P-MartCancer–Interactive Online Software to Enable Analysis of Shotgun Cancer Proteomic Datasets

The use of mass spectrometry (MS)-based technologies for global protein profiling of cancer-related tissues and bodily fluids has become a major focus in research centered on biomarker discovery to better detect and treat cancer. Global proteomic technologies are of interest in the field of cancer research because they provide a vital source of information regarding biological functions at the protein level (1–4). Global MS-based proteomics often allow hundreds of thousands of peptides mapping to tens of thousands of proteins to be measured, offering scientists an unprecedented view into processes involved in cancer development and progression. However, as with many newer high-throughput molecular technologies, the data are complex with multiple sources of variability (1); consequently, the postpeptide identification and quantification (i.e., downstream) analyses of these datasets require significant specialized expertise due to inherent data challenges, such as missing values and isoform quantification (5–7). Data processing is generally performed via custom and unpublished scripts, leading to continued issues in reproducibility of results (8). Thus, challenges in both access to the data and statistical methods in an easy to understand format for biomedical researchers has led to underutilization of publicly available databases, such as those generated through the Clinical Proteomics Tumor Analysis Consortium (CPTAC) funded through the NCI (Rockville, MD). The development of software that enables continued exploration and evaluation of existing data could increase the potential for new discovery from these comprehensive public datasets.

Existing software capabilities associated with CPTAC can broadly be placed in four categories for global MS-based data (https://proteomics.cancer.gov/proteomics/background): (i) spectral preprocessing; (ii) peptide and protein identification; (iii) specific methods for qualitative and quantitative differential statistics; and (iv) network analyses. P-MartCancer broadly fits in the third category; however, unlike these targeted analyses on data that have already been processed into a particular form, P-MartCancer offers a holistic approach to data analysis, allowing all steps of analysis from quality control through pattern discovery to be performed in a workflow-based manner. In addition, the manner that P-Mart offers direct access to the data for statistical analysis at multiple levels (peptide, gene, and protein), with the clinical information aligned, is unique within the set of tools supporting CPTAC. P-MartCancer provides an open, web-based interactive platform for performing quality control processing, statistics, protein quantification, and exploratory data analysis tasks in a manner that is reproducible (see Supplementary Video S1).

P-MartCancer functions are developed in R or Rccp, Rserve (https://www.rforge.net/Rserve/) is used to communicate between R and the web-service, and the interface is developed in Java. A subset of the R functions is currently available via GitHub (https://github.com/pmartR/), and the web service can also be installed via Docker Hub (https://hub.docker.com/r/pnnl/pmart-web/). Developed in this manner, adding functionality to P-MartCancer is straightforward through a standardized pipeline.

P-MartCancer currently accesses multiple proteomic datasets generated through the CPTAC available on the Data Portal (https://cptac-data-portal.georgetown.edu/cptacPublic/). Data are available at the peptide, protein, and/or gene levels where the protein and gene data are based on a defined Common Data Analysis Pipeline (CDAP; ref. 9). P-MartCancer offers flexibility to the user to either perform statistical processing on the peptide data (before or after parsimony) with P-MartCancer functions for gene or protein quantitation, or to use the data as supplied by CPTAC via the CDAP. Currently, various numbers of datasets are available for ovarian cancer and breast cancer (2); however, new datasets are being added as they become available. Each of these datasets contains metadata about the experiment and the user selects the clinical variable of interest (e.g., vital status, tumor stage) for data processing, allowing various hypotheses to be explored.

P-MartCancer is a modular workflow tool (Fig. 1A) with four primary capabilities: (i) quality control processing; (ii) gene or protein quantification; (iii) statistics; and (iv) exploratory data analysis. These are further divided into six key modules under which new functionality can be easily added through the R/Rserve to Java framework described previously. The modules and functions available depend upon the type of data being evaluated: peptide, protein, or gene. Each function called is displayed visually either via tables or figures in sequential order highlighted by the green text on top of the P-MartCancer screen where the current module is blue (Fig. 1B, top). Finally, the entire process is documented for the user at the end of the workflow, Fig. 1C, and all datasets and statistical results are available for download as .CSV files.

A challenge with proteomics data is preprocessing in a manner that does not ignore the different sources of variability that contribute to the complexity of these datasets. For example, peptides are not uniformly identified for all samples, and thus, large quantities of missing values are common, due to both random and nonrandom mechanisms. P-MartCancer offers a suite of preprocessing capabilities that handle issues of peptide and protein coverage, as well as the identification of potential outlier samples (7, 10, 11). At the peptide level, researchers can choose to (i) remove proteins with inadequate coverage (peptide filter); (ii) remove samples with outlier behavior (sample outlier filter; ref. 10); (iii) remove peptides too sparse for statistical analysis (peptide coverage filter); (iv) remove peptides with extremely high variability in the context of a coefficient of variation (CV filter); and (v) perform normalization. The quality control processing generates a high-quality dataset for continued statistical analysis. The gene- and protein-level datasets will allow functions 2–4 above, evaluating outlier behavior and removing genes or proteins that are too sparse or variable to add value to downstream statistics.

Statistical analysis of peptide, gene, or protein-level data is currently focused on quantitative ANOVA-based methods and qualitative G-test methods (11). The ANOVA method allows the comparison of any number of groups, performing a multiple test correction when more than two groups are compared (e.g., clinical variable “tumor residual disease” is separated into four categories by size). A Tukey adjustment is performed when the user compares all groups with one another, and a Dunnett adjustment is performed when the user compares back to a single control group. Data are not imputed; statistical results are generated based only on the observed data to assure that accurate estimates of variance are being utilized for these tests. To identify qualitative changes, a G-test is also performed for each biomolecule to evaluate whether the number of nonmissing observations in one group is more than expected by chance. Multiple test adjustments for the G-test are performed using a Holm–Bonferoni correction. The total significance is given in the context of a bar graph, and to facilitate exploratory data analysis capabilities a P value threshold (default of 0.05) is used to move forward only a subset of the peptides, genes, or proteins for further evaluation.

There are numerous approaches to quantify proteins from the measured peptide-level data (12). P-MartCancer currently offers a standard reference–based approach that scales all peptides to the most abundant, or most reproducible, peptide and gives the median signal (13).

P-MartCancer offers two exploratory data analysis capabilities. The first is probabilistic principal component analysis, which allows P-MartCancer to perform PCA without imputing missing values, demonstrated to be valuable in proteomics (7). The resulting scores are plotted using a standard scatter plot of the scores from the first two principal components that most cancer and biomedical researchers are accustomed to, allowing visual exploration of clustering across samples.

P-MartCancer also offers an interactive and customizable plotting capability called Trelliscope that allows sorting and querying across the peptides, genes, or proteins, Fig. 1B. Trelliscope uses the statistical results to plot each peptide, gene, or protein via either a boxplot of differential abundance or a bar graph of the number of observations, to view quantitative and qualitative changes, respectively. The entire space of the biomolecules being explored can be reduced by selecting various thresholds, such as P value or fold change, or the user can search for specific genes or proteins of interest. For each plot, the gene and protein information can be selected, and the associated information can be viewed in webpages, http://www.genecards.org and http://www.uniprot.org, respectively. In the example in Fig. 1B, the user has searched for BRAF and rapidly views the associated protein-level quantified information based on the clinical variable selected, “macroscopic disease,” and other information, such as that BRAF has a P value of approximately 0.04 for the specific comparison selected.

P-MartCancer offers a new online platform to access CPTAC datasets to enable new analyses. There is a wealth of capabilities that could be extremely useful to the proteomics community, many of which are under active development. For example, proteoform discovery, which is the identification of proteins with multiple forms, is also an important component of protein quantification (6). Additional future work is focused on adding new capabilities in statistical testing, machine learning, and gene set enrichment analysis, as well as the development of a user-upload capability to enable all researchers with MS-based peak-intensity data to create reproducible statistical downstream processing pipelines.

No potential conflicts of interest were disclosed.

Conception and design: B.-J.M. Webb-Robertson, J.L. Jensen, K.G. Stratton

Development of methodology: B.-J.M. Webb-Robertson, L.M. Bramer, J.L. Jensen, M.A. Kobold, K.G. Stratton

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): B.-J.M. Webb-Robertson, L.M. Bramer, K.G. Stratton, A.M. White

Writing, review, and/or revision of the manuscript: B.-J.M. Webb-Robertson, L.M. Bramer, K.G. Stratton, K.D. Rodland

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): L.M. Bramer, M.A. Kobold, K.G. Stratton, K.D. Rodland

P-MartCancer was developed at Pacific Northwest National Laboratory (PNNL), a multiprogram national laboratory operated by Battelle for the U.S. Department of Energy under contract DE-AC06-76RL01830.

This work was supported by NCI grant U01-1CA184783 to B.-J.M. Webb-Robertson.