{"paper_id":"fddeb4ef-13cf-4a17-8e8f-1702b856a825","body_text":"Cancer is a disease of cellular dysfunction.\nAs described by Hanahan\nand Weinberg, \n , \n  these dysfunctions typically fit\ninto one of a few broad categories called hallmarks. After the discovery\nthat cancer was caused by DNA mutation,  early molecular analysis of tumors focused on DNA sequencing to\nidentify common driver mutations. Extensive genomic characterization\nof tumors has identified DNA driver mutations, which help identify\ndisease subtypes and guide precision therapy. \n , \n  Although\nuseful, this catalog of DNA mutations does not describe the altered\ncellular state that defines the tumor phenotype. For this reason,\nmany cancer studies create proteogenomic data \n , \n  to\nbetter link somatic mutations to changes in mRNA and/or protein abundance.\nAlthough mRNA is only a temporary intermediary between genome and\nprotein, it is frequently measured and studied as a proxy for the\ncellular state due to its convenient and robust data acquisition methods.\nHowever, a large body of evidence shows that mRNA abundance has a\nlow correlation to protein abundance, \n − \n \n \n  due to translational and post-translational\nregulation. \n − \n \n \n \n \n  For proteogenomic data that contains both mRNA and protein abundance,\na common data integration exercise is to predict protein abundance\nfrom mRNA,  often with only modest accuracy.\nIn cancer studies, \n − \n \n \n  although many genes have a correlation >0.6, a large number of\ngenes\nhave correlations <0.2, including negative correlation values.\nPathway analyses consistently discovered that these patterns of mRNA/protein\ncorrelation – including high and low correlations –\nare biologically coherent; metabolic and signaling pathways are enriched\nin genes with a high correlation while ribosomes and oxidative phosphorylation\nare enriched in genes with a low/negative correlation.\nIntegrative\ndata analysis of mRNA and proteins can also be used\nto investigate biological hypotheses. For example, does the relationship\nof mRNA and protein change in a disease? Is there an altered protein\nhomeostasis in cancer? As cancer is marked by regulatory dysfunction,\nit is intriguing to evaluate whether the regulatory relationship between\nmRNA and proteins has changed in tumor cells. Indeed, a fundamental\nand unexplored question is whether mRNA/protein quantitative relationships\nare static or dynamic. A static relationship implies that the mRNA/protein\ncorrelation is the same in all tissues and disease states. This is\nan implicit assumption of algorithms that learn to predict protein\nabundance from mRNA, as they are trained on data from a limited set\nof conditions. Although some have suggested that this relationship\nis fixed and static, \n , \n  several studies demonstrate that\nthis can be a fluid relationship. \n − \n \n  For example, Takemon\net al. discovered age-dependent changes in protein concentration in\nthe absence of corresponding changes in mRNA.  Indeed, a change in any post-translational regulatory mechanism\nwould create a dynamic mRNA/protein relationship. The extent of such\nchanges in cancer is unknown.\nThe stability of the mRNA/protein\nrelationship remains underexplored,\nlargely due to the absence of a sufficiently powered data set. To\ncharacterize the correlation, a data set must have sufficient samples\nin multiple tissues and measure both protein and mRNA for thousands\nof genes. The Clinical Proteomic Tumor Analysis Consortium (CPTAC)\nhas created one such data set, with proteogenomic data for ∼100\ntumor samples per tumor type. \n , \n  It also contain matched\nnormal tissues for many tumor samples. Given the complex regulatory\nmechanisms utilized by cells and the growing evidence of regulatory\nplasticity in cancer cells, we sought to understand the mRNA/protein\nrelationship in cancer tissues. Specifically, we examined whether\nthis relationship is static or fluid and which cancer related genes\nand pathways are affected. We discover major regulatory changes across\nthousands of genes that lead to the activation of pathways associated\nwith hallmarks of cancer.\n\nAll of the data analysis\nscripts used to generate figures and metrics in this publication are\npublicly accessible in our GitHub repository at  https://github.com/PayneLab/pancancerProteinMRNA . Individual scripts are listed in various sections below.\nWe utilized the CPTAC data from the pan-cancer\nresource API.  Specifically, we used the\ndata for transcriptomics, proteomics and DNA somatic mutation, accessed\nthrough the publicly available CPTAC Python API.  Data generation and data analysis methods for these are\ndetailed in the linked publications. Briefly, the transcriptomics\ndata is RNASeq gathered on all samples (tumor and normal), and TMT-labeled\nglobal proteomics data was collected to characterize the proteome\n(tumor and normal). A direct link to the harmonized data tables stored\nat the Proteome Data Commons is  https://pdc.cancer.gov/pdc/cptac-pancancer . According to the published studies, all of the studies were performed\nwith proper ethics.\nTo identify the relationship\nbetween mRNA and proteins in the cancer cohorts, we calculated the\nSpearman correlations between transcriptomic and proteomic data by\npatient, separately, for tumor and normal samples. A minimum sample\nsize cutoff of 15 data points per gene was enforced, with the exception\nof endometrial cancer, which has a cutoff of 10 due to having a smaller\nnumber of normal samples. The exact implementation of this correlation\ncan be found in our GitHub repository at Implementation is here (∼/notebook_steps/data/Make_Tumor-Normal_Correlation_Dataframe.ipynb).\nData corresponding to the mRNA/protein correlation of each healthy\nand tumor genes for each of the cancer cohorts are here (∼/notebook_steps/data/tumor_normal_correlation_df.csv).\nAdditionally, using CPTAC’s clinical data, we can analyze\nthe cancer stage for each sample. We combined the stage clinical information\nand joined it with each patient’s proteomic and transcriptomic\ndata. We then grouped samples by cancer, gene, and stage in a single\ndata frame. Then, we did a Spearman correlation between normal and\ntumor gene transcriptomics and proteomics of the genes that are present\nin each stage. We set a cutoff for stages that did not have 10 or\nmore genes per stage to keep better consistency. We organized the\ncorrelations in each stage and plotted the data into a Kernel density\nestimation plot ( Figure  \n ). For implementation see:\nDistribution of mRNA/protein correlations. For\neach of the five\nCPTAC cohorts, the Spearman correlation between mRNA and protein was\nperformed for all genes. The distribution of these correlation coefficients\nis plotted for normal samples (blue) and tumor samples.\n∼/notebook_steps/data/Make_Cancer_stages_correlations_df.ipynb\nand ∼ /notebook_steps/Make_Figure_1_Correlation_Change_by_Cancer_Stage.ipynb\nTo quantify the change in regulation\nfor a given gene between tumor and normal tissue, we calculated the\ndifference in spearman correlation coefficients, a term we named Δ_corr.\nFor each gene with a minimum of  n  = 15 ( n  = 10 for endometrial) tumor and normal samples with proteomic and\ntranscriptomic data, we first calculated the tumor and normal correlation\ncoefficients (ρ). We then calculated the Δ_corr value\nusing the following equation: \n Δ _ corr = ρ tumor − ρ normal\nA positive Δ_corr\ntranslates to a higher correlation coefficient in tumor samples, while\na negative Δ_corr translates to a lower correlation coefficient\nin tumor samples. For implementation, see the readme found in the\nGItHub repository at ∼ /notebook_steps.\nIn order to calculate probability\nvalues for Δ_corr, we used label permutation. The advantage\nof label permutation is that it allows us to explicitly create a distribution\nof values under our null hypothesis and then compare the test statistic\nto this null distribution. Our null hypothesis is that the mRNA/protein\ncorrelation does not change between normal and tumor tissue. This\nis equivalent to saying that the label of a sample as a tumor or normal\nis not meaningful when creating an mRNA/protein correlation. Thus,\na direct instance of this null hypothesis is one where the data series\nhas a randomized label. By performing many randomized label experiments\n(i.e., label permutations), we can create a null distribution that\nexactly matches our null hypothesis.\nUsing tissue for both normal\nand tumor samples and its matching gene readings for individual patients,\nwe did a label permutation to create a false correlation. Since all\npatients have proteomic and transcriptomic gene information divided\ninto normal and tumor samples, we join them in a single data frame\nto be accessed from. Then, we randomly assigned the labels of normal\nand tumor to each measurement, keeping the same amount of tumor and\nnormal data points. To provide a more precise procedure, we set a\ncutoff for genes that have 15 or more valid data points (except endometrial,\nwhich has a cutoff of 10 or more). From each data frame that has been\npermuted, we do a Spearman correlation from the permuted data frame.\nThis gives us a false Δ_corr for permuted data frames as seen\nin  Supplemental Figure 3B . We then repeat\n10,000 permutations, each resulting in a false Δ_corr. In  Supplemental Figure 3C , we can see how different\nthe original Δ_corr from the permuted delta correlations is.\nFrom the list of Δ_corr we obtain a  z -score\nwith respect to the original nonpermuted Δ_corr. This  z -score is transformed into a  p -value showing\nthe significance of the change in terms of a gene's mRNA/protein\ncorrelation in a tissue. The p-values were corrected through a Benjamini-Hochberg\nProcedure to account for the False Discovery Rate (FDR). We do this\nfor all genes and all 5 selected tissues on the BYU supercomputer.\nFor implementation see:\n∼/notebook_steps/data/Scripts_to_Make_Cancer_Delta_Corr_and_P_Value_Dataframe\nTo determine differential expression\nof a gene’s mRNA and protein abundance between tumor and normal\ntissue, we performed a Wilcoxon rank-sums test and calculated the\nmean log2 fold-change for each gene in each cancer type.  P -values from the Wilcoxon rank-sums test were corrected for multiple\nhypothesis testing via Benjamini–Hochberg correction. Genes\nwere considered to be differentially expressed if they had a BH p-value\n<0.05 and |log2 fold-change| > 1, following cutoffs used in.  For implementation see ∼ /notebook_steps/data/Make_Proteomics_and_Transcriptomics_Differential_Expression_Dataframes.ipynb\nTo determine the effect\nof mutations on Δ_corr, we first identified the most frequently\nmutated genes in each tumor cohort. For each cohort, we identified\nthe top 10 most frequently mutated genes with at least  n  = 15 ( n  = 10 for endometrial) tumor mutant and\ntumor wild-type samples for statistical significance. For tumor cohorts\nthat did not contain at least 10 mutated genes that met the cutoff,\nall genes meeting the cutoff were used. For implementation, see ∼\n/notebook_steps/data/Scripts_to_make_transmutation_effects_dataframes/Find_Most_Mutated_Genes_and_Write_transmutation_scripts.ipynb\nFor each of the most frequently mutated genes, we next calculated\nthe Δ_corr for each transgene between the tumor-mutated and\ntumor-wildtype samples. To determine the significance of the Δ_corr,\nwe performed a label based permutation (see Permutation Test) with\n10,000 permutations. The permutation p-values for each mutated gene\ngroup were then corrected for multiple hypothesis testing via Benjamini–Hochberg\ncorrection. A mutation was considered to have a significant effect\non a transgene if the Δ_corr permutation BH  p -value <0.05. For implementation see ∼ /notebook_steps/data/Scripts_to_make_transmutation_effects_dataframes/transmutation_effects.py\nPermutation tests were performed on a BYU supercomputer. For bash\nscripts ran on the supercomputer, see the readme found in ∼\n/notebook_steps\nWe used the gProfiler python API to\nperform a gene set enrichment analysis of significant Δ_corr\ngenes. For the enrichment analysis, we used the genes that have a\nFDR corrected  p -value of less than 0.05 and absolute\nthreshold of 0.2. The genes used for background are the set of genes\nfound in the corresponding cancer. We evaluated our genes with the\nKyoto Encyclopedia of Genes and Genomes (KEGG) library of pathways.\nThis was repeated for each of our 5 tissues. For implementation, see\n∼ /notebook_steps/Make_Figure_5_Cancer_Hallmarks_Pathways.ipynb\n\nThe Clinical Proteomic Tumor Analysis Consortium\n(CPTAC) creates\nproteogenomic data for cancer discovery. A typical CPTAC cohort contains\n100 tumor samples, which are characterized with genomics, transcriptomics,\nand proteomics, as well as clinical demographic and treatment information.\nFive CPTAC cohorts have also characterized normal tissue: lung adenocarcinoma\n(LUAD), lung squamous cell carcinoma (LSCC), clear cell renal cell\ncarcinoma (ccRCC), endometrial cancer (ENDO), and head and neck squamous\ncell carcinoma (HNSCC). Therefore, we examined these five tumor types\nto identify changes in the mRNA-protein relationship between healthy\nand cancer tissue.\nWe performed a gene-wise correlation\ncomparison between the abundance of mRNA and protein for each gene\nacross all samples in a cohort. Each gene’s mRNA/protein relationship\nis unique  including positive, negative,\nand noncorrelations ( Figure S1 ). It is\noften assumed that the lack of correlation is an artifact of poor\nmeasurement;  however, the wide range\nof correlation values has been observed in proteogenomic analysis\nfor the past 20 years through multiple iterations of both mRNA and\nprotein measurement technologies. \n , , \n  Moreover, as shown in  Figure  \n , the correlation value is dynamic and changes\nacross tissues and disease states. The large shift in the distribution\nof correlation values between tumor and normal tissue is the result\nof a changing mRNA-protein relationship across thousands of genes.\nChanges to any translational or post-translational regulatory processes\nwill change this quantitative relationship. We assert that changes\nin this correlation between conditions such as health and disease\nare a result of regulatory changes within the cell \n , \n  and are not solely an artifact of measurement technology.\nWe investigated the change in a gene’s mRNA/protein relationship\nby observing a change in correlation between normal and cancer tissues\n( Figure  \n ). Our metric\nfor a change in correlation is simply comparing the Spearman correlation\ncoefficient between cancer and normal tissue, called Δ_corr.\nA positive Δ_corr means that the correlation coefficient in\ntumor samples is larger than in normal samples, i.e., a protein’s\nabundance tracks more closely to mRNA abundance. A negative Δ_corr\nmeans that the correlation in tumors is smaller than in normal samples.\nWe emphasize that Δ_corr is not the same as the differential\nexpression. A change in correlation may be coincident with differential\nexpression for some genes, as seen in ARID1A in lung adenocarcinoma\n( Figure  \n , middle row).\nHowever, a change in correlation is not synonymous with differential\nexpression. For example, the protein and mRNA abundance of ARID1B\ndoes not change between normal and tumor tissue; there is no differential\nexpression. However, there is a substantial change in the mRNA/protein\nrelationship, as denoted by a Δ_corr value of ∼0.5378.\nIn normal lung tissue, the mRNA/protein correlation is ∼0.027,\nbut in LUAD the correlation is ∼0.565. ( Figure  \n , bottom row). Thus, the Δ_corr highlights\na change in cellular regulation, distinct from a potential change\nin abundance.\nChange in correlation is distinct from change in expression.\nThree\ngenes from the LUAD data set are shown to demonstrate the independence\nof Δ_corr and differential expression. The left panel is a scatter\nplot and regression analysis of each sample’s mRNA and protein\nabundance (tumor samples in orange, normal tissue in blue). The middle\ncharts are a box plot of the protein abundance with a Wilcoxon rank-sums\ntest for differential expression between tumor and normal tissue.\nSimilar plots on the right are for the RNA abundance. NKTR, top row,\nhas a similar mRNA/protein correlation for both tumor and normal tissue\n(Δ_corr ∼ 0.056; permutation test BH adjusted  p  ∼ 0.681). However, both the protein and mRNA are\ndifferentially expressed (Wilcoxon BH adjusted  p  ∼\n2.707 × 10 –24 ). ARID1A, middle row, has a significant\nchange in the correlation (Δ_corr ∼ 0.430; permutation\ntest BH adjusted  p  ∼ 0.003); normal tissue\nhas a noncorrelation (flat blue line), whereas tumor tissue shows\na strong positive correlation. Both protein and mRNA are differentially\nexpressed (Wilcoxon BH adjusted  p  ∼ 3.703\n× 10 – \n 23 ). ARID1B, bottom row, has\na change in the correlation between tumor and normal (Δ_corr\n∼ 0.538; permutation test BH adjusted  p  ∼\n0.001). Normal samples have a noncorrelation (flat blue line), and\ntumor samples have a positive correlation (positive sloped orange\nline). This change in correlation is not associated with differential\nexpression (middle and right plots, Wilcoxon BH adjusted  p  > 0.05). The three genes were selected to showcase the diverse\ncases\nfound in the data.\nTo understand which Δ_corr values are statistically\nsignificant,\nwe calculated a  p -value using a label permutation\ntest ( Figure S3  and  Methods ). After applying a 5% p-value cutoff (BH corrected) and an absolute\nthreshold of 0.2, we find that a changing mRNA/protein relationship\nis a prominent feature of cancer. Although the numbers vary depending\non cancer type, thousands of genes exhibit statistically significant\nΔ_corr. Additionally, a majority of genes undergoing this change\nin correlation do not exhibit changes in abundance ( Figure  \n ).\nOverlap between differential\nexpression and Δ_correlation.\nFor each gene in each cancer type, we compute both the differential\nexpression and Δ_corr (see  Methods ).\nThe bar graph shows the overlap between the two metrics. Note that\na majority of genes with a statistically significant Δ_corr\ndo not display a statistically significant differential expression.\nGiven the prominence of somatic mutations in driving the cancer\nphenotype, we sought to understand the potential role of mutations\nin the dynamics of the mRNA/protein relationship. We specifically\nwanted to test whether the observed relationship changes could be\nattributed to the mutation status of oncogenes and tumor suppressors.\nFor each tumor type, we identified the most frequently mutated genes\nand compared the mRNA/protein correlation between mutant and wildtype\ntumors by computing Δ_corr and its associated  p -value (see  Methods ). In addition to identifying\nthe effect on the mutated gene, the cis effect, we also calculated\nthese metrics for all trans genes. After multiple hypothesis correction,\nvery few somatic mutations showed an effect, typically fewer than\n4 trans genes per mutated oncogene ( Figure S2 ). When compared with the thousands of genes that show a statistically\nsignificant change in Δ_corr between tumor and normal tissue,\nwe conclude that the impact of mutation status on the mRNA/protein\nrelationship is negligible.\nMoving beyond single-cancer-type\nanalyses, we next examined this relationship across multiple cancer\ntypes to see how a gene behaves in different contexts. For all five\ncancer types, we calculate the Δ_corr and associated  p -values for all genes. Genes most commonly exhibit a higher\nmRNA/protein correlation in tumor tissue, with a positive Δ_corr.\nHowever, this is both gene- and tissue-dependent; it is not universal.\nReinforcing the idea that the mRNA/protein relationship is fluid and\ncontext dependent, genes display a wide variety of patterns across\ncancer types ( Figure  \n ).\nPan-cancer changes in mRNA/protein relationships. For each gene,\nwe calculate the mRNA/protein correlation in all tissue types (5 tumor\nand 5 normal tissues). We plot the correlation coefficient and visually\nconnect the tumor/normal for each cohort. Solid lines represent statistically\nsignificant Δ_corr, and dashed lines are not statistically significant.\nGenes often display a variety of changes. The 9 genes in this figure\nwere selected to showcase the different changes in relationship from\nnormal to tumor and how they are different between tissues.\nFigure  \n  visually\nrepresents the diverse patterns of Δ_corr across various genes\nand cancer types. The BRAF gene has a weak mRNA/protein correlation\nin all five normal tissues corresponding to a general unresponsiveness\nof protein abundance to the mRNA levels. For the five cancer tissues,\nthe correlation increases. Additionally, in lung squamous cell carcinoma\n(LSCC), lung adenocarcinoma (LUAD), and head and neck squamous cell\ncarcinoma (HNSCC), the correlation change is statistically significant.\nSome genes also exhibit a decrease in the correlation in tumor tissues.\nSPTA1 has a negative Δ_corr in four of the tissue types we examined;\nonly in ccRCC is the Δ_corr positive. This diversity in regulation\nis present across all genes and tissues.\nAfter identifying specific genes with a significant change in mRNA/protein\ncorrelation via the Δ_corr metric, we examined whether these\ngenes were enriched for cancer-associated biological functions. The\nanalysis of somatic mutation data often identifies a set of genes\nwithin pathways related to cell signaling and cell cycle.  Similarly, differential expression analysis of\nCTPAC data often identifies enrichment for cell cycle, signaling and\nimmune pathways, etc.  As our Δ_corr\nmetric is a new way to describe and identify regulatory changes, we\nanticipate that these might be concentrated in cancer related genes.\nTherefore, we performed a gene set enrichment of genes in each cancer\nto identify pathways and cancer hallmarks that were enriched in changing\ngenes, defined as genes whose Δ_corr had an FDR corrected p-value\n<0.05 and an absolute threshold of 0.2 for each of the tissues.\nWe observed that the most commonly enriched pathways were metabolic\n( Figure  \n ). The most\nprominent feature of this enrichment analysis is that it identifies\nmetabolic pathways. This is an exciting result as changes in metabolism\nare well-known phenotype of tumors, \n , \n  yet this hallmark\nis rarely seen in somatic mutation based analyses.\nGene set enrichment analysis\nof Δ_corr genes. We performed\ngene set enrichment analysis for genes with a statistically significant\nΔ_corr value in each tumor type. Plotted are gene sets that\nwere significant across multiple cancers. We note that many metabolic\npathways are represented.\n\nIn cancer, cells live in an altered regulatory\nstate. Prior research\nfocused on either the presence of mutations or the relative abundance\nof mRNA or proteins. Here, we introduce a new computational method\nfor integrating transcriptomic and proteomic data. Our metric, Δ_corr,\nhighlights when the quantitative relationship between mRNA and protein\nchanges between conditions, such as a tumor and normal tissue. This\nmetric is distinct from simple differential expression and instead\nlikely points to changes in cellular regulation. We note that the\nmechanisms that govern these changes are likely to be specific to\neach gene. Just as gene transcription is controlled by a variety of\ntranscription factors and protein localizations and interactions,\nwe anticipate that the mechanisms controlling the protein-mRNA relationship\nare multifaceted. As an example of potential mechanisms, we note that\nthe abundance of P53 protein oscillates in the absence of any changes\nto mRNA abundance, and that these oscillations are coincident with\nP53 post-translational modification and a negative feedback loop with\nthe ubiquitin ligase MDM2. \n , \n  Moreover, these dynamics\nare variable by tissue type.  Additionally,\nin early development of C. elegans embryos and the formation of cell\nidentity, translational repression of mRNA transcripts inherited from\nthe egg is cell type dependent and utilizes mechanisms such as transcript\nlocalization, sequestration, and inhibition via RNA binding proteins.\nThe original hallmarks of cancer \n , \n  categorized\ncellular dysfunction into several broad biological processes, e.g.,\ncell cycle, angiogenesis, etc. The recent characterizations of tumors\nvia mutation profiles or gene expression often identify genes involved\nin only some of the hallmarks, most often those related to the cell\ncycle and cell signaling. The cancer hallmark of altered cellular\nmetabolism has rarely been found. However, the application of our\nnew metric to CPTAC pan-cancer data predominantly identifies genes\ninvolved in metabolism.\nIn addition to creating a new method\nfor integrating proteogenomic\ndata and identifying subtle changes in cellular regulation, understanding\nthe mRNA-protein relationship also has implications on future cancer\ntherapeutics. With the successful creation of mRNA therapies for COVID,  many researchers are beginning to explore mRNA\ntherapies for cancer.  In examining the\nmRNA/protein relationship in numerous tumor and normal tissues, we\nnote that the utilization of mRNA (i.e., the amount of quantified\nprotein relative to the amount of quantified mRNA) is highly variable\nacross tissues. Thus, the dosing of vaccines may need to be carefully\nexamined across tumor types.","source_license":"CC-BY-4.0","license_restricted":false}