Precision medicine programs in oncology almost universally incorporate genomic data, usually via targeted sequencing panels, or more rarely, whole genome sequencing, to detect somatic variants that aid diagnosis or choice of therapy1. Some precision medicine initiatives also concurrently measure the transcriptome, with the transcriptomic measurements used to identify cancer subtypes or the upregulation of genes whose translated proteins correspond to known anti-cancer drug targets2,3,4. However, few precision medicine initiatives routinely incorporate proteomic data at scale5,6. This is a potential issue because most targeted anti-cancer agents that are recommended for personalised treatments directly inhibit a protein or pathway in a cancer cell7.

Due to the paucity of proteomic data generated in precision oncology, ultimately transcriptomic measurements must be relied upon as a proxy for protein abundances, which typically remain unmeasured in a patient’s cancer. However, transcriptomic measurements are generally only moderately correlated with protein abundances, with correlations approximating 0.4–0.58,9,10,11,12. Both technical and biological factors can influence mRNA–protein correlations, such as post-transcriptional regulation, including protein degradation and differing rates of mRNA translation13. Several approaches have sought to address technical discrepancies, including pathway-level aggregation of expression data to reduce platform-related variation14, gene expression data harmonization through uniform shaping15, and multi-omics integration strategies16. Having an accurate knowledge of how well mRNA expression levels predict protein abundances for any given gene would enhance decision-making in precision medicine. Here we present a freely available precision oncology tool called OnCorr to interrogate mRNA–protein correlations across cancer types, available at https://oncorr.aws.procan.org.au/.

To assess pan-cancer mRNA–protein correlations, we first interrogated a multi-omic dataset of 949 human cell lines spanning 6692 proteins across over 40 cancer types8. This dataset (called ProCan-DepMapSanger) includes proteomics from data-independent acquisition mass spectrometry, acquired in a standardised laboratory17,18 across several technical replicates per cell line. The transcriptomics are from RNA-sequencing data available at Cell Model Passports19,20. Spearman’s rank correlation was used to quantify mRNA–protein concordance, as it is well suited to the variability and non-linear relationships that may be present in multi-tissue expression data, and is most commonly used in similar proteogenomic studies. The median Spearman’s mRNA–protein correlation was calculated for each gene–protein pair in this dataset across 19 tissue types that comprised measurements in more than 10 cell lines (Fig. 1a). mRNA–protein correlations in each tissue type ranged from 0.31 to 0.45, with a cohort-wide median gene-wise mRNA–protein correlation of 0.42, across 6205 genes with both mRNA and protein measurements (Fig. 1b). Next, mRNA–protein correlations from the ProCan-DepMapSanger cell line dataset were compared with two other publicly available pan-cancer studies with both transcriptome and proteome data. These include a series of cancer tissue cohorts from the Clinical Proteomic Tumor Analysis Consortium (CPTAC) (n = 1030 cancer samples available from LinkedOmics21), and a pan-cancer cell line dataset from the Cancer Cell Line Encyclopedia (CCLE) (n = 375 cell lines)9. Both datasets show a similar range in mRNA–protein correlations across tissue types (0.32–0.49; Supplementary Fig. 1a and Supplementary Fig. 1b), with cohort-wide median correlations of 0.38 and 0.46, respectively (n = 14,465 and 10,437 genes; Fig. 1b). We also compared our pan-cancer correlations with a publicly available healthy tissue dataset from Wang et al.22, comprising 29 tissue types profiled across 30 individuals with matched transcriptomic and proteomic measurements. This dataset showed a cohort-wide median mRNA–protein correlation of 0.35 (n = 9870 genes; Fig. 1b), which was lower than the pan-cancer correlations observed in cancer tissues and cell lines, but consistent with previous studies reporting reduced mRNA–protein coupling in non-malignant tissues13.

Fig. 1: mRNA–protein correlations across tissues and cancer-related pathways.
Fig. 1: mRNA–protein correlations across tissues and cancer-related pathways.
Full size image

a Distribution of mRNA–protein correlations across tissues in the ProCan-DepMapSanger dataset. Only tissues with more than 10 cell lines are shown. b Histogram of mRNA–protein correlations across each cohort for ProCan-DepMapSanger (left), Clinical Proteomic Tumor Analysis Consortium (CPTAC; middle left), Cancer Cell Line Encyclopedia (CCLE; middle right) and healthy human tissue (Wang et al.22; right) datasets, with median indicated. c Boxplot showing median mRNA–protein correlations by tissue in the ProCan-DepMapSanger dataset for key biological and selected cancer-related pathways from Sanchez-Vega et al.27 and KEGG28, displayed similarly to data shown in Ghoshdastider et al.13. d Median mRNA–protein correlations of pathways from (c) for CPTAC with ProCan-DepMapSanger (upper) and with CCLE (lower). Each data point represents the median mRNA–protein correlation across tissue types in common between the datasets.

While cancer cell lines do not always capture the underlying biology of patient tissue7, they do have the advantage of not being impacted by issues relating to tumour purity, which could have a small effect on mRNA–protein correlation13. The cell line datasets selected for this work were all measured in a single study, and, in the case of the ProCan-DepMapSanger dataset, the proteomics measurements were performed on a single instrument platform, which allows for robust comparisons of correlations between tissue types. Increasing confidence in these datasets, both cancer cell line studies showed consistent mRNA-protein correlations across key biological pathways and selected cancer-related pathways, as has been previously described in cancer tissues13, including those from CPTAC (Correlation of CPTAC with ProCan-DepMapSanger of r = 0.85 and with CCLE of r = 0.74; Fig. 1c, d, Supplementary Fig. 1c).

mRNA–protein correlations were next interrogated across genes that often act as cancer drivers, using a set of 1161 cancer genes from OncoKB23. This set combines genes from cancer sequencing panels, the COSMIC Cancer Gene Census24 and Vogelstein (2013)25. For robust analyses, mRNA–protein correlations were calculated only for the subset of these cancer driver genes in which the encoded protein was measured in more than 10 samples in a minimum of 15 tissue types (i.e., missing in fewer than five tissues; n = 261 genes). Using the ProCan-DepMapSanger dataset, hierarchical clustering revealed five clusters of driver genes with differing ranges of mRNA–protein correlations (Fig. 2a). Cancer drivers had slightly higher median correlations than genes not identified as cancer drivers (median r = 0.45 and 0.42, respectively; P < 0.01 by unpaired Student’s t-test) (Fig. 2b). Among cancer drivers, one cluster of 35 genes had the highest mRNA–protein correlations across tissue types (Cluster 3; median r = 0.68) and is enriched for genes from pathways involved in focal adhesion and cell-substrate junction (Fig. 2c). Another cluster of 35 genes had the poorest mRNA–protein correlations (Cluster 2; Median r = 0.16) and is comprised of genes from pathways involved in protein-DNA complex and chromatin (Fig. 2c). Other clusters had genes with mRNA–protein correlations that varied considerably between tissue types. These patterns of mRNA–protein correlations among cancer driver genes could be replicated in the cancer tissue dataset from CPTAC (Supplementary Fig. 2a) and the cancer cell line dataset from CCLE (Supplementary Fig. 2b).

Fig. 2: mRNA–protein correlations of cancer driver genes from OncoKB in the ProCan-DepMapSanger Dataset.
Fig. 2: mRNA–protein correlations of cancer driver genes from OncoKB in the ProCan-DepMapSanger Dataset.
Full size image

a Heatmap of mRNA–protein correlations across tissue types for cancer driver genes from OncoKB, with clusters identified by hierarchical clustering using Euclidean distance. b Violin plot showing mRNA–protein correlations of cancer driver and non-driver genes with median and standard deviation indicated. ** indicates P < 0.01 by unpaired Student’s t-test. c Distribution of mRNA–protein correlations for genes in each cluster identified in (a). Gene ontology terms with q-value < 0.05 are shown for each cluster, with all detected proteins with RNA measurements used as the list of background genes. For all plots, only genes observed in more than 10 samples in a minimum of 15 tissues (i.e., missing in fewer than five tissues) are considered.

The extent of correlation between any given gene and the protein that it encodes can be driven by both technical and biological factors12,13. Upadhya et al.26 developed a reproducibility rank to quantify the accuracy with which a protein can be measured by mass spectrometry. This ranking was shown to have a strong relationship with mRNA–protein correlation, suggesting that technical factors can influence the interpretation of such correlations. This association can also be observed in the ProCan-DepMapSanger (r = 0.49, Fig. 3a), CPTAC (r = 0.40, Supplementary Fig. 3a) and CCLE (r = 0.29, Supplementary Fig. 3b) datasets. Therefore, this reproducibility rank is a useful aid in the interpretation of mRNA-protein correlations, as an indicator of confidence in protein measurement accuracy across datasets and platforms.

Fig. 3: Integration of protein reproducibility ranks and gene-specific mRNA–protein correlations.
Fig. 3: Integration of protein reproducibility ranks and gene-specific mRNA–protein correlations.
Full size image

a Association between aggregated protein reproducibility ranks from Upadhya et al.26 and mRNA–protein correlations in the ProCan-DepMapSanger dataset. Aggregated protein reproducibility ranks are binned (upper), with Spearman’s correlation from unbinned data shown (lower). b Scatter plot of mRNA–protein correlation against aggregated protein reproducibility rank, with cancer drivers indicated by triangle data points. Labels indicate genes highlighted in the text and shown in (c). c For each gene indicated in (b), plots show the distribution of mRNA-protein correlations across tissues in the ProCan-DepMapSanger (pink solid line), Clinical Proteomic Tumor Analysis Consortium (CPTAC; light dashed line) and Cancer Cell Line Encyclopedia (CCLE; dark dashed line) datasets. Background (grey solid line) indicates the median mRNA-protein correlation from all genes in the ProCan-DepMapSanger dataset. Only tissues with data from more than ten samples are included. d mRNA–protein correlation of CTNNB1 for each tissue type in the ProCan-DepMapSanger dataset. e Distribution of CTNNB1 mRNA expression across samples from the ProCan-DepMapSanger dataset, with outliers (>3 standard deviations above the mean) indicated. f Scatterplot showing mRNA expression and protein abundance for CTNNB1 (upper) and distribution of CTNNB1 protein abundance (lower) across samples in the ProCan-DepMapSanger dataset. Outliers indicated in f are those calculated from mRNA expression data in (e).

To demonstrate the utility of mRNA–protein correlations for precision medicine, 474 cancer driver genes with robust measurements (i.e., observed with more than 10 measurements from at least one tissue type) were interrogated in detail. Of these, 129 cancer driver genes had an mRNA–protein correlation above 0.5 and a good protein reproducibility rank (>0.4; Fig. 3b). For these genes, which include ALDH2, IDH2, and EGFR (Fig. 3c), their transcriptome measurements should typically be considered a good predictor of protein abundance, and therefore high RNA expression could be reliable in a clinical setting.

In contrast, 30 cancer driver genes had low mRNA–protein correlations (<0.25) while maintaining a good protein reproducibility rank (>0.4; Fig. 3b). For these genes, which include CTNNB1 (Fig. 3c), transcriptome measurements will typically be a poor predictor of protein abundance. In a clinical setting, proteins translated from such genes may be unreliable drug targets when identified from transcriptomic measurements alone within a precision oncology pipeline. To explore this further with CTNNB1 as an example, a poor mRNA-protein correlation (median r = 0.10) is observed in most tissues (Fig. 3d). This suggests that in many tissues, outliers identified by RNA-sequencing (>3 standard deviations above the mean; Fig. 3e) are unlikely to be highly expressed at the protein level (see outlier samples by RNA-seq in Fig. 3f). Finally, other genes, such as KMT2D (Fig. 3c), have both a low mRNA–protein correlation and a poor reproducibility rank. These genes have uncertain implications for precision oncology, as a low mRNA-protein correlation may be a result of either biological factors, such as posttranslational modifications, or technical factors that influence protein measurement accuracy.

To make these mRNA–protein correlations accessible for precision medicine programs, the OnCorr web portal was built (https://oncorr.aws.procan.org.au/). This platform enables the interrogation of mRNA-protein correlations alongside reproducibility ranks within an easy-to-navigate user interface (Fig. 4). The tool integrates the ProCan-DepMapSanger, CPTAC and CCLE pan-cancer datasets, allowing the user to select their dataset of interest, as well as to filter and view results across all samples and by tissue type. This web portal is freely available and designed to be interpretable by oncologists, researchers and data curators.

Fig. 4: Example output from the OnCorr web tool.
Fig. 4: Example output from the OnCorr web tool.
Full size image

A screenshot from the OnCorr web tool showing mRNA–protein correlations for EGFR, with tissue-specific data from the lung.

This study and associated data interpretation have several limitations. For some genes, different mRNA-protein correlations are observed across datasets. This could be a result of differences in cell culture conditions, proteomic data acquisition methods or differences in the cancer types that make up each tissue annotation. Also, many of the datasets included in this study compare RNA-sequencing data acquired from a temporally or spatially different specimen than was used for the mass spectrometry analyses, leading to a false reduction in the measured mRNA–protein correlation. Finally, some genes, such as CDK4 (Supplementary Fig. 3c), show patient outliers in mRNA expression (Supplementary Fig. 3d) that remain highly expressed at the protein level (Supplementary Fig. 3e) despite what would be suggested by the overall mRNA–protein correlation and reproducibility rank (CDK4: r = 0.15 correlation and 0.51 reproducibility rank). Finally, due to limited sample sizes within individual cancer subtypes and according to specific genomic alterations, analyses were performed at the tissue level to increase statistical power, which may consequently mask subtype-specific patterns for some genes. Therefore, inclusion of mRNA-protein correlations in precision oncology should be considered as one piece of information alongside several other metrics that, together, will drive the choice of targeted agents from a patient’s molecular profile.

In summary, OnCorr can be used in precision oncology initiatives to augment transcriptomic findings in the absence of proteomic data to inform a patient’s personalised treatment options. It can also be adopted for research to interrogate mRNA-protein correlations across three independent pan-cancer datasets. OnCorr is available at https://oncorr.aws.procan.org.au/.

Methods

Data sources

All data analysed in this study are publicly available from four independent datasets. The proteomics data from ProCan-DepMapSanger are available from Gonçalves et al.8. The matched transcriptomic data were obtained from Cell Model Passports19,20, which provides a harmonised RNA-sequencing dataset integrating data from the Wellcome Sanger Institute and the Broad Institute. Data from the Clinical Proteomics Tumor Analysis Consortium (CPTAC) were downloaded from the LinkedOmicsKB database21. Data from the Cancer Cell Line Encyclopedia (CCLE) are available from Nusinow et al.9. Tissue type annotations were harmonised as in Supplementary Table 1. Correlation values for healthy tissues were obtained from the supplementary materials of Wang et al.22.

Correlation analysis

To calculate mRNA–protein correlations, a Spearman’s rank correlation was performed for each mRNA–protein pair across all samples. Analyses were restricted to proteins with measurements in more than 10 cell lines for both RNA-seq and proteomics datasets to reduce any biases arising from proteins with measurements only in very few samples. For correlation analysis of mRNA–protein pairs within each tissue, cancers were grouped based on their tissue of origin and analysis was restricted to tissues that contained more than 10 cell lines in the relevant dataset.

Functional enrichment

For biological pathways shown in Fig. 1d and Supplementary Fig. 1c, pathways and their associated genes were retrieved from Sanchez-Vega et al.27 and KEGG28 pathways, where Homo sapiens genes were retrieved from the msigdbr package (https://cran.r-project.org/web/packages/msigdbr/index.html). The median Spearman correlation rank was calculated for genes in each pathway per tissue type.

Analysis of cancer driver genes

A list of 1169 cancer driver genes was obtained from the OncoKB portal23. For heatmap and clustering analyses, genes were retained if they had non-missing mRNA–protein correlations in ≥15 tissue types (i.e., missing in fewer than 5 out of 19 tissues). The heatmap in Fig. 2a was plotted using the pheatmap package (https://cran.r-project.org/web/packages/pheatmap/index.html), which uses complete hierarchical clustering and Euclidean distance as a similarity measure. The function cutree with the option k = 5 was used to retrieve the driver genes in each cluster. Gene ontology analysis for each cluster was performed using clusterProfiler29, using all detected proteins with RNA measurements as the list of background genes. Only ontologies that had an FDR-adjusted p-value < 0.05 were retained.

Correlation analysis with reproducibility ranks

Aggregated protein–protein reproducibility ranks were retrieved from Upadhya et al.26. Ranks were binned for plotting, and Spearman’s correlations were reported from unbinned data.

Statistical analyses

All analyses were carried out in Python version 3.12.1, and R version 4.3.2, using the pandas library (Python), and dplyr, tidyverse and reshape2 (R). All plots were made using ggplot2 (https://cran.r-project.org/web/packages/ggplot2/index.html), and extensions ggdist (https://cran.r-project.org/web/packages/ggdist/index.html), ggpubr (https://cran.r-project.org/web/packages/ggpubr/index.html), and ggbeeswarm (https://cran.r-project.org/web/packages/ggbeeswarm/index.html).

Shiny app development

An interactive web application was developed in R (version 4.3.2) using the Shiny framework to enable exploration of mRNA–protein correlations across the datasets described above. The app uses dplyr, tidyverse and reshape2 for data handling, plotly (https://cran.r-project.org/web/packages/plotly/index.html) and ggiraph (https://cran.r-project.org/web/packages/ggiraph/index.html) for interactive visualisation, and an SQLite database accessed through DBI to efficiently query only the gene-dataset-tissue subsets requested by the user, minimising memory load. A lightweight lobstr-based logger was used to monitor memory at key steps. All plots and tables update reactively based on user inputs within a tab-based interface, providing a responsive platform for interrogating large-scale expression and correlation data.