced by TNF and attenuated by insulin. Our analysis showed that predictive tensor modeling of proteomic and transcriptomic data sets can uncover pathway crosstalk that produces specific patterns of gene expression in cells receiving multiple stimuli.

INTRODUCTION

A goal of systems biology is to identify regulatory properties of biological networks by integrating experiments and kinetic models around a defined biological “parts list” (1). For signal transduction, such approaches have clarified the molecular pathways and regulatory events that can explain signaling thresholds (2–4), adaptation (5–8), and oscillations (9–15), as well as signal-induced cell -fate transitions (16–18). A complementary use of systems methods is for large-scale profiling and analysis of cellular signals in pursuit of integrative understanding or new molecular mechanisms (19–22). These applications favor data-driven models based on statistics or bioinformatics, which propose new links within the network that can be tested experimentally (23–28). Testing of a data-driven hypothesis about mechanism requires reductionist experimental approaches, even when the hypothesis itself emerges from network-level data (29, 30). Consequently, systems-level studies can yield findings that are immediately relevant to signal transduction biology (31).

Receptor-mediated signaling pathways are densely connected and exhibit nonadditive behaviors when two ligands activating different receptors are applied simultaneously (32–34). Downstream of receptor activation and signal transduction, complexity increases even further when considering the consequences of signaling on gene regulation (35–37). Fortunately, new signaling synergy or antagonism does not typically emerge with more than two inputs (38–41), suggesting that stimulus pairs are sufficient to assess potential crosstalk. For systems-level discovery, this type of in-depth cellular profiling must be complemented by data-driven modeling and extensive mechanistic follow-up.

Here, we examined crosstalk between a proinflammatory stimulus, tumor necrosis factor (TNF), and one of two growth factors, insulin or epidermal growth factor (EGF), with a focus on signaling -network activation and downstream regulation of the transcriptome. To examine the interface between signal -transduction and transcription networks at the systems level, we built upon a preexisting compendium of signaling events collected in the adenocarcinoma cell line HT-29, which were initially exposed to interferon- γ (IFNγ) and then stimulated with TNF, EGF, and insulin in various combinations (42, 43). Because TNF, EGF, and insulin trigger widespread and time-dependent changes in transcript abundance (44–47), we collected dynamic mRNA profiles after stimulation with single or paired stimuli and integrated these data with the previously measured intracellular signaling events (42, 43).

Because each signal and transcript was measured at every time point and stimulus condition, we could organize the signaling and transcriptomic data cubes as three-way tensors to build a model constrained by the structure of the data. Model analysis together with bioinformatics predicted a link between Akt–glycogen synthase kinase 3 (GSK3) signaling and the endodermal transcription factor GATA6. We found that GSK3 phosphorylation of Ser³⁷ on the long form of GATA6 (GATA6_L) accelerated its degradation in cells. The increased turnover reduced transcriptional repression by GATA6_L of genes induced by TNF. Collectively, our results showed that GATA6_L integrated growth factor–induced signaling activity and inflammatory transcriptional regulation.

RESULTS

Cytokine combinations elicit complex changes in signaling and transcript abundance

For the signaling pathway measurements, we used data from our own previous studies (42, 43), which enabled us to combine these data with the transcriptomic data generated here under the same experimental conditions. In the previous studies, HT-29 cells were exposed to IFNγ, which primes HT-29 cells to apoptose upon stimulation with TNF (48), and subsequently stimulated with saturating or subsaturating doses of TNF, EGF, or insulin alone or in combination. Lysates were profiled at 13 time points over 24 hours for 19 intracellular signaling events measured by kinase assay, immunoblot, or antibody array (Fig. 1, A and B) (49–51). These data provided quantitative, systematically collected information on phosphorylation-mediated regulatory events, changes in protein abundance, and cleavage-dependent protein activation.

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.

To determine how the signaling events altered gene expression, we complemented the signaling compendium with a matched set of transcriptomic profiles (Fig. 1A). IFNγ-pretreated HT-29 cells were exposed to the same combinations of TNF-EGF-insulin and analyzed at a subset of the time points from the previous signaling studies (Fig. 1B). We collected transcriptomic profiles by microarray at intermediate-to-late times after cytokine simulation, thereby avoiding the bursts of immediate-early transcripts that are already well characterized (46, 47, 52). We selected the time points for transcriptomic analysis according to earlier modeling of the signaling compendium, which showed that signaling from 4 to 16 hours did not predict apoptosis accurately (42). We reasoned that the loss of predictive ability was because prolonged cytokine stimulation had transmitted the relevant information to the downstream transcriptional network.

The microarray data revealed extensive transcriptional alterations with time and stimulus condition (Fig. 1C). Among 14,541 probe sets identified as present in at least one sample, we identified significant changes in 10,319 with time, 4948 upon TNF stimulation, 75 upon EGF stimulation, and 15 upon insulin stimulation after correction for multiple hypothesis testing [four-way analysis of variance (ANOVA), false -discovery rate = 5%; file S1]. One unanticipated complication was that many transcript abundances changed with time in the mock stimulation condition lacking TNF, EGF, and insulin (Fig. 1C, leftmost column). We attributed these background transcriptional dynamics to the ongoing IFNγ exposure. Extensive background drifts in transcript abundance can confound interpretations from standard analyses of differentially expressed genes that focus on time-dependent changes (53, 54). Therefore, alternative methods were required to identify meaningful changes in transcriptional regulation and link them to the upstream signaling network.

Models of dynamic, multivariate data sets are properly structured as data tensors

Systematic biology experiments monitor the same signaling events or transcripts over multiple time points and across multiple stimulus conditions (19, 55). Each data point thus contains information about the various “modes” of its acquisition, for example, which stimulus was added (mode 1), the time after stimulus (mode 2), and the signal or transcript measured (mode 3). Additional modes are possible if multiple pharmacologic perturbations (56, 57) or cell types (58) are profiled systematically along with the modes listed above.

For systematic measurements, the acquisition modes create a data structure that is very powerful mathematically because it conveys how different data points are related to one another. This structure vanishes when, for example, a data cube is sliced along one of its modes and “unfolded” end-to-end as a series of matrices (Fig. 2A). When matrix-based algorithms are applied to unfolded data, each unfolded measurement variable is treated independently, and modes 2 and greater are lost. Using Fig. 2A as an example, Akt measurements at 2 and 4 hours after stimulation (same signal, two time points) are not handled any differently than Akt and epidermal growth factor receptor (EGFR) measurements at 2 hours after stimulation (two signals, same time point). The result of unfolding is a model that is less interpretable because of too many fitted regression coefficients (59).

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 n_s signals and n_t time points. Using the unfolded matrix, data-driven modeling approaches (20) treat each time point of each signal as a separate predictor variable, yielding n_s × n_t 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 n_s + n_t regression coefficients and thus a more parsimonious model. (C) A dependent third-order transcriptomic tensor (Y) structured by stimulus, n_c gene clusters, and n_t2 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 (w_j and w_k or q_l and q_m). 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.

The alternative to unfolding is to retain data sets as cubes (three modes) or hypercubes (four or more modes) in the form of data “tensors” which are the higher-dimensional generalization of vectors (one mode) and matrices (two modes). For example, the TNF-EGF-insulin signaling compendium naturally organizes as a third-order tensor defined by stimulus, time point, and measured signaling event (Fig. 2B). The transcriptomic profiles likewise arrange as a third-order tensor according to stimulus, time point, and transcript or cluster of transcripts (Fig. 2C). Tensors reduce the parameterization of a data-driven model because free regression coefficients remain fixed across the other acquisition modes of each tensor (Fig. 2B) (59, 60). In this instance, the stimulus–time point–signaling tensor (the “regressor” tensor) is linked to the stimulus–time point–transcript tensor (the “regressand” tensor) by the regression coefficients.

Biological data tensors have been used successfully for unsupervised purposes, such as singular value decomposition (27), to analyze transcriptional kinetics during DNA replication origin firing (28) and to identify consistent copy number changes across different array-based comparative genomic hybridization platforms (61). Here, we sought a supervised method that could connect the signaling tensor to the transcriptomic tensor and predict gene expression patterns from signaling network dynamics. This application is ideal for the tensor generalization of partial least squares regression (PLSR), a matrix implementation that has been used widely to model signaling networks (20, 25, 58, 62–76).

Tensor PLSR [equivalently, “multilinear PLS” (59)] is an established method that creates a data-driven model by jointly factorizing an independent “predictor” tensor (X; here, the signaling tensor) and a dependent “predicted” tensor (Y; the transcriptomic tensor). X and Y are factorized as an element-by-element product of vectors, where the number of vector elements multiplied is equal to the number of dimensions in the data tensor (text S1). Thus, if X is a third-order tensor, then X(1,1,1) [the tensor element in X occupying the first position in mode 1 (stimulus), the first position in mode 2 (time point), and the first position in mode 3 (signal)] is factorized as: X(1,1,1) = t(1)• w_j(1)•w_k(1) (Fig. 2D, purple). In the factorization, t(1) is the first element of a “scores” vector (t) that relates to the stimulus conditions that are shared with the Y tensor. w_j(1) and w_k(1) are the first elements of two “weight” vectors (w_j and w_k) that relate to modes 2 and 3 of the tensor (here, time and signal). A similar calculation is performed for Y by factorizing it into its own scores (u) and weight (q_l and q_m) vectors. X and Y are linked by an “inner relationship” between their respective scores vectors: u = bt, where b is a linear regression coefficient determined by the model. The inner relationship implies that how a stimulus projects on t [through the signaling (w_k) that occurs over time (w_j)] is directly proportional to its projection on u and thus how that stimulus changes gene expression (q_m) with time (q_l).

The factorization of the two tensors is posed as a numerical optimization that seeks to capture as much of the inner relationship between X and Y as possible (details on the covariance maximization algorithm are described in text S1). The best first set of scores and weight vectors defines the first “latent variable” of the tensor PLSR model. Residual information (covariation) in X and Y not captured by the first latent variable is then subjected to a second factorization, which is optimized to capture as much covariance in the residual as possible (Fig. 2D). By repeating the algorithm, latent variables are iteratively calculated until there are no predictive inner relationships remaining between the X and Y data tensors (20, 62).

Predictions with a tensor PLSR model use w_j and w_k from each latent variable to project an X-like observation onto t (Fig. 2E). Then, the u = bt inner relationship is used to predict u, which is backprojected with q_l and q_m to yield a predicted set of values in the form of Y (time-dependent gene expression). The project-predict-backproject sequence is important for making independent predictions with new data and for cross-validation of the model to identify the optimum number of latent variables (20, 31, 62, 77).

Tensor PLSR modeling identifies predictive links between signaling and transcriptional dynamics

We first constructed a tensor PLSR model of three latent variables that predicted the 14,541 probe set fluorescence intensities of the transcriptomic data set. Although cross-validated predictions of the model were 99% accurate (fig. S1A), the model was strongly biased toward the differences in fluorescence intensities across probe sets. Consequently, changes in probe set intensities across treatment conditions were overlooked, and the resulting components of the model were uninterpretable (fig. S1, B to E). To focus on recurrent stimulus-dependent changes in transcript abundance shared by multiple genes, we condensed the transcriptomic data set by using the unbiased CLuster Identification via Connectivity Kernels (CLICK) algorithm (78). Among the transcripts profiled, CLICK identified nine separable clusters composed of dozens to hundreds of genes (file S2), the mean trajectories of which were organized as the Y data tensor (Fig. 2C). Using the entire signaling compendium as X, we constructed a tensor PLSR model of four latent variables that predicted gene -cluster dynamics to within 72% (Fig. 3A and fig. S2A). Although the model did not predict certain cytokine-induced changes for some gene clusters (Fig. 3A, see EGF and insulin stimuli of cluster #3), we considered the overall accuracy of predictions remarkable considering that the model involved ~10-fold fewer parameters than previous PLSR models of TNF-induced apoptosis (42, 62, 64).

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 XXX.

Our principal motivation for building the tensor PLSR model was to use the model to reveal undiscovered mechanisms of how signaling alters gene expression. To identify which latent variables captured both signaling and gene -cluster dynamics, we analyzed the time weights (w_j and q_l) and inner regression coefficients for X and Y (Fig. 3B). The leading two latent variables (LV#1 and LV#2) harbored signaling time weights (w_j1 and w_j2) that were nearly constant from 0 to 24 hours, indicating that time-dependent changes in signaling did not determine the projection of X along LV#1 or LV#2. Accordingly, time weights for the gene clusters (q_l1 and q_l2) were time-variant and derived from the stimulus-independent transcriptional changes of clusters #1, #2, and #7 (Fig. 3, A and B), presumably resulting from IFNγ pretreatment. Because latent variables are calculated iteratively (Fig. 2D), LV#1 and LV#2 eliminated the TNF-, EGF-, and insulin-independent transcriptional changes, revealing paired signaling and gene -cluster dynamics in the third and fourth latent variables (LV#3 and LV#4). LV#3 harbored time weights of late-phase signaling (w_k3) and sustained transcriptional activation (q_l3). Conversely, LV#4 was weighted with early-phase signaling (w_k4) and late-phase transcriptional regulation (q_l4). The inner regression coefficient for LV#4 was negative (Fig. 3B, orange), implying a link between early-phase signaling and downstream transcriptional repression.

Focusing on LV#3 and LV#4, we evaluated the relationship between the treatment scores (t₃ and t₄; Fig. 3C). Relative to mock treatment, saturating TNF stimulation projected almost entirely along LV#3, suggesting that LV#3 represented a TNF “axis” In contrast to TNF, we found that EGF and insulin projected in opposite directions along LV#4, indicating that this latent variable distinguished between the two growth -factor stimuli. Combinatorial stimulations exhibited intermediate scores that approximately averaged the scores of the individual stimuli. For example, TNF+EGF projected positively along LV#3 (like TNF) and negatively along LV#4 (like EGF). The interpolated response observed here for gene regulation contrasts with previous work on apoptosis in which EGF and insulin each nonlinearly antagonized TNF-induced cell death (62).

To connect specific signals and gene clusters with the prevalent cytokine-induced dynamics, we evaluated the signaling and gene -cluster weights along LV#3 (TNF; late-phase signaling axis: w_k3 and q_m3) and LV#4 (EGF-insulin; early-phase signaling axis: w_j4 and q_m4) (Fig. 3, D and E). Multiple signals, such as Ser⁶³⁶-phosphorylated insulin receptor substrate-1 [pIRS1 (Ser⁶³⁶)], c-Jun N-terminal kinase (JNK) activity, and mitogen-activated protein kinase (MAPK)–activated protein kinase-2 (MK2) activity, were negligibly weighted (Fig. 3D), implying that these early-phase TNF-induced signals (Fig. 1B) were statistically uninformative for predicting Y (the transcriptional response). Gene cluster #1 was also unweighted along LV#3 and LV#4, because its dynamics were almost entirely captured by LV#1 and LV#2 (Fig. 3, A, B, and E, and fig. S2B). To filter the weight vectors further, we randomly shuffled the signaling, gene -cluster, and time information within each cytokine stimulation (mode 1 slice) of X and Y (79). With hundreds of shuffled tensor PLSR models, we estimated a null projection for the weight vectors of LV#3 and LV#4 (Fig. 3, D and E, gray line). We considered signals and gene clusters outside 1 SD (σ, gray dashed lines) of the null projection as weighted strongly enough to warrant further analysis (Fig. 3, D and E).

Among the strongest signaling weights, we found clear agreement with known mechanisms of signal transduction (Fig. 3D). For example, cleavage of apoptotic caspases [negative weighting for procaspase-8 and procaspase-3 and positive weighting of cleaved caspase-8] was strongly aligned along LV#3, which is consistent with late-phase caspase activation triggered by TNF (48, 80). Along LV#4, insulin stimulation coincided with positive weights for three complementary measures of Akt activation, a recognized effector pathway (81). Likewise, multiple measures of EGFR phosphorylation were weighted in a direction that corresponded to EGF stimulation. Also strongly associated with EGF signaling was phosphorylated insulin receptor substrate-1 [pIRS1 (Tyr⁸⁹⁶)], consistent with reports that this site may be directly phosphorylated by active EGFR (42, 82). Not all signaling events were associated with individual stimuli. For instance, the weights associated with inhibitor of nuclear factor-κB kinase (IKK) activation mapped not only to TNF but also to insulin stimulation, possibly because Akt signaling can activate IKK in certain contexts (83, 84). Similarly, phospho-EGFR (Tyr¹⁰⁶⁸) projected with early-phase EGF and also late-phase TNF signaling. The latter is probably due to autocrine signaling by transforming growth factor-α, an EGFR ligand that is released after TNF stimulation (43, 85). Together, the weights of LV#3 and LV#4 provided a condensed map of the signaling compendium that was optimized for predicting the observed transcriptomic profiles.

The inner regression coefficient (b₃) connecting X and Y along LV#3 was a positive value, indicating gene activation, whereas the inner regression coefficient b₄ was negative, indicating that signaling along LV#4 resulted in gene repression (Fig. 3, B and E). Contrary to that of the signaling compendium, the projection of gene clusters along LV#3 and LV#4 was surprising (Fig. 3E). Amidst thousands of time-dependent transcriptional changes, few clusters were weighted toward specific stimuli. Clusters #3 and #6 were primarily weighted along LV#3, indicating an association with TNF stimulation. Accordingly, promoter analysis (86) of the transcripts in these two clusters revealed a strong overrepresentation of binding sites for nuclear factor κB (NF-κB) (fig. S2C). Cluster #7 mapped along LV#4 because of the mild suppression of transcripts observed with saturating insulin alone (Fig. 3, A and E). Only cluster #9 projected strongly along both latent variables, indicating that the transcripts in this cluster were induced by TNF and repressed by insulin (fig. S3). TNF antagonism of insulin function has been well documented in adipocytes (87, 88), but there are few reports of insulin antagonizing TNF (89). Given this predicted TNF and insulin “crosstalk cluster” we used the tensor PLSR model to investigate its mechanism of regulation by the upstream signaling network.

With respect to its latent variable projections, cluster #9 was cartographically most similar to IKK (Fig. 3, D and E). If these shared projections were indicative of mechanism, however, it would imply that early-phase IKK activity (downstream of TNF and insulin signaling) represses transcription of the crosstalk cluster, whereas late-phase IKK (downstream of TNF signaling) promotes it. Repress-then-activate kinetics are opposite of the prevailing view of IKK signaling (90). Accordingly, we found that TNF-induced responses of 85% of transcripts in cluster #9 were not significantly affected when a phosphorylation- and degradation-resistant mutant of IκBα was ectopically expressed in HT-29 cells (log-transformed Welch’s t test, false discovery rate = 15%; fig. S4). We therefore considered alternatives that were consistent with the tensor PLSR model.

One possible explanation was that the crosstalk cluster integrated two distinct signaling inputs. An activating input could arise from a TNF-specific signal that was either not measured or not projected strongly on LV#3 and LV#4. In parallel, the cluster could be transcriptionally inhibited by an insulin-specific signal, such as Akt (Fig. 3D), or a downstream effector pathway of Akt (91).

Promoter and signaling bioinformatics suggest a link between GSK3 and the crosstalk cluster through GATA6

First, we determined the reliability and generality of TNF-insulin crosstalk among transcripts in cluster #9. We repeated the stimulation experiments with an independently obtained vial of HT-29 cells and assayed individual transcripts by quantitative reverse transcription polymerase chain reaction (qRT-PCR). This analysis confirmed the expression of more than 90% of the 22 cluster #9 transcripts (SPRR1B and PPARD from fig. S3 were false positives), and we detected an antagonistic interaction between TNF and insulin from 2 to 8 hours after stimulation (fig. S5). At individual time points for specific genes, we observed instances of antagonism represented by significant interaction P value (P_int < 0.05, two-way ANOVA; Fig. 4, A and B), an indication that TNF and insulin have nonadditive effects on the expression of those genes. We also observed nonlinear significant differences in gene expression for other transcripts even when the P_int was not significantly different (Fig. 4, C and D). The qRT-PCR data thus confirmed the microarray results and the tensor PLSR model, showing an early-phase suppression of TNF-induced cluster #9 genes by insulin (Fig. 3, B to E). Furthermore, these data indicated that the TNF-insulin crosstalk cannot be predicted by adding the effect of insulin to the TNF response.

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) suggest 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) Phospho-mass spectrometry identifies 11 phosphorylation sites on GATA6_L. 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.

To identify candidate mediators of TNF-insulin crosstalk, we analyzed the expression-verified transcripts of cluster #9 with three orthogonal promoter -analysis algorithms (86, 92, 93). Only two transcription factors were suggested as candidate regulators by all three algorithms: T cell factor 4 (TCF4) and GATA (Fig. 4E). HT-29 cells harbor a truncating mutation in APC (adenomatous polyposis coli), a protein that inhibits the β-catenin pathway, and this truncation renders β-catenin and its transcriptional partner TCF4 constitutively active (94). Moreover, no changes in β-catenin localization were observed upon TNF simulation with or without insulin (fig. S6). Therefore, we focused on GATA, a family of six transcription factors that are important for development and differentiation (95). Using qRT-PCR (96, 97), we quantified the relative copy numbers of the GATA family and found that GATA6 was the most abundant isoform (Fig. 4F and fig. S7). Copies of GATA6 transcript also remained high during the early phase of TNF-only and TNF + insulin stimulation, whereas GATA2 and GATA3 were reduced two- to fourfold (Fig. 4, G to I). Notably, bioinformatic analysis (98) of the full-length GATA6 protein sequence uncovered a cluster of highly conserved serine residues that were candidate phosphorylation sites for GSK3 (Fig. 4J). Because GSK3 is a recognized substrate of Akt (81), these results suggested that GATA6 could be an insulin-dependent regulator of the crosstalk cluster.

Full-length GATA6 (GATA6_L) is distinguished by a long N-terminal extension that is conserved across vertebrates but missing in other GATA family members (fig. S8). In addition, leaky ribosome scanning (99) onto an in-frame methionine (Met¹⁴⁷) gives rise to a short form of GATA6 (GATA6_S) that is comparable in size (~45 kD) to the other GATA isoforms. GATA6_S lacks the candidate GSK3 phosphorylation sites that reside at the N terminus of GATA6_L (Fig. 4J), raising the possibility of selective regulation of GATA6_L by insulin.

Compared to GATA6_S, GATA6_L has been understudied because of early misannotations of nonhuman genomes and a lack of suitable reagents. Common plasmid repositories, including Addgene (100), the human ORFeome (101), and the Mammalian Gene Collection (102), have only the short form or lack the gene entirely. The unusually slow electrophoretic mobility of GATA6 has created additional confusion because commercial antibody vendors mistakenly label GATA6_S as “GATA6” implying that the detected protein is the full-length form. As a result, the GATA6 literature is incredibly ambiguous, with many papers inadvertently focusing on GATA6_S.

To determine whether the predicted GSK3 phosphorylation sites of GATA6_L could be phosphorylated in HT-29 cells, we cloned the full-length gene with N-terminal FLAG and C-terminal AU1 tags into a doxycycline inducible lentivector (see Materials and Methods) (103). Stable HT-29 transductants were induced with doxycycline for 24 hours before lysis, and FLAG immunoprecipitates were subjected to phosphorylation analysis by mass spectrometry. We achieved 92% coverage of the GATA6_L sequence and identified 11 phosphorylation sites, including 7 that had not been previously reported (Fig. 4K) (104). Among sites in the N-terminal extension specific to GATA6_L, two (Ser³³ and Ser³⁷) were consistent with the bioinformatic predictions of GSK3 phosphorylation (Fig. 4J). Two N-terminal sites (Thr³⁴ and Ser³⁷) were also corroborated by a proteomics study of proline-directed phosphorylation in 293T cells (105). We concluded that Ser³³, Thr³⁴, and Ser³⁷ were the leading candidates for GATA6_L phosphorylation–mediated regulation by GSK3.

Perturbation of basophilic kinases differentially affects GATA6_L and GATA6_S

Studies of GATA6 phosphorylation have largely focused on its posttranslational regulation by MAPKs (106–108). However, GATA6 is reportedly phosphorylated on Ser⁴³⁶ (Ser²⁹⁰ in GATA6_S) and stabilized upon prolonged mechanistic target of rapamycin complex 1 (mTORC1) inhibition with rapamycin and subsequent feedback activation of Akt2 in vascular smooth muscle cells (VSMCs) (109). Our mass spectrometry data on GATA6_L did not include peptides containing Ser⁴³⁶; thus, we could not rule out a role for Akt as a GATA6 kinase activated by insulin stimulation.

To determine whether Akt-catalyzed stabilization of GATA6 was relevant to our study, we treated HT-29 cells with the mTORC1 inhibitor rapamycin for 3 hours and monitored targets by quantitative immunoblotting (51). As expected, rapamycin eliminated phosphorylation of ribosomal protein S6 and increased phosphorylation of Akt by about twofold; however, the abundances of GATA6_S and GATA6_L were essentially unchanged (Fig. 5A). In the context of IFNγ pretreatment and TNF-insulin stimulation, we found that rapamycin treatment for 3 hours altered the distribution of GATA6 forms by decreasing the abundance of GATA6_L relative to the abundance of GATA6_S (fig. S9). The mechanism reported in VSMCs (109) might be restricted to mesodermal tissues, so we repeated the experiment in AC16 ventricular cardiomyocytes (110). In these cells, S6 phosphorylation disappeared without subsequent feedback activation of Akt, and yet, GATA6_S abundance increased modestly after 3 hours as reported in VSMCs (Fig. 5B) (109). Critically, under the same conditions in AC16 cells, we did not observe any alterations in the abundance of GATA6_L. These experiments illustrated that Akt feedback activation could be uncoupled from GATA6 stabilization, as could the posttranslational regulation of its long and short forms.

Fig. 5. GATA6_L 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 GATA6_S, but not GATA6_L, 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.

GSK3-dependent phosphorylation of Ser³⁷ accelerates GATA6_L turnover

To assess the importance of the GATA6_L phosphorylation sites, we transfected single alanine mutants of Ser³³, Thr³⁴, and Ser³⁷ or the triple mutant (3×SA) into 293T cells, cells in which GATA6_L phosphorylation has been detected previously (105). Compared to the wild-type allele, we noted a pronounced electrophoretic downshift in the major FLAG-immunoreactive band of the Ser³⁷ and 3×SA mutants (Fig. 6A). The mobility shift was larger than that expected for a single phosphorylation site, suggesting that Ser³⁷ phosphorylation was required for other phosphorylation events within GATA6_L. Iterative phosphorylation-dependent phosphorylation is characteristic of many GSK3 substrates, such as glycogen synthase (GS) (111).

Fig. 6. Extensive phosphorylation of GATA6_L is blocked by S37A mutation, reversed by TNF stimulation, and stabilized in HT-29 cells.

(A) Electrophoretic mobility of FLAG-tagged GATA6_L 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 GATA6_L. (C and D) Phos-tag electrophoresis (C) and quantification (D) of the upper and lower forms of GATA6_L in response to IFNγ sensitization for 24 hours, pretreatment with 1 μM CT99021 for 1 hour, and stimulation with TNF or insulin for one1 hour. Data are median proportion ± range of n = 3 biological replicates. (E and F) Doxycycline (DOX)–inducible addback in HT-29 cells replaces endogenous GATA6_S with epitope-tagged GATA6_L. Cells were treated with doxycycline (1 μg/ml) for 48 hours. (G and H) The less phosphorylated form of wild-type (WT) GATA6_L 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.

Careful inspection of endogenous GATA6_L immunoreactivity in HT-29 extracts revealed a slower migrating species that was similar to the electrophoretic shifts observed in 293T cells transfected with the phosphorylation-deficient mutants. We isolated this species from the faster migrating GATA6_L through Phos-tag electrophoresis (112) followed by Gaussian mixture modeling of the densitometric traces (see Materials and Methods). Acute TNF treatment reduced the upper form of GATA6_L but with a concomitant increase in the lower form, such that total GATA6_L abundance was not altered (Fig. 6B). Costimulation with insulin or pretreatment with the GSK3 inhibitor CT99021 (113) did not alter the GATA6_L downshift, despite insulin increasing GSK3 phosphorylation and CT99021 decreasing GSK3 activity (Fig. 6, C and D, and fig. S10). Because GATA6 mRNA was not induced by TNF (Fig. 4I), these results indicated that GATA6_L is dephosphorylated on some residues in response to TNF stimulation.

Our next goal was to evaluate the specific impact of Ser³⁷ phosphorylation on GATA6_L in HT-29 cells. One challenge was that the endogenous abundance of GATA6_S was high compared to GATA6_L (Fig. 5A), which could confound interpretations of ectopically expressed proteins. Therefore, we inducibly knocked down endogenous GATA6 with short hairpin RNA and added back epitope-tagged versions of wild-type or S37A GATA6_L so that the abundance was comparable to total endogenous GATA6 (Fig. 6, E and F; see Materials and Methods). Doxycycline-induced addback in HT-29 cells recapitulated the electrophoretic mobilities of GATA6_L that were observed in transfected 293T cells. The data suggested that the GATA6_L modifications are not artifacts of overexpression, enabling use of the addback cells to examine the consequences of Ser³⁷ phosphorylation.

Because GSK3 phosphorylation often accelerates substrate turnover (114), we combined the Phos- tag analysis with the HT-29 addback lines to estimate half-lives of wild-type and S37A GATA6_L. We combined inhibition of protein synthesis with TNF stimulation to enrich for the lower migrating form of wild-type GATA6_L in the addback cells. Under these conditions, we found that the half-life of more phosphorylated GATA6_L was more than twice that of the wild-type GATA6_L form that was less phosphorylated (Fig. 6, G and H). Surprisingly, the half-life of the S37A mutant was comparable to that of the more phosphorylated form of GATA6_L, suggesting that phosphorylation of Ser³⁷ without subsequent additional phosphorylation renders GATA6_L unstable. Ser³⁷ resides in the middle of a proline-glutamate-serine-threonine (PEST) degradation motif of GATA6_L, and this motif scores more strongly as a PEST motif than those in other well-known unstable proteins (Fig. 4J and table S1) (115, 116). Ser³⁷ phosphorylation might activate or expose the PEST sequence for rapid proteolytic degradation of GATA6_L, whereas additional phosphorylation at other sites could inhibit substrate recognition (115).

To monitor Ser³⁷ phosphorylation specifically, we raised and affinity-purified a phospho-specific antibody against a monophosphorylated peptide fragment of the PEST sequence in GATA6_L (see Materials and Methods; fig. S11). If Ser³⁷ modification were a prerequisite for subsequent phosphorylation, then the antibody would capture this initial phosphorylation event of GATA6_L, with the caveat that phosphorylation on Ser³³ and Thr³⁴ would ultimately disrupt the antibody epitope. To evaluate the phospho-Ser³⁷ antibody, we reassessed the immunoreactivity of total GATA6. The predicted molecular weights of GATA6_S and GATA6_L are 45.4 and 60 kD, respectively. However, extensive posttranslational modifications (Fig. 4K) cause most GATA6_S and GATA6_L to run at an apparent molecular weight of ~54 kD and ~69 to 75 kD depending on electrophoresis conditions (Fig. 7A). Multiply phosphorylated GATA6_S (~54 kD) can be misinterpreted as unmodified GATA6_L (60 kD). Upon long exposure with a total GATA6 antibody, we revealed an additional immunoreactive band at ~60 kD that was eliminated by GATA6 knockdown and reconstituted with addback of wild-type GATA6_L and the S37A mutant (Fig. 7A). In contrast to the 75-kD form (Fig. 6E), the ~60-kD form of wild-type GATA6_L was significantly less abundant than the S37A mutant (Fig. 7B), consistent with decreased stability. We interpreted the ~60-kD band as the unmodified form of GATA6_L.

Fig. 7. Phosphorylation and destabilization of endogenous GATA6_L at 60 kD.

(A) Knockdown (shGATA6) and FLAG-tagged addback of GATA6_L 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 GATA6_L compared to the S37A-mutant addback cells. (C and D) Endogenous phospho-GATA6_L (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 GATA6_L upon serum starvation. Specificity was confirmed by preincubation of cells with 20 μM CT99021 for 1 hour before serum starvation. (G and H) Phospho-GATA6_L (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% input%) 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.

Endogenous Ser³⁷ phosphorylation of the 75 and 60-kD GATA6_L forms was not detectably altered in response to TNF treatment for 1 hour or CT99021 treatment for 6 hours (Fig. 7, C and D). However, the endogenous 60-kD phospho-GATA6_L signal was barely above the detection limit and thus highly variable (coefficient of variation, ~40%), yielding only ~50% statistical power for detecting a 1.5-fold change. To evaluate phosphorylation of endogenous GATA6_L on Ser³⁷, we serum- starved the HT-29 cells to produce a stronger activation of GSK3 and confirmed specificity of the phosphorylation events with CT99021 pretreatment for 1 hour (Fig. 7, E and F). As expected, serum starvation reduced GSK3 phosphorylation (increasing GSK3 activity) and increased phosphorylation of GS (Fig. 7F). CT99021 reduced GS phosphorylation and total GSK3 abundance but also transiently increased total GS. Notably, within 1 hour of serum withdrawal, we observed a robust increase in the 60-kD form of phospho-GATA6_L, which coincided with the timing of GSK3 dephosphorylation and was blocked by CT99021 in a dose-dependent manner (Fig. 7E and fig. S12). By contrast, phospho-Ser³⁷ immunoreactivity of the 75-kD form of GATA6_L was not altered by serum starvation or CT99021 treatment, suggesting that Ser³⁷ was already stably phosphorylated in this form of GATA6_L.

Total abundances of the different forms of GATA6_L also showed dynamic changes. GATA6_L at 75 kD and GATA6_S decreased significantly with CT99021 treatment, whereas GATA6_L at 60 kD increased compared to uninhibited control cells that were serum-starved (P < 0.05, two-way ANOVA). The time-dependent changes in 60-kD GATA6_L abundance mirrored the changes in total GS and inversely correlated with changes in 60-kD GATA6_L phosphorylation at Ser³⁷ (Fig. 7, E and F). These experiments provide further evidence that Ser³⁷ is a site phosphorylated by GSK3 and that phosphorylation at this site promotes turnover of GATA6_L in the absence of phosphorylation at additional sites.

With greater confidence in the phospho-GATA6_L (Ser³⁷) antibody, we revisited the original biological context of IFNγ-pretreated HT-29 cells stimulated with TNF and insulin. To enable detection, we immunoprecipitated cell extracts with phospho-GATA6_L (Ser³⁷) antisera and immunoblotted for total GATA6 (Fig. 7G). An extended electrophoresis was required to separate the 60-kD form from the heavy chain of the immunoprecipitating antibody, causing a smear of immunoreactivity rather than a discrete band (see Materials and Methods). In response to TNF alone, we repeatedly observed a drop in 75-kD, but not 60-kD, GATA6_L phosphorylation (Fig. 7H), corroborating the dephosphorylation previously noted by Phos-tag electrophoresis (Fig. 6B). Moreover, the reduction in 75-kD phospho-GATA6_L (Ser³⁷) was blocked by insulin costimulation, indicating a specific point of crosstalk between TNF and insulin. Insulin, by contrast, independently elevated the abundance of 60-kD GATA6_L phosphorylation, suggesting that insulin delays the turnover of this form beyond its effect on Ser³⁷ phosphorylation. We conclude that the phosphorylation of endogenous GATA6_L at Ser³⁷ is consistent with the antagonism and linear superposition in abundance of cluster #9 transcripts observed upon TNF + insulin stimulation (Fig. 4, A to D, and fig. S5).

GSK3-dependent phosphorylation of Ser³⁷ alleviates GATA6_L repression of transcripts in the crosstalk cluster

We investigated the role of GATA6_L phosphorylation in the regulation of transcripts subjected to TNF-insulin crosstalk. We inducibly overexpressed wild-type GATA6_L in HT-29 cells (Fig. 8A) before stimulation with TNF for 2 hours and transcriptomic profiling by microarray (see Materials and Methods). GATA6 can act as either a transcriptional activator or repressor (117), but we found that, in the GATA6_L-overexpressing HT-29 cells, the effects of GATA6_L were predominantly repressive. Without stimulation, GATA6_L overexpression induced 51 transcripts and repressed 136 transcripts at a 5% false -discovery rate (P < 10⁻¹⁰, binomial test) (file S4). TNF stimulation increased the number of genes affected by GATA6_L overexpression, but the bias toward repression persisted (317 induced transcripts versus 438 repressed transcripts; P < 10⁻⁵, binomial test) (file S4). Notably, the same repressive bias was observed for transcripts in the crosstalk cluster (Fig. 8B), and those with the strongest GATA6_L-associated repression were among the clearest examples of TNF-insulin crosstalk (Fig. 4, A to C). Using chromatin immunoprecipitation (ChIP), we confirmed binding of GATA6_L to consensus sites within the promoters of many TNF-insulin crosstalk genes (fig. S13), suggesting that repression is direct.

Fig. 8. S37A mutation of GATA6_L mimics and competes with the repression of transcript abundance in the TNF-insulin crosstalk cluster.

(A) Doxycycline-inducible overexpression of WT GATA6_L 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 GATA6_L 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 GATA6_L mimics insulin stimulation (green) and antagonizes TNF-insulin crosstalk (purple). qRT-PCR data for cluster #9 genes in WT and S37A mutant (S37A) GATA6_L 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 GATA6_L regulation by TNF and insulin and its relation to the crosstalk cluster of transcripts. Ovals annotate the figure subpanels supporting the links depicted.

If GATA6_L mediates the insulin-stimulated repression of the crosstalk cluster and is stabilized by inhibition of the GSK3 pathway, then phosphorylation of Ser³⁷ would provide a mechanism for TNF-insulin crosstalk. Furthermore, the S37A mutant of GATA6_L should mimic the effect of insulin on TNF-stimulated gene expression for transcripts in the crosstalk cluster and dampen the crosstalk observed when insulin is added to S37A mutant cells. We tested this prediction with the wild-type and S37A addback lines (Fig. 6, E and F), inducing GATA6_L and then stimulating with TNF, insulin, or both. By qRT-PCR, we identified multiple instances in which S37A addback reduced transcript abundance similar to that observed in wild-type addback cells stimulated with insulin (Fig. 8C, green). We also found many examples in which TNF + insulin–induced transcript abundance was higher in S37A addback cells compared to wild-type addback cells (Fig. 8C, purple) and more similar to TNF-treated cells, suggesting reduced crosstalk. Analysis of the entire cluster #9 data set revealed significant interactions between GATA6 and TNF or insulin (interaction P < 10⁻³, five-way ANOVA), indicating that the Ser³⁷ genotype alters the transcriptional response to both stimuli. Together, our data support a model whereby TNF promotes and insulin inhibits the formation and degradation of GATA6_L monophosphorylated on Ser³⁷ (Fig. 8D). This phosphoregulation is ultimately reflected by the abundance of transcripts in the crosstalk cluster.

Ser³⁷ phosphorylation of different GATA6_L forms is observed in diverse cell types

The model of GATA6_L phosphorylation–mediated regulation (Fig. 8D) may be specific to HT-29 cells or could occur in other cell types. We immunoblotted various cell lines with the phospho-Ser³⁷ antibody in comparison to affinity-purified antisera binding the nonphosphorylated peptide surrounding Ser³⁷ (see Materials and Methods) and to other commercial GATA6 antibodies (100, 118). In HCT-8 and DLD-1 colorectal cancer lines, AC16 cardiomyocytes (110), and MCF10A-5E breast epithelial cells (119), we observed the 60 and 75-kD forms of GATA6_L, which were phosphorylated to variable extents according to phospho-Ser³⁷ immunoreactivity (Fig. 9, A and B). Multiple GATA6 antibodies also recognized another species at ~100 kD (Fig. 9, A to D), suggesting that an even more phosphorylated form of GATA6_L may remain to be characterized. The aggregate number of reported phosphorylation sites on GATA6_L now exceeds 20 (Fig. 4K).

Fig. 9. Diversity of GATA6_L 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 MCF10A-5E and AC16 samples are on an immunoblot; HT-29, HCT-8, and DLD-1 are on another blot. The blots have been scaled to match. Data are representative of n ≥ 3 independent experiments.

DISCUSSION

Our study here introduces and implements tensor PLSR as an approach for structured biological data sets. Considering that signaling dynamics often occur in discrete temporal phases (49, 80, 120, 121), tensor PLSR provides an attractive means to deconstruct time -course data in a systematic manner. Although the mathematics has been established for decades (59), data types that can exploit the tensor framework are relatively new to cell signaling. For tensor generation, a multiplex technique that simply measures many genes or proteins is insufficient. The method must also be cost-effective, reproducible, and scalable for repeated use across multiple treatments, time points, and perturbations. The newest technologies rarely meet these criteria, prompting our use of long-established methods at a scale not typically considered.

We applied tensor PLSR with the goal of discovering molecular mechanisms that connect signaling to transcriptional regulation. Ideally, the mechanisms would involve proteins not originally included in the systematic data set. Systems-level studies rarely uncover these “hidden nodes” and validate them experimentally like we achieved here for GATA6_L (25, 122). Testing model- or bioinformatics-derived predictions requires a skill set entirely different from the one needed to perform the analysis. Our findings argue for the benefits of dual training, where computationalists work at the bench and experimentalists use quantitative models, gaining an appreciation for the thematic similarities in each approach. For example, just as modeling assumptions should be subjected to falsification (31), we sought to challenge the prevailing biological assumptions about GATA6 and its different forms.

The deceptive electrophoretic mobilities of GATA6_S and GATA6_L have important implications for biological function. Although GATA6_L is generally less abundant than GATA6_S in most cell types, there is evidence that GATA6_L is the more potent transcriptional regulator (99). GATA6 promotes the expression of the stem cell marker LGR5 in colorectal cancer (118, 123). Neither paper clarified whether the regulation occurs through GATA6_L, GATA6_S, or both. However, insulin-like growth factor inhibits GSK3 and promotes expansion of Lgr5⁺ stem cells in mice (81, 124). Our results indicate that one mechanism for this expansion is the stabilization of GATA6_L.

The phosphorylation of Ser³⁷ adds a GATA6_L-specific mode of regulation to reports of posttranslational modifications that would presumably target both long and short forms (107–109). Although we were unable to reproduce the mechanism exactly (109), modification of Ser⁴³⁶ by Akt2 should coincide with the loss of Ser³⁷ phosphorylation to stabilize GATA6_L synergistically in contexts where both pathways operate. GATA6_L phosphorylation–mediated regulation could prove important in endothelial cells, a TNF- and insulin-responsive cell type in which GSK3 and GATA6 interact as a complex (125). Our mass spectrometry study also uncovered other GATA6_L-specific proline-directed modification sites (Thr⁶² and Ser¹³⁷) that could function with the Ser²⁶⁶ site phosphorylated by extracellular signal–regulated kinase (ERK) (107, 108). Such complex layers of regulation should be expected of a transcription factor that is central to embryonic development and cell specification (95).

By coupling systematic experiments with statistical modeling approaches, such as tensor PLSR, one can identify relationships that would otherwise go unnoticed. Although it remains atypical to collect transcriptomic data as tensors, we expect widespread systematization of transcriptomics as expression -profiling costs drop. A model is just the first step, however, because the most surprising data-derived connections will require the identification of previously unrecognized mechanisms to explain them. These, in turn, require hypothesis-driven experiments with the best molecular -genetic and pharmacologic perturbations available. For understanding how gene expression is controlled by complex stimuli, the integration of molecular biology and systems biology has yet to be fully exploited.

MATERIALS AND METHODS

Cell culture

HT-29, 293T, DLD-1, and HCT-8 cells were obtained from the American Type Culture Collection (ATCC) and cultured according to their recommendations. The 5E clone of MCF10A cells was cultured as described previously (119). AC16 cells (110) were purchased from M. Davidson (Columbia University) and cultured in Dulbecco’s modified Eagle’s medium/F-12 medium (Life Technologies) with 12.5% tetracycline-free fetal bovine serum (Clontech) and penicillin-streptomycin (Gibco).

Cell stimulation

HT-29 cells were plated at 50,000 cells/cm² for 24 hours, sensitized with human IFNγ (200 U/ml) for 24 hours (Roche), and then treated with TNF (100 ng/ml; PeproTech), insulin (500 ng/ml; Sigma), or both for the indicated times. HT-29 cells engineered to express GATA6_L inducibly were treated with doxycycline (1 μg/ml) for 24 hours (overexpression) or 48 hours (addback) before cytokine stimulation.

Plasmids

Wild-type GATA6_L was amplified by PCR from HT-29 RNA that had been reverse-transcribed with a GATA6-specific primer (CAAAAGCAGACACGAGTGGA). An N-terminal 3×FLAG tag and a C-terminal 3×AU1 tag were added by PCR before cloning into the Bam HI and Sal I sites of pBabe puro (126) or the Mfe I and Spe I sites of pEN_TTmiRc2 (103). The pEN_TT donor vector containing GATA6_L was then recombined with the pSLIK Neo destination vector (103) by using LR Clonase (Invitrogen). The shGATA6 sequence (CCCAGACCACTTGCTATGAAA; #TRCN0000005390 from the RNAi Consortium) was cloned into tet-pLKO-puro (127) as described previously (97). S33A, T34A, S37A, and 3×SA point mutants were prepared by site-directed mutagenesis (QuikChange II XL, Agilent). RNA interference–resistant mutants of wild-type and S37A GATA6_L were prepared by introducing four silent mutations into the sequence targeted by shGATA6, which were replaced with rare mammalian codons that would minimize ectopic expression. The phosphorylation- and degradation-resistant IκBα super-repressor plasmid has been previously described (128). All DNA constructs were verified by sequencing.

Production and purification of phospho-GATA6_L (Ser³⁷) antibody

The peptide sequence Ac-CREPSTPPpSPIS-amide was conjugated to keyhole limpet hemocyanin and used to immunize rabbits according to the manufacturer’s recommendations (Covance). Serum samples were tested by immunoblotting with positive and negative controls for phospho-GATA6_L (Ser³⁷). Serum pooled from the production and terminal bleeds was negatively selected on a CREPSTPPpSPIS peptide–conjugated N-hydroxysuccinimide (NHS)–Sepharose column. The bound immunoglobulin G (IgG) was eluted as the non–phospho-GATA6_L custom antibody while the flow through was exposed to a second CREPSTPPpSPIS peptide–conjugated NHS Sepharose column. The bound IgG was eluted as the phospho-GATA6_L (Ser³⁷) antibody and used for detection by immunoblotting.

Lentiviral packaging and transduction

Lentiviruses were prepared in human embryonic kidney (HEK) 293T cells (ATCC) by calcium phosphate transfection of the lentivector together with psPAX2 and pMD.2G (Addgene). Lentiviral transduction of HT-29 cells was performed as described previously (96). Transduced cells were selected in growth medium containing 2 puromycin (2 μg/ml) or G418 (600 μg/ml) until control plates had cleared. For addback experiments, viral titers were reduced to ensure single-virion transductants that matched the endogenous protein abundance as closely as possible.

Microarray profiling

HT-29 cells were plated at 50,000 cells/cm² for 24 hours and sensitized with IFNγ (200 U/ml; Roche) for 24 hours before stimulation with TNF (0, 5, or 100 ng/ml), EGF (0, 1, or 100 ng/ml), and insulin (0, 5, or 500 ng/ml) for 4, 8, or 16 hours. RNA isolation was performed with the RNeasy Mini Kit (Qiagen), and integrity of purified RNA was confirmed on a Bioanalyzer (Agilent). Preparation of labeled complementary RNA, hybridization to GeneChip Human Genome U133A Arrays (Affymetrix), microarray scanning, and microarray processing were performed as previously described (129).

For inducible GATA6_L overexpression, stably transduced HT-29 cells were plated at 50,000 cells/cm² for 24 hours, induced with doxycycline (1 μg/ml), and sensitized with IFNγ (200 U/ml Roche) for 24 hours before stimulation with TNF (100 ng/ml) for 2 hours. RNA was purified as described above and amplified with the Illumina TotalPrep-96 RNA Amplification Kit (Life Technologies) before hybridization to a HumanHT-12 v4 Expression BeadChip.

Hierarchical and CLICK clustering

One-way hierarchical clustering of the signaling and transcriptomic compendia was performed in MATLAB with the clustergram function using Euclidean distance and Ward’s linkage after row standardization. CLICK clustering was performed as described (78) with the default homogeneity parameter.

Tensor PLSR

Tensor PLSR was performed in MATLAB with version 2.02 of the NPLS Toolbox (130). The signaling compendium was structured by stimulus condition (mode 1), time point (mode 2), and measured signal (mode 3). The transcriptomic profiles were structured by stimulus condition (mode 1), time point (mode 2), and CLICK gene cluster (mode 3). Both data tensors were mean-centered along mode 1 and variance-scaled along modes 2 and 3 before calculation of latent variables (131). The scores and time weights of LV#4 of the signaling tensor were both multiplied by –1 to improve model interpretability. Randomized models were constructed in MATLAB with the shuffle matrix function applied within each stimulus condition before preprocessing and calculation of latent variables. The code for the tensor PLSR models is included in file S3.

Bioinformatic analyses of crosstalk cluster

The 20 transcripts from cluster #9 confirmed present by qRT-PCR were submitted to three promoter analysis algorithms. First, the proximal promoter of each transcript [defined as 2000 base pairs (bp) upstream and 500 bp downstream of the transcription start site] was collected from National Center for Biotechnology Information (NCBI) and used as an input set for MEME, which uses expectation maximization to define recurrent motifs in a set of sequences (132). The top five enriched motifs were searched against a database of 843 binding specificities (133) using TOMTOM (134) to identify known transcription factor recognition sequences. A GATA motif was also enriched when using 2000 bp of upstream sequence alone or 1500 bp of upstream sequence and 500 bp of downstream sequence.

Expression-verified cluster #9 transcripts were additionally analyzed with DiRE (86), which uses interspecies sequence conservation to define motifs that are searched against the TRANSFAC 10.2 database of roughly 400 transcription factor binding motifs. In DiRE, the occurrence metric reflects the overall frequency of a conserved binding motif in the input data set, whereas the importance metric reflects the specificity of the binding motif to the input data set compared to a background data set of 5000 randomly selected genes. The top 20 motifs based on occurrence were used as the DiRE predictions.

Last, X2K (92) was used to identify bioinformatic connections between cluster #9 transcripts and signaling pathways. X2K integrates the ChEA database (135) of transcription factor binding sites detected by ChIP, the JASPAR and TRANSFAC position weight matrices, as well as various protein-protein interaction and kinase-substrate databases to connect kinase signaling events to gene expression patterns. The top 20 transcription factors linked to signaling and cluster #9 transcripts in a 2011 analysis were used as the X2K predictions.

Quantitative RT-PCR

RNA from cultured cells was isolated with the RNeasy Plus Mini Kit (Qiagen) according to the manufacturer’s protocol. First-strand complementary DNA synthesis and qRT-PCR were performed as described (63). Parental HT-29 samples were normalized to the geometric mean of GAPDH, HINT1, PPIA, and PRDX6. GATA6_L addback samples were normalized to the geometric mean of GAPDH, HINT1, PPIA, PRDX6, B2M, and GUSB. Primer sequences are available in table S2.

Mass spectrometry

HT-29 cells stably expressing doxycycline-inducible 3×FLAG-GATA6_L were induced with 1 doxycycline (1 μg/ml) for 24 hours and lysed in NP-40 lysis buffer plus protease and phosphatase inhibitors (51). Protein extract (60 mg) in 6-ml volume was first cleared with 50 μl of mouse IgG–agarose beads (Sigma) for 1 hour at 4°C on a nutator. The cleared lysates were subjected to immunopurification using 80 μl of anti-FLAG M2 affinity gel (Sigma) for 3 to 4 hours followed by two washes with NP-40 lysis buffer, one wash with 500 mM NaCl, and one wash with tris-buffered saline. Immunoprecipitates were eluted with 3×FLAG peptide (500 ng/ml; Sigma) in 100 μl for 30 min at 4°C on a nutator. The eluate was concentrated using an Amicon ultra centrifugal filter (Millipore), and samples were prepared in dithiothreitol-containing Laemmli sample buffer and separated by SDS polyacrylamide gel electrophoresis on an 8% polyacrylamide gel followed by Coomassie brilliant blue staining. The stained bands were cut and subsequently reduced, alkylated, and digested with trypsin, chymotrypsin, or pepsin. Peptides from each enzymatic digestion were acrylamide-extracted and subjected to liquid chromatography–mass spectrometry on a Thermo Electron Orbitrap Velos ETD mass spectrometer system. The data were analyzed using the Sequest search algorithm against the International Protein Index human proteome database and the predicted GATA6 protein sequence. Full mass spectrometry details are available in texts S1 and S2.

Immunoblotting

Quantitative immunoblotting was performed as described previously in detail (51) with primary antibodies recognizing the following proteins or epitopes: phospho-GATA6_L (Ser³⁷, Covance, 1:1000 for crude antiserum and 1:500 after affinity purification), non–phospho-GATA6_L (Ser³⁷, Covance, 1:500), GATA6 (D61E4, Cell Signaling Technology #5851, 1:2000), GATA6 (H-92, Santa Cruz Biotechnology #9055, 1:600), phospho-GS (Ser⁶⁴¹, Cell Signaling Technology #3891, 1:1000), GS (Cell Signaling Technology #3893, 1:1000), GSK3α (Cell Signaling Technology #9338, 1:1000), phospho-GSK3α (Ser²¹, Cell Signaling Technology #9316, 1:1000), phospho-Akt (Ser⁴⁷³, Cell Signaling Technology #4060, 1:1000), Akt (Cell Signaling Technology #9272, 1:1000), phospho-S6 (Ser^240/244, Cell Signaling Technology #5364, 1:1000), S6 (54D2, Cell Signaling Technology #2317, 1:1000), FLAG (M2, Sigma #F1804, 1:10,000), β-actin (Ambion #4302, 1:5000), vinculin (Millipore #05-386, 1:10,000), GAPDH (Ambion #4300, 1:20,000), tubulin (Abcam #89984, 1:20,000), and p38 (C-20, Santa Cruz Biotechnology #535, 1:5000). Membrane blocking, antibody probing, and near-infrared fluorescence detection were performed as described (51), except for phospho-GATA6 (Ser³⁷) immunoblotting, where blocking with 5% nonfat skim milk and use of tris-buffered saline buffers were required.

Phos-tag immunoblotting was performed on a 6% polyacrylamide gel containing 10 μM Phos-tag acrylamide AAL-107 (Wako Chemicals) and 0.1 μM MnCl₂. Gels were run with WIDE-VIEW prestained protein markers under constant current (40 mA) for 170 min. Before electrophoretic transfer, gels were incubated with 1 mM EDTA in modified Towbin’s transfer buffer (51) for 15 min. Membrane blocking, antibody probing, and near-infrared fluorescence detection were then performed as described (51).

For Gaussian mixture modeling of GATA6_L forms, raw 16-bit pixel intensities were integrated horizontally across each lane and then plotted along the vertical dimension. Using the fit function in MATLAB, the vertical trace was fit by nonlinear least-squares to the following function:

$$f(x)=b+w_1\text{exp}\left(-\frac{x-{\mathrm{\mu }}_1}{2{\mathrm{\sigma }}^2}\right)+w_2\text{exp}\left(-\frac{x-{\mathrm{\mu }}_2}{2{\mathrm{\sigma }}^2}\right)$$

where f(x) is height of the vertical trace as a function of the vertical position x, b is a fixed background, w₁ and w₂ are the relative weights of the two bands, μ₁ and μ₂ are the mean vertical positions of the two bands, and σ² is a shared variance for the two bands. Normalized versions of w₁ and w₂ were taken as the relative band densities for the two forms.

Phospho-GATA6_L (Ser³⁷) immunoprecipitation

HT-29 cells were plated at 75,000 cells/cm², pretreated with human IFNγ for 24 hours (Roche), and then treated with TNF (100 ng/ml; PeproTech), insulin (500 ng/ml; Sigma), or both for 1 hour. Cells were lysed in NP-40 lysis buffer (51) supplemented with 10 mM sodium pyrophosphate and 30 mM sodium fluoride, and ~4 mg of cellular extract (adjusted according to total GATA6_L abundance based on immunoblotting) was incubated with 10 μl of phospho-GATA6_L (Ser³⁷) antiserum overnight on a nutator at 4°C. The following day, 30 μl of Protein A/G Plus UltraLink resin (Thermo) was added to the immune complexes for 1 hour on a nutator at 4°C. Beads were washed twice with ice-cold supplemented NP-40 lysis buffer and twice with ice-cold phosphate-buffered saline (PBS) before elution in Laemmli sample buffer (136). Samples were electrophoresed on a 10% polyacrylamide gel for 3 to 5 hours (until the 50-kD marker reached the bottom of the gel) before proceeding with quantitative immunoblotting as described (51). Densitometry was performed by integrating the pixel intensity (without lane-specific background subtraction) of the 75-kD band and, for the 60-kD form, the intensity between the 75-kD band and the upper shoulder of the immunoprecipitating antibody heavy chain (file S5).

Chromatin immunoprecipitation

Five million wild-type and S37A GATA6_L-addback HT-29 cells were seeded in 10-cm culture plates for 24 hours before inducing knockdown-addback with doxycycline (1 μg/ml) for 48 hours. Cells were fixed for 7 to 10 min by adding a 37% formaldehyde stock to the culture medium to a final concentration of 1%. Fixation was quenched with 1_/20 volume of 2.5 M glycine for 7 to 10 min at room temperature. Cells were washed twice with cold PBS, scraped into 1 ml of PBS, and centrifuged at 400 relative centrifugal force (rcf) for 3 min. The cell pellets from four 10-cm plates were combined and lysed in ChIP lysis buffer (96) to a final volume of 1.5 ml. Lysates were incubated on ice for 10 min and then sonicated using a Branson digital sonifier for 5 min at 40% amplitude with 0.7 s “on” and 1.3 s “off” pulse cycles. After centrifugation at 14,000 rcf for 20 min, the supernatant was collected, and 20 μl of the soluble chromatin was retained as the input fraction. Soluble chromatin was diluted 10-fold in dilution buffer (96), precleared with 100 μl of mouse IgG–conjugated agarose beads (Sigma) for 4 hours at 4°C with constant agitation, and then incubated with 100 μl of anti-FLAG M2 affinity gel (Sigma) or mouse IgG–conjugated agarose beads overnight at 4°C with constant agitation. Agarose beads were collected and washed as previously described (96). DNA from the beads and the input fraction was eluted by reversing methylene cross-links with 500 μl of elution buffer (96) at 65°C for 5 hours. Samples were then treated with ribonuclease (100 μg/ml) for 30 min at 37°C and proteinase K (200 μg/ml) for 90 min at 50°C, followed by phenol-chloroform extraction. The aqueous fraction was ethanol-precipitated, washed once in 70% ethanol, air-dried, and dissolved in nuclease-free water. The samples were diluted 10-fold in nuclease-free water and quantified by PCR with primers designed for proximal promoter regions of selected crosstalk genes (123). Primer sequences are available in table S3.

Statistical analysis

Microarray data were analyzed by four-way ANOVA (factors: TNF, EGF, insulin, and time) in MATLAB with the anovan function at a false discovery rate of 5% (file S1). qRT-PCR data of cluster #9 genes were analyzed by four-way ANOVA (factors: transcript, TNF, insulin, and time) or, for GATA6_L addback, by five-way ANOVA (factors: GATA6_L genotype, transcript, TNF, insulin, and time) in MATLAB with the anovan function after log transformation. Half-lives of GATA6_L forms were estimated by nonlinear least-squares curve fitting to the following function:

$$g(t)=cexp\left(-\frac{ln(2)}{{\mathrm{\tau }}_{1/2}}t\right)+b$$

where g(t) is the relative band intensity as a function of time t, c is the scaling coefficient, b is a fixed background, and τ_1/2 is the half-life. Differences in means were assessed by Welch’s t test, and differences in geometric means were assessed by Welch’s t test after log transformation. One- or two-sidedness was based on previous evidence or expectation for a directional change. Tests for enrichment were performed by binomial test. Differences between immunoblotting time courses were assessed by two-way ANOVA.

Supplementary Material

Supplement

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 GATA6_S to GATA6_L independently of TNF or insulin treatment.

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

Fig. S11. Phospho-GATA6_L (Ser³⁷) antiserum is specific for mobility-shifted wild-type GATA6_L but not the S37A GATA6_L mutant.

Fig. S12. GATA6_L phosphorylation on Ser³⁷ and GS phosphorylation on Ser⁶⁴¹ are lost in a dose-dependent manner upon treatment with the GSK3 inhibitor, CT99021.

Fig. S13. GATA6_L 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 GATA6_L overexpression with or without TNF stimulation.

File S5. Raw files and densitometric analysis of phospho-GATA6_L (Ser³⁷) immunoprecipitation and total GATA6 immunoblots.

Text S1. Detailed description of tensor PLSR.

Text S2. Detailed description of GATA6_L mass spectrometry.