Transcript and protein signatures of pan-cancer cell lines organize by E-cadherin expression

The Cancer Cell Line Encyclopedia (CCLE) contains 1037 cancer cell lines with profiled transcriptomes [11], and the MD Andersen Cell Line Project (MCLP) contains 736 cancer cell lines profiled by RPPA [10]. Out of these cancer cell lines, 381 have both available RPPA and microarray data. RPPA measurements are available for 450 proteins and phospho-proteins, of which 311 genes can be matched to mRNAs measured in CCLE. In the RPPA data, 79 proteins are measured both at the basal expression and phosphorylation levels (Fig 1A). To our knowledge, this data set, although far from genome-wide at the protein level, represents the largest collection of cancer cell line data measured at the transcriptional, translational, and post-translational levels.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Pan-cancer cell line data from CCLE transcriptomic and reverse phase protein arrays (RPPA) cluster by tissue of origin and E-cadherin expression but not by prior metastasis classification.

(A) Overlap of available RPPA and CCLE data with regard to cancer cell lines (left), measured transcripts and proteins (middle), and proteins measured for both basal expression and phosphorylation levels (right). The colored areas indicate data used to calculate and compare Euclidean distances between cell lines. (B) t-SNE plots of overlapping cancer cell lines based on protein, transcript, and equally weighted combined data. Each point represents a cell line and is colored by the tissue of origin (top), E-cadherin expression (middle), or tumor classification (bottom). NS: not specified. (C) Comparing pairwise distances between all cell lines using a linear model at the mRNA or protein levels. The red points show the top-100 highest residuals of cell line pairs, and the blue points the top-100 lowest residuals. (D) Dendrograms of breast cancer cell lines mapped for transcriptomic and RPPA data. The leaves of the trees were arranged to minimize the number of crossing lines between leaves of the two trees. L1-5 represents clusters found within the luminal subtype of breast cancer cell lines.

https://doi.org/10.1371/journal.pcbi.1005911.g001

Transcriptional profiling of human tumor samples accurately predicts the tissue of origin for common cancer types [15]. This suggests that despite oncogenic transformation, cancer cells retain cellular identity and molecular features of their ancestral cell lineage, which is a key confounding factor in pan-cancer analyses. To assess how cancer cell lines relate based on transcript and protein expression, we visualized distances between cell lines using the t-Distributed Stochastic Neighbor Embedding (t-SNE) method [16]. As expected, the cell lines are clustered predominantly by their tissue of origin for both RPPA and transcriptomic data (Fig 1B, top). Cancer types with ill-defined or multiple clusters included breast and ovary as well as cell lines from the most common targets of metastasis: liver, lung, and bone. Nonetheless, most cell lines were correctly classified by nearest neighbor classification (S1 Fig), even when the t-SNE perplexity parameter was varied widely. In addition, independently from the t-SNE analysis, Gap statistics [17] from the average linkage hierarchical clustering at different tree cuts resulting in clusters of varying cardinalities displays similar grouping of cell lines. Furthermore, the inflection points in the number of clusters formed at different thresholds demonstrate that the cell lines are organized into distinct clusters based on expression vector similarity (S3 Fig). Next, we asked: to what extent the RPPA and transcriptomic data is concurrent. Although the cell line distances for protein and mRNA data were correlated (r = 0.58), they were surprisingly different for particular pairs of cell lines (Fig 1C). To quantify these differences and rank cell lines with the most characteristic protein or transcript signatures, we calculated the residuals of a linear regression between the protein and mRNA cell line distances. According to this model, we found that the 49 breast cancer cell lines had the largest distances at the protein vs. mRNA levels compared with other sets of cells from other tissues of origin, suggesting that the RPPA measurements better distinguish breast cancer subtypes. Interestingly, hierarchical clustering based on the RPPA data supports three luminal breast cancer subtypes compared with two subtypes identified by transcriptomic data (Fig 1D) here and elsewhere [12]. More broadly, combining data from microarray and RPPA data strengthened the cancer type clustering of cell lines, further suggesting that these measurements of global cellular states are complementary. Overall, clustering of cell lines by transcriptomic and RPPA data is consistent with some cancer types being well-defined and others spanning a wide spectrum of molecular states, while retaining few but important distinguishing differences at both the cancer type and subtype level.

To test the hypothesis that EMT governs the molecular states of cell lines across cancer types, we colored the z-scores of E-cadherin expression on the points on the t-SNE maps. For both transcripts and proteins, the cancer cell lines were globally organized by a gradient of E-cadherin expression (Fig 1B, middle). This organization indicates a central role for EMT in characterizing the molecular states of cancer cell lines. Most cancer types associated with common carcinomas had cell lines that spanned this E-cadherin gradient, with lung and breast cancer displaying the largest span. In contrast, cell lines from skin, bone, blood, and kidney were exclusively found in regions with low E-cadherin expression; whereas cell lines from pancreatic and large intestine cancers were found mostly in regions with high E-cadherin expression with only few cell lines expressing E-cadherin at low levels. To ensure the robustness of these findings, we ran independent t-SNE analyses by varying the perplexity parameter, which recapitulated both the E-cadherin gradient and the cancer type-specific clusters (S2 Fig). In comparison, principal component analysis (PCA) yielded less separation of cancer types and a less prominent gradient of E-cadherin expression (S4 Fig). Using cell line annotations from the Catalogue of Somatic Mutations in Cancer (COSMIC), we found no obvious association to whether the cell lines were derived from primary or metastatic tumors (Fig 1B, bottom). This suggests that the arrangement of cell-lines on the t-SNE plots, and thus global expression at the mRNA and protein levels, is dominated by tissue of origin much more than metastatic status. Nonetheless, we propose that collections of pan-cancer cell lines can be used to study aspects of EMT related to E-cadherin expression, which is also clearly bimodal (Fig 2A).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Bimodal protein expression and phosphorylation detected across cancer types associate with known oncogenic processes including EMT.

(A) Two-component Gaussian mixture model fit to E-cadherin protein expression. The lines indicate the probability density contribution from the low (-) and high (+) expression components. The histogram represents the RPPA measurements for the cell lines. (B) By comparing a two- versus one-component fit using the Bayesian Information Criterion (BIC), 260 out of 450 RPPA measurements supported bimodal expression. (C) Heat map of the posterior probabilities of each cell line belonging to the low (-, blue) or high (+, red) mixture component for the top-20 most bimodal proteins. The posterior probabilities can be thought of as soft assignments for the cell lines to low or high expression. Shannon entropy of the tissues assigned to low and high expression quantify the tissue diversity giving rise to the bimodal fits. (D) Overview of classification approach of proteins in terms of bimodality, tissue diversity (Shannon entropy), and frequency of cell lines assigned to the fitted distributions. (E) Significant GO terms for common bimodal proteins that were not found to be significant for non-bimodal proteins (p < 0.05, Benjamini-Hochberg).

https://doi.org/10.1371/journal.pcbi.1005911.g002

Bimodal protein expression and phosphorylation indicate oncogenic transitions

Oncogenesis is a multi-step process by which cells acquire cancerous traits, which often mimic physiological cellular processes such as embryogenesis [14]. Such processes are governed by molecular switches that turn on or off coordinated cellular programs. Hence, analyzing the bimodality of protein expression can potentially illuminate cellular states of oncogenesis. To evaluate this idea, we fit a univariate two-component Gaussian mixture model to each RPPA measurement using the expectation-maximization (EM) algorithm. We evaluated bimodality against unimodal distributions using the Bayesian Information Criterion (BIC). Out of the 450 antibody-based RPPA measurements, 260 were bimodal across 736 pan-cancer cell lines (Fig 2B, S1 Table). Among the most bimodal proteins were E-cadherin, Claudin7, and Rab25, all of which have been previously associated with EMT or MET [18]. However, because of the preponderance of tissue-specific signatures among pan-cancer cell lines, bimodal protein expression could more simply be explained by cell type-specific expression. For example, LCK was highly expressed only in a subset of blood cancer cell lines (Fig 2C) concordant with its specific roles in T cell development [19]. To account for residual effects of ancestral cell types, we quantified the tissue diversity of the cell lines assigned to the low- and high expression states by the Shannon entropy of the tissue distributions. We then excluded the lower tertile of the minimum tissue entropy of the low- and high expression states. This approach yielded 172 bimodal proteins and phosphosites (Fig 2B, S1 Table). Out of these, 90 had balanced bimodal distributions including E-cadherin, Claudin7, and Rab25, indicating common pan-cancer oncogenic switches, while 82 were classified as rare transitions (Fig 2D). This filtering and classification is likely prone to false positives due to other confounding factors such as different stages of the circadian clock at time of measurement. Compared to non-bimodal proteins, the proteins associated with the common switches were uniquely enriched for 107 Gene Ontology terms (p < 0.05, after Benjamini-Hochberg correction) many of which can be linked to metastasis and invasion (Fig 2E).

Coupled bimodality suggests transcript- and protein-based regulatory basis for oncogenic switches

To identify whether the observed bimodal protein expression across cancer cell lines correlate with transcriptional regulation, we evaluated the bimodality of matching transcripts from CCLE (Fig 3A). We then defined bimodal coupling coefficients between mRNA and protein measurements as the Spearman’s correlation between the posterior probabilities of the mixture model. Overall, 14.0% of proteins measured in MCLP had highly coupled (r_b > 0.5) bimodal expression of mRNA and protein. Slightly fewer proteins (10.8%) were uniquely bimodal only at the protein level, including important cancer-related proteins such as MEK1, mTOR, E2F1, TTF1, EIF4G, and JAB1. Hence, these proteins are bimodally expressed due to post-transcriptional regulatory mechanisms such as protein translation and degradation. In addition, we compared the bimodality of proteins and their phosphosites as measured by antibody binding in the RPPA data (Fig 3B). Here, we found weaker bimodal coupling, indicating that phosphosignaling leading to bimodal phosphorylation is mostly independent from basal protein expression. Interestingly, bimodal HER2 phosphorylation at Y1248 was moderately coupled to HER2 protein expression (r_b = 0.46), most likely due to autophosphorylation on increased dimerization at higher expression [20]. The bimodal EMT proteins E-cadherin, Claudin7, and Rab25 all had high bimodal mRNA-protein coupling (Fig 3C), confirming that these EMT switches are mostly determined by transcriptional regulation. Nonetheless, 30 cancer cell lines had high expression of the E-cadherin transcript but low protein expression (Fig 3D and 3E), suggesting that E-cadherin could be translationally or post-translationally controlled in some cellular contexts. Among these cell lines, 3 out of 4 CDH1 genotyped cell lines in COSMIC had either nonsense (MDA-MB-453 and HT115) or frameshift (MDA-MB-134-VI) mutations in CDH1, which validate our ability to identify effects on E-cadherin translation. Inactivating mutations in CDH1 are frequently observed in breast and gastric cancers with cancer type-specific mutational patterns and are associated with loss of cell-cell adhesion and increased cell motility [21]. The nature of the low E-cadherin protein expression in the other 26 cell lines remains unknown, but likely includes inactivating mutations and possibly translational or post-translational regulation.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Bimodal coupling between regulatory layers.

(A) Comparison of mRNA and protein bimodality colored by the coupling (Spearman’s correlation) between posterior probabilities of two-component Gaussian mixture models. The red percentage indicates the fraction of compared genes with coupled (r_b > 0.5) bimodalities at the transcript- and protein level. (B) Bimodal coupling of phosphosites and protein expression. (C) High confidence assignments (p < 0.1) to low or high expression for selection of bimodally coupled mRNA-protein pairs. (D) Scatter plot of E-cadherin mRNA and protein expression, indicating in red the 30 cell lines assigned to high mRNA but low protein expression (+–). r is Pearson’s correlation and r_b the bimodal coupling coefficient. (E) Tissue of origin and CDH1 (E-cadherin) mutational status of 30 cell lines with high E-cahderin mRNA but low protein expression. Of these cell lines, 3 out of 4 cell lines genotyped for CDH1 in COSMIC, all had mutations in the coding sequence. fs: frameshift, *: missense.

https://doi.org/10.1371/journal.pcbi.1005911.g003

Transcriptional mechanisms that determine molecular switches are regulated by upstream signaling, such as phosphorylation cascades, which leads to coordinated expression of multiple genes. To detect candidates for such signaling and further characterize the EMT-related states in cancer cell lines, we analyzed the network of bimodal coupling coefficients among bimodal protein and phosphosites associated with high tissue diversity. We first trimmed the protein network by including only significant bimodal coupling coefficients (FDR < 5%, Bonferroni) with |r_b| > 0.3. This yielded a network of 172 protein nodes connected by 507 edges, from which network communities were defined based on the leading non-negative eigenvector [22]. In total, we detected 8 protein communities that likely reflects shared underlying signaling or cellular events (Fig 4A). One community (EMT1) was clearly linked to EMT, containing E-cadherin, β-catenin, Fibronectin and Twist among several other EMT-related proteins (Fig 4A). E-cadherin was connected to EPPK1, INPP4B, Stathmin, Jagged1, UGT1A, and PDCD1L1, indicating that these might be involved in EMT. The strong bimodal coupling to EPPK1 could help explain why loss of E-cadherin is associated with migratory phenotypes and not just loss of cell-cell adherence; in mice keratinocytes, EPPK1 knockout cells exhibit faster migration and increased wound healing [23]. E-cadherin was also positively coupled to the phosphosite EGFR pY1068, which was in turn positively coupled to SRC pY416 and STAT3 pY705, suggesting a role for phosphorylation of these sites in EMT. Strikingly, all detected communities contained multiple proteins with known mechanisms linking them to EMT but also identified potentially undiscovered components (Fig 4A). The dispersion of these EMT-related proteins among the identified protein communities suggests that they are either part of separate biological processes, or that their involvement in EMT depends on cancer subtypes. Another intriguing possibility is that the multiple protein communities associated with EMT reflect partial cellular states in-between epithelial and mesenchymal phenotypes (Fig 4B). In support of this idea, P-cadherin has previously been suggested as a marker of metastable EMT states [24]. Here we find P-cadherin in the EMT2 community (Fig 4A). In addition, Claudin7 is highly coupled to E-cadherin (r_b = 0.70), but found in a separate community (EMT3), along with Rab25 and N-cadherin. Looking closer at this correlation, cell lines had high E-cadherin and low Claudin7 expression but not conversely (Fig 4C). Two other communities are identified, EMT4 and EMT5. The EMT4 cluster contains key cell-cycle transcription-factors such as FoxM1, Cyclin-B1, and Elk1, together with protein kinases that are known to positively regulate their activity, including PLK1, MAPK, and MEK1. Consequently, this cluster indicates changes in cell proliferation regulation. The EMT5 cluster contains a clique made of 3 protein kinases known to regulate the protein translation machinery: RICTOR, P70S6K, and PDK1; and S6 a key protein in the 40S ribosomal subunit. Hence, this cluster likely represents changes in protein translation activity related to overall cell growth. It should also be noted that highly studied proteins and phosphoproteins such pAKT, Cyclin D1, PTEN, and PKC are known to be central to many other pathways, not just to EMT. Hence, labeling all identified clusters as EMT clusters needs to be considered with such general functions in mind. Altogether, it is possible that the protein communities EMT1 and EMT3 may reflect a two-step transition (Fig 4B). In summary, the quintessential EMT marker E-cadherin was found centrally in a large protein and phosphosite network community with clear associations to known EMT markers. For these reasons, we focused subsequent analyses around the expression of E-cadherin, arguing that this approach reflects core aspects of EMT that are invariant across cancer types.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Proteins and phosphosites with coupled bimodality form network communities associated with EMT and intermediate transitions.

(A) Pan-cancer protein communities detected by Spearman’s correlation of the posterior probabilities of cell lines having low or high expression (|r_b| > 0.3). Only RPPA measurements associated with bimodal fits with high tissue diversity were included. Network communities were detected by calculating the leading non-negative eigenvector according to Newman’s method. Only edges within identified communities are shown, colored by the magnitude of the bimodal coupling coefficients. The size and color of the nodes represent the fitted mixing parameters from the Gaussian mixture models, quantifying whether underlying switches are common or rare in cancer cell lines. Each community was manually named according to plausible biological mechanisms by conducting a literature search for their protein members. Asterisks (*) indicate proteins with reported mechanisms linked to EMT. (B) Proposed interpretation of a two-step transition from the endothelial–E, to the mesenchymal–M states through two identified modules: EMT1 and EMT3. (C) Supporting protein expression data showing that Claudin7 and E-cadherin are correlated. Cell lines are colored by the tissue of origin (see Fig 3 for tissue labels).

https://doi.org/10.1371/journal.pcbi.1005911.g004

E-cadherin and bimodally coupled mRNA define genome-wide EMT profile

Because E-cadherin is primarily transcriptionally controlled, we next sought to characterize the coordinated transcriptional program associated with E-cadherin down-regulation. First, we defined an EMT signature based on the bimodal coupling (|r_b| > 0.5) between E-cadherin protein expression and transcriptomic measurements from CCLE, resulting in 239 transcripts—215 positively and 24 negatively coupled (Fig 5A, S2 Table). To our knowledge, these 239 genes include many novel epithelial and mesenchymal markers, while recovering many known EMT markers previously described (Fig 5C), for example, Axl which was reported for non-small cell lung carcinoma [25], or KPNA2 for ovarian carcinoma [26]. The preponderance of positively coupled transcripts suggests that the EMT signature is predominantly characterized by down-regulation of genes governing epithelial traits rather than by gain of mesenchymal traits. Nonetheless, the bimodal coupling coefficients were shifted towards negative values (Fig 5B) and we did find negatively coupled mesenchymal markers such as ZEB1/2 [4].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Pan-cancer bimodal coupling between E-cadherin protein expression and genome-wide transcripts defines an EMT signature, predicting EMT- and MET-inducing small-molecules.

(A) Top-25 transcripts in CCLE with the strongest positive and top 24 negative bimodal coupling coefficients (r_b) to E-cadherin protein expression. Red squares indicate previous EMT signature genes in non-small cell lung carcinoma published by Byers et al. [51]. To define an EMT signature, we considered transcripts with |r_b| > 0.5, resulting in 215 epithelial and 24 mesenchymal markers. (B) Distribution of bimodal coupling coefficients, showing that E-cadherin coupling coefficients are shifted towards negative values compared to all measured proteins. (C) Overlap of EMT signature with previously published transcriptomic EMT signatures. The ‘mesenchymal’ bar plot is for the inversely correlated (coupled) genes and the ‘epithelial’ for the positively correlated genes. (D-E) Gene set enrichment analysis of epithelial part of the EMT signature. The TF enrichment analysis used ChIP-seq data to predict TFs involved in the regulation of the epithelial genes. The pie charts indicate the fraction of the signature genes associated with significantly enriched terms or TFs. (F) Small-molecule perturbations predicted to induce EMT and MET based on L1000 cell line data and the L1000CDS² method. The top-50 signatures are shown with results from multiple cell lines or concentrations aggregated by boxplots. PK: protein kinase.

https://doi.org/10.1371/journal.pcbi.1005911.g005

To further characterize the EMT signature, we performed enrichment analysis on the epithelial markers (Fig 5B). Enrichment analysis [27, 28] for Gene Ontology (GO) cellular components and biological processes clearly demonstrated epithelial phenotypes (Fig 5D). We also found enrichment for localization to the perinuclear region, which is a cytosolic region next to the nuclear envelope with largely unknown composition and biological function. This suggests that the epithelial markers can be used to prioritize spatial cellular regions not widely considered to be affected by EMT such as the perinuclear region, where components of endocytosis aggeregate, although it is well established that endocytosis is central to cell migration. Enrichment for TF binding, using aggregated results from ChIP-seq studies [29], identified SOX2, SMAD2-4, TP63, GATA3-4, and GATA6, which likely act to down-regulate epithelial genes during EMT (Fig 5E). The identified enriched TFs OCT4, SOX2, NANOG, KLF4, and ESRRB are all known to be essential for maintaining pluripotency of human and mouse embryonic stem cells [30]. These TFs bind to super-enhancer regions and through the Mediator complex [31]. Therefore, large parts of the observed transcriptional bimodality could be explained by TF co-operation at super-enhancers resulting in switch-like regulation at numerous genomic loci. At the CDH1 loci, ENCODE ChIP-seq data supports the involvement of super-enhancers since the loci is marked by high H3K27ac correlated with E-cadherin expression (S5 Fig). Previously, super-enhancers have been proposed to control partial EMT through the putative master regulator TFs ETS2, HNF4A, and JUNB [32], the first two of which we also identified through the TF enrichment analyses. Taken together, pan-cancer bimodality uncovers oncogenic states and regulatory mechanisms of EMT and MET.

EMT and MET can potentially be induced by small molecules

Gene expression-based, high-throughput screening is a promising approach to identifying small-molecule candidates that can reverse or mimic changes in expression observed in transition to a disease state [33]. To detect small molecules that would maximally push cells toward the EMT or MET expression state, we queried the EMT signature against signatures from ~20,000 small-molecule perturbations of ~50 human cell lines generated by the library of network-based cellular signatures (LINCS) project L1000 dataset [34]. We searched for small molecules that down-regulate epithelial genes and up-regulate mesenchymal genes, resulting in candidate EMT inducers (Fig 5E). Small molecules with the opposite effects were interpreted as MET inducers. Strikingly, most small molecules predicted to induce EMT were HDAC inhibitors, whereas most small molecules predicted to induce MET were kinase inhibitors. The identified HDAC inhibitor Trichostatin A has been shown to induce EMT in prostate cancer cells through modification of H3 near promoters of EMT-related genes [35]. Of the candidate MET inducers, Selumetinib, Trametinib, and PD-0325901 are thought to inhibit MEK, while Saracatinib and Dasatinib to inhibit SRC among other kinome targets. In agreement with these findings, a prior high-content chemical screen aimed at identifying inhibitors of EMT has predominantly identified other similar kinase inhibitors based on cell growth and migration assays [36]. Hence, in summary, caution should be placed in utilizing HDAC inhibitors as therapeutics due to their putative potential to enhance EMT as predicted by chemogenomics screening.

Causal protein and phosphorylation models identify drivers of cancer signaling and progression

The bimodal coupling model we implemented to analyze EMT is essentially correlative and hence not causal. However, establishing causal interactions based on RPPA data is challenging without time-series or direct perturbation data such as gene knock-downs or knock-outs [37]. With sufficient sample size and coverage of diverse cell lines, it is in principle possible to identify causal, regulatory interactions between measured signaling components. Despite not satisfying the observation that cell signaling regulatory networks contain feedback loops [38], learning Bayesian network learning algorithms can be applied to construct causal models of cellular regulatory networks, including cell signaling networks, from observational data [39–41]. To infer causal relationships among proteins and phosphosites measured by RPPA, we used a Fast Greedy Search algorithm to estimate a Bayesian network over all 450 RPPA measurements (Fig 6). Based on the resulting directed acyclic graph, we calculated betweenness centrality, subgraph centrality, in-degree, and out-degree for each analyte. Bimodal phosphosites overall had higher betweenness centrality (p = 0.036, t-test). By considering measures of network influence, several proteins and phosphosites were identified as promising candidate drivers (Fig 6A and 6B). In particular, SRC pY416 had the highest out-degree. This phosphosite is known to be highly predictive of patient survival [42].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Bayesian networks of proteins and phosphosites inferred from pan-cancer cell lines identify drivers of EMT and correlate to tumor networks.

All Bayesian network structures were inferred by a Fast Greedy Search algorithm. (A) Network centrality statistics of the directed causal graph over all measured proteins pertaining to the influence of proteins on cancer signaling. (B) In and out degree distributions of proteins and phosphosites from the inferred network. (C) Causal neighborhood (1st neighbors) of EMT markers E-cadherin, Rab25, and Claudin7. Tissue-specific correlations in support for each edge are shown as bars along the edges. The layout was determined using the hierarchical Sugiyama algorithm with all edges oriented downwards. (D) Network comparisons between cell line and tumor data. (E) Distribution of average connectivity in bootstrapped Bayesian networks. (F) Distribution of networks of bimodal coupling coefficients.

https://doi.org/10.1371/journal.pcbi.1005911.g006

Furthermore, we analyzed the network neighborhood of E-cadherin, Claudin7, and Rab25, which spanned the proposed two-step transition: EMT3 followed by EMT1 (Fig 6C). Using a hierarchical network layout algorithm, E-cadherin was found downstream of Claudin7 and Rab25 concurrent with EMT3 preceding EMT1 in cancer cell lines. In addition, several proteins and phosphosites upstream of the EMT markers were plausible oncogenic drivers for specific cancer types, supported by prior reports. For example, MACC1 is associated with pancreatic EMT and metastasis [43], which the analysis found for both pancreatic and pleural cancer cell lines, while suggesting the opposite effect in lung, endometrium, and upper digestive tract cancers. Lastly, we also found LKB1, CHK1, and Stathmin upstream of EMT markers. The inactivation of LKB1, which is frequently mutated in lung adenocarcinomas, induces EMT in lung cancer cells through activation of ZEB1 [44], whereas CHK1 mediates DNA damage response as part of EMT by stabilizing ZEB1 [45]. Inhibiting the microtubule destabilizer Stathmin impedes EMT by increased microtubule formation [46]. In conclusion, inferred Bayesian protein networks based on pan-cancer cell lines can potentially identify key drivers of EMT.

To validate the causal models, we carried out bootstrapping for 200 iterations and considered the average network for both cell line and tumor data (Supporting Information, S3 Table, Fig 6E). Although the reproducibility of particular edges in the ensemble of Bayesian networks was relatively low for the cell line data, the connection from Claudin7 to Rab25 was present in 48% of the networks, and in all the networks inferred from tumor samples. In contrast, the statistical reproducibility of the tumor networks was higher, most likely due to the larger sample size (n = 3,161). Overall, the average Bayesian networks were significantly correlated between cell line and tumor data (Fig 6D). The bimodal coupling coefficients, however, were lower in the tumor data, indicating that bimodal expression is less pervasive in tumors compared to cell lines. This result might indicate that tumors, more-so than cell lines, contain mixtures of cell types that are in multiple cellular states. Overall, the learning Bayesian Network strategy employed here is exploratory and needs further evaluation, parameter tuning and validation.

Matching RPPA and CCLE cell line data

The RPPA data for 736 cancer cell lines were generated at the MD Anderson Cancer Center. The selection criteria of the 474 measured proteins were based on the aim to cover known cancer-related signaling pathways. We excluded antibodies with missing values across cell lines by requiring that each RPPA measurement is present in at least 40 cell lines. This resulted in a dataset with 450 antibody-based measurements. The CCLE mRNA data and cell line annotations of 1,037 cancer cell lines were retrieved from the CCLE portal at: https://portals.broadinstitute.org/ccle. We used the gene-centric RMA-normalized data.

Cell line distances and clustering

For all methods relying on geometric distances, Euclidean distances were computed considering only pairwise complete features. Sparse RPPA measurements were excluded, requiring that each protein is measured in at least 100 cell lines, which resulted in the inclusion of 263 protein measurements. To reduce the dimensionality of the RPPA and mRNA data, we used t-SNE implemented in the R package ‘tsne’ with perplexity value of 30 and at 5,000 iterations, and all other arguments at their default values [16]. Only cell lines with available RPPA and mRNA data were included. For the combined RPPA and CCLE mRNA embedding, the distance matrices for each data set were weighted by the sum of all distances. In this way, each data type contributed equally to the combined analysis. To more rigorously assess the number of clusters supported by the RPPA and CCLE data sets, we calculated the Gap statistic [17] from the average linkage hierarchical clustering at tree cuts resulting in clusters of varying cardinalities. PCA was performed using the R package ‘pcaMethods’ from data that were centered and scaled to unit variance, while imputing missing values with the ‘svdImpute’ method. Furthermore, we analyzed patterns in the classification and misclassification of the tissue of origin for the RPPA and mRNA data using 3-nearest-neighbor classification according to a leave-one-out cross-validation scheme. Linear regression was carried out between the RPPA and CCLE mRNA distance matrices with the RPPA distances considered the target variable. The residuals of the regression thus quantify the deviation from expected distance for RPPA data for each pairwise cell line distance. To compare the clustering of breast cancer cell lines, we computed tanglegrams using the dendextend R package. The tanglegram method uses a random search to rotate tree nodes minimizing the overlap of lines drawn between leaves of two trees.

Statistical models of bimodal expression

We fit univariate two-component Gaussian distributions using the expectation-maximization (EM) algorithm implemented in the ‘mixtools’ R package with default parameters. To compare the fitted distribution to unimodal Gaussian distributions, we calculated the difference between the Bayesian Information Criterion (BIC). The data were determined to be bimodal if the BIC difference was larger than 2. Based on the fitted Gaussian mixture model, we calculated, using Bayes’ theorem, the posterior probabilities of measurements being generated from the high expression component. Note that the probability of belonging to the low component is 1-p. To estimate the tissue diversity of each bimodal fit, we first calculated the frequencies of tissues assigned to the low (p < 0.5) and the high (p > = 0.5) component. We then calculated the Shannon entropy of the tissue distributions associated with the low- and the high components. The bimodal RPPA measurements were classified into groups of low, medium, and high tissue diversity by the tertiles of the minimum tissue entropy associated with low- and high expression. The bimodal expression was considered common if the fitted mixture coefficients were above 1/4 and rare if below. Based on the posterior probabilities of the bimodal fits associated with high tissue diversity, we calculated a network of bimodal coupling coefficients defined as Spearman’s correlations between the posterior probabilities. To detect robust communities in this network, we set a cutoff of |r_b| > 0.3 and calculated the leading non-negative eigenvectors using the igraph R package. The network was visualized in Cytoscape with node size proportional to the mixing parameter of the two-component Gaussian fit and with edge coloring based on the coupling coefficients.

EMT signature enrichment and queries

The coupling coefficients between the E-cadherin RPPA measurements and matched CCLE transcript data were used to define an EMT signature (r_b > 0.5). Enrichment analysis was performed with Enrichr [28]. L1000CDS2 was used to query small molecules as potential inducers or reversers of EMT [34]. We summarized the EMT and MET small-molecule predictions by reporting the top-50 small molecules identified using boxplots to aggregate small molecules with multiple experimental conditions such as cell lines, dosage, or timing.

Causal modeling

The causal models of the RPPA measurements across cancer cell lines were inferred using the Fast Greedy Search algorithm [49] implemented by the BD2K Center for Causal Discovery [50]. We used the rcausal R package version 0.99.5 to run the Java implementation with penalty discount 4 and depth 3. To visualize causal neighborhoods, we computed graph cuts and rendered the subnetwork in R using a Sugiyama layout of the directed acyclic graph. The tissue-specific correlations were layered on top of the edges as histograms. To estimate the robustness of the resulting causal network, we ran the algorithm several times in a bootstrap scheme (M = 200) by sampling with replacement.