Proteogenomic Network Analysis of Context-Specific KRAS Signaling In Mouse-to-Human Cross-Species Translation

SUMMARY

The highest frequencies of KRAS mutations occur in colorectal carcinoma (CRC) and pancreatic ductal adenocarcinoma (PDAC). The ability to target downstream pathways mediating KRAS oncogenicity is limited by an incomplete understanding of the contextual cues modulating the signaling output of activated K-RAS. We performed mass spectrometry on mouse tissues expressing wild-type or mutant Kras to determine how tissue context and genetic background modulate oncogenic signaling. Mutant Kras dramatically altered the proteomes and phosphoproteomes of pre-neoplastic and neoplastic colons and pancreases in a context-specific manner. We developed an approach to statistically humanize the mouse networks with data from human cancer and identified genes within the humanized CRC and PDAC networks synthetically lethal with mutant KRAS. Our studies demonstrate the context-dependent plasticity of oncogenic signaling, identify non-canonical mediators of KRAS oncogenicity within the KRAS-regulated signaling network, and demonstrate how statistical integration of mouse and human datasets can reveal cross-species therapeutic insights.

Graphical Abstract

To better understand how tissue context contributes to cancer-specific signaling by the same protein, Brubaker and Paulo et al. performed mass spectrometry based analysis of total and phosphoproteomic datasets from four murine tissues with mutationally activated K-RAS and developed computational techniques to translate mouse signaling insights to patient tumors.

INTRODUCTION

KRAS is the most frequently mutated oncogene (Cerami et al., 2012; Forbes et al., 2017). Missense substitutions in the KRAS gene, which encodes a monomeric GTPase, activate K-RAS protein by inhibiting its ability to hydrolyze GTP or by promoting the rapid exchange of GDP for GTP (Haigis, 2017). Activating KRAS mutations are most common in colorectal carcinoma (CRC) and pancreatic ductal adenocarcinoma (PDAC), with substitution of glycine 12 for aspartic acid (G12D) being most common (Haigis, 2017). Activated K-RAS cooperates with mutations in specific tumor suppressor genes to promote progression beyond a pre-neoplastic state in these cancers (Haigis et al., 2008; Hingorani et al., 2003). In PDAC, KRAS^G12D is associated with shorter overall survival and poorer prognosis than other forms of KRAS (Bournet et al., 2016). In CRC, KRAS mutations are associated with anti-EGFR therapy resistance and KRAS^G12D is associated with reduced overall survival (Knickelbein and Zhang, 2015).

K-RAS has been a high value therapeutic target since the discovery of its role in cancer, but efforts to directly target it have been largely unsuccessful. The failure to inhibit K-RAS has motivated the search for alternative therapeutic strategies targeting its downstream signaling effectors (Simanshu et al., 2017; Velho and Haigis, 2011). However, because activated K-RAS engages multiple downstream pathways, single agents targeting canonical effectors are not clinically effective (Blumenschein et al., 2015). Since mutant KRAS effector utilization is likely context-dependent, a better understanding of context-dependent signaling may help identify therapeutically relevant variations in mutant KRAS effector pathways.

We characterized the contextual cues that impact oncogenic signaling by measuring signaling responses of primary mouse tissues to mutational activation of Kras. We profiled colonic epithelium, colonic tumors, pancreas, and pancreatic tumors from genetically defined mouse models (+/− mutant Kras) with multiplexed quantitative mass spectrometry (MS). Analysis of total and phosphoproteomic datasets revealed tissue-specific and cross-tissue signaling dysregulation driven by mutant Kras. When translating insights from mouse tumor models to patients, the added complexity of human tumors with multiple mutations potentially obscures the effect of specific oncogenic events. We developed an approach to statistically humanize proteomic insights from mouse models by integration with total proteomics and mutation data from human CRC and PDAC cohorts. We then compared these humanized networks with genetic screening data from human cancer cell lines to identify genes with KRAS^G12D-specific synthetic lethality. This integrated analysis of mouse and human datasets provides a deep characterization of context-dependent signaling by mutant KRAS.

RESULTS

Generating and translating a compendium of KRAS^G12D driven signaling

Our goals were to characterize the downstream signaling of KRAS mutant tissues by generating proteomic and phosphoproteomic datasets from mouse tissues +/− mutant Kras and to translate this characterization to humans. Clinically, CRC and PDAC both present with mutations in a tumor suppressor gene (APC for CRC, TP53 for PDAC) as well as KRAS mutation. We mimicked this profile by incorporating the combination of a Kras^G12D mutation with Apc in mouse colonic tumors and Tp53 in mouse pancreatic tumors. We crossed mice with the Kras^LSL-G12D allele to different Cre drivers and conditional alleles of Apc and Tp53 to create a variety of contexts to evaluate K-RAS signaling. We performed multiplexed MS with tandem mass tag (TMT) labeling to generate proteomic and phosphoproteomic datasets for colonic epithelium (CE), colonic tumors (CT), whole pancreas (WP), and pancreatic tumors (PT) (Figure 1A, Table S1). Within each tissue, Kras^G12D mutants were compared to Kras^WT from the same genetic background. Datasets from each tissue were analyzed for differential regulation, pathway and kinase enrichment, and the results integrated in a systemic view of K-RAS signaling.

Figure 1. Overview of experimental and computational analyses.

A. Tissues from mouse models with Kras^G12D mutations and appropriate controls were subjected to TMT labeled LC-MS/MS analysis to generate total and phosphoproteomics datasets. Differential regulation, pathway enrichment and kinase enrichment were performed on the individual total and phosphoproteomics datasets. Kinase-pathway network analysis was performed on combined total and phosphoproteomics datasets. B. Overview of the Statistical Humanization methodology for integrating proteomic datasets from mouse models with patient tumor datasets. C. Hierarchical clustering of the 4,085 proteins measured across tissues. D. Hierarchical clustering of the 2,357 phosphosites measured across tissues (heatmaps normalized by z-score). (See also Figures S1 and S2, Table S1).

Translating mouse proteomic dysregulation to patient tumors with a KRAS^G12D mutation is of high clinical importance. We developed an inter-species translation technique termed Statistical Humanization (Figure 1B), motivated by the analogous challenge of evaluating a statistical hypothesis through integration of p-values from differently powered studies. While it is often infeasible to pool raw data from multiple studies, information can be combined considering study power by a weighted Z-test (Zaykin, 2011). In mouse-to-human translation, pooling of raw, multi-omic datasets from different species and sequencing platforms is similarly not feasible. Therefore, statistical meta-analysis techniques aggregating evidence around a hypothesis may provide an avenue to address inter-species, trans-omic discrepancies.

Our translational objective was to assess whether a protein or phosphorylation signaling event was associated with KRAS^G12D-driven signaling, accounting for differences in mutational burden between mouse and human tumors. We used the weighted Z-test to statistically humanize the mouse insights and prioritize therapeutic targets by integrating the findings of the mouse models with The Cancer Genome Atlas (TCGA) CRC and PDAC proteomics and mutation profiling data (Cancer Genome Atlas, 2012; Cancer Genome Atlas Research Network, 2017). We determined each variable’s significance by a weighted combination of p-values from mouse and human analyses, with weights proportional to the dataset sample sizes, giving greater weight to human evidence from the larger TCGA cohorts.

Principal component analysis (PCA) of each mouse dataset demonstrated that all mice within a tissue separated by Kras genotype (Figure S1–S2). Hierarchical clustering of the commonly measured proteins (4,085) and phosphosites (2,357) revealed context-specific responses to Kras^G12D (Figure 1C–D). The Tp53 mutation in the pancreatic tumors appeared to determine the clustering, but differences in activated cell types within the pancreatic tumor datasets are likely responsible. The Tp53-Kras^WT mice had no tumors and thus had fewer activated stromal cells than the Tp53-Kras^G12D mice.

The KRAS-regulated protein repertoire is context-dependent

We identified proteins and phosphosites differentially regulated between Kras mutant samples and controls in each tissue context (Figure 2A–2B, Tables S2 and S3) (See STAR Methods). Most phosphosite changes did not correspond to changes in protein abundance, suggesting that most differential phosphorylation was functional (Figure 2C). 93 proteins and 43 phosphosites were differentially regulated in response to Kras^G12D mutation across tissues. We mapped these identifiers onto the STRING protein-protein interaction network (PPIN), filtering for validated interactions to compile a “core K-RAS network” (Homo sapiens, Confidence: 0.20) (Figure 2D) (Szklarczyk et al., 2015). This included known regulators of K-RAS (NF1 and RASA1), known K-RAS effectors (MAPK3/ERK1 and RALA), and protein families not previously associated with K-RAS activation including up-regulated annexin family members (ANXA3, ANXA8, ANXA11) and down-regulated aldehyde dehydrogenases (ALDH1L1 and ALDH1L2) and heat shock proteins (DNAJC1 and DNAJC3) (Figure 2D). PANTHER pathway enrichment of the species commonly regulated by Kras^G12D identified Insulin/IGF and MAP kinase (Fisher Exact p = 5.90*10⁻⁵, FDR q = 4.81*10⁻³), EGFR (Fisher Exact p = 1.67*10⁻⁵, FDR q = 2.72*10⁻³), and integrin signaling (Fisher Exact p = 7.62*10⁻⁴, FDR q = 4.14*10⁻²) as conserved Kras^G12D responsive pathways (Thomas et al., 2003; Thomas et al., 2006).

Figure 2. Differential regulation analysis of proteomic datasets.

A. Differential regulation analysis of total proteomic datasets by Wilcoxon Mann-Whitney (WMW) test with Benjamini Hochberg False Discovery Rate Correction (FDR) (p < 0.05, q < 0.25). “Top candidates” from whole pancreas (p < 0.10, q < 0.25) were selected for comparison since no proteins reached significance. B. Differential regulation analysis of phosphoproteomic datasets (WMW p < 0.05, FDR q < 0.25). “Top candidates” from whole pancreas (p < 0.10, q < 0.25) were selected for comparison since no phosphosites reached significance C. Differentially abundant and phosphorylated proteins. D. STRING protein-protein interaction network (PPIN) (Homo sapiens, Confidence: 0.20)) of validated interactions between 93 differentially regulated (circles) and 43 differentially phosphorylated (triangles) proteins regulated by Kras^G12D in all tissue contexts. Disconnected proteins were removed. E. STRING PPIN (Homo sapiens, Confidence: 0.80) of validated interactions between proteins differentially regulated or phosphorylated by Kras^G12D showing differential activation in pancreatic and colonic mouse tumors. F. PANTHER protein class enrichment of proteins and phosphopeptides differentially regulated by Kras^G12D in all contexts. PANTHER protein class enrichment of proteins and phosphosites differentially regulated in colon or pancreas tissues and oncogenic or healthy tissues. (See also Tables S2–S4).

We mapped the 506 proteins and phosphosites differentially regulated in colonic or pancreatic tumors onto the STRING PPIN, filtering for validated interactions (Homo sapiens, Confidence: 0.80), to characterize Kras^G12D differential oncogenic signaling (Figure 2E) (Szklarczyk et al., 2015). The highest confidence proteins and interactions formed 20 subnetworks, most down-regulated in colonic tumors and up-regulated in pancreatic tumors (Figure 2E). The largest subnetwork, containing proteins up-regulated by Kras^G12D in pancreatic tumors, contained several proteasome subunits and the cell cycle regulator RB1. RB1 phosphorylation has been implicated in CRC progression, but that analysis did not stratify by KRAS alleles, a factor that our previous work shows can drive distinct downstream signaling (Poulin et al., 2019; Vasaikar et al., 2019). Our results suggest that RB1 phosphorylation has KRAS-allele-specific variability within tumors.

We employed PANTHER protein class enrichment to assess the diversity of the Kras^G12D regulated protein repertoire (Figure 2F, Table S4) (Ashburner et al., 2000; Thomas et al., 2003; Thomas et al., 2006). Though protein class distributions suggest that signaling processes ubiquitously regulated by Kras^G12D fall into distinct functional categories at the total and phosphoproteomic level, cross-tissue Kras^G12D driven pathways contain both protein and phosphosite evidence. Though the particular Kras^G12D-regulated proteins and phosphosites were context-specific, functional classes were more similar (Figure 2F). The top 7 total protein functional classes were the same in colonic and pancreatic tissues, the largest group coming from populations comprising less than 4% of total protein abundance (Figure 2F). Across tissues, K-RAS activation resulted in nearly identical distributions of differentially regulated phosphoproteomic functional classes (Figure 2F).

The rare populations included signaling molecules, receptors, defense/immunity proteins, transcription factors, calcium binding, membrane traffic, and transporter proteins. Changes in the types of proteins in the “Rare Populations” category suggest differences in the mode of regulation by Kras^G12D in different biological contexts. For example, a shift occurs in the fraction of transcription factors with differences in total protein amount and phosphorylation (Figure 2F). The shift in proportion of transcription factors at the total protein level to nearly 10% of all differentially phosphorylated proteins suggests that the primary mode of Kras^G12D regulation of transcription factors is by modifying phosphorylation state. Distributions of protein classes regulated by Kras^G12D were similar between normal and tumor contexts, with nucleic acid binding, mixed rare populations, and transcription factor proteins accounting for more than 65 percent of all differential phosphorylation (Figure 2F). This suggests that activated K-RAS regulates similar functional output despite context-specificity in differentially regulated species.

The mechanism of KRAS^G12D signaling varies between contexts

The large number of context-specific differentially regulated species suggests distinct Kras^G12D effector signaling. We performed gene set enrichment analysis (GSEA) using the Hallmarks gene sets on the total proteomics data and kinase enrichment analysis on the phosphoproteomics data to assess the signaling consequences of K-RAS activation (Liberzon et al., 2015; Lyons et al., 2018; Subramanian et al., 2005). Kinase enrichment is analogous to GSEA, with kinase enrichment inferred by differential phosphorylation of known substrates (Hornbeck et al., 2015). In the total proteomics data, Kras^G12D drove significant enrichment of 12 pathways in colonic epithelium, 16 in colonic tumors, 18 in bulk pancreas, and 30 in pancreatic tumors, with the majority of pathway enrichment events being tissue or tumor-specific (Figure 3A–3B). Across tissues, Kras^G12D associated with enrichment of Angiogenesis, Epithelial-Mesenchymal Transition (EMT), Interferon Alpha Response, Interferon Gamma Response, and KRAS Signaling Up (a set of genes transcriptionally up-regulated by activated KRAS) pathways (Figure 3A). These common pathways were never coordinately dysregulated. An example of this was EMT signaling, down-regulated in colonic epithelium and up-regulated in the other contexts (Figure 3A).

Figure 3. Pathway and kinase enrichment analysis.

A. GSEA of total proteomic data using the Hallmark Gene Sets with significant pathways (q< 0.05) marked with a (*) and plotted by normalized enrichment score. B. Venn diagram summary of differentially regulated pathways across contexts. C. Kinase enrichment analysis of phosphoproteomic datasets with significant kinases (q< 0.05) marked with a (*) and plotted by normalized enrichment score. D. Volcano plots of kinase enrichment results.

Tissue-specific Kras^G12D-signaling was also apparent in the kinase enrichment analysis (Figure 3C–3D). The tissues with significantly enriched kinases were colonic epithelium and colonic tumors, with CDK2 the only commonly dysregulated kinase, but activated by Kras^G12D in epithelium and inactivated in tumors (Figure 3C). Though no kinases were significantly enriched (q < 0.05) in pancreatic tissues, a more permissive threshold (q < 0.25) revealed enrichment of CDK2 (p < 0.001, q = 0.232) in pancreatic tumors and AMPKA1 (p = 0.012, q = 0.151) and DYRK2 (p = 0.015, q = 0.179) in whole pancreas. These observations suggest that Kras primarily modifies protein abundance in pancreas and has broader effects on total and phospho-signaling in colon.

Kinase-pathway network analysis links total and phosphoproteomic signaling insights

Analysis of total and phosphoproteomics data in isolation likely overlooks more complex signaling effects of Kras^G12D mutation. We developed an integrative framework, Kinase-Pathway Network Analysis (KiPNA), to combine insights from kinase and pathway enrichment to characterize the combined total and phosphoproteomic signaling network regulated by Kras^G12D. KiPNA links proteins and phosphosites that drive enrichment (leading edge subset) to significant pathways and kinases in a network integrating both analyses. KiPNA encodes biological information that can be mined in subnetworks to reveal functional pathway integration of kinase signaling cascades.

We constructed KiPNA networks for colon tumor and epithelium, tissues with both enriched kinases and pathways (Figure 4A–4B). Activated WNT signaling from loss of Apc results in a Kras^G12D-regulated network of up and down-regulated pathways in colonic tumors, rather than solely suppressive signaling as in colonic epithelium. In colonic epithelium, Kras^G12D drives down-regulation of several pathways that are connected to leading edge phosphosites of enriched kinases. In colonic tumors, we found up-(TGFB, EMT, IL2-STAT5 Signaling) and down-regulated (Oxidative Phosphorylation, Fatty Acid Metabolism) pathways connected to enriched kinases (Figure 4B). One differentially regulated pathway was “KRAS Signaling Up,” activated by Kras^G12D in tumors and suppressed in epithelium. KiPNA network topology shows how tissue context can result in many differences in signaling by the same oncoprotein.

Figure 4. Kinase-Pathway Network Analysis (KiPNA) of colonic datasets.

A. Colonic epithelial KiPNA network. B. Colonic tumor KiPNA network. C. Colonic tumor EGFR kinase-pathway subnetwork. D. CDK2 kinase-pathway subnetworks in colonic tumors and epithelium. (See also Figure S3).

Since kinase inhibition is one of the principle therapeutic strategies in cancer, a better understanding of the interaction between kinase inhibition and the broader cellular network may suggest therapeutic strategies. Therefore, a biologically motivated subnetwork motif in the KiPNA framework is the set of subnetworks linking an enriched kinase to enriched pathways. These subnetworks are mined by a three step directed walk from an enriched kinase to its associated pathways (See STAR Methods). These subnetworks include proteins, kinases, and substrates functionally linked to Kras^G12D driven signaling and suggest hypotheses of how mutant Kras rewires cellular signaling.

We mined kinase-pathway subnetworks from colonic epithelium and tumor KiPNAs, finding Kras^G12D-associated rewiring of EGFR and CDK2 signaling (Figure 4C–4D, Figure S3). The colonic tumor EGFR kinase-pathway subnetwork revealed EGFR activity and abundance increased by Kras^G12D (Figure 4C). In the colonic epithelium subnetworks, EGFR was a leading edge protein or substrate of other enriched kinases, but did not have increased activity (Figure S3). Therefore, EGFR signaling in colonic tumors is likely due to the combination of Apc and Kras mutations.

CDK2 also changes activity in colonic tissues in the presence of a Kras^G12D mutation, but the status of Apc determines directionality (Figure 4D). In colonic tumors, Kras^G12D resulted in decreased CDK2 activity and dephosphorylation of MARCKS (Figure 4D). The CDK2 subnetwork linked MARCKS dephosphorylation to activation of TNFA Signaling via NFKB, consistent with MARCKS function as a negative regulator of inflammatory signaling (Mancek-Keber et al., 2012). This shows how Kras^G12D results in down-regulation of an anti-inflammatory regulator to promote oncogenic signaling by TNFA and NFKB in colonic tumors. In colonic epithelium, CDK2 activity increased and connected to cell cycle checkpoint signaling (E2F Targets, G2M Checkpoint pathways). In CRC, CDK2 phosphorylation of RB1 associates with increased proliferation and reduced apoptosis, making CDK2 a potential therapeutic target (Vasaikar et al., 2019). Our results suggest that CDK2 signaling is KRAS-allele specific and targeting RB1 through inhibiting CDK2 may not be effective in KRAS^G12D mutant CRC.

Statistical humanization translates insights from mouse models to patients

The extent to which signaling dysregulation in mice translates to KRAS^G12D mutant patient tumors is unknown. The TCGA characterized proteomic and genomic features of primary CRCs and PDACs (Cancer Genome Atlas, 2012; Cancer Genome Atlas Research Network, 2017). The mutation data can be used to prioritize genes with KRAS^G12D-co-occuring mutations as features of interest and help account for the heterogeneity of tumors. The proteomics data can be similarly useful, however there are limitations. The PDAC samples were analyzed by low coverage reverse-phase protein arrays (RPPA) and the CRC analysis performed using a label-free approach less quantitative than our TMT approach. Since TCGA lacks phosphoproteomics, the mouse and human data provide complementary characterizations of the effects of KRAS^G12D. In order to obtain better characterization of KRAS^G12D in humans, we used statistical humanization to aggregate proteomic and mutation information from mouse and human contexts into a network of KRAS^G12D-regulated signaling. Protein, kinase, and pathway significance for KRAS^G12D-driven signaling dysregulation in humans was determined by a weighted combination of p-values from mouse proteomics, mouse phosphoproteomics, human proteomics, and human mutation sequencing data, with p-value weights proportional to dataset sample sizes (see STAR Methods). We then constructed predicted human KiPNA networks to integrate the statistical humanization insights into a biologically interpretable framework.

In colonic tumors, the 2,093 nominally significant (p < 0.05) mouse proteins were prioritized to 458 proteins, including APC, PIK3CA, and CDK2 (Figure 5A, Table S5). These proteins were enriched for pathways including N-acetylglucosamine metabolism, Insulin/IGF pathway-protein kinase B signaling cascade, PI3 kinase pathway, and p53 pathway (Table S6). Two KRAS^G12D-co-mutated kinases, EPHA2 and CDK2, were prioritized for relevance in human CRC, with EPHA2 only identified after statistical humanization. We used phosphosites from mouse kinase enrichment to estimate human kinase-pathway regulation (Figure 5B). The predicted human CRC KiPNA network included EPHA2 signaling to Apical Junction and Xenobiotic Metabolism pathways, connections not in the mouse KiPNA network (Figure 5B). The sparser up-regulated region of the humanized CRC KiPNA included TNFA Signaling via NFKB, TGFB, EMT, Apical Junction, and Mitotic Spindle signaling linked to CDK2 and EPHA2.

Figure 5. Statistical humanization of mouse tumor proteomics.

A. Statistical humanization of mouse colonic tumor proteomic insights for human CRC. B. Predicted kinase-pathway network for human KRAS^G12D mutant CRC. C. Statistical humanization of mouse pancreatic tumor proteomic insights for human PDAC. D. Predicted kinase-pathway network for human KRAS^G12D mutant PDAC. Proteins and pathways are colored by fold change in human samples (KRAS^G12D vs. others), phosphosites are colored by mouse fold change (KRAS^G12D tumors vs. controls). (See also Table S5–S6 and Figure S5).

Statistical humanization of the mouse pancreatic tumor dataset prioritized 356 proteins potentially relevant to KRAS^G12D–driven signaling in PDAC and confirmed our finding of no kinase enrichment in pancreatic tumors (Figure 5C, Table S5). The 356 humanized PDAC proteins were enriched for 8 PANTHER pathways including DNA Replication, RAS, EGF receptor signaling, and Integrin signaling pathway (Table S6). The wider range of enriched pathways in PDAC confirmed results from the mouse, where many more pathways were enriched in pancreatic relative to colonic tumors (Figure 3B).

We constructed the predicted human PDAC KiPNA with CDK2 for comparison with the predicted human CRC KiPNA (Figure 5D). CDK2 was connected to more pathways in the CRC KiPNA than the PDAC KiPNA, likely due to coverage differences between MS proteomics and RPPA. Despite the mismatch in coverage, we found pathway crosstalk rewiring of EMT Signaling, with KRAS^G12D mutant PDAC exhibiting pathway crosstalk between EMT, UV Response DOWN, Inflammatory Response, Complement, and Coagulation, signaling, pathways that are decoupled in CRC.

Integrating humanized KiPNA with genetic screening identifies KRAS^G12D synthetic lethal candidates

We compared the humanized KiPNA networks to Project Achilles data to identify KRAS allele-specific synthetic lethal targets. Achilles quantifies the dependency of human cell lines on expression of a gene through shRNA and CRISPR/Cas9 based genetic screens (Aguirre et al., 2016; Cowley et al., 2014; Tsherniak et al., 2017). We searched for genes with higher essentiality in KRAS^G12D mutant cell lines by comparing essentiality probabilities in KRAS^G12D mutant cell lines to those with different KRAS alleles (WMW p < 0.05, probability fold change > 2, Humanized KiPNA overlap). Since different alleles can drive different signaling (Poulin et al., 2019), we reasoned that different KRAS alleles were likely to have distinct lethality dependencies and treated KRAS^G12D as a distinct entity for comparison with other alleles. We identified two genes in the humanized CRC KiPNA network – Adenylosuccinate Lyase (ASL) and Carnitine Acetyltransferase (CRAT) – and four genes in the humanized PDAC KiPNA network – SMAD family member 3 (SMAD3), Mesenchymal-Epithelial Transition factor (MET), Jun Proto-oncogene (JUN) and Serine and Arginine Rich Splicing Factor 1 (SRSF1) – that had higher essentiality in KRAS^G12D mutant cell lines. ASL was significantly humanized in CRC (p = 0.027) and SRSF1 in PDAC (p = 0.0030), with the remaining synthetic lethality genes being in the leading-edge of enriched humanized pathways (Table S5).

CRC humanized kinase-pathway subnetworks of EPHA2 and CDK2 showed that ASL and CRAT were linked to pathways associated with these kinases (Figure 6A). In the EPHA2 subnetwork, ASL is a leading-edge protein in the Xenobiotic Metabolism pathway. Previous studies show that ASL inhibition causes G2M cell cycle arrest and inhibit growth of colon cancer cells in cell lines and mice (Huang et al., 2017). CRAT was a linked to Peroxisome and Fatty Acid Metabolism pathways in the CDK2 kinase-pathway subnetwork (Figure 6A). In the humanized CRC KiPNA, crosstalk between Xenobiotic and Fatty Acid Metabolism pathways appears to integrate signaling by the synthetic lethal genes (ASL and CRAT) and kinases (CDK2 and EPHA2) (Figure 5B).

Figure 6. Synthetic lethality of genes in the humanized KiPNA networks.

A. CDK2 and EPHA2 kinase-pathway subnetworks in the humanized CRC KiPNA including ASL and CRAT, synthetic lethal partners in CRCs expressing KRAS^G12D. B. CDK2 Kinase-pathway subnetwork in the humanized PDAC KiPNA including JUN, MET, SMAD3, and SRSF1, synthetic lethal partners in PDACs expressing KRAS^G12D. Wilcoxon Mann-Whitney p-values (p < 0.05) and fold changes (FC) are shown for gene essentiality probabilities in colon cell lines (KRAS^G12D vs. and other KRAS alleles). C. Quantitative real-time PCR (qRT-PCR) analysis of knock-down efficiency with siRNA against indicated genes in colon (n = 4) and pancreas (n = 4) cell lines. Data presented as a mean fold-change relative to scrambled control ± SEM (n = 3 technical replicates in each of 2 biological replicates). D. Caspase 3/7 activity 96 hours after siRNA-mediated gene knock-down in KRAS^G12D and KRAS^WT or KRAS^G13D colon cell lines. Bar plot shows mean of caspase 3/7 activity (normalized to scrambled control) in colon cell lines with KRAS^G12D mutations (LS513, LS180) or other KRAS allele [HT115 (WT), DLD1 (G13D)]. Bars represent the average of means from cell lines in each group ± SEM (n = 3 technical replicates in each of 2 biological replicates). Data were compared by Welch’s t-test [p > 0.05 (n.s.); p ≤ 0.05 (*); p ≤ 0.01 (**); p ≤ 0.001 (***); p ≤ 0.0001 (****)]. E. Caspase 3/7 activity 96 hours after siRNA-mediated gene knock-down in KRAS^G12D and KRAS^G12R pancreas cell lines. Bar plot shows mean of caspase 3/7 activity (normalized to scrambled control) in pancreas cell lines with G12D mutations (SUIT2, KP4) or with G12R mutations (PSN1, TCCPAN2). Bars represent the average of means from cell lines in each group ± SEM (n = 3 technical replicates in each of 2 biological replicates). Data were compared by Welch’s t-test [p > 0.05 (ns); p ≤ 0.05 (*); p ≤ 0.01 (**); p ≤ 0.001 (***); p ≤ 0.0001 (****)]. (See also Figure S4).

The humanized PDAC CDK2 kinase-pathway subnetwork showed that, while SMAD3 and SRSF1 were adjacent to pathways regulated by CDK2, JUN and MET were a few steps away from the subnetwork (Figure 6B). SMAD3 was a leading-edge protein in the G2M Checkpoint and UV Response Down signaling pathways, while MET is a leading-edge protein in the UV Response Down and Inflammatory Response pathways. SMAD3 and MET have been implicated in metastatic pancreatic cancer (Basilico et al., 2018; Liu et al., 2013). JUN is a leading-edge protein in the EMT pathway and SRSF1 is a leading-edge protein in the G2M Checkpoint and E2F Targets pathways (Figure 6B).

Validation of K-RAS^G12D-associated essential genes

Achilles data analysis identified genes that had higher dependency in KRAS^G12D mutant cell lines relative to other KRAS alleles. We sought to validate these synthetic lethality hits by siRNA-mediated knockdown in a panel of colon and pancreas cell lines of defined KRAS genotype and measured apoptosis and cell viability. We saw a >50% knock-down of all genes in 7 out of 8 cell lines (Figure 6C). siRNA-mediated knockdown of ASL, but not CRAT, in KRAS^G12D mutant CRC cell lines resulted in significant induction of apoptosis and decreased cell viability, not affecting cell lines with other KRAS alleles (Figures 6D, S4). Although not predicted as KRAS^G12D synthetic lethal in PDAC (p = 0.22, probability FC = 2.5), knockdown of ASL in pancreatic cancer cell lines also led to an increase in caspase 3/7 activity and decrease in cell viability in a KRAS^G12D-specific manner (Figures 6E, S4). This suggests that ASL synthetic lethality crosses tissue boundaries.

Our analysis of the Achilles data for PDAC cell lines implicated SMAD3 and MET as more essential in KRAS^G12D mutant cell lines. Upon siRNA knockdown of these genes in PDAC cell lines, we observed a slight increase in caspase 3/7 activity in a KRAS^G12D-specific manner, but no concomitant decrease in cell viability (Figure 6E, S4). We observed that knockdown of SMAD3 and MET had a greater effect on overall viability in cell lines not expressing KRAS^G12D. This indicates that SMAD3 and MET knockdown induces cell death in a KRAS^G12D-dependent manner, but also affects general viability, perhaps via proliferation, in other KRAS allele contexts.

DISCUSSION

Activated oncoproteins promote cancer by transmitting signals to downstream pathways that control cellular behaviors within the confines of a tissue’s basal signaling network. An activated oncoprotein can elicit different signaling effects in different tissues because the tissue-specific basal signaling network is wired to promote its unique physiological function. Here, we characterized the signaling consequences of KRAS^G12D in colonic and pancreatic cancers and their preneoplastic counterparts and developed a battery of tools with broader applications in characterizing cellular signaling dysregulation. We identified a core K-RAS network of proteins and phosphorylation sites dysregulated by endogenous levels of K-RAS^G12D in four tissue contexts. This network included MAPK and RAL signaling components, canonical RAS effector pathways (Figure 2D). Most of the core network was composed of proteins that have not been previously connected to K-RAS. KRAS^G12D up-regulated the expression of several kinases – SLK and MAP2K3 – that have uncertain roles in cancer, but which could serve as therapeutic targets. Many proteins commonly associated with K-RAS signaling (e.g. AKT1 and RASSF2) were detected in all of the contexts, but were not commonly dysregulated. Since proteomic coverage varies between MS runs, our core network is likely incomplete. Despite differences in coverage, however, we characterized the KRAS^G12D-differential oncogenic network in CRC and PDAC.

The majority of the effects of KRAS^G12D were context-specific. We combined total and phosphoproteomics insights and to identify mechanistic hypotheses for how kinase signaling associates with pathway dysregulation (KiPNA). Comparison of the colonic epithelium and tumor KiPNA networks revealed rewiring of Kras^G12D signaling that results from adding an Apc mutation. Consistent with prior studies (Haigis et al., 2008), kinases within the canonical MAPK pathway (MEK1, ERK1, and ERK2) were activated by Kras^G12D in the colonic epithelium, but were not in colonic tumors (Figure 4). TGFB signaling was induced by Kras^G12D in colonic tumors (Figure 4), and this pathway was previously shown to regulate metastasis in Kras mutant mouse colonic tumors (Boutin et al., 2017). Finally, multiple pathways were discordantly regulated by mutant Kras in different contexts, including Angiogenesis and EMT pathway activation by Kras^G12D in colonic tumors and suppression in colonic epithelium (Figure 4). We identified rewiring of EGFR and CDK2 signaling in the KiPNA subnetworks, providing further examples of how secondary mutations (Apc) affect the signaling output of mutant Kras.

Because K-RAS signaling depends upon genetic background, it was unclear how signaling in genetically defined mouse tissues related to heterogeneous human tissues. To address this, we statistically humanized the results of our mouse analysis to infer kinase-pathway signaling processes in human KRAS^G12D mutant CRC and PDAC. A caveat to this approach is that the human data coverage provides an upper-bound on the amount of information that can be translated. We partially address this limitation of the human PDAC proteomics by integration with mutation data. If no proteomics data were available, an alternative approach would be to statistically humanize mouse proteomics with human transcriptomics or other biological network information. Here, we weighted human p-values 2–5 times more than mouse, a range we believe is reasonable for broader use. We characterized the theoretical behavior of combining mouse and human p-values and found that weighting human evidence at 2x the mouse, as if few human samples were available, meant that combining human and mouse p-values of 0.1 and 0.1 would give a significant humanized p-value (Figure S5). Weighting human p-values 5x the mouse showed that a mouse p-value of 0.1 and human p-value of 0.2 resulted in a significant humanized p-value. When few human samples are available, weighting human evidence 5x the mouse may be useful for discovery-based translation.

Statistical humanization characterized the tissue-specific signaling dysregulation associated with KRAS^G12D and identified CDK2 as a dysregulated kinase downstream of KRAS^G12D in both CRC and PDAC (Figures 5 and 6). The evidence for CDK2 as a dysregulated kinase and its tissue-specific signaling characteristics suggest broader differences in cell cycle signaling and therapeutic response potential between KRAS mutant PDAC and CRC. EPHA2 was up-regulated in the humanized network and was not identifiable from the mouse data alone (Figure 5). EPHA2 overexpression correlates with poor prognosis in other cancers and is a potential therapeutic target (Song et al., 2017; Tandon et al., 2011). Our results suggest further investigation of EPHA2 and CDK2 is warranted based on their roles in KRAS-associated signaling dysregulation.

In PDAC, CDK2 signals through STMN1 to G2M Checkpoint and E2F transcription factor pathways (Figure 5). The control of both the transition from G1 to S phase via E2F signaling and G to M phase signaling by CDK2 in PDAC suggests that this kinase drives cell cycle activity and progression (Morris et al., 2000). Nevertheless, therapies targeting CDK2 and other cell cycle regulators are toxic, killing both healthy and tumor cells, making them potentially problematic therapeutic targets (Zeitouni et al., 2016). Moreover, the lack of kinase enrichments in PDAC may explain why therapeutics targeting kinases have little efficacy in PDAC (Di Marco et al., 2016).

Comparing our humanized networks to genetic screening data identified a handful of synthetic lethal partners falling within the KRAS-regulated signaling networks. In CRC, the predicted synthetic lethal targets ASL and CRAT are members of pathways down-regulated by KRAS^G12D (Figure 6A), suggesting that therapeutic targets are not restricted to pathways that are activated by oncogenic KRAS. Although neither of these metabolic enzymes have a clear prior connection to KRAS or to CRC, ASL knockdown resulted in robust induction of cell death specifically in the context of KRAS^G12D in CRC and PDAC (Figure 6D, E). Although the humanized PDAC network was smaller, we identified more synthetic lethal genes (JUN, MET, SMAD3, and SRSF1). Like the CRC network, none of these has a clear prior connection to KRAS. Inhibition of MET – the most therapeutically tractable of these four potential targets – has shown efficacy in preclinical PDAC models (Brandes et al., 2015), but results in severe toxicity in patients in combination with standard-of-care (Zhen et al., 2016). The identification of synthetic lethal genes within KRAS-regulated networks provides a validation of the computational melding of mouse and human data, but further studies will be necessary to determine how these targets mediate the oncogenic function of KRAS.

KRAS^G12D drives tissue and tumor-specific proteomic and phosphoproteomic signaling in mouse models and that this specificity is conserved in human CRC and PDAC. Our finding that the same KRAS mutation drives divergent signaling processes depending on tissue and genetic context underscores the complexity of oncoprotein signaling and suggests that characterizing KRAS signaling across cancer contexts will lead to a more comprehensive understanding of oncogenic signaling. Our work shows that context-specific signaling explains differences in synthetic lethality relationships between cells and may determine the efficacy of precision medicine approaches.

STAR METHODS

LEAD CONTACT AND MATERIALS AVAILABILITY

Further information and requests for resources and reagents should be directed to and will be fulfilled by lead contact Kevin Haigis (khaigis@bidmc.harvard.edu).

Materials Availability Statement

This study did not generate new unique reagents.

EXPERIMENTAL MODEL AND SUBJECT DETAILS

Animal models

We used four mouse strains, 8–12 week-old Fabp1-Cre and Fabp1-Cre; Kras^LSL-G12D/+ mice (colonic epithelium) Villin-Cre^ER; Apc^2lox14/+ and Villin-Cre^ER; Apc^2lox14/+ Kras^LSL-G12D/+ mice (colonic tumors) (Haigis et al., 2008), Ptf1a-Cre and Ptf1a-Cre; Kras^LSL-G12D/+ mice (whole pancreas), and 3–4 month-old Pdx1-Cre or Pdx1-Cre; Tp53^LSL-R172H/+ mice and Pdx1-Cre; Tp53^LSL-R172H/+; Kras^LSL-G12D/+ mice (pancreatic tumors). Mice were fed ad libitum, housed in a barrier facility with a temperature-controlled environment and twelve-hour light/dark cycle, and maintained on a primarily C57BL/6 background. Animal studies were approved by the Institutional Animal Care and Use Committee (IACUC) at Beth Israel Deaconess Medical Center under animal protocol 081–2017. IACUC guidelines on the ethical care and use of animals were followed.

Cell lines

We performed our siRNA-knockdowns and subsequent apoptosis and cell viability assays in a panel of human colorectal (n = 4) and pancreatic cancer (n = 4) cell lines. The LS180 cell line was obtained from ATCC, maintained in EMEM supplemented with 10% FBS and 1% Penicillin/Streptomycin. The LS513 and DLD-1 cell lines were obtained from ATCC, maintained in RPMI supplemented with 10% FBS and 1% Penicillin/Streptomycin. The HT115 cell line was obtained from Sigma Aldrich and maintained in DMEM supplemented with 15% FBS, 1% Penicillin/Streptomycin, and 1% GlutaMAX. The KP4 and TCC-PAN2 cell lines were obtained from the laboratory of Andrew Aguirre, and maintained in RPMI supplemented with 10% FBS and 1% Penicillin/Streptomycin. The PSN-1 cell line was obtained from ATCC and maintained in RPMI supplemented with 10% FBS and 1% Penicillin/Streptomycin. The SUIT-2 cell line was obtained from the laboratory of Nada Kalaany and maintained in RPMI supplemented with 10% FBS and 1% Penicillin/Streptomycin. All cell lines were cultured in a humidified incubator at 37°C and 5% CO₂.

METHOD DETAILS

Mouse tissue collection

We used 36 biological replicates (colonic epithelium: 4 mice per genotype, colonic tumors: 5 mice per genotype, whole pancreas: 3 Kras^G12D and 4 Kras^WT mice, pancreatic tumors 5 Kras^G12D and 6 Kras^WT mice), each sample comprised of at least 3 pooled technical replicates. Epithelial surfaces of colons from 8–12-week-old Fabp1-Cre or Fabp1-Cre; Kras^LSL-G12D/+ mice were scraped with a razor blade and tissue from six individual mice was pooled and four samples per genotype were analyzed. Colonic tumorigenesis was induced by local administration of 4-hydroxytamoxifen (1 mg per mouse in 200 μL of 100% ethanol) via enema to 8–12-week-old Villin-Cre^ER; Apc^2lox14/+ or Villin-Cre^ER.; Apc^2lox14/+; Kras^LSL-G12D/+ mice (Haigis et al., 2008). Mice were sacrificed one week later and tumor lawns were collected by scraping with a razor blade. Tissue from 3–6 individual mice was pooled and five samples per genotype were analyzed. Whole pancreases from three 8–12-week-old Ptf1a-Cre or Ptf1a-Cre; Kras^LSL-G12D/+ mice were pooled, and three (G12D) or four (WT) samples per genotype were analyzed. For the pancreas tumor dataset, whole pancreases from three individual 4-month-old Pdx1-Cre or Pdx1-Cre; Tp53^LSL-R172H/+ mice were pooled and three samples per genotype were analyzed. Five individual pancreatic tumors were analyzed from Pdx1-Cre; Tp53^LSL-R172H/+ ; Kras^LSL-G12D/+ mice, collected when mice became moribund due to tumor burden (70–119 days). Tissues were collected and immediately flash frozen for proteomics analysis.

We controlled for different Cre promoters by generating Kras^WT mice within each promoter group. Different promoters were used due to how each expressed the desired mutations and number of mice per genotype we were able to breed. Fabp1-Cre was used for colonic epithelium because it was locally expressed and did not interfere with nutrient absorption like Villin-Cre. Fabp1-Cre was not suitable for colonic tumors because Fabp1-Cre combined with an Apc mutation is embryonic lethal, so we used to the inducible Villin-Cre. We used Ptf1a-Cre for whole pancreas and intended to use it for the tumors, but we were unable to breed a sufficient number of mice carrying both the Kras^G12D and Tp53 alleles, so we used Pdx1-Cre. Proteomic changes were compared within Cre promoters between Kras^G12D and Kras^WT mice, isolating the effect of Kras^G12D.

Tissue homogenization and cell lysis

Tissues were homogenized in lysis buffer (8 M urea, 200 mM EPPS pH 8.5, 1X Roche Protease Inhibitors, 1X Roche PhosphoStop phosphatase inhibitors) at a tissue concentration of approximately 10–15 mg/mL using a polytron tissue grinder. Next, the homogenate was sedimented by centrifugation at 21,000 × g for 5 min and the supernatant was transferred to a new tube. Protein concentrations were determined using the bicinchoninic acid (BCA) assay (ThermoFisher Scientific, Waltham, MA). 100 μg of protein was aliquoted from each fraction for subsequent reduction and alkylation. Proteins were subjected to disulfide bond reduction with 5 mM tris (2-carboxyethyl)phosphine (room temperature, 30 min) and alkylation with 10 mM iodoacetamide (room temperature, 30 min in the dark). Excess iodoacetamide was quenched with 10 mM dithiotreitol (room temperature, 15 min in the dark). Chloroform-methanol precipitation of proteins was performed prior to protease digestion. In brief, four parts neat methanol was added to each sample and vortexed, one-part chloroform was added to the sample and vortexed, and three parts water was added to the sample and vortexed. The sample was centrifuged at 4,000 RPM for 15 min at room temperature and subsequently washed twice with 100% methanol and vacuum centrifuged to dryness.

Samples were resuspended in 200 mM EPPS, pH 8.5 and digested at room temperature for 13 h with Lys-C protease at a 100:1 protein-to-protease ratio. Trypsin was then added at a 100:1 ratio and the reaction was incubated 6 h at 37°C. This reaction was then quenched with 1% formic acid, subjected to C18 solid-phase extraction (SPE) (Sep-Pak, Waters) and subsequently vacuum-centrifuged to near-dryness. Peptides were resuspended in 200 mM EPPS pH 8.5 and the Pierce Quantitative Colorimetric Peptide Assay (cat. #. 23275) was used to quantify the digest. From each sample, 100 μg of peptide was removed for protein level analysis (diluted to 1 mg/mL in 200 mM EPPS pH 8.5) and the remaining amount of the tissue digests were used for phosphoprotein enrichment.

Phosphopeptide enrichment

Phosphopeptides were enriched using a method based on that of Kettenbach and Gerber (Kettenbach and Gerber, 2011). In brief, Titanosphere TiO₂ 5 μm particles (GL Biosciences, Tokyo, Japan) were washed three times with 2 M lactic acid/ 50% acetonitrile. Peptides were resuspended in 2.5 mL of 2 M lactic acid/50 % acetonitrile. For ~10 mg of peptide digest, 40 mg beads were added and incubated with gentle rotation for 1 hr at room temperature. Beads were washed twice with 2.5 mL of 2 M lactic acid/ 50% acetonitrile, then twice with 2.5 mL of 50% acetonitrile/ 0.1% TFA, and finally twice with 2.5 mL of 25% acetonitrile/ 0.1% TFA. Enriched phosphopeptides were eluted twice with 500 μL of 50 mM K₂HPO₄ pH 10 and vacuum centrifuged to dryness. The dried peptides were then desalted via C18 solid-phase extraction (SPE) (Sep-Pak, Waters) and subsequently vacuum-centrifuged to near-dryness.

Tandem mass tag (TMT) labeling

In preparation for TMT labeling, desalted peptides (both for protein expression and phosphopeptide level analyses) were dissolved in 100 mM HEPES, pH 8.5. Peptide concentrations were determined using the microBCA assay. Approximately 100 μg of peptides from each sample were labeled with TMT reagent. TMT reagents (0.8 mg) were dissolved in anhydrous acetonitrile (40 μL) of which 10 μL was added to the peptides (100 μL) along with 40 μL of acetonitrile to achieve a final acetonitrile concentration of >30% (v/v). Following incubation at room temperature for 1 hr, the reaction was quenched with hydroxylamine to a final concentration of 0.3% (v/v). The TMT-labeled samples were combined at a 1:1 ratio in all channels. The sample was acidified, vacuum centrifuged to near dryness and subjected to C18 solid-phase extraction (SPE) (Sep-Pak, Waters).

Off-line basic pH reversed-phase (BPRP) fractionation

We fractionated the pooled TMT-labeled peptide sample using BPRP HPLC (Wang et al., 2011). We used an Agilent 1200 pump equipped with a degasser and a photodiode array (PDA) detector (set at 220 and 280 nm wavelength) from ThermoFisher Scientific (Waltham, MA). Peptides were subjected to a 50-min linear gradient from 5% to 35% acetonitrile in 10 mM ammonium bicarbonate pH 8 at a flow rate of 0.6 mL/min over an Agilent 300Extend C18 column (3.5 m particles, 4.6 mm ID and 220 mm in length). The peptide mixture was fractionated into a total of 96 fractions. To investigate the total proteome, these fractions were consolidated into 24, (12 non-adjacent samples) (Paulo et al., 2016a), while for phosphopeptide analysis, the samples were consolidated into 12 (8 samples pooled down column of a 96-well plate). Samples were subsequently acidified with 1% formic acid and vacuum centrifuged to near dryness. Each consolidated fraction was desalted via StageTip, dried again via vacuum centrifugation, and reconstituted in 5% acetonitrile and 5% formic acid for LC-MS³ processing.

Liquid chromatography and tandem mass spectrometry

All samples were analyzed on an Orbitrap Fusion or Fusion Lumos mass spectrometer (Thermo Fisher Scientific, San Jose, CA) coupled to a Proxeon EASY-nLC liquid chromatography (LC) pump (Thermo Fisher Scientific). Peptides were separated on a 100 μm inner diameter microcapillary column packed with 35 cm of Accucore C18 resin (2.6 μm, 150 Å, ThermoFisher). For each analysis, we loaded approximately 2 μg onto the column. Peptides were separated using a 150min gradient of 3 to 25% acetonitrile in 0.125% formic acid with a flow rate of 450 nL/min. Each analysis used an MS3-based TMT method (McAlister et al., 2014; Ting et al., 2011), which has been shown to reduce ion interference compared to MS2 quantification (Paulo et al., 2016b). Prior to analysis, we perform two injections of trifluoroethanol (TFE) to elute peptides that may be bound to the analytical column from prior injections to limit carry over. The scan sequence began with an MS1 spectrum (Orbitrap analysis, resolution 120,000, 350−1400 Th, automatic gain control (AGC) target 5E5, maximum injection time 100 ms). The top ten precursors were then selected for MS2/MS3 analysis. MS2 analysis consisted of: collision-induced dissociation (CID), quadrupole ion trap analysis, automatic gain control (AGC) 2E4, NCE (normalized collision energy) 35, q-value 0.25, maximum injection time 120 ms, and isolation window at 0.7. For phosphorylation analysis, multistage activation was used, and the neutral loss was set at 97.9763. Following acquisition of each MS2 spectrum, we collected an MS3 spectrum in which multiple MS2 fragment ions are captured in the MS3 precursor population using isolation waveforms with multiple frequency notches. MS3 precursors were fragmented by HCD and analyzed using the Orbitrap (NCE 65, AGC 1.5E5, maximum injection time 150 ms (250 ms for phosphorylation analysis), resolution was 50,000 at 400 Th). For MS3 analysis, we used charge state-dependent isolation windows: For charge state z=2, the isolation window was set at 1.3 Th, for z=3 at 1 Th, for z=4 at 0.8 Th, and for z=5 at 0.7 Th.

siRNA Transfection, RNA extraction and Real-time PCR (RT-PCR)

siRNA transfections were performed in 96-well plates using reverse-transfection of cell lines (4 colonic, 4 pancreatic) with Lipofectamine RNAiMAX (Invitrogen). Complexes were prepared in the 96-well plates (25 nM siRNA diluted in Opti-MEM with 0.2 uL Lipofectamine RNAiMAX diluted in Opti-MEM, per well) and incubated at room temperature for 20 minutes. Cells were washed with PBS, trypsinized using TrypLE, counted, and plated at a density of 5,000 cells/well with the appropriate media lacking antibiotics. A four-molecule pool of oligonucleotides targeting GAPDH (D-001830-10-05, Dharmacon), KIF11 (L-003317-00-0005, Dharmacon), ASL (L-008840-00-005, Dharmacon), CRAT (L-009524-00-0005, Dharmacon), SMAD3 (L-020067-00-0005, Dharmacon), MET (L-003156-00-0005, Dharmacon), or a non-targeting scrambled control pool (D-001810-10-05, Dharmacon) was used for each transfection. Each siRNA transfection in each cell line was performed in triplicate, and cells were plated for RNA isolation (for knockdown validation) and the ApoLive Glo Assay in parallel. Total RNA was extracted using the Quick-RNA 96 kit (Zymo) 48 hours after siRNA transfection. 30 ng total RNA prepared from the siRNA-transfected cells described above was used for first strand cDNA synthesis with the High-Capacity cDNA Reverse Transcription Kit (Applied Biosystems, ThermoFisher Scientific) following the manufacturer’s protocol. Quantitative PCR (qPCR) was performed using TaqMan gene expression assays (18S: Hs99999901_s1, GAPDH: Hs02758991_g1, KIF11: Hs00189698_m1, ASL: Hs00902699_m1, CRAT: Hs00912963_m1, SMAD3: Hs00969210_m1, MET: Hs01565584_m1). TaqMan Universal Master Mix (Applied Biosystems, ThermoFisher Scientific) was used to prepare reactions, and samples were run and analyzed in quadruplicate. The relative expression of genes was calculated with the 2^−ΔΔCt method, with normalization to 18S RNA levels. In all figures, the data presented are an average of 3 technical replicates presented with error bars +/− SEM.

Cell Viability and Apoptosis Assay

Cell viability and apoptosis upon siRNA-mediated knockdown of each gene in our panel of colon and pancreas cell lines was measured using the ApoLive-Glo Multiplex Assay (Promega), according to the manufacturer’s protocol. The assay was performed 96 hours after siRNA transfection, and results were analyzed in triplicate. In all figures, the data presented are representative of 3 technical replicates and 2 cell lines per condition.

QUANTIFICATION AND STATISTICAL ANALYSIS

MS data analysis

Mass spectra were processed using a Sequest-based in-house software pipeline. MS spectra were converted to mzXML using a modified version of ReAdW.exe. Database searching included all entries from the mouse Uniprot database (October 21, 2014), which was concatenated with a reverse database composed of all protein sequences in reversed order. Searches were performed using a 50 ppm precursor ion tolerance for total protein level analysis and 20 ppm for phosphopeptide analysis. Product ion tolerance was set to 1 Da. TMT tags on lysine residues and peptide N termini (+229.1629 Da) and carbamidomethylation of cysteine residues (+57.0215 Da) were set as static modifications, while oxidation of methionine residues (+15.9949 Da) was set as a variable modification. For phosphorylation analysis, +79.9663 Da on serine, threonine, and tyrosine was also set as a variable modification.

Peptide spectral matches (PSMs) were filtered to a 1% FDR (Elias and Gygi, 2007, 2010). PSM filtering was performed using linear discriminant analysis, as described previously (Huttlin et al., 2010), while considering the following parameters: XCorr, ΔCn, missed cleavages, peptide length, charge state, and precursor mass accuracy. For TMT-based reporter ion quantitation, we extracted the signal-to-noise (S/N) ratio for each TMT channel and found the closest matching centroid to the expected mass of the TMT reporter ion.

The search space for each reporter ion was limited to a range of 0.002 Th to prevent overlap between the isobaric reporter ions. For protein-level comparisons, peptide-spectral matches were identified, quantified, and collapsed to a 1% FDR and then collapsed further to a final protein-level FDR of 1%. Furthermore, protein assembly was guided by principles of parsimony to produce the smallest set of proteins necessary to account for all observed peptides.

Proteins and phosphorylation sites were quantified by summing reporter ion counts across all matching PSMs using in-house software, as described previously (Huttlin et al., 2010). Briefly, a 0.002 Th window around the theoretical m/z of each reporter ion was scanned for ions, and the maximum intensity nearest the theoretical m/z was used. PSMs with poor quality, MS³ spectra with more than seven TMT channels missing, less than 100 TMT reporter summed signal to noise ratio, or no MS³ spectra at all were excluded from quantitation (McAlister et al., 2012). Protein quantitation values were exported for further bioinformatics analysis.

Data quality assessment

Datasets were visualized using unsupervised hierarchical clustering (Euclidean distance metric) and principal component analysis (PCA) to assess quality of experimental grouping by Kras^G12D and WT status (Figure S1 and S2). In both analyses, samples separated by genotype and no samples were excluded from downstream analysis. Analysis was completed using MATLAB_2018a.

Differential regulation analysis

Differential protein and phosphopeptide activity was assessed with the Wilcoxon Mann-Whitney (WMW) test with Benjamini-Hochberg False Discovery Rate (FDR) correction (Figure 2, Table S2). Nonparametric methods were selected due to the small sample sizes. A protein or phosphosite was considered differentially active at WMW p<0.05 and BH-FDR q<0.25. However, no species in the whole pancreas datasets reached this level of statistical significance. For the whole pancreas data, proteins or phosphosites with p<0.10 and q<0.25 were selected as “top candidates” for comparison purposes with other datasets. Analysis was completed using MATLAB_2018a.

PANTHER protein class and pathway enrichment analysis

Differentially regulated or phosphorylated proteins were analyzed using PANTHER protein class and pathway enrichment analysis to estimate the abundances of different types of proteins (Thomas et al., 2003; Thomas et al., 2006). Protein lists from statistically humanized mouse proteomics data were also analyzed using the PANTHER pathway enrichment analysis tools (Figure 2, Tables S4 and S6).

Pathway and kinase enrichment analysis

Gene Set Enrichment Analysis (GSEA) was performed on total proteomics datasets using the gene set permutation method and the Hallmarks Gene Sets (Figure 3) (Liberzon et al., 2015). Kinase enrichment analysis on the phosphoproteomics data was performed using the GSEA software and kinase-substrate “gene sets” assembled from PhosphositePlus as previously described (Hornbeck et al., 2015; Lyons et al., 2018). Pathway and kinase were considered significantly enriched at FDR q < 0.05.

Kinase-Pathway Network Analysis (KiPNA)

We developed a network analysis framework for relating kinase signaling with pathway dysregulation in our total and phosphoproteomics datasets. In this framework, a network is constructed using four types of nodes: (1) nodes for significantly enriched pathways, (2) nodes for significantly enriched kinases, (3) nodes for phosphosites in the leading-edge subset of a significantly enriched kinase, and (4) nodes for proteins either in the leading edge subset of significantly enriched pathways or that have an associated leading edge phosphosite. These nodes are connected by edges such that (1) enriched kinases connect to leading edge phosphosites, (2) leading edge phosphosites connect to target proteins, and (3) proteins connect to enriched pathways. We performed topological analysis of these networks to identify biologically motivated motifs linking enriched kinases to enriched pathways using a 3-step directed walk from enriched kinases to leading edge phosphosites, leading edge phosphosites to associated proteins, and from these proteins to enriched pathways (Figures 4–6). Networks were visualized using Cytoscape and deposited at the Network Data Exchange (Pillich et al., 2017; Pratt et al., 2017; Pratt et al., 2015; Shannon et al., 2003).

Statistical humanization

Insights from mouse total and phosphoproteomics datasets were statistically humanized using a modified version for the weighted Z-test in which we integrated information from multiple molecular data types and species (Table S5) (Zaykin, 2011). P-values from each dataset were weighted by the square root of the sample size for each data stream, with more weight naturally given to evidence from human contexts. From CRC tumors, 53 proteomic and 224 mutation profiles were used (Cancer Genome Atlas, 2012; Zhang et al., 2014). From PDAC tumors, 46 proteomic and 150 mutation profiles were used (Cancer Genome Atlas Research Network, 2017). Two-tailed p-values from proteomics and enrichment analyses were converted to one-sided p-values with the direction chosen as indicated by the mouse data. The analysis was restricted to proteins that were measured by the mouse proteomics analysis and in at least one human dataset. Mouse protein identifiers were converted to homologous human identifiers using the mouse genome informatics database (Blake et al., 2017; Eppig et al., 2015).

Integration of Project Achilles data

We obtained gene dependency probability data for PDAC (n = 22) and CRC (n = 27) cell lines from Project Achilles to investigate the essentiality of genes in the humanized KiPNA networks and identify potential therapeutic targets (Aguirre et al., 2016; Cowley et al., 2014; Meyers et al., 2017; Tsherniak et al., 2017). We investigated whether each gene had higher probability of cell line dependency in the KRAS^G12D mutant cell lines relative to all other cell lines of that cancer type. A Wilcoxon Mann-Whitney test was applied to assess whether probability of dependency was significantly different between cell lines expressing KRAS^G12D and the rest of the population, nonparametric methods being used since the probabilities were non-normally distributed. Genes with higher dependency probability in KRAS^G12D mutant cell lines were selected for comparison with the humanized CRC and PDAC KiPNA networks (p < 0.05) (Figure 6).

Analysis of siRNA Knockdown Data

Raw read values from the ApoLive-Glo assay were analyzed in PRISM 8.0 (GraphPad Software). Background reads were subtracted from each sample read, and values from each sample were normalized relative to scrambled control. Replicates were averaged to calculate means, standard error of means, and p-values. We assumed variances of cell line treatment groups were unequal and selected Welch’s t-test to determine significance (p < 0.05 significant) (Figure 6). In all figures, the data presented are representative of 3 technical replicates and 2 cell lines per condition.

DATA AND CODE AVAILABILITY

The mass spectrometry mouse proteomics datasets have been deposited to the ProteomeXchange Consortium via the PRIDE partner repository with the dataset identifier PXD013922. The KiPNA networks and analysis tools are deposited at the Network Data Exchange. Mouse Colonic Epithelium KiPNA (UUID: 2ca790f5-8087-11e9-848d-0ac135e8bacf), Mouse Colonic Tumor KiPNA (UUID: c464e471-8086-11e9-848d-0ac135e8bacf), CRC Humanized KiPNA (UUID: 3d0b4688-8087-11e9-848d-0ac135e8bacf), and PDAC Humanized PDAC KiPNA (UUID: 4c71141b-8087-11e9-848d-0ac135e8bacf). The weighted Z-test code is available on Mathworks FileExchange, ID 71671.

Supplementary Material

3

Table S1. Related to Figure 1. Coverage of total and phosphoproteomic datasets in each tissue context.

4

Table S2. Related to Figure 2. Statistics and fold changes for differential regulation analysis of mouse total and phosphoproteomics datasets. Proteins are labeled by name and phosphosites by “Protein_SiteNumber”.

5

Table S3. Related to Figure 2. Venn diagram regions of differentially regulated proteins and phosphosites in each tissue. (Colon Epithelium (CE), Colon Tumors (CT), Whole Pancreas (WP), Pancreas Tumors (PT)). Naming: “Protein_Direction” or “Protein_Site_Direction”.

6

Table S4. Related to Figure 2. PANTHER protein class distributions for differentially regulated proteins and phosphosites across all tissues and in tissue-specific or tumor-specific contexts.

7

Table S5. Related to Figure 5. Statistical humanization results for integration of patient CRC And PDAC proteomics and mutation data with mouse total and phosphoproteomics data.

8

Table S6. Related to Figure 5. Pathways enriched in statistically humanized CRC and PDAC proteins.

9

Table S7. Related To Figure 6. Reagents used in experiments.

KEY RESOURCES TABLE

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Biological Samples
The Cancer Genome Atlas Pancreatic Ductal Adenocarcinoma Reverse Phase Protein Array and Mutation Data	(Cancer Genome Atlas Research Network, 2017)	PAAD
The Cancer Genome Atlas Colorectal Carcinoma Proteomics	(Zhang et al., 2014)	COADREAD
The Cancer Genome Atlas Colorectal Carcinoma Mutations	(Cancer Genome Atlas, 2012)	COADREAD
Project Achilles Gene Dependency Data	(Cowley et al., 2014)	https://portals.broadinstitute.org/achilles/datasets/21/download/gene_dependency.csv
Chemicals, Peptides, and Recombinant Proteins
HI FBS	Gibco	16140–071
GlutaMAX Supplement	Gibco	35050–061
Antibiotic Antimycotic Solution (100x) - Penicillin, Streptomycin, Amphotericin B	Sigma Aldrich	A5955–100
Opti-MEM Reduced Serum Medium	Gibco	31985062
Critical Commercial Assays
Bicinchoninic acid (BCA) assay	ThermoFisher Scientific	Cat. No. 23225
Pierce Quantitative Colorimetric Peptide Assay	ThermoFisher Scientific	Cat. No. 23275
C18 solid-phase extraction (SPE)	Waters	Cat. No.: WAT054925
Tandem Mass Tag	ThermoFisher Scientific	Cat. No. A34808
ApoLive-Glo Multiplex Assay	Promega	G6411
TaqMan Universal PCR Master Mix	Applied Biosystems	4304437
High-Capacity cDNA Reverse Transcription Kit	Applied Biosystems	4368814
Quick-RNA 96 Kit	Zymo Research	R1053
Lipofectamine RNAiMAX Transfection Reagent	Invitrogen	13778075
High-Capacity cDNA Reverse Transcription Kit	Applied Biosystems	4368814
Deposited Data
Mouse Total and Phosphoproteomics Data	This Work	ProteomeXchange PXD013922
Mouse Colonic Tumor KiPNA	This Work	NDeX: c464e471-8086-11e9-848d-0ac135e8bacf
Mouse Colonic Epithelium KiPNA	This Work	NDeX: 2ca790f5-8087-11e9-848d-0ac135e8bacf
Humanized CRC KiPNA	This Work	NDeX: 3d0b4688-8087-11e9-848d-0ac135e8bacf
Humanized PDAC KiPNA	This Work	NDeX: 4c71141b-8087-11e9-848d-0ac135e8bacf
Experimental Models: Cell Lines
Human: LS180 Cells	ATCC	CL-187; RRID:CVCL_0397
Human: LS513 Cells	ATCC	CRL-2134; RRID:CVCL_1386
Human: DLD-1 Cells	ATCC	CCL-221; RRID:CVCL-0248
Human: HT115 Cells	Sigma Aldrich	85061104; RRID:CVCL_2520
Human: SUIT-2 Cells	Laboratory of Nada Kalaany	N/A
Human: KP4 Cells	Laboratory of Andrew Aguirre	N/A
Human: PSN-1 Cells	ATCC	CRL-3211; RRID:CVCL_1644
Human: TCC-PAN2 Cells	Laboratory of Andrew Aguirre	N/A
Experimental Models: Organisms/Strains
Pdx1-Cre inbred mouse (80–95% C57BL/6 Background)	The Jackson Laboratory	Strain 014647
Fabp1-Cre inbred mouse (80–95% C57BL/6 Background)	NCI Mouse Repository	Strain 01XD8
Villin-CreER12 inbred mouse (80–95% C57BL/6 Background)	The Jackson Laboratory	Strain 020282
Ptf1a-Cre inbred mouse (80–95% C57BL/6 Background)	The Jackson Laboratory	Strain 023329
Oligonucleotides
See Table S7
Software and Algorithms
MATLAB	Mathworks	R2018b
Weighted Z-Test	This Work	Mathworks File Exchange: 71671
Sequest	Eng, al et 1994	http://proteomicsresource.washington.edu/protocols06/sequestphp

Highlights.

• Total and phosphoproteomic characterization of mutationally activated K-RAS

• Translation of murine signaling insights to patients by statistical humanization

• Identify KRAS^G12D allele specific synthetic lethality with ASL and SMAD3