Deciphering Phenotypes from Protein Biomarkers for Translational Research with PIPER

Abstract

Liquid chromatography-multiple reaction monitoring mass spectrometry (LC-MRM) has widespread clinical use for detection of inborn errors of metabolism, therapeutic drug monitoring, and numerous other applications. This technique detects proteolytic peptides as surrogates for protein biomarker expression, mutation, and post-translational modification in individual clinical assays and in cancer research with highly multiplexed quantitation across biological pathways. LC-MRM for protein biomarkers must be translated from multiplexed research-grade panels to clinical use. LC-MRM panels provide the capability to quantify clinical biomarkers and emerging protein markers to establish the context of tumor phenotypes that provide highly relevant supporting information. An application to visualize and communicate targeted proteomics data will empower translational researchers to move protein biomarker panels from discovery to clinical use. Therefore, we have developed a web-based tool for targeted proteomics that provides pathway-level evaluations of key biological drivers (e.g., EGFR signaling), signature scores (representing phenotypes) (e.g., EMT) and the ability to quantify specific drug targets across a sample cohort. This tool represents a framework for integrating summary information, decision algorithms and risk scores to support Physician-Interpretable Phenotypic Evaluation in R (PIPER) that can be reused or repurposed by other labs to communicate and interpret their own biomarker panels.

Graphical Abstract

Introduction

Molecular diagnostics for lung cancer have become increasingly complex. Multiplexed genomics assays can be complemented with the proteomic phenotypes to further support clinical decision-making, yet these increases in data produce additional barriers to the effective use of this information in the clinic. While individual lab test results can be communicated effectively from the clinical lab to the physician and patient, cancer researchers and treating physicians are now faced with a large variety of therapeutic options that have specific companion diagnostics. To make the information available and to support ranking of therapeutic options to maximize expected patient benefit, complex proteomics and genomics data must be related to the physicians in an accessible way that enables thorough review. Using lung cancer as an example, detailed molecular characterization of each individual tumor is necessary to rank therapies according to their expected benefit to the patient. Numerous treatment options are now available, including chemotherapy, targeted therapy, and immunotherapy as well as novel approaches in clinical trials. In addition, some tumor subtypes (e.g., adenocarcinomas driven by ALK fusions) are observed only in a small proportion of the patient population. Targeted proteomics and genomics panels including drug targets and known biomarkers can be used to characterize genes and proteins and help direct treatment for each individual cancer patient. However, the results often indicate multiple options could be effective, so better methods are needed to triage patients for FDA-approved companion diagnostics, specific treatment regimens, and clinical trials.

Genomics and transcriptomics both already have strongly established clinical utility in cancer treatment. Genomic approaches (e.g., the Illumina TruSight Tumor 170) assess mutation status, fusions due to chromosomal translocation, amplifications, insertions/deletions, and other aberrations that produce tumor drivers or provide additional clinically relevant information. These data are reported as positive or negative results (e.g., the EGFR L858R mutation is present in a tumor), which are straightforward to communicate on a one-by-one basis but may be hard to evaluate in aggregate if multiple drivers are detected. For clinical use, transcriptomics uses full gene expression profiles, RNA sequencing data (RNAseq), or large targeted panels to derive signatures that assess risk. While multiple signatures can be determined from these multiplexed panels, the clinical reporting is frequently condensed to a single score, as in the PAM50 and the 70 gene MammaPrint signature (provided by Agendia) for prediction of breast cancer recurrence ^1-3, the (22 gene signature called Decipher-Prostate⁴ that predicts prostate cancer metastasis (offered by Veracyte Labs) and the 70 gene signature Myeloma Prognostic Risk Score or MyPRS (from Signal Genetics) ⁵. While this strategy has been effective for clinical transcriptomics, proteomics is routinely called upon to provide additional value over the existing techniques. In particular, proteomics can provide both information on specific targets (e.g., protein levels) as well as signature scores. In addition, targeted proteomics quantifies biomarkers as continuous variables that may need to be combined based on our knowledge of tumor biology and clinical performance of different therapies. Therefore, a proteomics report should utilize multiplexed biomarker data to bridge the current gaps by providing easily interpreted scores, quantification of specific drug targets, and pathway integration to help assess which therapy will provide the most benefit for each patient.

Targeted proteomics has been successful in developing assays for routine clinical use, which provide advantages or additional data when compared to traditional antibody-based techniques. Examples include Thyroglobulin for detection of thyroid cancer recurrence ^6-8, HER2 quantification for selection of breast cancer therapy ^{9, 10}, PTEN as a biomarker for breast cancer ¹¹, and cytokeratins for evaluating tumor histology ^{12, 13}. In much the same way that targeted therapy has moved beyond the one drug-one target thought process, targeted proteomics in cancer must also examine tumor phenotypes in more detail to provide context to the biomarker measurements. As an example, receptor tyrosine kinase quantification needs to be performed as a panel to evaluate the potential for compensatory signaling as well as evaluating whether the tumor has undergone epithelial-to-mesenchymal transition, which provides another mechanism for therapeutic escape. To that end, a panel with 97 protein biomarkers has been developed for tumor assessment, which includes measurements for tissue quality control (QC) ^{14, 15}, cancer signaling proteins ^16-18, immune cell surface markers, clinically actionable immune checkpoint proteins ^{19, 20}, and antibody-drug conjugate targets. This platform was applied to tumor proteomes from 108 lung squamous cell carcinomas that were previously profiled with proteogenomics ²¹.

Using this example dataset, we illustrate the use of a framework for visualizing and ultimately interpreting proteomics panel measurements with the translational goal of determining clinical utility. The development of biomarkers using proteomics typically follows the path illustrated in Figure 1. Discovery of biomarkers and/or tumor phenotypes occurs from diverse strategies, including biological studies, proteomics analyses, and clinical trial correlates. Several tools are already available for identifying candidate biomarkers. However, verification, analytical validation, and clinical validation require extensive collaboration with clinical faculty (including but not limited to molecular pathologists and treating physicians) as well as prospective data collection for verification cohorts or in a clinical trial to test the efficacy of the biomarker for patient selection. To support this process, we have started the development of an application, Physician-Interpretable Phenotypic Evaluation in R (PIPER), which enables the user to review the targeted proteomics profile of each individual tumor in the context of the overall population for patient classification. PIPER has been developed to complement previous software for visualization of raw and processed proteomics data ^22-25 and targeted proteomics data analysis tools like Skyline ²⁶. A variety of visualization approaches were implemented to enable multiple stakeholders to review and interpret this complex, multi-analyte data. With the current LSCC dataset and PIPER v. 1.0 reported here, protein biomarkers can be linked to tumor phenotypes (Figure 1). We provide a framework in R to demonstrate different ways to evaluate these datasets that could be adapted to suit the needs of other investigators at different stages of candidate biomarker evaluation. For instance, the analysis of samples collected before and after treatment would enable linkage between protein biomarkers and patient outcomes, which will be complemented by the addition of a patient outcomes layer in the future development of the PIPER software. Currently, we envision that PIPER can use protein biomarker data from targeted proteomics to visualize clinically relevant tumor phenotypes and that ultimately the software can be developed by integration of patient outcomes to help select therapies for each individual patient.

Figure 1: Development of Protein Biomarkers from Discovery to Clinical with PIPER Software Support for Data Interpretation.

Example pathways for biomarker development start with biological studies or clinical correlates in treatment trials for discovery. Targeted proteomics provides an ideal tool for translate these discoveries to clinical use. Clinical studies using targeted proteomics can link the biomarkers to phenotypes and ultimately responses using prospective cohorts, leading to a biomarker-guided clinical trial. While unified software to take protein biomarkers from discovery to clinical utility is the goal, the current version of PIPER software makes targeted proteomics datasets more accessible for translational researchers during the verification and validation process. The steps shown in bold indicate the current applications for this software; future addition of the layer of patient response will also support interpretation of the biomarkers’ clinical utility (italics).

Experimental

Example Dataset

Based on extensive discussions with different groups of physicians at Moffitt, 97 protein biomarker targets were selected based on their utility to pathologists and oncologists for tumor phenotyping as well as direction of either targeted therapy or immunotherapy. Peptide selection was performed using spectral libraries created from in-house label-free expression proteomics experiments of different human tumors; consistency of detection and magnitude of ion signal intensity were used to select doubly protonated peptides that were good candidates for LC-MRM analysis. An in-house dataset quantifying 97 proteins in 108 lung squamous cell carcinoma samples was generated with liquid chromatography-multiple reaction monitoring mass spectrometry analysis of 1 microgram of total tumor protein digest spiked with 20 fmol of each stable isotope-labeled standard peptide. Both the cohort of 108 lung squamous cell carcinomas as well as the proteomics sample preparation have been described in a previous publication describing proteogenomics; ²¹ de-identified patient number is the same in both studies to enable data comparison. The LC-MRM data acquisition parameters have also been previously published ¹⁸. Skyline software was used for quality control, peak selection, and relative quantification. Data were exported from Skyline²⁶ and imported into Excel to calculate protein expression based on comparison of ion signal for the proteolytic peptide from the endogenous protein to the stable isotope-labeled standard peptide. Values were log₂ transformed prior to utilization in PIPER (see Supplementary Table 1), and missing data (ND, not detected) were not imputed. Data have been uploaded to Panorama Public²⁷ (https://panoramaweb.org/Moffitt-LSCC_MRM.url) and ProteomeXchange²⁸ (PXD040680, available at https://proteomecentral.proteomexchange.org/cgi/GetDataset?ID=PXD040680).

Visualization Library

An R-based software package was developed (piper, source: https://github.com/steveneschrich/piper) using R 4.2.0. Visualizations in the piper library were implemented using Bioconductor (3.15), ggplot2 (3.3.6) ²⁹, ComplexHeatmap (2.12.1), and ggradar(0.2). The software was developed for the LC-MRM data used as demonstration, but it can be used with other data as well. Missing values are managed differently, depending on visualization: imputed as 0^th percentile in the spider plot, marked as “ND” (0 value) in lollipop plot, and indicated at 0 value in the rug plot.

Visualization Tool (PIPER)

An R-based Shiny (v1.7.2) application (PiperApp, https://github.com/steveneschrich/PiperApp) was developed to interactively visualize the LC-MRM data using the piper library. The software was developed under R 4.2.0. Data from the LC-MRM cohort were loaded, and the PIPER library is used for visualizations. Panel measurement statistics are computed on-the-fly for each peptide measurement and include the maximum, minimum and mean expression values (excluding missing values). The empirical cumulative distribution function is computed, and quantiles of measurement are calculated using only those samples that had detectable levels of the protein biomarker (see Supplementary Table 1).

Results and Discussion

Selection of therapy in cancer provides an excellent analytical challenge that also needs bioinformatics support for data visualization and interpretation. Different pillars of treatment are available for cancer patients: surgery, radiotherapy, targeted therapy, and immunotherapy, as well as emerging areas like metabolic therapy. Here, the focus is put on targeted therapy and immunotherapy, where targeted proteomics measurements can have an immediate impact. Examples of different known biomarkers and candidate phenotyping proteins are illustrated in Figure 2. The targets were chosen based on known cancer biology as well as established clinical value, particularly in lung cancer, that fits the needs of current diagnostic strategies. The inner circle is separated into three categories: Phenotyping, Targeted Therapy, and Immunotherapy. The second ring indicates selected panels that could be used in each category. For example, Phenotyping includes quality control, histology verification, epithelial-to-mesenchymal transition, proliferation, and metabolism (exemplified by the glycolytic enzyme, GAPDH). Targeted Therapy includes cancer signaling proteins with FDA-approved inhibitors available for lung cancer. Immunotherapy includes immune cell markers, checkpoint blockade targets with FDA-approved antibody therapies, targets of antibody-drug conjugates, and cancer antigens that could be attacked with cellular therapies. The outer ring then contains sets of individual protein biomarkers included in each panel for the dataset used to illustrate the utility of PIPER. These panels can be configured to include additional biomarkers, as needed.

Figure 2: Support for Selection of Appropriate Cancer Therapy Requires Measurement of Large Biomarker Panels.

Current strategies for evaluation of protein biomarkers in tumor tissue rely on a battery of parallel tests (e.g., multiple immunohistochemistry experiments on individual tissue sections). In addition to monitoring specific biomarkers with highly multiplexed targeted proteomics strategies, the context of tumor phenotype can also be established. To fully examine the protein biomarkers for available therapies, a comprehensive panel would include diverse proteins contributing to tumor phenotypes (blue), cancer signaling and resistance to targeted therapies (green), as well as immune markers and mechanisms of immune escape (orange). Example proteins are listed here for lung adenocarcinoma, and the list would need to be tailored for each tumor type.

Specific goals for utilization of the targeted assay platform range from tissue quality control to detection of proteins that can be inhibited with targeted therapy or serve as antigens for immunotherapy. In addition to measuring the individual biomarkers, panels and pathways can be derived to provide additional information for selection of therapy. Building on the example listed in the Introduction, epidermal growth factor receptor (EGFR) inhibitor treatment would be expected to be effective in lung adenocarcinoma with an epithelial subtype, so the panel should indicate high EGFR expression as a potential biomarker with confirmation of high cellularity, adenocarcinoma as histology, epithelial subtype, and low expression of compensatory tyrosine kinases, For immunotherapy, the combination of high amounts of infiltrating immune cells detected with cellular markers (e.g., CD4 or CD8) with high expression of a cancer antigen could predict positive response to cellular immunotherapy, because the tumor would be classified as “immune hot.” ^30-32 If developed effectively, these data should be able to help triage diverse treatment options, select patients for clinical trials, measure correlates in clinical trials, and provide a molecular rationale for predicting responses to treatment and categorizing patients for the therapy that best meets their needs.

To accomplish these goals, an application must assist with navigation through highly multiplexed datasets, enable comparison of an individual patient’s tumor to a larger sample population, and reduce complex biomarker data to straightforward scores (Figure 3). Selected data from an individual patient have been displayed as a list with a color scale to show high levels (red) to low levels (blue) of protein biomarker expression (Figure 3A); this panel can be readily reviewed by a physician to chart the expression of specific biomarkers and their quantile ranking in one patient’s tumor. For example, the patient’s tumor in Figure 2A has high levels of EGFR expression and relatively high levels of mitogen-activated protein kinases compared to the other tumors in this cohort; expression of Bcl-2 and Mcl-1 are also expressed in high quantiles : 89 and 99, respectively. Initial insights can be made at this level for individual biomarkers for a given patient. The overall dataset is displayed as a heat map similar to the visualization strategy typically employed for proteomics publications (Figure 3B and Supplemental Figure 1); this view is highly complex and obscures the individual data but does enable the viewer to understand how the protein biomarkers can be used for patient stratification or explore relationships between protein biomarker(s) and either molecular subtype or mutation status. Reduction to (pathway) scores and cutoffs similar to the metrics used for predicting patient risk based on cholesterol levels are also needed to effectively communicate complex biomarker data (Figure 3C). PIPER is being developed to enable translational biomarker research, to view the data at different levels with selection by patient as well as by biomarker panel, and to assist with the development of scores and manual determination of thresholds for biomarker expression that could ultimately support clinical decisions once integrated with outcome data (see Figure 1). The current version of the software focuses on linking protein biomarkers to phenotypes, while future additions to PIPER will incorporate patient outcomes to support the next steps of protein biomarker development.

Figure 3: Protein Biomarker Data from Targeted Proteomics Should be Accessible at Multiple Levels to Support Translational Research Prior to Evolution of Medical Decision-Making Tools.

Panel results from an individual tumor (A) provide the measurements and quantile of each individual protein biomarker (see also Supplemental Table 1). Visualization of the entire dataset (A) for a patient cohort (n=108) enables categorization by molecular subtype (B) and Supplemental Figure 1, but the data may also need to be compressed into scores, like the simple readout of a cholesterol test (C), using the analogous strategy already employed by transcriptomics. This gap needs to be bridged for proteomics to play an effective role in precision medicine for cancer patients.

The PIPER application (https://steveneschrich.shinyapps.io/Piper/) was developed in Shiny and works in different browsers; Edge, Chrome, Safari, and Internet Explorer have been tested. The landing page architecture is illustrated in Figure 4. The dropdown boxes enable selection by patient and biomarker panel, and the tabs provide the different mechanisms for visualization. The subway map view provides another visualization of the biomarker panels to complement their illustration in Figure 2. Scores are overlaid on each stop of the subway map for a summary view using color coded values for high to low expression or pathway scores. In this example for the tumor from Patient 3, the tissue quality control or cellularity metric, is high, while the tumor has low expression of proliferation markers and GAPDH as a Metabolism marker. Although this tumor scores high for immune cells, the expression levels for antibody-drug conjugate targets and cancer antigens are low. This “at-a-glance” interpretation of the overall tumor phenotype can be tested to see if it is effective in supporting selection of therapy and then more fully developed with cutoff values for protein expression to inform clinical decision-making. Additional specific details can be obtained by selecting a specific node and using one of the other visualization tools. In this example, the lollipop plots for the tumor from Patient 3 has 80^th percentile expression of beta actin, and 40^th and 41^st percentile for hemoglobin alpha and beta, indicating high cellularity and lower amounts of blood than most other tumors in the cohort.

Figure 4: PIPER Software Architecture Enables Navigation from Individual Patients and Biomarkers to Visualization of Pathways and Entire Datasets.

As shown in Figure 2, different views of the data need to be available to switch between individual biomarker measurements, pathways, and the overall dataset. Data can be displayed for an individual patient and then further selected by the group of target biomarkers. The Overview tab describes the assay platform, and the Panel Measurements tab provides the data for each individual patient. Different strategies for data visualization are included that can be selected for each biomarker panel based on the user’s preference (radar, donut, lollipop, and rug plots are currently available). The overall data are available in the Heatmap tab. The different biomarker panels are presented as a subway map to enable selection of each category and an overall readout of the scores.

The derivation of scores for these biomarker panels remains ongoing but includes several different strategies that can be tailored to each biomarker panel. For example, tissue quality control will examine a balance between cellular proteins and blood proteins and could be summarized as a ratio (e.g., beta actin to hemoglobin alpha protein levels) to determine whether the tumor specimen has sufficient cellularity for highly confident data interpretation. EGFR inhibitor selection would use several metrics, as described above, but score could be derived as a ratio of EGFR expression to total receptor tyrosine kinase expression to indicate the level of compensatory signaling expected due to the levels of MET^33-35 and UFO/AXL³⁶ expression. Cancer signaling pathways (e.g., MAPK or PI3K/AKT/MTOR) could be integrated using all members, though inhibitory elements in the pathway could be added as metrics for the intact pathway or subtracted as limiting to the signaling, depending on the biology and the clinical evidence of their value. Immune cell markers should be evaluated in total for scoring similar to “immune hot” and “immune cold” metrics calculated from gene expression signatures but could be further refined by levels of T cell infiltrates and B cell inclusion in the tumor from targeted proteomics data; the expression level of immune cell markers (e.g. CD3, CD4, CD8, and CD20) can be directly interpreted to indicate the levels of immune cells in the tumor and establish the phenotype of the tumor-immune microenvironment as “hot” or “cold.” ^30-32 Cancer antigens need only be presented first as detectable or not detectable and subsequently in rank order of highest expression level within each tumor. The following paragraphs will explore these test cases in additional detail.

Similar to the different methods for calculating the scores, different visualization strategies are necessary for different elements of the dataset and will have different levels of accessibility to translational researchers and physicians. Pathway-level displays are shown for the tissue quality control biomarkers in Figure 5. Both lollipop and rug plots enable comparison of an individual sample to the entire cohort. In this case, high levels and quantiles for cellular proteins and low levels and quantiles for albumin and hemoglobin indicate high cellularity for this tumor sample compared to other LSCC tumors that were analyzed. This metric could be the first stop on many decision trees for tumor phenotyping and selection of therapy, because it confirms sufficient cellularity to produce effective measurements of the cellular protein biomarkers in the assay panel. This type of display is also amenable to the comparison of individual biomarkers and pathway scores, as shown in Figure 6. The tumor from patient 15 in this cohort was PTEN null at the protein level (left), while the tumor from patient 95 had high levels of PTEN expression compared to the rest of the cohort (right). Furthermore, levels of protein expression across the PI3K/AKT/MTOR pathway were also higher for the tumor from patient 95 compared to the tumor from patient 15; these data could be exploited to determine whether certain patients would benefit from targeted therapy based on the pre-treatment signaling phenotypes interpreted from these protein expression panels. These visualizations could fuel evaluation of differences between responders and non-responders by providing more information than individual biomarkers.

Figure 5: Illustration of Pathway Level Data Display with Tissue Quality Metrics.

To evaluate related biomarkers, multi-panel views can be organized as either lollipop plots (A) and rug plots (B). In both graphs, log2 transformed values for protein expression in attomoles per microgram of total protein are displayed on the x axis. The lollipop plots also include the quantile (the value inside the circle) for the expression level for this tumor to show the relationship to the rest of the cohort. The rug plots show each individual measurement (black tick marks) for a given protein biomarker with the individual patient’s data highlighted (paired red triangles). Both views enable comparison of the results for an individual tumor to be compared across the patient cohort. This example displays the readout of high cellularity (e.g., ACTB) and low blood content (e.g., HBA) for this sample compared with the other LSCC tumors in this cohort. Based on these metrics, the sample would pass quality control and be expected to have fewer missing values in the rest of the dataset than a sample that had low cellularity.

Figure 6: Combined Evaluation of an Individual Biomarker, PTEN, and the PI3 Kinase Pathway.

Two different patients’ tumors in the cohort can be distinguished as PTEN null (top) and PTEN high (bottom). PTEN status should be confirmed with genomics, when available. Furthermore, the data for expression of related proteins in the PI3 kinase-AKT-MTOR pathway can also be displayed to show differences in the expression level of proteins throughout the pathway, leading to differences in the individual biomarker and the pathway score.

Other phenotyping comparisons could benefit from different readouts. While tumors have been previously classified as immune hot and immune cold, the different cluster of differentiation (CD) protein markers can be used to indicate the presence of different cell types. CD3, CD4, and CD8 are included to monitor T cells, while CD20 is used to evaluate the presence of B cells. Radar plots enable the identification of immune hot tumors with high expression of all markers (Figure 7A) and immune cold tumors with low expression levels of these markers (Figure 7B). The tumor in the top panel shows distinct populations of CD4⁺ T cells, CD8⁺ T cells, and CD20⁺ B cells, while the tumor in the bottom panel shows only low level detection of the pan-T cell marker, CD3. Tumors with high levels of particular tumor-infiltrating lymphocytes can be matched to different immune therapies, while “immune cold” tumors with fewer infiltrating lymphocytes may not be as effectively treated with those same strategies.

Figure 7: Radar Plots Enable Evaluation of Individual Biomarkers Related to Immune Cell Infiltration in the Tumor.

Quantile values are plotted for CD20, a B cell marker, and selected T cell markers, including CD3, CD4, and CD8. One tumor would be classified as immune hot (A) by detection of markers for all cell types, whereas the other tumor could be classified as immune cold, because CD3 was the only marker detected (B).

Tumor antigens provide another example for visualization; these proteins can be assessed in parallel in the multiplexed LC-MRM data and ranked for each tumor based on expression level and the rank of the tumor’s level of expression compared to the rest of the sample set (Figure 8). This strategy can be used to prioritize targets including melanoma-associated antigens and cancer-testis antigens to support selection of an appropriate cellular therapy. As an example, autologous T cell therapy targeting MAGE A4 (Afamitresgene autoleucel or afami-cel) has been evaluated in solid tumors. ³⁷ In this cohort, thirty-five (of 108) tumors expressed detectable levels of MAGE A4 (MAGA4_HUMAN); this group could be selected as a match for therapy. The detected values for this peptide range from 117 to 4,966 amol/μg total protein, representing a >42-fold difference in target expression that could have an impact on the efficacy of therapy. To leverage the quantitative data for more detailed selection by tumor phenotype, a cutoff for patient selection could be set either by MAGE A4 protein amount or by quantile. A similar strategy can be employed for ranking targets for antibody-drug conjugates, like TACD2 targeted by Sacituzumab Govitecan, which was detectable in 102/108 LSCC tumors with a range of expression from 102 to 7,143 amol/μg total protein, representing >70-fold difference across the cohort that could result in a broad distribution in chemotherapy payload delivery. Selection based on positive detection would include a large cohort; further investigation could be used to implement a cutoff for TACD2 expression based on tumor response and subsequent patient outcomes. The tumor-immune microenvironment phenotype interpreted from expression levels of immune cell markers may also be helpful in interpreting outcomes from Sacituzumab Govitecan. These data could be interpreted for each patient either at the individual biomarker level from the panel results tab (Figure 3A) or using the rug plots to compare the expression level in the individual tumor to the distribution of expression across the population (Figure 8).

Figure 8: Parallel Evaluation of Cancer Antigens as Potential Immunotherapy Targets Helps Rank Treatment Options.

A rug plot of cancer antigen quantification can rank expression levels in each tumor and to show the amount of protein detected in that tumor to select the best targets for personalized immunotherapy. In this example, melanoma-associated antigen A4 emerges as the potential target with the highest expression level in this tumor.

In each of these cases, protein quantification can provide actionable information, but interaction with the data is needed to optimize the clinical value. One initial challenge is that protein biomarker measurements are not simply positive or negative data like detection of mutations or high level scores for a gene signature in transcriptomics. The phenotype and its associated score can provide more connection between the biology and the biomarker and may prove to be more accessible to the diverse audience members that participate in biomarker development. This data review tool can provide a means to develop the thresholds needed for clinical implementation of a biomarker or biomarker panel.

Conclusions

PIPER has been developed to visualize complex panels of proteomics data; it can be further adapted for data mining of transcriptomics and genomics data. Different visualization strategies were developed to enable translational researchers to select a display that suits their needs and the selected panel of biomarkers that they want to review. The goal for this strategy is to provide more connection to the proteomics data than existing better-established techniques used for molecular signatures of cancer classification; this strategy should promote the use of clinical proteomics and support protein biomarker development. The current PIPER iteration has many areas for potential growth, including integration of multiple omics datasets (e.g., proteogenomics or proteometabolomics). In addition, survival outcomes with Kaplan-Meier curves based on protein biomarker levels and pathway scores would help determine how each patient could respond to specific therapies; the addition of this layer of information can inform interpretation of clinical value (Figure 1), which is the key next step in protein biomarker development for this panel. The addition of literature review to support the rules-based clinical decisions would also make PIPER more useful to physicians. Comparisons between samples and hierarchical clustering of the tumor data for each individual patient into the cohort could be used to stratify the patient for a therapy where molecularly similar tumors had maximum benefit. For comparison between biomarkers and mapping across panels, co-occurrence and mutual exclusivity analysis could be used to discover relationships between biomarkers that were not expected to be related as well as potential rule out criteria for therapy. PIPER can also help recognize the benefit of adapting the multiplexed targeted assay panel as either a better or a different understanding of tumor biology or clinical need emerges. Finally, we plan to deploy PIPER in different clinical vignettes (e.g., KRAS mutant adenocarcinomas, response to EGFR or MET targeted therapy) to help communicate the complex data acquired from patient specimens with stronger connections to patient outcome. Feedback on these communication strategies and suggestions for additional novel approaches are welcome to make the framework and the PIPER application more adaptable to the needs of the community.