OncoProExp: An interactive shiny web application for comprehensive cancer proteomics and phosphoproteomics analysis

Abstract

Cancer research has been revolutionized by mass spectrometry (MS)-based proteomics, enabling large-scale profiling of proteins and post-translational modifications (PTMs) to identify critical alterations in cancer signaling pathways. However, the lack of comprehensive, user-friendly platforms for integrative analysis limits efficient data exploration, biomarker discovery, and translational insights. To address this, we developed OncoProExp, a Shiny-based interactive web application for in-depth cancer proteomic and phosphoproteomic analyses. OncoProExp offers robust workflows for data preprocessing, interactive visualizations (PCA, hierarchical clustering, heatmaps, gene set enrichment analysis (GSEA)), and functional annotation of gene expression data. Differential expression analysis facilitates biomarker and therapeutic target discovery, while survival analysis identifies proteins whose expression stratifies overall survival, and pan-cancer exploration integrates clinical proteomic and phosphoproteomic datasets. OncoProExp also incorporates state-of-the-art predictive modeling, including Support Vector Machines (SVMs), Random Forests, and Artificial Neural Networks (ANNs) to classify cancer types from proteomic and phosphoproteomic profiles. These models were enhanced by SHapley Additive exPlanations (SHAP) for interpretability. To enhance its translational utility, OncoProExp supports user-uploaded data, protein-protein interactions, pathway enrichment, drug relevance evaluation, and clinical annotation analysis. OncoProExp is deployable via Docker containers, ensuring flexible and scalable integration into individual servers. Its utility has been demonstrated using Clinical Proteomic Tumor Analysis Consortium (CPTAC) datasets. OncoProExp is freely accessible at https://oncopro.cs.ut.ee/ without login requirements, offering a comprehensive resource for translational cancer research.

Highlights

• ●OncoProExp is a Shiny-based web tool for cancer proteome and phosphoproteome analysis.
• ●Supports interactive preprocessing, visualization, and differential expression analysis.
• ●Integrates machine learning models and SHAP interpretation for biomarker prediction.
• ●Enables KEGG pathway mapping, survival, and pan-cancer analysis across CPTAC datasets.
• ●Freely accessible and Docker-deployable with exportable plots and tables.

1. Introduction

Cancer is a highly heterogeneous disease driven by complex molecular alterations, including genetic mutations, epigenetic modifications, and proteomic dysregulation [1], [2], [3]. Although genomic and transcriptomic studies have significantly advanced our understanding of tumor biology, they often fail to capture dynamic protein activities and post-translational modifications (PTMs), which are essential regulators of cellular function and disease progression [4], [5]. PTMs, such as phosphorylation, acetylation, and ubiquitination, occur post-synthesis and are not encoded by mRNA. They directly impact protein stability, interaction, and signaling pathways. Moreover, recent studies have identified new microproteins encoded by previously classified non-coding regions that act as oncogenic drivers or tumor suppressors [6], [7]. These findings underscore the critical role of proteomics in cancer research, providing a functional layer of information that complements genomic and transcriptomic data [8], [9], [10], [11].

Mass spectrometry (MS)-based proteomics has revolutionized cancer research by enabling large-scale profiling of proteins and PTMs, identifying key drivers of tumorigenesis, drug resistance, and immune evasion [12], [13], [14], [15], [16]. Large-scale initiatives, such as the Clinical Proteomic Tumor Analysis Consortium (CPTAC), have generated extensive proteomic datasets that facilitate cancer biomarker discovery and therapeutic development. However, translating these datasets into actionable insights remains challenging because of their computational complexity, which requires missing data imputation, complex normalization strategies, and high-dimensional analyses. Moreover, the lack of accessible integrated bioinformatics tools and the need for specialized expertise further hinder the effective utilization of these datasets.

AI and Machine Learning (ML) have emerged as powerful approaches for extracting meaningful insights from multi-omics datasets [17]. ML techniques such as Support Vector Machines (SVMs), Random Forests (RFs), and Artificial Neural Networks (ANNs) have been applied to cancer subtype classification, prognosis, biomarker discovery, and patient stratification [18], [19]. However, most existing bioinformatics platforms lack integrated ML-based predictive modeling, interactive visualization, and survival analysis, limiting their utility in translational oncology. Although platforms like UALCAN and TCPA provide visualization and basic statistical analyses, they do not offer comprehensive workflows that integrate robust preprocessing, differential expression analysis, AI-driven predictive modeling, pan-cancer analysis, interactive visualization, survival analysis, and drug target identification [20].

To address this gap, we developed OncoProExp, a Shiny-based web application for comprehensive cancer proteomics and phosphoproteomics analysis, which addresses the need for advanced bioinformatics tools in precision oncology. Built on CPTAC datasets, OncoProExp supports analyses across eight cancer types, including CCRCC, COAD, HNSCC, LSCC, LUAD, OV, PDAC, and UCEC, along with survival analyses for GBM and BRCA. The platform streamlines data preprocessing through missing value imputation, feature scaling, and outlier detection, thereby ensuring high-quality inputs for downstream analyses. Users can explore the relationships between tumor and normal samples using interactive visualizations, including Principal Component Analysis (PCA), Multidimensional Scaling (MDS), Uniform Manifold Approximation and Projection (UMAP), and heatmaps.

OncoProExp integrates ML-based predictive models (SVM, RF, ANN) with traditional bioinformatics workflows, achieving a high classification accuracy of above 95 %, while using SHapley Additive exPlanations (SHAP) for model interpretability. This platform supports Differential Expression Analysis (DEA), Gene Set Enrichment Analysis (GSEA), pathway enrichment, and protein-protein interaction (PPI) networks, offering functional insights into dysregulated proteins. Additionally, integration with drug databases (e.g., DrugBank) and survival analysis tools enhances its application in biomarker discovery and therapeutic target research. OncoProExp provides a powerful and user-friendly solution that seamlessly integrates AI-driven analytics with bioinformatics workflows, thereby accelerating cancer proteomic research. By bridging the gap between large-scale omics data and translational oncology, OncoProExp minimizes computational barriers, facilitates biomarker discovery, enhances functional annotation, predicts cancer type, and supports precision medicine.

This paper is organized as follows. The Materials and Methods section cover data collection, preprocessing, visualization, differential expression analysis, enrichment analysis, survival analysis, pan-cancer analysis, feature selection, AI-based predictive models, model interpretation with SHAP, Shiny-based web server development, and data security and compliance. The Results section provides an overview and construction of OncoProExp, data processing and visualization, differential expression and functional insights, pan-cancer patterns and prognostic analysis, machine learning predictions and interpretability, and a comparison of OncoProExp with existing tools. The Discussion section addresses the strengths, limitations of the platform. Finally, the Conclusion summarizes the key contributions of OncoProExp and outlines the future directions.

2. Materials and methods

2.1. Data collection

Proteomic and phosphoproteomic data were obtained from the CPTAC pan-cancer cohort using LinkedOmics [21]. The datasets included Tandem Mass Tag (TMT)-based quantitative profiles, with protein and phosphoprotein abundance expressed as log2-transformed ratios relative to the reference sample. The inclusion criteria required ≥ 10 normal samples and ≥ 60 patient samples per cancer type. Glioblastoma Multiforme (GBM) and Breast Cancer (BRCA) were included for survival analysis despite lacking normal samples due to their clinical relevance. All data consisted of log2-transformed MS1 intensities, with paired tumor and normal samples prioritized for direct comparisons. A summary of the datasets, including sample size, mean age, and proportion of females, is provided in Table 1.

Table 1.

Overview of datasets used in the current study from CPTAC, comprising proteomic and phosphoproteomic datasets.

Cancer Code	Cancer Type	Average Age	Female (%)	Tumor Samples	Normal Samples
CCRCC	Clear Cell Renal Cell Carcinoma	60.79	25.24	103	80
COAD	Colon Adenocarcinoma	64.43	57.8	109	100
HNSCC	Head and Neck Squamous Cell Carcinoma	61.44	12.96	108	62
LSCC	Lung Squamous Cell Carcinoma	65.87	20.37	108	99
LUAD	Lung Adenocarcinoma	62.6	34.55	110	101
OV	Ovarian Cancer	58.92	100	83	19
PDAC	Pancreatic Ductal Adenocarcinoma	64.49	43.81	105	44
UCEC	Uterine Corpus Endometrial Carcinoma	63.47	100	95	18
GBM	Glioblastoma Multiforme	57.89	44.44	99	None
BRCA	Breast Cancer	59.99	100	122	None

2.2. Data preprocessing

Proteome and phosphoproteome data were processed using a common pipeline to ensure consistency. Data quality control excluded features with > 50 % missing values for the proteome and >70 % for the phosphoproteome (according to user-adjustable thresholds). Missing values were filled in by the Random Forest algorithm, using missForest (v1.5 in R) [22], with default parameters until convergence. For both the proteome and phosphoproteome datasets, feature identifiers were mapped to standardized gene symbols using the Ensembl database via the biomaRt package (v2.58.2) [23]. In cases where multiple phosphosites or protein entries were mapped to the same gene, their expression values were averaged to obtain a single gene-level representation. Finally, to ensure that every protein entered the tumor versus normal comparison on an equivalent basis, we retained only those genes with at least 30 % non-missing values in both conditions by intersecting the filtered protein lists from each cohort.

2.3. Visualization

The top 1,000 most variable proteins or phosphoproteins, selected based on the Median Absolute Deviation (MAD), were subjected to dimensionality reduction using Principal Component Analysis (PCA), Multidimensional Scaling (MDS), and Uniform Manifold Approximation and Projection (UMAP). Scree plots were used to determine the number of principal components that explained most of the variance. Box and density plots revealed expression distributions, aiding outlier detection, while heatmaps with hierarchical clustering summarized the expression patterns. When applicable, Gene Set Enrichment Analysis (GSEA) (using fgsea v4.7–12 in R with default parameters) [24] highlighted significantly enriched pathways among the primary data partitions, offering biological insights into functional processes and pathway activation.

2.4. Differential expression analysis

Differentially expressed proteins (DEPs) and phosphoproteins (DEPPs) between tumor and normal samples were identified using limma with default parameters (v3.58.1) [25], applying significance thresholds of |log₂ fold change (FC)| > 0.8 and false discovery rate (FDR) < 0.01. Volcano plots depict log₂ fold change (FC) versus –log₁₀ p-values, while box plots and heatmaps illustrate DEP/DEPP expression differences between tumor and normal samples.

2.5. Enrichment analysis

Enrichment analysis for DEPs/DEPPs was performed using gprofiler2 v0.2.3, [26] to identify functional annotations from Gene Ontology (GO), Kyoto Encyclopedia of Genes and Genomes (KEGG), and Reactome annotations (FDR<0.05). KEGG pathway diagrams were generated using the Pathview package with default parameters to visualize the relationships between DEPs and DEPPs within enriched pathways. Protein-protein interaction (PPI) networks were constructed using the STRING database (STRINGdb v2.14.3, using default parameters) [27] to identify key regulatory proteins and their interaction hubs. To evaluate the therapeutic relevance, dysregulated proteins were cross-referenced with public drug databases, including DrugBank [28] and the Cancer Drugs Database [29].

2.6. Survival analysis

Cox proportional hazards models [30] facilitating both univariable and multivariable analyses were implemented using the default parameters of survival (v3.7) and survminer (v0.4.9) packages [31]. To assess survival differences, patients were stratified into high- and low-expression groups based on median expression levels, with adjustments for age, sex, and tumor stage. Survival differences between the high- and low-expression groups were evaluated using log-rank tests with FDR adjustment (alpha=0.01). Kaplan-Meier curves were used to visualize survival probabilities over time and hazard ratios (HR) with 95 % confidence intervals (CI) [32].

2.7. Pan-cancer analysis

A comparative analysis of proteomic and phosphoproteomic datasets from eight cancer types (CCRCC, COAD, HNSCC, LSCC, LUAD, OV, PDAC, and UCEC) identified common and distinct dysregulated features between tumor and normal samples. Differential expression metrics and p-values highlighted inter-cancer expression patterns, and related survival analyses identified biomarkers with prognostic significance across multiple cancer types.

2.8. Feature selection

A two-step feature selection process was used to identify the diagnostic biomarkers. First, Random Forest-based importance scores were calculated using the randomForest package (v4.7–1.2) [33] to capture non-linear relationships between the variables, defined as proteomic features, with scores based on Gini impurity reduction and a reference provided. Features with importance scores above a predefined threshold and high variability (based on the MAD) were retained. Second, t-distributed Stochastic Neighbor Embedding (t-SNE) projections (Rtsne package, v0.17 with default parameters) were used to evaluate the ability of the selected features to differentiate sample clusters, where the method constructs a low-dimensional embedding by converting pairwise similarities in high-dimensional space into probabilities and optimizes a map that preserves the local neighborhood structure through Kullback–Leibler divergence minimization [34].

2.9. Class imbalance correction

To prevent bias from unequal tumor: normal ratios, we applied the Synthetic Minority Oversampling Technique (SMOTE) using the DMwR package, following our previous implementation [35]. SMOTE generates new minority class examples by interpolating existing samples, yielding a balanced training set for downstream model fitting.

2.10. AI-based predictive models and evaluation

Three machine learning algorithms, SVM (e1071, v1.7–16) [36], Random Forest (randomForest), and ANN (keras, v2.15.0) [37], were applied to predict the cancer type from the proteomics/phosphoproteomics data. For SVM and Random Forest, the dataset was split into 80 % training and 20 % testing. The 20 % test set was the independent test set used for the final evaluation. Within the 80 % training set, 10-fold cross-validation (CV) was applied to optimize the hyperparameters and assess the preliminary performance. SVM used a linear kernel with cost= 3, while Random Forest comprised 100 trees. For the ANN, the dataset was split into 80 % training, 10 % validation, and 10 % testing. The 10 % test set was the independent test set for the final evaluation, whereas the 10 % validation set was used during training to monitor performance and trigger early stopping (after five epochs of no improvement in validation loss). The ANN had four layers (200, 100, 30, and 16 neurons) with Rectified Linear Unit (ReLU) activation, dropout (40 %, 30 %, 10 %), L2 regularization in hidden layers, and an Adaptive Moment Estimation (Adam) optimizer (learning rate=0.0009). To further assess the statistical significance and robustness of the predictive performance, a permutation test with 10 iterations was performed by randomly shuffling class labels within the training set, particularly to validate the feature importance for biomarker identification. The model performance was evaluated using accuracy, sensitivity, specificity, precision, F1-score, and AUC, which were calculated using MBMethPred (v0.1.4.2) [35].

2.11. Model interpretation with SHAP

SHapley Additive exPlanations (SHAP) were applied to Random Forest classifiers (scikit-learn, v1.6.0, with default parameters in Python) [38]. TreeExplainer computes the SHAP values per feature, revealing their contribution to feature-level predictions. The mean absolute SHAP values ranked among the top 10 biomarkers, and summary plots showed how high or low expression values nudged classifications, thereby enhancing the interpretability of AI-based models.

2.12. Shiny-based web server development

The OncoProExp dashboard and backend were built in R Shiny (v1.10.0) [39] and deployed via Docker on a cloud host with five dedicated CPU cores. To support large CPTAC-scale inputs (e.g., 10 000 + features), we perform server-side chunking of uploaded matrices (default: 100 features per chunk), parallelize each chunk across cores, and display a real-time progress bar. User uploads are limited to 50 MB (approximately 1 min processing for a 10 MB file, 5–10 min near the limit), with guidance on expected runtimes shown under “Technical Limitations” on the home page. Detailed Docker installation and resource-customization instructions are provided in README.md, enabling local deployment with an increased upload cap or additional cores for users with heavier workloads.

2.13. Data security and compliance

All data transfers occur over HTTPS (SSL/TLS), and uploaded files are renamed with random identifiers that are never linked to user identities. At the end of the session, every file was purged from the server. Optional cookies used solely for session persistence can be disabled by the user, and no personal information or payload data is stored. Usage metrics are captured only in an anonymized form to guide performance tuning. Our data-handling practices fully comply with the General Data Protection Regulation (GDPR), ensuring confidentiality and user control over their data.

3. Results

3.1. Overview and construction of OncoProExp

OncoProExp is a comprehensive, user-friendly, open-source bioinformatics platform for analyzing cancer proteomics and phosphoproteomic data without the need for programming skills. This versatile tool is freely available as a web server for the quick analysis of small datasets and can also be deployed locally using a Docker container or Shiny-based R package to ensure adaptability, scalability, and reproducibility. To improve accessibility, OncoProExp provides downloadable CPTAC datasets for reproducibility and detailed video tutorials for guidance on using the tool.

The interface comprises four components, as shown in Fig. 1. (1) Data Processing, (2) Differential Analysis, (3) Pan-Cancer and Survival Analysis, and (4) Machine Learning Analysis. These components have been tailored to facilitate the seamless exploration of cancer proteomes and phosphoproteomes, enabling researchers to identify biomarkers, potential therapeutic targets, and prognostic indicators. The subsequent sections provide a comprehensive examination of each of these aforementioned components.

Fig. 1.

Overview of the OncoProExp platform for cancer proteomics and phosphoproteomic analyses OncoProExp consists of four primary modules: (1) Data Input Module—integrates CPTAC cancer datasets or user-uploaded data covering multiple cancer types, such as GBM, HNSCC, BRCA, LUAD, LSCC, PDAC, UCEC, OV, CCRCC, and COAD. (2) Differential Analysis Module: This module detects protein expression differences between normal and tumor samples, providing visualizations such as volcano plots, PCA scatter plots, heatmaps, and protein interaction networks, alongside enrichment analyses, KEGG pathway mapping, and drug relevance evaluation using DrugBank and Cancer Drugs Database. (3) Pan-Cancer and Survival Analysis Module: assesses protein expression across various cancer types and its clinical relevance using Kaplan-Meier survival curves and violin plots. (4) Machine Learning Module: applies SVM, Random Forest, and ANN algorithms for biomarker discovery and predictive modeling, facilitating robust and accurate data-driven insights.

3.2. Data processing and visualization

OncoProExp enables users to efficiently preprocess proteome and phosphoproteome data by accepting uploaded data in CSV format or CPTAC data, with a “Data Privacy and Security” section detailing HTTPS, GDPR Regulation compliance, post-session deletion, random identifiers, opt-out cookies, and anonymized analytics. Users can provide their data as tab-separated values (TSV) or comma-separated values (CSV) files, with rows representing genes/proteins or phosphorylation sites and columns representing samples with log2-transformed ratios or intensities (Supplementary File 1). After uploading, OncoProExp applies filters to remove samples with high proportions (>50 % for proteomic data and >70 % for phosphoproteomic data) of missing values. For both the proteome and phosphoproteome datasets, identifiers, including phosphosite-level entries, were mapped to gene symbols using the Ensembl database, and expression values corresponding to the same gene were averaged, resulting in approximately 11,000 unique gene symbols for the CPTAC datasets. After data imputation, the preprocessed dataset was saved as a CSV file, making it ready for further exploration and analysis. To ensure data privacy, all uploaded files were deleted after the session. Dimensionality reduction techniques, such as PCA, MDS, and UMAP, enable users to explore sample clustering and batch effects (Fig. 2A-B). Box plots and density plots provide insights into expression distributions, aiding outlier detection and quality control (Fig. 2C-D). A heatmap with hierarchical clustering further summarized these expression patterns (Supplementary Figure 1).

Fig. 2.

Visualization of CCRCC Tumor and Normal Samples. (A) The Principal Component Analysis (PCA) plot shows distinct separation between tumor and normal samples along the first two principal components, explaining 65.02 % and 12.75 % of the variance, respectively. (B) Multidimensional Scaling (MDS) plot showing the clear clustering of tumor and normal samples. (C) Density plot highlighting the distribution differences between the two groups. (D) Box plot displaying lower median protein expression levels in tumor samples than in normal samples.

3.3. Differential expressed proteins and phosphoproteins

Users can conduct differential expression analysis of DEPs and DEPPs across cancers using limma, calculating fold changes and adjusted p-values with user-defined cutoffs to pinpoint significantly dysregulated features. For example, CPTAC data analysis revealed FGFR2 and FGFR4 downregulation in LUAD (protein and phosphoprotein levels, respectively), consistent with previous findings linking FGFR4 expression in LUAD to tumor stage and differentiation [40], [41]. Additionally, JAK2 is significantly downregulated at both the protein and phosphoprotein levels, supporting its role in cancer stem cell maintenance and radioresistance [42].

Additional findings across cancer types, including key dysregulated proteins and their functional implications, are summarized in Table 2 and Supplementary Tables 1–2, highlighting OncoProExp’s ability to reveal cancer-specific dysregulation and functional implications of each dysregulated feature (for a step-by-step case study, see Supplementary Tutorial 1).

Table 2.

Differentially expressed proteins and phosphoproteins across cancer types in CPTAC data. This table presents the differentially expressed proteins and phosphoproteins identified across multiple cancer types using CPTAC data. The table includes the log fold-change (logFC), average expression levels, adjusted p-values, and functional implications of each dysregulated feature.

Type	Cancer	Gene	logFC	Average expression	Adjusted p-value	Function/Implication
Proteome	LUAD	FGFR2	-0.803	16.87	2.98E-62	Linked to tumor stage and differentiation [40].
	COAD	JAK2	-1.154	20.851	6.94E-40	Cancer stem cell maintenance, radioresistance [42].
	CCRCC	CA9	1.922	25.478	1.70E-49	Hypoxia biomarker [43].
	LSCC	TP53	0.98	22.335	6.19E-13	Tumor progression, therapy resistance [44].
	OV	EGFR	-0.891	23.651	5.07E-12	Uncertain prognostic biomarker [45].
	PDAC	CD44	-1.19	27.588	4.34E-14	Tumor plasticity, invasion, drug resistance [46].
	UCEC	KRT17	2.23	25.605	2.50E-10	HIF−1α/VEGF pathway, migration, angiogenesis.
	UCEC	WNT2	-1.389	21.346	7.46E-12	Context-dependent tumor biology [47].
Phosphoproteome	COAD	JAK2	-1.097	20.77	3.82E-41	Cancer stem cell maintenance, radioresistance [42].
	LSCC	PLK1	1.525	15.992	8.47E-77	Tumor progression, cell cycle dysregulation [48].
	PDAC	T TR	1.148	19.186	1.90E-12	Tumor metabolism, systemic response [49].
	LUAD	CHRNA3	1.14	13.618	1.26E-65	Lung cancer susceptibility, nicotine oncogenic pathways [50].
	LUAD	FGFR2	-1.769	13.77	4.86E-146	Linked to tumor stage and differentiation [40].
	LUAD	FGFR4	-1.332	19.54	4.20E-77	Linked to tumor stage and differentiation [51].
	UCEC	KRT17	1.25	19.613	1.03E-14	HIF−1α/VEGF pathway, migration, angiogenesis [52].

Volcano plots with interactive features enable users to examine DEPs and DEPPs by displaying the fold-change against statistical significance, facilitating the easy recognition of crucial molecular components (Fig. 3A). Heatmaps offer an additional perspective by grouping dysregulated features and exposing co-expression patterns that align with biological processes or pathways (Fig. 3B).

Fig. 3.

CCRCC differential proteomics visualization: (A) An interactive volcano plot displaying -log10(adjusted P-value) on the y-axis and log2 fold change on the x-axis, highlighting the differential expression patterns. Vertical dashed lines represent log fold-change thresholds, whereas the horizontal dashed red line denotes the significance threshold. Hovering over the data points reveals gene symbols, enabling a detailed examination of the significant differentially expressed proteins (DEPs) and phosphoproteins (DEPPs). (B) Heatmap of top user-selected genes, with genes represented as rows and samples as columns. Hierarchical clustering of the top groups’ samples revealed co-expression patterns aligned with biological processes or pathways.

3.4. Enrichment analysis of DEPs and DEPPs, pathway visualization, PPI and Drug interaction analysis

OncoProExp integrates a comprehensive gene enrichment analysis implemented using the gProfileR package for DEPs and DEPPs. This allows users to explore biological pathways and functional annotations across multiple databases (GO, KEGG, REACTOME, TF, miRNA, HPA, CORUM, and WikiPathways). Users can analyze upregulated, downregulated, or combined DEPs/DEPPs to tailor the results to specific research questions. A key feature is KEGG pathway enrichment analysis, which maps dysregulated proteins onto cancer-relevant pathways using the Pathview tool, providing an intuitive and interactive way to interpret pathway-level changes in the context of cancer biology. In the PDAC datasets, OncoProExp identified the activation of the pancreatic secretion pathway, reflecting its role in tumor progression (Fig. 4A). Additionally, OncoProExp identified the upregulation of ITPR3 (Fig. 4B), a key regulator of intracellular Ca²⁺ release, supporting its role in tumor cell survival, proliferation, and endoplasmic reticulum stress [53]. Meanwhile, the downregulation of PMCA and ATP-related genes suggests reduced calcium efflux capacity and metabolic strain, consistent with the finding that glycolytic ATP is critical for maintaining calcium homeostasis in PDAC [54]. These findings demonstrate the tool’s ability to uncover critical molecular mechanisms and provide pathway-level insights into cancer biology.

Fig. 4.

Analysis of Pathways, Protein Interactions, and Drug Targets in Pancreatic Ductal Adenocarcinoma (PDAC) using OncoProExp. (A) KEGG pathway analysis identified pancreatic secretion as the most enriched pathway in the PDAC datasets. (B) Pathview visualization showing ITPR3 upregulation, indicating increased calcium signaling, whereas PMCA and ATP-related genes were downregulated. (C) The protein–protein interaction (PPI) network, generated using OncoProExp in the LUAD proteome dataset, revealed PITX2 as a central gene in a four-connection network (p-value = 0.0304). (D) Boxplot analysis demonstrated significant FN1 overexpression in pancreatic ductal adenocarcinoma (PDAC) compared to normal tissues. Data from CancerDrugs_DB show that dacarbazine (DrugBank ID: DB00851) targets FN1, which is notably upregulated in PDAC.

Users can perform protein-protein interaction (PPI) analysis to identify key interaction hubs and explore the connectivity among dysregulated proteins. In LUAD, analysis of the top 10 DEPs from CPTAC revealed four significant interactions (p-value = 0.0304) (Fig. 4C). The resulting network highlighted PITX2, SALL3, SOX8, and IER2 as the crucial nodes. PITX2 acts as an oncogene in LUAD by activating WNT3A and driving Wnt/β-catenin signaling to promote tumor progression [55]. In contrast, SALL3 functions as a tumor suppressor, and its epigenetic silencing is linked to poor prognosis in HNSCC [56]. Similarly, SOX8 is involved in maintaining cancer stem-like cells, contributing to tumor progression in triple-negative breast cancer [57]. The immediate early gene IER2 further underscores its significance in cancer biology by enhancing tumor cell motility and invasiveness, facilitating metastasis, and correlating with poor survival in colorectal cancer patients, making it a promising prognostic biomarker and therapeutic target [58]. Thus, the interconnections observed among these proteins point to a functional unit that may contribute to vital cancer-related processes, including cellular differentiation, tumor growth and metastasis.

Users can further explore drug target information integrated from CancerDrugs_DB, linking dysregulated proteins to potential therapeutic interventions. For instance, boxplot analysis of FN1 in PDAC revealed its significant upregulation in tumor samples compared to normal tissues. FN1, identified as a target of the chemotherapeutic agent dacarbazine, highlights the translational potential of OncoProExp in linking molecular alterations to actionable drug targets (Fig. 4D). These findings underscore the utility of OncoProExp in linking molecular changes to biological functions, thereby supporting hypothesis generation, drug repurposing, and targeted experimental design.

3.5. Pan-cancer analysis

The pan-cancer module of OncoProExp allows users to investigate DEPs and DEPPs across various cancer types using CPTAC datasets. For instance, users can examine EPCAM expression in lung adenocarcinoma (LUAD), ovarian cancer (OV), and clear cell renal cell carcinoma (CCRCC), revealing distinct patterns between tumors and normal tissues (Supplementary Figure 2). In LUAD, EPCAM is frequently overexpressed, although its prognostic relevance remains unclear [59], [60]. Conversely, in CCRCC, EPCAM often correlates with favorable tumor features [61]. In contrast, ovarian cancer overexpresses EPCAM in chemoresistant cancer stem cell–like populations, contributing to poorer clinical outcomes [62]. The platform also provides a comprehensive table of DEPs and DEPPs for all CPTAC cancer types, filtered for significance (FDR < 0.05 for DEPs/DEPPs and p < 0.05 for survival biomarkers). Additionally, users can explore survival biomarkers and assess the prognostic significance of specific proteins across multiple cancer types, enhancing biomarker discovery and precision oncology applications.

3.6. Feature selection

The selected features, derived through a two-step process combining Random Forest importance scores and MAD-based filtering, demonstrated robust discriminatory power. The t-SNE projections revealed distinct clustering of tumor and normal samples across all cancer types in both the proteome (Fig. 5A) and phosphoproteome (Fig. 5B) datasets. The clear separation between the tumor and normal groups, as well as the distinct cancer-specific clusters, highlights the biological relevance of the top 250 proteins (Supplementary File 2) and 100 phosphoproteins (Supplementary File 3). These findings confirm that the selected features effectively captured meaningful variations for accurate classification using downstream machine learning models.

Fig. 5.

t-SNE analysis of (A) proteome and (B) phosphoproteome data, illustrating the separation between tumor and normal samples across multiple cancer types (CCRCC, COAD, HNSCC, LUAD, LSCC, OV, PDAC, and UCEC). Each point is colored by cancer type and sample origin, revealing distinct clusters for different cancers and a clear separation between tumor and normal samples within each cancer type. The clustering patterns indicated significant differences in protein and phosphoprotein expression profiles, underscoring the potential of both data types to effectively distinguish tumors from normal tissues.

3.7. AI-based predictive models and their performance

OncoProExp integrates advanced machine learning algorithms, including Support Vector Machines (SVM), Random Forests (RF), and Artificial Neural Networks (ANN), to predict cancer types from proteomic and phosphoproteomic data. Users can train these frameworks in both multi-cancer and single-cancer modes, achieving high accuracy and robustness across the test and validation sets. Before training our classifiers, we applied the Synthetic Minority Oversampling Technique (SMOTE) to balance the tumor versus normal sample ratio in the three cancer cohorts (UCEC, PDAC, and OV; see Methods, Class Imbalance Correction). Supplementary Figure 3 illustrates the class counts before (panel A) and after (panel B) oversampling. Originally, UCEC had 95 tumors versus 18 normal samples (5:1), OV had approximately 83 tumors versus 19 normal samples (4:1), and PDAC had approximately 105 tumors versus 44 normal samples (2:1). After SMOTE, each cohort was resampled to a 1:1 ratio, with approximately 80 tumor and 80 normal samples, thereby stabilizing model training and ensuring that no class dominated the learning process.

3.7.1. Multi-cancer mode

In multi-cancer classification using proteomics data, the SVM framework achieved perfect predictive performance on the test set, with accuracy, precision, sensitivity, F1-score, specificity, and AUC all equal to 1 (Supplementary Tables 3–4). To assess potential overfitting, we conducted a permutation test by shuffling class labels prior to training, which reduced the accuracy to ∼0.88 and the AUC to ∼0.57 (Supplementary Table 5), confirming that the framework captures biologically meaningful patterns rather than random noise. On the independent validation set, the SVM framework maintained a strong performance, particularly for HNSCC and PDAC, achieving high accuracy across metrics and an AUC of 0.996 (Supplementary Tables 6 and 7). Although these results indicate robust generalization, further validation using external datasets is warranted to confirm clinical applicability.

The RF framework showed a perfect classification on the test set, with accuracy, precision, sensitivity, F1-score, and specificity all at 1.0 for most classes (Supplementary Tables 8–9). Minor drops in the Normal PDAC and Tumor PDAC accuracy and F1-score were observed, but they remained near perfect. On the validation set, the RF framework maintained high performance across all metrics (Supplementary Tables 10–11).

The ANN framework also exhibited excellent test set performance, achieving perfect scores in most cancer types and maintaining high precision and recall (Supplementary Tables 12–13). The training and validation loss curves remained well-aligned throughout (Supplementary Figure 4), indicating minimal overfitting. On the validation set, the ANN framework results mirrored those of the RF framework, with consistently high accuracy and robust metrics for both HNSCC and PDAC (Supplementary Tables 14–15).

3.7.2. Single-cancer mode

In the single-cancer classification mode, the SVM, RF, and ANN frameworks demonstrated highly accurate performance for HNSCC and PDAC using independent validation sets sourced from TCGA datasets, which share compatible normalization methods with CPTAC proteomics data. For HNSCC, all frameworks achieved identical performance with 99.4 % accuracy, 100 % precision, and 98.4 % sensitivity, resulting in an F1 score of 0.992. The AUC for all frameworks was 0.995, highlighting their strong discriminatory power (Supplementary Tables 16–17). For PDAC, each framework performed equally well, achieving an accuracy of 97.2 %, with 96 % precision, 96 % sensitivity, and an F1 score of 0.96. The AUC for all frameworks was 0.969, further emphasizing their robust classification ability (Supplementary Tables 18 and 19). Overall, all three frameworks exhibited consistent and high performance in distinguishing tumor and normal samples in both HNSCC and PDAC, with HNSCC classification reaching near-perfect accuracy and PDAC classification demonstrating strong predictive reliability.

3.7.3. Phosphoproteome data

The evaluation of machine learning frameworks on phosphoproteome data demonstrated exceptional classification performance in both multi-cancer and single-cancer classification modes. External validation of phosphoproteomics data was not feasible because of differences in normalization methods between the TCGA and CPTAC cohorts, as the standard values required to reverse TCGA’s normalization were unavailable. In the multi-cancer classification mode, which included CCRCC, COAD, HNSCC, LSCC, LUAD, OV, PDAC, and UCEC, the SVM, RF, and ANN frameworks achieved perfect classification accuracy across all cancer types. The confusion matrices for these frameworks showed no misclassifications, with each framework accurately distinguishing between normal and tumor samples. The key performance metrics, including accuracy, precision, sensitivity, F1 score, specificity, and AUC, all reached 1.0, underscoring the reliability of the framework (Supplementary Tables 20–25).

Similarly, in the single-cancer classification mode for CCRCC, all three models (SVM, RF, and ANN) achieved perfect differentiation between normal and tumor samples, as confirmed by the confusion matrices (Supplementary Table 26). The performance metrics were uniformly optimal (accuracy = 1.0 across all frameworks), reinforcing the frameworks’ ability to classify CCRCC with complete accuracy (Supplementary Table 27). Overall, the analysis confirmed that the models excelled in classifying cancer and normal samples using proteome and phosphoproteome data, showcasing their robustness and precision in both multi-cancer and single-cancer settings.

3.8. Model explanation using SHAP

To enhance the interpretability of AI-based predictions, OncoProExp incorporates SHapley Additive exPlanations (SHAP). Users can explore the contributions of individual proteins and phosphoproteins to model predictions, gaining insights into the molecular features that drive cancer classification.

3.8.1. Multi-cancer model

In the multi-cancer model for proteome data (Fig. 6A), genes such as ITGA7 and LILRB5 showed mixed impacts, with ITGA7 having a broader influence and MYO1C significantly enhancing the model performance. In the phosphoproteome data (Fig. 6B), XPO4 and COLGALT1 displayed varied effects, with XPO4 having a wider influence and SYN1 showing a strong positive impact. Other genes, such as SEPSECS, MCCC2, and PLIN4, contributed both positively and negatively, whereas GYPC and LRP1 exhibited diverse impacts, with LRP1 leaning positively. FAM83F and CD34 exhibited mixed effects, with CD34 exhibiting a broader range.

Fig. 6.

Interpretation of model predictions using SHAP values. SHAP summary plots showing the impact of top proteins (A) and phosphoproteins (B) on model predictions in the multi-cancer mode. Each point represents a sample, with its x-position indicating the SHAP value for a given feature (impact on model output), whereas the vertical axis lists the features. The color denotes the feature value (red: high, blue: low). The spread of points along the x-axis indicates the range of the feature's impact. These plots show that most features contribute to both increases and decreases in the model's predicted values. Panels (C) and (D) highlight the features with more consistent negative or positive impacts in the single-cancer mode, respectively.

3.8.2. Single-cancer model

In the single-cancer model for CCRCC proteome data (Fig. 6C), AIF1L and NOL3 had strong negative effects on the model output. A similar negative trend, although less pronounced, was observed for ACLY, ERO1A and NDRG1. For the same cancer type in the phosphoproteome data (Fig. 6D), genes such as SYN1 and IGFBP3 showed strong negative impacts, whereas CHP1 exhibited a positive impact. Higher expression levels (represented by pink for proteome and red for phosphoproteome) correlated with stronger impacts, highlighting the diverse roles of genes in model performance.

3.9. Survival analysis

OncoProExp allows users to identify biomarkers with prognostic value in various cancer types. For instance, users can investigate the effects of CA12 or MGMT proteome expression in UCEC or GBM, respectively, by creating Kaplan-Meier plots and determining hazard ratios (HR). The application automatically selects the optimal expression threshold for each biomarker, ensuring balanced group comparisons and accurate HR calculations. In the case of CA12 in UCEC, elevated expression was linked to significantly poorer survival outcomes (HR: 0.422, p = 0.01), aligning with its established role in facilitating tumor progression via pH regulation and hypoxia adaptation [63]. Likewise, reduced MGMT expression in GBM was associated with decreased survival (HR: 1.8, p = 0.03), reflecting its involvement in treatment resistance [64], [65]. These results, depicted through user-friendly Kaplan-Meier graphs (Supplementary Figure 5), underscore the clinical significance of these biomarkers and their potential as therapeutic targets. analyses to address specific research questions.

3.10. Comparison of OncoProExp with existing tools

OncoProExp provides an integrated platform for cancer proteomics and phosphoproteomics analysis, combining machine learning (ML) with statistical methods to enhance analytical capabilities. Unlike PhosMap [66], cProSite [67], TCPA [68], CPPA [69], UALCAN [70], and iProPhos [71] which primarily use basic statistical methods. OncoProExp offers a range of functionalities that are not concurrently available in these tools. OncoProExp employs advanced ML classifiers such as Random Forest (RF), Support Vector Machines (SVM), and Artificial Neural Networks (ANN), in combination with limma for differential expression analysis. A unique feature of OncoProExp is its incorporation of SHAP (SHapley Additive exPlanations) for CPTAC data, providing explainable AI insights that are absent in other tools. OncoProExp represents a significant advancement over existing cancer proteome tools by incorporating advanced dimensionality reduction techniques (PCA, MDS, UMAP), pathway enrichment methods (GSEA, KEGG, Reactome, PPI), and robust imputation of missing values. Its integration with Cox models, Kaplan-Meier survival analysis, and drug relevance evaluation via DrugBank and Cancer Drugs Database makes it a comprehensive tool. Unlike UALCAN and iProPhos, OncoProExp offers enhanced accessibility through a web-based dashboard and Docker deployment. We demonstrate through our HNSCC case study (Supplementary Tutorial 1) that OncoProExp delivers practical advantages, such as clearly separating tumor and normal samples using PCA/UMAP, identifying cancer-relevant DEPs/DEPPs (e.g., PAGE2, PPM1D), and linking them directly to KEGG pathways and drug targets. These real-data applications illustrate unique biological insights that existing platforms do not readily provide. Thus, our case study demonstrates practical advantages and unique biological insights, which are not directly supported by alternatives such as UALCAN or iProPhos. A detailed comparison of OncoProExp's capabilities with those of other tools is provided in Table 3.

Table 3.

Comprehensive comparison of OncoProExp's with those of current Phosphoproteome and proteome tools.

Feature	OncoProExp	PhosMap	TCPA	cProSite	UALCAN	iProPhos
Machine Learning vs. Basic Statistical Methods	ML + Stats (advanced classifiers (RF, SVM, ANN) plus limma)	Stats only (limma, ANOVA, etc.)	Stats only (correlation, t-test, univariate K–M)	Stats only (basic group comparisons, correlations)	Stats only (basic group comparisons, correlation, survival analyses)	Stats only (differential expression, correlation, KSEA)
SHAP / Explainable AI Integration	Yes (Single and pan-cancer)	No	No	No	No	No
Dimensionality Reduction	Yes (PCA, MDS, UMAP)	Yes (PCA, t-SNE, UMAP)	No (clustered heatmap, correlation network)	No (heatmaps, correlation)	No (primarily boxplots, correlation scatter, survival curves)	No (focus on differential expression and correlation analyses)
Pathway / Network Enrichment	Yes (GSEA, KEGG, Reactome, PPI network)	Limited (KSEA, motif-x; no direct KEGG/Pathview)	No (only RPPA-based correlations)	No	No	Yes (Over-Representation, GSEA, and PPI network analyses)
Advanced Missing-Value Imputation	Yes (Using percentage of NAs, RF, mean, median, mod)	Yes (Min, BPCA, LLS, kNN, etc.)	No	No	No	Yes (filters >50 % missing then applies kNN)
Survival Analysis	Yes (univariate and multivariable Cox, Kaplan-Meier plot)	Yes (multivariable CoxPH available)	No (univariate Kaplan-Meier only)	No	Yes (univariate Kaplan–Meier with some multivariable options)	Yes (Kaplan–Meier/log-rank tests with user-specified cutoffs)
Plotting & Visualization	Volcano plots, Heatmaps, PCA/UMAP/MDS, Scree plot, Box and density plots, GSEA plot, Enrichment plot, KEGG image, PPI network, Boxplots for drug-targetable genes, Kaplan–Meier curves, SHAP	Volcano plots, Heatmaps, Correlation scatter, PCA/t-SNE/UMAP, Survival curves	Clustered heatmap, Correlation network, Basic survival plots	Boxplots, Barplots, Heatmaps, Tumor vs. Normal fold-change plots	Boxplots, Correlation scatter, Kaplan–Meier curves, Pan-cancer comparisons	Volcano plots, Boxplots, Correlation scatter, KSEA barplots, Kaplan–Meier curves, PPI network maps
Therapeutic Relevance Evaluation	Yes (dysregulated proteins are cross-referenced with DrugBank and the Cancer Drugs Database; box plots illustrate differential expression of drug-targetable genes)	No	No	No	No	Partial (identifies potential drug targets and upstream kinases, but does not explicitly cross-reference with public drug databases)
User Data Upload	Yes (Allows upload of proteomics/phosphoproteomics data along with their metadata)	No	No	No	No	Yes (Allows upload of custom proteomics/phosphoproteomics data)
Cancer Database Integration	Yes (integrates CPTAC proteomics/phoproteomics)	No (requires user-provided data)	Yes (built on TCGA RPPA data plus CCLE cell lines)	Yes (CPTAC proteomics/phosphoproteomics and GDC/TCGA mRNA)	Yes (integrates TCGA transcriptome, miRNA, lncRNA, promoter methylation and CPTAC proteomics)	Yes (integrates CPTAC proteome & phosphoproteome from 12 cancer types plus matched transcriptome)
User Interface (Web vs. Local vs. CLI)	Web + Docker (interactive dashboard)	Local (Docker container, R-based)	Web-based portal	Web-based portal	Web-based portal	Web-based (R Shiny), no Docker
Docker Functionality	Yes	Yes	No	No	No	No (R Shiny web server; no Docker image)
License / Source Code	Open source (Entire code on GitLab link)	Open source (GitHub: BADD-XMU/PhosMap)	Free access (hosted; not fully open-sourced)	Free (hosted by NCI-CCR; no open-source statement)	Free (hosted by UAB; partial code available on GitHub)	Free (hosted by Zhejiang University; code to be available on GitHub upon publication)

4. Discussion

This study presents OncoProExp, a web-based, user-friendly platform that integrates MS-based proteomic and phosphoproteomic data with advanced computational tools, enabling researchers to explore cancer-related dysregulation and its association with cancer progression and therapy. By combining clinical, proteomic, and phosphoproteomic data, OncoProExp addresses key gaps in existing tools, offering robust preprocessing, statistical analyses (e.g., differential expression analysis, gene set enrichment analysis), and machine learning (ML) models (e.g., SVM, RF, ANN) to facilitate biomarker discovery and translational research. Its intuitive interface and support for CPTAC datasets make high-throughput proteomics accessible to researchers without advanced bioinformatics knowledge.

Although genetic and epigenetic approaches have dominated cancer biomarker research, large-scale proteomics remains underexplored in pan-cancer studies [72], [73], [74], [75], [76], [77]. Research efforts, such as Álvez et al. ML-based blood proteome profiling [3], and Gonçalves et al. mapping of 949 cancer cell lines [78], underscores the potential of proteomics in cancer research. Currently available tools such as PhosMap, TCPA, cProSite, UALCAN, and iProPhos primarily focus on visualization and basic analysis, complementing research efforts. However, the above-mentioned platforms often lack integrated preprocessing, user-uploaded data support, customizable visualizations, AI-driven modeling, and translational features like drug target evaluation [71],

OncoProExp outperforms existing platforms by integrating pan-cancer analysis with AI-based predictive modeling, achieving high classification accuracy (99.4 % in HNSCC and PDAC) and leveraging SHAP-based model interpretability. Interactive visualizations, such as PCA, UMAP, and heatmaps, further aid hypothesis generation, and CPTAC integration enables cross-cancer comparisons of shared and unique signatures. By cross-referencing dysregulated proteins with drug databases [28], [29], OncoProExp supports drug repurposing, which is a key advantage for precision oncology, where actionable targets are critical. The platform’s predictive modeling spans pan- and single-cancer levels, supporting analytics from discovery to application. Its Docker deployment and data-sharing features ensure reproducibility and collaboration across research groups, thereby accelerating cancer proteomics research.

Despite its strengths, OncoProExp currently relies on CPTAC datasets, limiting the inclusion of other PTMs (e.g., glycosylation and acetylation) and diverse data types. Moreover, the platform employs an aggregation method for phosphosite data, which consolidates multiple phosphorylation sites into a single gene-level representation. The aggregation approach is computationally efficient; however, it presents biological limitations. Specifically, site-specific phosphorylation events play a crucial role in regulating distinct signaling pathways and protein interactions. Aggregating these events may lead to an oversimplification of phosphoproteomic data, potentially overlooking critical biological insights. Future versions of OncoProExp will strive to integrate site-level detail to achieve a balance between usability and biological specificity.

Although OncoProExp demonstrated high classification accuracy, future studies should validate its performance on independent, externally sourced proteomic datasets to confirm its generalizability. Furthermore, processing large datasets may also slow network-based or ML analyses, and although it excels in hypothesis generation, experimental validation remains essential for clinical translation. Future updates will expand the multi-omics datasets beyond CPTAC, incorporate additional PTMs, and optimize the performance for large-scale data processing.

5. Conclusion

OncoProExp fills a critical gap in proteomics research by integrating comprehensive computational methods, AI-driven predictive modeling, and interactive visualization into a single platform. Its application to CPTAC datasets demonstrated its potential for uncovering novel insights into cancer biology, enabling researchers to translate complex datasets into actionable insights for precision oncology. OncoProExp will be continuously maintained and updated to improve its analytical and visualization functionalities. Future plans include incorporating site-specific phosphoproteomics and kinase-level datasets for enhancements and exploring PRIDE and MassIVE datasets as external validation sets for its machine learning modules.

Ethics statement

Not applicable.

CRediT authorship contribution statement

Edris Sharif Rahmani: Writing – original draft, Visualization, Validation, Software, Methodology, Investigation, Formal analysis, Conceptualization. Prakash Lingasamy: Writing – review & editing, Writing – original draft, Formal analysis. Sergio Vela Moreno: Writing – review & editing. Andres Salumets: Writing – review & editing, Resources, Funding acquisition. Soheila Khojand: Writing – review & editing, Formal analysis. Ankita Lawarde: Writing – review & editing, Visualization. Vijayachitra Modhukur: Writing – review & editing, Writing – original draft, Visualization, Software, Resources, Project administration, Formal analysis, Conceptualization.

Consent to participate

Not applicable.

Consent for publication

Not applicable.

Funding

This research was supported by the Horizon Europe grant (NESTOR, no. 101120075) and the Estonian Research Council (grant no. PRG1076).

Declaration of Competing Interest

The authors declare no conflict of interest.

Appendix A. Supplementary material

Supplementary material

Supplementary material

Supplementary material

Supplementary material

Supplementary material

Data availability

All data used in this study are publicly available from the Clinical Proteomic Tumor Analysis Consortium (CPTAC) at https://proteomics.cancer.gov/programs/cptac. The source code for OncoProExp is accessible on GitLab at https://gitlab.cs.ut.ee/edris/oncoproexp.git, including a README with instructions for local deployment. The web application is freely available at https://oncopro.cs.ut.ee/, where users can access tutorials, upload their own data, and download analysis outputs. There are no restrictions on data access, ensuring the full reproducibility of the results presented in this study.