Abstract
Gene expression studies are fundamental in molecular biology, offering insights into development, disease progression, and therapeutic targets. To address the need for precise analysis of large datasets, we developed THRESHOLD, a novel tool that introduces the concept of gene saturation. Unlike traditional methods focused on absolute or binary expression levels, THRESHOLD quantifies the consistency of gene expression across patients, revealing co-regulation patterns critical for understanding disease mechanisms and stratifying patients by molecular signatures. The tool offers several features, including user-defined parameters, statistical comparisons, and interactive data visualization. THRESHOLD has uncovered compelling insights into disease progression using TCGA cancer datasets. For instance, bladder urothelial carcinoma demonstrated increasing upregulated gene saturation in progressive cancer stages (P < .00001). Moreover, THRESHOLD identified heightened gene saturation in patients with earlier onset of prostate adenocarcinoma (P < .0001) and revealed a critical fusion transcript, SLC45A2-AMACR, implicated in prostate adenocarcinoma progression, recurrence, and metastasis. Additionally, novel biomarkers and potential candidates for drug therapies were identified through protein–protein interaction networks and functional analyses of saturation data in colon adenocarcinoma and breast invasive carcinoma. THRESHOLD offers a new approach for studying gene expression dynamics and patient stratification. The tool is publicly available at Zenodo: https://zenodo.org/records/15287195.
Graphical Abstract
Graphical Abstract.
Introduction
Researchers’ and clinicians’ ability to accurately diagnose and treat diseases is aided by their capacity to understand disease heterogeneity [1] as it affects disease etiology and progression. Cancer is a complex disease and clinically challenging to treat due to the numerous pathways through which carcinogenesis can occur and the molecular variation within cancer subtypes [2]. Therefore, it is highly valuable for researchers and clinicians to identify significant differences in gene expression. These differences can help predict disease progression and determine whether a tumor is likely to respond to or resist various therapeutic approaches.
Cancer remains a leading cause of global mortality, with incidence rates increasing at an alarming pace. New cancer cases in the United States have increased 30.8% from around 1 500 000 in 2010 to 2 000 000 in 2024 [3, 4]. Changes in new cancer cases cannot be explained by population changes [5, 6]. The molecular complexity of cancer, coupled with the increase in cancer incidence, creates a critical need for new analytical approaches for analyzing clinical samples in a manner that aids in the development of new and effective therapeutic approaches [7]. Examining the subtle biological pathways through which illnesses arise and manifest can help clarify these complexities.
The field of genomics offers promise in healthcare in its capacity to aid in early detection of diseases, patient stratification, and personalized medicine, with gene expression analysis driving this innovation. Stratified analyses of upregulated and downregulated genes can shed light on potential pathways and therapeutic targets across patient groups of interest. The patterns in which genes are expressed can unveil significant insights into disease mechanisms, guide therapeutic responses, and help interpret complex biological interactions. Currently, differential gene expression (DGE) analysis is utilized as an existing computational approach to evaluate gene expression and compare differences in gene expression across relevant groups or conditions [8].
Although THRESHOLD offers a novel approach by focusing on gene saturation and consistency in gene expression across patient cohorts, several existing tools can be considered the closest in terms of application, though their methodologies differ. DESeq2 [9] is a widely used DGE analysis tool that emphasizes identifying statistically significant changes in gene expression across conditions but it may overlook consistent gene expression patterns across individuals— a key aspect that THRESHOLD is designed to capture. Similarly, edgeR [10] is another DGE tool that excels in detecting differentially expressed genes in RNA-seq datasets but, like DESeq2, focuses primarily on the magnitude of expression changes rather than consistent gene patterns. Limma-Voom [11] is a versatile tool that uses linear models for differential expression analysis of microarray and RNA-seq data, providing valuable insights into gene expression profiling but lacking the focus on gene expression consistency that THRESHOLD offers. Additionally, while CIBERSORT [12] is not a DGE tool, it focuses on deconvoluting immune cell populations from bulk RNA-seq data, and its statistical framework for analyzing gene expression across patient samples offers a useful point of comparison with THRESHOLD’s capabilities in evaluating gene expression patterns across populations. Despite these differences, these tools represent the most relevant comparisons in terms of analyzing gene expression in large datasets.
DGE’s underlying calculations rely on the identification of genes that demonstrate statistically significant changes in expression across patients as compared to some standard, revealing which genes are upregulated or downregulated in response to specific conditions. This approach provides value for detecting major expression shifts and associating them with biological processes or relevant disease states. However, DGE analysis is limited in its ability to capture the broader, more distinct gene expression patterns by patients that may be crucial for understanding complex disease pathology. Specifically, traditional DGE analysis emphasizes the magnitude of expression changes and relies on grouped averages, which can be skewed by individual patients, potentially overlooking genes that play consistent roles across individual patients but do not exhibit as significant changes in expression or tight variances. THRESHOLD is particularly well-suited for identifying patterns where genes are consistently, unusually expressed, even if only moderately. This is because the concept of saturation, though dependent on high differential ranks of expression, operates specifically on the number of times a gene occurs across patient groups and fundamentally quantifies a proportion of patients expressing recurring genes. These patterns are potentially relevant for highlighting co-regulated genes, including stable biomarkers or key nodes existing in regulatory networks. This holds particular pertinence in the context of highly complex diseases, in which there is high variability and propensity for convolution even between patients of the same labeled disease, such as cancer or genetically heterogeneous disorders. In these cases, THRESHOLD’s gene saturation concept that stresses the consistency of gene expression ranks across a patient cohort can aid in bridging this gap, helping to identify potentially overlooked disease mechanisms and facilitate representative patient stratification via novel saturation curves that do not have an existing parallel in current DGE analyses.
Additional existing tools include WGCNA (Weighted Gene Co-expression Network Analysis) and NMF (Non-negative Matrix Factorization), often used to co-expression networks and gene expression patterns. Each of these approaches require either correlation based clustering or matrix factorization, while THRESHOLD does not necessitate network interference or dimensionality reduction. Instead, THRESHOLD’s novel saturation curve perspective offers a complementary perspective that is able to produce consistent result measures for comparative analysis and evaluation with minimal oversight and interference for broader comparative analysis application. This makes THRESHOLD particularly well-equipped for highlighting significant recurrent expression signatures between heterogeneous cohorts.
To address existing limitations and complement analyses through a distinct form of calculation, we have developed a new tool called THRESHOLD, which offers a novel computational analysis to visualize both overarching and detailed patterns of gene expression across large patient samples.
THRESHOLD is a tool for analysis of transcriptomic data across large samples of patients to understand the cohesion of the most upregulated/downregulated genes in a given disease. The tool generates a novel metric called saturation, which quantifies the proportion of genes occurring frequently at similar ranks of relative upregulation. THRESHOLD offers several user-inputted parameters that allow researchers to shape their analyses to the desired saturation type, restriction factors, and rank type. Additionally, THRESHOLD offers in-tool statistical analyses to help understand significant stratification between patterns of gene expression. The tool outputs an interactive graph of saturation permitting the calculation of specific saturation thresholds and most saturated genes.
THRESHOLD includes two distinct analytical features: incremental saturation and overall saturation. Incremental saturation is designed to articulate the step-by-step changes in gene expression patterns by nth rank, whereas overall saturation offers a comprehensive view of changes in gene expression up to a specific nth rank. In addition, THRESHOLD can generate gene ranks based on high to low (most upregulated) or low to high (most downregulated) expression. This calibration allows researchers to easily investigate both the most upregulated or downregulated genes depending on their needs, expanding THRESHOLD’s utility. The primary aim of developing THRESHOLD is to facilitate our understanding of the multifaceted transcriptomic landscape. To that end, it plays an essential role in several key domains:
Disease sub-stratification: Beyond merely classifying diseases based on observable symptoms or common biomarkers, THRESHOLD provides a detailed molecular perspective, identifying unique genetic markers within different patient groups and enabling a more refined and precise disease categorization.
Drug target identification: THRESHOLD can be used to identify genes that are highly upregulated or downregulated across diverse patient populations. The highly upregulated or downregulated genes occurring across large numbers of patients are likely indicative of genes critical to the disease’s pathology and carcinogenesis. Thus, THRESHOLD’s analyses can aid in the identification of new potential therapeutic targets.
Biomarker discovery: In conditions such as cancer, where prompt diagnosis is crucial, THRESHOLD offers valuable assistance. THRESHOLD helps researchers identify genes that are frequently overexpressed, suggesting their possible role as indicators of disease and aiding in the development of early detection methods. Together, the THRESHOLD serves as a useful addition to genomics research by providing insights into disease mechanisms and supporting the advancements in personalized medicine.
Materials and methods
Data requirements
THRESHOLD requires the input of a file of patient transcriptomic data with bulk messenger RNA (mRNA)-seq expression data in the form of z-scores, comparing expression against a control population, or percentiles, ranking expression relatively within an individual patient’s transcriptomic profile.
While, THRESHOLD is primarily intended for bulk RNA-seq mRNA data, it can be applied to normal-referenced scRNA-seq data to perform comparative analyses between scRNA-seq clusters, given some of the following considerations. Dropout effects and sparsity often characteristic of scRNA-seq data primarily affects the outputs of reverse rank (downregulated) saturation analyses if impacting specific genes more often than others. Regardless, regular rank (upregulated) saturation analyses remain largely unaffected by dropout effects or sparsity, and given a sufficiently large sample size, can cut through the high variability associated with scRNA-seq data.
The input file for THRESHOLD should be in a tab-delimited ('\t’) format in a text (.txt) file. Each column should represent a patient sample, and each row should correspond to a specific gene or biomarker. The first row of the file should contain the header, including a required “Hugo_Symbol” (HUGO Gene Nomenclature Committee) title for the first column, a blank space on the second column, and the relevant patient identifier for each successive column. The first column (“Hugo_Symbol”) should include unique identifiers of each gene, the second column should remain empty, while each successive column should include expression data by each relevant patient. Missing expression data should be represented by “NA.”
Saturation analysis methods
THRESHOLD offers two primary functionalities: “overall saturation” and “incremental saturation.” The former offers a holistic perspective on gene expression by considering the saturation of all genes up to and including the nth level, while the latter focuses on saturation at each specific nth gene level. Both are original metrics developed to facilitate THRESHOLD’s novel outputs.
Overall saturation: The saturation of all the genes up to and including the nth level. Quantifies the count of genes up to the nth level that exceed the inputted restriction level, divided by the total count of genes up to and including the nth level (Fig. 2).
Figure 2.
Overall saturation visual example: Genes are ranked by most relatively upregulated for each patient group, represented by each column of shapes. Given a restriction level of one, any gene that occurs beyond the first occurrence by nth level is considered saturated. For overall saturation, saturation is calculated by the proportion of saturated genes at all levels up to and including a given nth level, n = 3 in this case.
Quantitatively, this is represented by the sum of each unique element in the multiset A’s non-negative greater value between the element count in multiset A minus the restriction level divided by the number of elements in the multiset A. Formula for overall saturation was presented in Fig. 1.
Figure 1.
Formula for overall saturation.
Visual example:
Find the overall saturation up to the third level (n = 3) given a restriction level of 1:
Incremental saturation: The saturation only at the incremental nth gene level. Quantifies the count of genes at the nth level that exceed the inputted restriction level up to and including that level, divided by the total count of genes at the nth level (Fig. 4).
Figure 4.
Incremental saturation visual example: Genes are ranked by most relatively upregulated for each patient group, represented by each column of shapes. Given a restriction level of one, any gene that occurs beyond the first occurrence by nth level is considered saturated. For incremental saturation, saturation is calculated by the proportion of saturated genes specifically present at the given nth level, n = 3 in this case.
Quantitatively, this is represented by the sum of each unique element in the multiset B’s non-negative lesser value between the difference in the element count in multiset A minus the restriction level and the element count in multiset B divided by the number of elements in multiset B. Formula for overall saturation was presented in Fig. 3.
Figure 3.
Formula for incremental saturation.
Visual example:
Find the incremental saturation of the third level (n = 3) given a restriction level of 1:
Software and algorithm
The THRESHOLD software was built using Python and Java and leveraged dependencies including NumPy, Matplotlib, SciPy, and PyQt6 to support calculations, visualizations, statistics, and GUI display, respectively. The core algorithm relies on a novel statistical metric developed called “Saturation,” which systematically analyzes gene expression data to detect highly expressed genes shared across different patient datasets.
To demonstrate the application of THRESHOLD, gene expression data from TCGA [13] datasets accessed via UCSC Xena [14] were used, and clinical features from cBioPortal [15] were incorporated for further comparative analyses. The datasets spanned several cancers including bladder urothelial carcinoma, prostate adenocarcinoma, breast invasive carcinoma, and colon adenocarcinoma, incorporating gene expression profiles of patients. The first three cancer datasets are from the TCGA Pan-Cancer Atlas, which describes preprocessing steps to correct for batch effects. An additional dataset, TCGA Colon Adenocarcinoma (UCSC Xena: IlluminaHiSeq percentile), is included in the supplementary file. This dataset, while not explicitly batch-effect corrected, was normalized via percentiles within each sample. This normalization technique allows for fair comparisons of relative expression between samples, even if there are differences in the magnitudes of measured expression.
THRESHOLD requires the input of a file of patient sample transcriptomic data with gene expression data in the form of z-scores or percentiles comparing expression against a control population or ranked relatively within an individual patient’s expression. THRESHOLD initially ranks all genes by relative expression for each patient, which is determined by each gene’s z-scores, percentiles, and ranks. These relative magnitudes, and thus comparisons between intra-sample genes, are maintained regardless of the normalization method chosen. Such data are publicly available in many TCGA datasets via UCSC Xena, cBioPortal, and more (Fig. 5).
Figure 5.
Overview of GUI workflow. An outline of THRESHOLD’s workflow, specifically how data are organized in order to calculate and understand gene saturation.
For most analyses, overall saturation provides the most consistent results, especially for smaller datasets, while incremental saturation offers insights into marginal changes in saturation but benefits from more sample sizes for consistency. Moreover, researchers should adjust restriction levels based on their specific dataset sensitivity. From our analyses, starting at ∼10% and adjusting higher for increased sensitivity or lower for lesser sensitivity until a discrete saturation curve is apparent. The nth level parameter, typically set between 25 and 100, determines the range of genes analyzed per sample, and its magnitude should be chosen based on how wide of a view a researcher wants to analyze for each sample.
Data preprocessing
Once uploaded to the THRESHOLD GUI, the data will automatically be cleaned to ensure the most relevant results. This will include removing pseudogenes including non-coding RNA such as snoRNA, sniRNA, snRNA, and microRNA. The removed pseudogenes will be exported to a resulting .txt file to validate appropriate removal. If a user wants to keep these pseudogenes, they can also select to include them. Gene expression is then automatically ranked for each patient by most upregulated or downregulated depending on the user’s selection within the tool. These ranks are recorded as JSON data for each patient. This rank leverages the z-score/percentile value assigned to each gene from the given dataset. Once saturation values are calculated for each level up to the user-inputted request, the saturation data are saved as a downloadable .txt file and displayed within an interactive graph. Users can calculate specific saturation values at given levels and also find when a saturation threshold is met. Both the graph and the aforementioned .txt file of the overall saturation data are easily exportable for further analysis or records.
Comparative analysis
Standardization
To compare saturation across datasets, restriction levels can be standardized. Increasing the restriction level by a factor of the data set size increases and produces the same saturation results and can be used for standardization. For example, a data set of 100 patients has a restriction level of two genes. To standardize a data set of 200 patients (twice the size), you would need to double the restriction level: 2 × 2 = 4.
It is important to mention that restriction levels in terms of a percentage are universal and do not need further standardization. However, due to round by patient sample size, there may be potentially ±1 gene differences in the restriction level.
Statistical analysis
THRESHOLD also facilitates the statistical comparison between two saturation data sets to assess whether there are statistically significant differences between samples to provide additional insights in user analyses. Ensure the files are in the proper, standard format (.txt) as exported by the THRESHOLD tool. The files should have three columns: “nth gene included,” “incremental saturation,” and “overall saturation.” The file should be the same size, i.e. the same number of rows. The test of choice is a two-tailed unpaired t-test, assessing the difference between each dataset’s respective saturation results. This test is suitable because subjects are independent, and the measured differences are normally distributed (per the central limit theorem, sample sizes ≤ 30 are normally distributed).
Once uploaded, THRESHOLD outlines the result of your analyses, describing whether the differences between the saturation types in each data set yielded statistically significant differences. Additional relevant statistical measures, including the calculated P-value, are also listed. A more detailed description of the statistical analyses can also be exported with a (.txt) file of relevant calculated statistical measures for each of the saturation types.
Furthered analysis
Saturation data, including saturation curves and the most saturated gene ascertained from the THRESHOLD tool were leveraged to facilitate further analyses and demonstrate broader use cases of THRESHOLD. g:Profiler [16] was utilized to conduct g:GOST functional profiling to understand the functional role of the most saturated genes identified. Proteinarium [17] was employed to construct protein–protein interaction networks, while Gephi [18] was utilized for visualizing these networks and applying various network parameters to better understand the critical pathways and communities driving cancer pathology. Moreover, network metrics such as modularity, separation testing [19], and Jaccard Similarity [20] were leveraged to understand patterns within and between the created protein–protein interaction networks.
GUI overview
THRESHOLD is present as a local GUI (Fig. 7). When run, a home page facilitates a new session or statistical comparison of existing analyses. To investigate a new data set, a .txt file can be uploaded and relevant saturation parameters to export a graph of saturation supporting results. To perform a statistical comparison, upload two reference .txt files of THRESHOLD’s exported saturation data to assess whether there are statistically significant differences between samples to provide additional insights in user analyses.
Figure 7.
Bladder urothelial carcinoma case study continued. (A) Analysis workflow. Datasets acquired from cBioPortal for Cancer Genomics, divided by sex and mRNA-seq data were compared utilizing the THRESHOLD tool. (B) Upregulated overall saturation of bladder urothelial carcinoma by sex. Overall saturation (restriction level 10%) was calculated for each of the cancer datasets by stage. Data were compiled into one graph elucidating differences in gene saturation by gene expression. (C) Top 10 most saturated upregulated genes by stage. The 10 most saturated genes from each of the bladder urothelial carcinoma datasets visualized indicating the percent of patients expressing the saturated gene within the nth gene rank included, 100 (Fig. 7B).
Results
Validation testing
Validation testing was conducted on a small test data set to verify the calculations outputted by THRESHOLD (Table 1). From our testing, the tool has demonstrated consistent accuracy (Table 2).
Table 1.
Saturation test dataset
| Hugo_symbol | - | Pat1 | Pat2 | Pat3 | Pat4 |
|---|---|---|---|---|---|
| A | - | 1 | 5 | 1 | 5 |
| B | - | −3 | 3 | 5 | 0 |
| X | - | 5 | 0 | −1 | −2 |
| W | - | −1 | −1 | −2 | 3 |
| Q | - | −2 | −3 | 3 | 1 |
| Y | - | 3 | −2 | −3 | −3 |
| T | - | −4 | 1 | 0 | −1 |
An abbreviated data set of random “genes” represented by letters and corresponding z-scores representing expression was generated. The data set size was chosen to best permit and illustrate manual calculation of saturation per expanding computational complexity with increased dimensions. This data can then be used to calculate manual saturation values to verify THRESHOLD’s outputs.
Table 2.
T-test test dataset
| Saturation data reference 1 | Saturation data reference 2 | ||||
|---|---|---|---|---|---|
| Nth gene included | Incremental saturation | Overall saturation | Nth gene included | Incremental saturation | Overall saturation |
| 1 | 0.2 | 0.2 | 1 | 0.1 | 0.1 |
| 2 | 0.3 | 0.2 | 2 | 0.4 | 0.3 |
| 3 | 0.4 | 0.3 | 3 | 0.5 | 0.4 |
| 4 | 0.6 | 0.4 | 4 | 0.5 | 0.4 |
| 5 | 0.5 | 0.4 | 5 | 0.4 | 0.4 |
| 6 | 0.6 | 0.4 | 6 | 0.7 | 0.4 |
| 7 | 0.7 | 0.4 | 7 | 0.8 | 0.5 |
| 8 | 0.8 | 0.8 | 8 | 0.9 | 0.5 |
| 9 | 0.9 | 0.8 | 9 | 0.95 | 0.5 |
| 10 | 0.9 | 0.8 | 10 | 0.9 | 0.5 |
Two arbitrarily generated saturation data references were generated to verify THRESHOLD’s statistical analysis outputs. The data sets mirror saturation data automatically exported by the THRESHOLD tool. References 1 and 2 were then compared to corroborate calculated statistically significant differences and corresponding statistical parameters.
Given the following test dataset:
THRESHOLD passed verified calculations all the following tests:
| - Regular rank (level = 1, n = 5) | - Reverse rank (level = 1, n = 5) |
| - Regular rank (level = 2, n = 5) | - Reverse rank (level = 2, n = 5) |
| - Regular rank (level = 3, n = 5) | - Reverse rank (level = 3, n = 5) |
| - Regular rank (level = 1%, n = 5) | - Reverse rank (level = 1%, n = 5) |
| - Regular rank (level = 60%, n = 5) | - Reverse rank (level = 60%, n = 5) |
| - Regular rank (level = 70%, n = 5) | - Reverse rank (level = 70%, n = 5) |
Furthermore, we assessed THRESHOLD’s paired t-test function. We compared the tool’s outputted statistical metrics versus our validated calculations to verify its results.
We compared the following two test data sets:
There was no difference found in the calculated mean differences, standard deviations, pooled standard deviation, standard error, and t-statistic. This validates THRESHOLD’s output statistical measures.
Application of THRESHOLD tool
Bladder urothelial carcinoma use case
The THRESHOLD tool was used to elucidate insights into variable gene expression between bladder urothelial carcinoma patients. Patients were divided by cancer stage (II, III, and IV), and mRNA-seq data were analyzed to assess differences in saturation through bladder urothelial carcinoma progression (Fig. 6A). Upon compiling each of the upregulated saturation curves (Fig. 6B), a statistically significant difference (P < .00001) was found between each of the upregulated overall curves, with progressive stages yielding heightened saturation values. These results were reciprocated in the top 10 most saturated genes from each of the respective stages, with the most saturated genes being present in growing proportions of patients by progressive stages (Fig. 6C). Particularly, greater expression of cystatins, including CST2 and CST4, which were the first and second most saturated upregulated genes, respectively, in Stage IV bladder urothelial carcinoma patients, was observed through progressive stages. In addition, including COL10A1 and COL11A, which were the third and fourth most saturated upregulated genes, respectively, in Stage IV bladder urothelial carcinoma patients, were also observed through progressive stages. This trend is further supported by the greater observed Jaccard Similarity between adjacent stages (Fig. 6E), suggesting the progressive expression of certain critical saturated genes throughout bladder urothelial carcinoma stage progression. Moreover, functional profiling (g:GOST) performed on the top 10 most saturated genes in Stage IV cancer yielded significant major biological pathways and molecular functions relevant to bladder cancer progression (Fig. 6D). Most notably, these genes were implicated in cysteine-type endopeptidase inhibitor activity (P = 1.311 × 10–4) and endopeptidase inhibitor activity (P = 1.105 × 10–2), in addition to collagen degradation (P = 1.318 × 10–4) activity. The THRESHOLD tool was used to elucidate insights into expression profiles of bladder urothelial carcinoma patients based on datasets divided by sex (Fig. 7A). THRESHOLD found no statistically significant difference (P > .05) in the saturation between the male and female overall saturation curve (Fig. 7B). Upon further investigation into the most saturated genes in each dataset, similar top 10 most saturated genes were identified regardless of dataset, including cystatins and collagen alpha one genes (Fig. 7C). However, greater cystatin expression was observed in the female dataset, with CST2 and CST4 were present in 67% and 66% of female patients, respectively, juxtaposed to the male dataset, in which CST2 and CST4 were present in 58% and 61% of male patients.
Figure 6.
Bladder urothelial carcinoma case study. (A) Analysis workflow. Datasets acquired from cBioPortal for Cancer Genomics, divided by cancer stage and mRNA-seq data were compared utilizing the THRESHOLD tool. (B) Upregulated overall saturation of bladder urothelial carcinoma by stage. Overall saturation (restriction level 10%) was calculated for each of the cancer datasets by stage. Data were compiled into one graph elucidating differences in gene saturation by gene expression. (C) Top 10 most saturated upregulated genes by stage. The 10 most saturated genes from each of the bladder urothelial carcinoma datasets visualized indicating the percent of patients expressing the saturated gene within the nth gene rank included, 100 (Fig. 6B). (D) Functional analysis of the most saturated upregulated genes for Stage IV bladder urothelial carcinoma. g:Profiler functional profiling (g:GOST) performed on the top 10 most saturated gene (Fig. 6C) from the Stage IV data set. (E) Jaccard Similarity of most saturated upregulated genes by stage. The top 10 saturated genes were compared using Jaccard Similarity to assess differences among high expression at each stage.
Prostate adenocarcinoma use case
To discern insights in the relationship between onset of prostate adenocarcinoma and gene expression, the THRESHOLD tool was leveraged to compare gene expression across patient samples clustered by diagnosis age. Patients were distributed into three age groups (41–57, 58–64, and 65–78), and compiled mRNA-seq data were analyzed to assess variations in saturation of different onsets of prostate adenocarcinoma (Fig. 8A). Comparative analysis of each of the overall upregulated saturation curves (Fig. 8B) yielded a statistically significant difference (P < .0001), with younger onsets yielding heightened saturation values. The most saturated upregulated gene in all age groups was SLC45A2, demonstrating progressive prevalence in younger age groups (Fig. 9A). AMACR was also a top 10 most saturated gene in all age groups with progressive prevalence in younger age groups. However, all disease onsets demonstrated similar degrees of overall downregulated saturation, with high degrees of cohesion in all age groups (Fig. 8C). The most saturated downregulated genes driving this high expression include olfactory receptor (OR) genes, including OR9Q1 and OR2AT4, in addition to a histone regulating gene, HIST1H4F, present in all age groups’ top 10 most saturated downregulated genes (Fig. 9B). Heightened levels of Jaccard Similarity between top 10 most saturated downregulated genes in each age group (Fig. 9D) further supports the notion of a high degree of downregulated gene saturation in prostate adenocarcinoma. Functional profiling (g:GOST) performed on the top 100 most saturated upregulated and downregulated genes (Fig. 8D and E) both yielded significant olfactory receptor activity, particularly in the downregulated gene set (P = 2.074 × 10–20). Moreover, the upregulated gene set indicated relevance in histidine synthase activity (P = 1.061 × 10–2) in addition to G protein-coupled receptor activity (P = 1.997 × 10–3).
Figure 8.
Prostate adenocarcinoma case study. (A) Analysis workflow. Datasets acquired from cBioPortal for Cancer Genomics, divided by diagnosis age and mRNA-seq data were compared utilizing the THRESHOLD tool. (B) Upregulated overall saturation of prostate adenocarcinoma by diagnosis age. Overall saturation (restriction level 10%) was calculated for each of the cancer datasets by age. Data were compiled into one graph elucidating differences in gene saturation by gene expression. (C) Downregulated overall saturation of prostate adenocarcinoma by diagnosis age. Overall saturation (restriction level 10%) was calculated for each of the cancer datasets by age. Data were compiled into one graph elucidating differences in gene saturation by gene expression. (D) Functional analysis of the most saturated upregulated genes for diagnoses between 41 and 57 in prostate adenocarcinoma. g:Profiler functional profiling (g:GOST) performed on the top 10 most saturated upregulated genes (Fig. 9A) from the diagnosis age 41–57 data set. (E) Functional analysis of the most saturated downregulated genes for patients diagnoses between the ages of 41 and 57 in prostate adenocarcinoma. g:Profiler functional profiling (g:GOST) performed on the top 10 most saturated downregulated genes (Fig. 9B) from the diagnosis age 41–57 data set.
Figure 9.
Prostate adenocarcinoma case study continued. (A) Top 10 most saturated upregulated genes by diagnosis age. The 10 most saturated upregulated genes from each of the prostate adenocarcinoma datasets are visualized indicating the percent of patients expressing the saturated gene within the nth gene rank included, 100 (Fig. 8B). (B) Top 10 most saturated downregulated genes by diagnosis age. The 10 most saturated downregulated genes from each of the prostate adenocarcinoma datasets are visualized indicating the percent of patients expressing the saturated gene within the nth gene rank included, 100 (Fig. 8C). (C) Jaccard Similarity of most saturated upregulated genes by diagnosis age. The top 10 saturated upregulated genes were compared using Jaccard Similarity to assess differences among high gene expression at each stage. (D) Jaccard Similarity of most saturated downregulated genes by diagnosis age. The top 10 saturated downregulated genes were compared using Jaccard Similarity to assess differences among low gene expression at each stage.
Comparative analysis against DGE (Limma-Voom)
To capture THRESHOLD’s distinct utility in the context of existing methods of DGE analysis, a prostate adenocarcinoma dataset was analyzed for top genes by leveraging THRESHOLD’s regular rank (top saturated upregulated genes) and subsequently Limma-Voom, a common DGE analysis. Common cancer diagnostic metrics were then quantified for each of the resulting outputs to evaluate their respective relevance as tools for biomarker identification, markers of survival, etc.
Above represents a snapshot of the overall 100 top annotated genes outputted by each analysis on the same dataset. The analysis of the top 100 gene sets reveals a small Jaccard Similarity (0.0101) between the two methods, indicating a largely distinct output. This highlights THRESHOLD’s novel, complementary outputs to conventional DGE analyses.
Kaplan–Meier Survival Analysis curves were generated to evaluate the number of genes elucidated by each method that have predictive potential over several common clinical features. For each gene highlighted by the respective methods, patients were separated into high and low expression groups. Log-rank p-values adjusted for false discovery rate (FDR) were used to evaluate whether the genes offered predictive potential in disease free survival (DFS), the length of time after primary treatment where he patient survives without symptoms of the disease, and progression free survival (PFS), the length of time after primary treatment where a patient survives without their disease progressing. Overall survival and disease-specific survival were not included due to a lack of variance in the clinical data. In each of these groups there were only 10 out of 494 and 5 out of 494 patients that reflected the poor outcome phenotype, respectively.
First, an FDR adjusted p-value less than 5% was used to select for genes highly predictive with DFS and PFS. In this data set, DGE and THRESHOLD mirrored each other in the count of significant genes identified for DFS, though DGE outperformed in the PFS investigation, Table 4. However, it was observed that the top genes in the THRESHOLD genesets rendered more significant FDR adjusted p-values or greater magnitude hazard ratios (HR or 1/HR). Thus, a further investigation with a tighter FDR was warranted.
Table 4.
Significantly correlated (FDR adjusted P < .05) genes elucidated by DGE and THRESHOLD
| Metric | DGE genes correlated | THRESHOLD genes correlated |
|---|---|---|
| DFS | 10 | 10 |
| PFS | 23 | 11 |
Kaplan–Meier Survival analysis was conducted for the DGE and THRESHOLD gene sets for DFS and PFS. Median split between high and low expression groups for each gene was performed, and subsequent FDR adjusted log-rank P-values (P < .05) were used to evaluate the number of predictive genes in each set by feature.
When the FDR adjusted p-value was adjusted to a more stringent P< .01, THRESHOLD outperformed DGE in the DFS and PFS categories, identifying a greater number of significant correlated genes. This highlights THRESHOLD’s capacity to highlight clinically relevant genes that DGE might otherwise overlook.
It also must be noted that THRESHOLD and DGE operate in fundamentally different manners, which offer context to the results. THRESHOLD was used to specifically evaluate the top 100 most differentially upregulated genes, while DGE in its method of calculation can draw from the most differentially upregulated, differentially downregulated, or consistently moderately expressed with tight variance, and thus has a larger pool of genes to draw from. Thus, in a binary 1–1 analysis of identifying quantities of significant genes, gives DGE an advantage in having more opportunities of top genes to draw from. This might explain why DGE was able to find a greater number of significantly correlated genes with the more lenient FDR adjusted P < .05, while THRESHOLD had more significantly correlated genes when FDR was tightened to P < .01.
THRESHOLD’s capacity to identify highly predictive and clinically significant gene can also be visualized below in the top 5 FDR adjusted P-value graphs below for each group and feature:
In each case, THRESHOLD’s most significant predictive genes as measured by FDR adjusted P-value were more significant than DGE’s and also demonstrated greater magnitude (HR or 1/HR) of hazard ratio illustrating stronger predictive potential with the corresponding feature. An interesting pattern of >1 hazard ratios was observed in the THRESHOLD annotated genes as compared to DGE with hazard ratios <1. This may be considering the fact that THRESHOLD in its default setting performs regular rank search on saturated genes, searching for the most upregulated saturated genes. This is in comparison to DGE, which may draw from genes that are differentially downregulated, at moderate ranges of expression with tight variance, in addition to genes that are unusually upregulated. Together this illustrates THRESHOLD’s capacity to highlight specific saturated genes with significant predictive potential that otherwise would not have been annotated by DGE in a given dataset.
THRESHOLD identified a slightly higher proportion of known OncoKB genes than DGE, Table 6. However, perhaps significant is that considering the low Jaccard Similarity between the DGE and THRESHOLD gene sets and the fact it can elucidate significant genes highlights THRESHOLD as an effective complement to DGE analysis, potentially identifying relevant genes that conventional approaches might otherwise overlook. Moreover, the relatively modest overlap with OncoKB in either method is expected, as both methods may highlight genes with relevant biological roles beyond necessarily being classified as an oncogene or tumor suppressor gene.
Table 6.
Proportion of overlapping genes by method
| Metric | DGE | THRESHOLD |
|---|---|---|
| Overlapping OncoKB genes | 7/100 | 9/100 |
DGE and THRESHOLD’s top 100 genes were compared against a known database (OncoKB) to evaluate differences in the number of oncogenes/tumor suppressor genes highlighted by each database.
Discussion
The THRESHOLD’s results demonstrate the numerous valuable analyses and insights into disease as a novel genomics tool. Through its results, we have validated the tool's capacity for patient stratification and biomarker and drug targeting identification and found new areas for researchers to explore further. Our bladder urothelial carcinoma analyses (Fig. 6) indicate that the THRESHOLD tool demonstrated a strong capacity to analyze gene expression underlying disease progression. THRESHOLD highlighted heightened unregulated gene saturation in progressive stages, suggesting greater gene cohesion as cancer develops within patients (Fig. 6B). This pattern of greater homogeneity of expression in progressive cancer stages was further supported by the supplementary lung adenocarcinoma data set, (Supplementary Fig. S4B), where later cancer stages too demonstrated statistically significant increases in overall saturation (Supplementary Fig. S5B). Further insights permitted the identification of the most critical genes driving this growth in saturation, including the identification of collagen alpha 1 genes and cystatins (Fig. 6C). Collagen alpha 1 genes including COL10A1 identified as the third most saturated gene in Stage IV bladder. Functional analysis of COL10A1 and co-expressed genes has indicated significant relevance in ECM-receptor interaction, protein digestion and absorption, and PI3K-AKT signaling pathways. The COL10A1 gene has also been implicated as a valuable prognostic and predictive biomarker [21] in bladder cancer with increased COL10A1 expression being related to poor overall patient survival.
Cystatins including CDC2 and CDC4 were also implicated as leading drivers of gene saturation progression in bladder urothelial carcinoma, representing the first and second most saturated genes in Stage IV Bladder Cancer B (Fig. 6C). Prior literature validates these results, demonstrating that within patients with elevated cystatin expression [22] in urine such as Cystatin B, there was a short mean time to disease recurrence and in grade/stage progression of Transitional Cell Carcinoma (Urothelial Carcinoma). These results again demonstrate THRESHOLD’s utility in identifying novel predictive biomarkers implicated in disease progression, grade, and recurrence.
Moreover, the THRESHOLD tool demonstrated insights in evaluating differences in gene expression by clinical features such as age of diagnosis in the prostate adenocarcinoma datasets (Fig. 8). THRESHOLD demonstrated heightened upregulated gene saturation in samples with earlier diagnosis ages, suggesting greater gene cohesion in patients with earlier onsets of prostate adenocarcinoma (Fig. 8B). The most saturated genes driving this trend included SLC45A2 and AMACR, which demonstrated heightened saturation in younger ages (Fig. 9A). This pair is significant as scientific literature has implicated the two genes as major fusion transcripts underlying prostate adenocarcinoma. In a study [23] of eight fusion transcripts including SLC45A2-AMACR, it was demonstrated that 91% of patients positive for any of the fusion transcripts experienced recurrence, metastasis, or prostate adenocarcinoma-associated death even after radical prostatectomy as compared to only 37% of patients not carrying the fusion transcripts. This suggests THRESHOLD could be used as a tool to identify fusion transcripts implicated in disease severity.
Additionally, prostate adenocarcinoma expression data yielded significantly downregulated gene saturation in all diagnosis age groups, suggesting pronounced suppression of several genes (Fig. 8C). The most downregulated saturated genes demonstrated an intriguing pattern of olfactory receptor genes such as OR9Q1 and OR2AT4 (Fig. 9B), and functional analyses also revealed significant activities concerning olfactory receptors (Fig. 8E). In prostate adenocarcinoma specifically, the olfactory receptor gene OR51E2 has been implicated [24] in activating ERK1/2 via the Gβγ-PI3Kγ-ARF1 pathway elucidating important insights into characteristic MAPK hyper-activation. These results validate THRESHOLD’s relevance in elucidating critical gene pathways underlying differential progression or manifestation of disease.
In the supplementary colon adenocarcinoma analyses, we demonstrated THRESHOLD’s utility in elucidating novel drug targets among developed networks of highly saturated genes (Supplementary Fig. S1). Upon generation of an overall upregulated saturation curve (Supplementary Fig. S1B), the most saturated genes (Supplementary Fig. S1C) were extracted for network analyses and functional profiling. Of these network community hub genes and most saturated genes included actins, such as ACTB and ACTG1. ACTB and ACTG1 have been implicated as a significant biomarkers and regulators of tumorigenicity in numerous cancers, including hepatomas, renal cell carcinoma, and colon adenocarcinoma as we investigated. ACTB regulates [25] F-actins, which play roles in chemoresistance and increased cell proliferation. Moreover, ACTB can increase membrane protrusions, focal adhesions, and myosin activity, which increase tumor migration. In parallel, ACTG1 increases [25] cell proliferation through the mitochondrial apoptotic pathway, Warburg effect, upregulation of CDKs, and ROCK signaling pathways. Collectively, these roles suggest actins, including ACTB and ACTG1, could serve as biomarkers and drug-targeting candidates in several cancers, including colon adenocarcinoma.
Extended networks were generated from these most saturated genes, and modularity testing yielded compelling communities of interest (Supplementary Fig. S1D and F). This included a community around HSPA8, the most interconnected node with degree ten (Supplementary Fig. S1G). HSPA8 (also known as Hsc70) [26–30] has been implicated as a key driver of cell proliferation [31] under many conditions, with its absence inhibiting the growth of tumors and via apoptosis and cell cycle arrest. As indicated by the network, HSPA8 also interacts with MAPK1 via its broader community (Supplementary Fig. S1F), a signaling pathway under intense investigation for its relevance in a multitude of cancers. For these reasons, HSPA8 has been recognized as a biomarker candidate for the early detection and diagnosis of cancers, including endometrial carcinoma [32] and colon adenocarcinoma [33]. This gene, among several other network hub genes, including YWHAZ and EEF1A1, could serve as novel drug and therapeutic target, representing critical hubs underlying colon adenocarcinoma pathology.
THRESHOLD evaluated if there were differences in expression among breast invasive carcinoma patients undergoing or not undergoing radiation treatment (Supplementary Fig. S2). After separation of patient populations and comparison of individually calculated incremental saturation curves, there was no significant difference between either curve (Supplementary Fig. S2B and C). This suggests that the radiation treatments administered did not significantly impact gene expression between the sample populations and was further supported by networks developed via the most upregulated genes that were highly overlapping in the interactome with an sAB of −0.84 (Supplementary Fig. S3F). Additionally, similarly to prostate adenocarcinoma, olfactory receptor genes were highly downregulated as derived from the downregulated incremental saturation curve (Supplementary Fig. S2E). Functional analysis also implicated olfactory activity as a highly significant (P = 2.310 × 10–30) function of the top 100 most saturated downregulated genes. Several of these olfactory receptor genes have been implicated in cancer metastasis, invasion, and proliferation through signaling pathways such as NF-κB/STAT [34] in breast cancers.
Furthermore, modularity testing of the protein-protein interaction networks formed from the top 100 most upregulated saturated in each treatment group (Supplementary Fig. S3B and D) revealed significant insights into biomarker candidates and drug targets for new therapies. For example, overexpression of highly interconnected hub genes such as RBBP7 regulating chromatin metabolism has been associated with poor overall patient survival [35] in cancers such as esophageal squamous cell carcinoma, being implicated with enhanced tumor migration and invasion. Additionally, in adjacent network communities, MAPKAPK2 (MK2) was present and has significant relevance as a breast cancer biomarker and therapeutic target. The p38MAPK-MK2 signaling pathway contributes to the stimulation of triple-negative breast cancer tumorigenesis by promoting AP1 activity [36], associated with aggressive cancer manifestations. In all, the breast invasive carcinoma use case demonstrated utility in identifying drug targets and biomarkers, in addition to establishing where homogeneity may exist between samples of varying clinical features.
THRESHOLD was further used to pursue insights into variable gene expression between thyroid carcinoma patients. Patients were first divided by sex, and their respective mRNA-seq data were analyzed to assess differences in saturation between females and males. The upregulated overall saturation curves were calculated and compiled into one graph to perform comparisons. A statistically significant difference (P< .05) was observed, particularly at later nth gene ranks included between the female and male groups (Supplementary Fig. S6B). That being said, while having less overall saturation at later nth gene ranks included, the male thyroid carcinoma group exhibited a more saturated top gene, with QSOX1 being expressed in 93% of the studied patients up to the nth rank included compared to QSOX1 at 84% in the female group (Supplementary Fig. S6C). Taken together, this suggests potentially a narrower range of genes underlying male manifestations of thyroid cancer when compared to females, if the female group saw greater overall saturation in spite of these top saturated gene differences.
The THRESHOLD analysis tool was leveraged to investigate disease outcome-based differences in kidney renal clear cell carcinoma data sets. Patients were first divided by those who had died or lived following kidney renal clear cell carcinoma diagnosis, and mRNA-seq data were analyzed to assess differences in saturation based on their outcomes. The upregulated overall saturation curves were then calculated and compiled into a single graph for comparison. A statistically significant difference was found between the two curves (P < .001), with greater overall saturation being observed in the living dataset over the top nth genes included (Supplementary Fig. S7B). Moreover, the top saturated gene profiles observed in these two groups were quite distinct, with ZNF395 and NDUFA4L2 being the most and second most saturated genes in the alive group, yet not appearing at all in the top 10 for the deceased group (Supplementary Fig. S7C). Functional profiling of the alive top saturated genes too supports these top genes as relevant to kidney renal clear cell carcinoma, being annotated as a relevant pathway to the disease (Supplementary Fig. S7D). Both these genes are often implicated in responses to hypoxia and could prove insights into biological mechanisms underlying distinct patient outcomes in kidney renal clear cell carcinoma. THRESHOLD was also utilized to evaluate differences in gene expression for uterine corpus endometrial carcinoma in regards to three distinct subtypes: microsatellite instability high (MSI-High), copy number low (CN-Low), and copy number high (CN-High). Respective mRNA-seq data were evaluated for each feature and used to compile a graph of their overall saturation curves to assess differences in saturation. The upregulated overall saturation curves all showed significant differences (P < .01), though less so between MSI-High and CN-Low curves given homogeneity following initial nth genes included. The CN-High subtype was quite distinct from the other two curves, exhibiting much less overall saturation (Supplementary Fig. S8B). Moreover, top saturated in each subtype were quite distinct, with each group exhibiting a different top saturated gene and only some overlap in the remaining top 10 saturated genes between the other subtypes (Supplementary Fig. S8C). Together, this suggests a different expression profile underlying each subtype and calling for distinct approaches to understanding each subtype’s distinct pathology and its necessary response accordingly.
We investigated a head and neck squamous cell carcinoma data set. First, the data were divided into three groups by cancer stage (stage I & II, stage III, and stage IV), and their respective mRNA-seq data were analyzed to assess differences in gene expression. Upregulated overall saturation curves were calculated and compiled into a single graph and investigated for differences. A statistically significant difference (P < .001) was found between each curve, particularly toward middle nth ranks of expression, where stage IV exhibited the highest overall saturation (Supplementary Fig. S9B). Top-ranked saturated genes in each stage exhibited similar genes across top ranks of saturation, including SUN3 and homeobox genes such as HOXA13 and HOXB9 (Supplementary Fig. S9C). Data were also investigated for differences based on differing diagnosis age of head and neck squamous cell carcinoma. The mRNA-seq data were divided into three groups by diagnosis ages: ≤55, 56–65, and >65. Resulting overall saturation curves were compiled and compared in a single graph to evaluate differences in overall saturation. A statistically significant difference was found between all curves (P< .01), particularly between the ≤55 curves and the remaining, and less so between the 56–65 and >65 curves (Supplementary Fig. S10B). This suggests greater gene homogeneity of expression in earlier manifestations of the disease. Top-ranked saturated genes were also investigated, returning a similar pattern of SUN3 and homeobox genes (Supplementary Fig. S10C).
The preceding analysis represents THRESHOLD’s utility in a liver hepatocellular carcinoma analysis of mRNA-seq expression data divided by disease outcome. Overall saturation curves were compiled for both the living and deceased outcome groups and compared to evaluate differences in gene expression. A statistically significant difference (P < .0001) was observed between the living and deceased, with increased overall saturation exhibited in the deceased curve (Supplementary Fig. S11B). Top saturated gene analysis of each group highlighted sources of these differences in saturation. Increased saturation of homeobox genes HOXD4 was observed in addition to TERT, a gene critical for the generation of the enzyme telomerase (Supplementary Fig. S11C). Functional profiling of these top genes in the deceased group underscores the potential relevance of TERT-RMRP complex in the expression profile of liver hepatocellular carcinoma (Supplementary Fig. S11D).
Finally, THRESHOLD’s comparative analysis against a common DGE method, Limma-Voom, illustrates THRESHOLD’s potential for distinct analysis with significant clinical value. Each tool was leveraged on a prostate adenocarcinoma data set, and top genes identified by either method were compared for their differences and clinical significance. The gene sets identified by THRESHOLD and DGE demonstrated a Jaccard Similarity of 0.0101 (Table 3), illustrating largely distinct results. Importantly, THRESHOLD’s outputted genes in a Kaplan–Meier survival analysis of two clinical features, DFS and PFS, indicated a capacity to highlight highly significant genes (Table 5) associated with greater predictive value (Fig. 10) than the DGE gene set outputted. Together, this underscores THRESHOLD’s capacity for distinct and complementary analysis to DGE, which might otherwise overlook highly significant genes of interest with predictive clinical value.
Table 3.
Comparative analysis: top genes by method
| Top n gene | THRESHOLD | DGE |
|---|---|---|
| 1 | SLC45A2 | APOBEC3C |
| 2 | AMH | QPRT |
| 3 | AMACR | DLX2 |
| 4 | KISS1R | CA14 |
| 5 | XPO6 | SERPINA5 |
| 6 | IL1F10 | SNCG |
| 7 | MON1B | ANGPT1 |
| 8 | OR4N4 | PPARGC1A |
| 9 | GALR3 | GPX2 |
| 10 | OR52R1 | KCNJ15 |
| Jaccard Similarity (top 100) | 0.0101 | |
The top 10 genes of the overall 100 genes outputted by each method are represented earlier. Jaccard similarity between the overall top 100 gene sets is included to investigate differences in analysis outputs.
Table 5.
Significantly correlated (FDR adjusted P < .01) genes elucidated by DGE, and THRESHOLD
| Metric | DGE genes correlated | THRESHOLD genes correlated |
|---|---|---|
| DFS | 0 | 3 |
| PFS | 1 | 3 |
Kaplan–Meier survival analysis was conducted for the DGE and THRESHOLD gene sets for DFS and PFS. Median split between high and low expression groups for each gene was performed, and subsequent FDR adjusted log-rank P-values (P < .01) were used to evaluate the number of predictive genes in each set by feature.
Figure 10.
Kaplan–Meier comparative analysis. (A) PFS Kaplan–Meier Curves. Top 100 genes outputted by DGE and THRESHOLD tested for predictive value in PFS clinical data. Top five curves and displayed for each method selected by lowest FDR adjusted P-value. (B) DFS Kaplan–Meier curves. Top 100 genes outputted by DGE and THRESHOLD tested for predictive value in DFS. Top five curves displayed for each method by lowest FDR adjusted P-value.
Conclusion
The THRESHOLD bioinformatics tool serves as a robust and innovative approach for deciphering shared gene expression patterns across patient populations. Its dual functionalities, incremental saturation and overall saturation, provide both granular and overarching perspectives on gene expression and facilitate relevant comparative analyses. This makes THRESHOLD not only pivotal for disease research, drug repurposing studies, and personalized medicine but also for understanding the stepwise changes in gene expression as diseases progress or respond to treatment. Moreover, adapting the THRESHOLD to focus on suppressed genes could further extend its utility in exploring downregulated or silenced pathways in various conditions. Collectively, the tool offers significant potential in advancing our understanding of molecular signatures, streamlining patient stratification, and fostering the development of tailored therapeutic interventions.
The THRESHOLD tool is freely available at GitHub (https://github.com/alperuzun/THRESHOLD) and Zenodo (https://zenodo.org/records/15287195).
Supplementary Material
Acknowledgements
We would like to thank Chair of Pathology and Laboratory Medicine Jonathan D. Curtis, MD, PhD, The Warren Alpert Medical School, Brown University for his encouraging support and for providing resources that made the publication of this work possible. The results published here are in part based upon data generated by the TCGA Research Network: https://www.cancer.gov/tcga.
Author contributions: Finan M Gammell (Data curation [lead], Formal analysis [lead], Investigation [equal], Methodology [equal], Software [lead], Validation [equal], Visualization [lead], Writing—original draft [equal], Writing—review & editing [equal]), Jennifer Li (Data curation [supporting], Software [supporting]), Christopher Elco (Funding acquisition [lead], Writing—review & editing [supporting]), Jessica Plavicki (Funding acquisition [lead], Writing—review & editing [supporting]), Alper Uzun (Conceptualization [lead], Funding acquisition [supporting], Investigation [equal], Methodology [lead], Project administration [lead], Resources [lead], Software [equal], Supervision [lead], Writing—original draft [equal], Writing—review & editing [equal]).
Contributor Information
Finán Gammell, East Greenwich High School, East Greenwich, RI 02818, United States.
Jennifer Li, Department of Computer Science, Brown University, Providence, RI 02906, United States.
Christopher Elco, Department of Pathology and Laboratory Medicine, BioMed, Brown University, Providence, RI 02903, United States.
Jessica Plavicki, Department of Pathology and Laboratory Medicine, BioMed, Brown University, Providence, RI 02903, United States.
Alper Uzun, Legorreta Cancer Center, Brown University, Providence, RI 02903, United States; Department of Pathology and Laboratory Medicine, BioMed, Brown University, Providence, RI 02903, United States; Department of Pediatrics, Warren Alpert Medical School, Brown University, Providence, RI 02903, United States; Center for Clinical Cancer Informatics and Data Science (CCIDS), Brown, Providence, RI, 02912, United States.
Supplementary data
Supplementary data is available at NAR Cancer online.
Conflict of interest
None declared.
Funding
Pathology and Laboratory Medicine (PLM) Summer Internship Program, Brown University.
References
- 1. Genes, Behavior, and the Social Environment Hernandez LM, Blazer DG Moving beyond the nature/nurture debate. Institute of Medicine (US) Committee on Assessing Interactions Among Social, Behavioral, and Genetic Factors in Health. 3. 2006; Washington (DC)National Academies Press (US)50–57. [PubMed] [Google Scholar]
- 2. Grizzi F, Chiriva-Internati M Cancer: looking for simplicity and finding complexity. Cancer Cell Int. 2006; 6:4. 10.1186/1475-2867-6-4. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3. Jemal A, Siegel R, Xu J et al. Cancer statistics, 2010. CA Cancer J Clin. 2010; 60:277–300. [DOI] [PubMed] [Google Scholar]
- 4. Siegel RL, Giaquinto AN, Jemal A Cancer statistics, 2024. CA Cancer J Clin. 2024; 74:12–49. [DOI] [PubMed] [Google Scholar]
- 5. U.S. Census Bureau U.S. Census Bureau Announces 2010 Census Population Counts Apportionment Counts Delivered to President 2010. https://www.census.gov/newsroom/releases/archives/2010_census/cb10-cn93.html.
- 6. U.S. Census Bureau U.S. and world population clock 2024. https://www.census.gov/popclock/.
- 7. Heiser LM, Sadanandam A, Kuo W-L et al. Subtype and pathway specific responses to anticancer compounds in breast cancer. Proc Natl Acad Sci USA. 2012; 109:2724–9. 10.1073/pnas.1018854108. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8. McDermaid A, Monier B, Zhao J et al. Interpretation of differential gene expression results of RNA-seq data: review and integration. Briefings Bioinf. 2019; 20:2044–54. 10.1093/bib/bby067. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9. Love MI, Huber W, Anders S Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol. 2014; 15:550. 10.1186/s13059-014-0550-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10. Robinson MD, McCarthy DJ, Smyth GK edgeR: a Bioconductor package for differential expression analysis of digital gene expression data. Bioinformatics. 2010; 26:139–40. 10.1093/bioinformatics/btp616. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11. Ritchie ME, Phipson B, Wu D et al. limma powers differential expression analyses for RNA-sequencing and microarray studies. Nucleic Acids Res. 2015; 43:e47. 10.1093/nar/gkv007. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12. Chen B, Khodadoust MS, Liu CL et al. Profiling tumor infiltrating immune cells with CIBERSORT. Methods Mol Biol. 2018; 1711:243–59. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13. NCI The Cancer Genome Atlas Program (TCGA) [2024]. https://www.cancer.gov/ccg/research/genome-sequencing/tcga.
- 14. Goldman MJ, Craft B, Hastie M et al. Visualizing and interpreting cancer genomics data via the Xena platform. Nat Biotechnol. 2020; 38:675–8. 10.1038/s41587-020-0546-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15. Cerami E, Gao J, Dogrusoz U et al. The cBio cancer genomics portal: an open platform for exploring multidimensional cancer genomics data. Cancer Discov. 2012; 2:401–4. 10.1158/2159-8290.CD-12-0095. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16. Kolberg L, Raudvere U, Kuzmin I et al. g: Profiler—interoperable web service for functional enrichment analysis and gene identifier mapping (2023 update). Nucleic Acids Res. 2023; 51:W207–12. 10.1093/nar/gkad347. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17. Armanious D, Schuster J, Tollefson GA et al. Proteinarium: multi-sample protein–protein interaction analysis and visualization tool. Genomics. 2020; 112:4288–96. 10.1016/j.ygeno.2020.07.028. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18. Bastian M, Heymann S, Jacomy M Gephi: an open source software for exploring and manipulating networks. International AAAI Conference on Weblogs and Social Media. 2009; 194:1031–37. 10.1016/j.juro.2015.04.079. [DOI] [Google Scholar]
- 19. Menche J, Sharma A, Kitsak M et al. Uncovering disease-disease relationships through the incomplete interactome. Science. 2015; 347:1257601. 10.1126/science.1257601. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20. Hancock JM Jaccard distance (Jaccard index, Jaccard similarity coefficient). [Google Scholar]
- 21. Wang X, Bai Y, Zhang F et al. Prognostic value of COL10A1 and its correlation with tumor-infiltrating immune cells in urothelial bladder cancer: a comprehensive study based on bioinformatics and clinical analysis validation. Front Immunol. 2023; 14:955949. 10.3389/fimmu.2023.955949. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22. Feldman AS, Banyard J, Wu C-L et al. Cystatin B as a tissue and urinary biomarker of bladder cancer recurrence and disease progression. Clin Cancer Res. 2009; 15:1024–31. 10.1158/1078-0432.CCR-08-1143. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23. Yu YP, Ding Y, Chen Z et al. Novel fusion transcripts associate with progressive prostate cancer. Am J Pathol. 2014; 184:2840–9. 10.1016/j.ajpath.2014.06.025. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24. Xu X, Khater M, Wu G The olfactory receptor OR51E2 activates ERK1/2 through the Golgi-localized Gβγ-PI3Kγ-ARF1 pathway in prostate cancer cells. Front Pharmacol. 2022; 13:1009380. 10.3389/fphar.2022.1009380. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25. Suresh R, Diaz RJ The remodelling of actin composition as a hallmark of cancer. Transl Oncol. 2021; 14:101051. 10.1016/j.tranon.2021.101051. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26. Chen S, Brown IR Neuronal expression of constitutive heat shock proteins: implications for neurodegenerative diseases. Cell Stress Chaperones. 2007; 12:51–8. 10.1379/CSC-236R.1. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27. Hubbi ME, Hu H, Kshitiz AI et al. Chaperone-mediated autophagy targets hypoxia-inducible factor-1alpha (HIF-1alpha) for lysosomal degradation. J Biol Chem. 2013; 288:10703–14. 10.1074/jbc.M112.414771. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28. Li PP, Itoh N, Watanabe M et al. Association of simian virus 40 vp1 with 70-kilodalton heat shock proteins and viral tumor antigens. J Virol. 2009; 83:37–46. 10.1128/JVI.00844-08. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29. Lu HAJ, Sun TX, Matsuzaki T et al. Heat shock protein 70 interacts with aquaporin-2 and regulates its trafficking. J Biol Chem. 2007; 282:28721–32. 10.1074/jbc.M611101200. [DOI] [PubMed] [Google Scholar]
- 30. Welsch T, Younsi A, Disanza A et al. Eps8 is recruited to lysosomes and subjected to chaperone-mediated autophagy in cancer cells. Exp Cell Res. 2010; 316:1914–24. 10.1016/j.yexcr.2010.02.020. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31. Ying B, Xu W, Nie Y et al. HSPA8 is a new biomarker of triple negative breast cancer related to prognosis and immune infiltration. Dis Markers. 2022; 2022:8446857. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32. Shan N, Zhou W, Zhang S et al. Identification of HSPA8 as a candidate biomarker for endometrial carcinoma by using iTRAQ-based proteomic analysis. Onco Targets Ther. 2016; 9:2169–79. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33. Ruzza A, Zaltron E, Vianello F et al. HSPA8 and HSPA9: Two prognostic and therapeutic targets in breast, colon, and kidney cancers?. Biochimica et Biophysica Acta (BBA)-Molecular Basis of Disease. 2025; 167827. [DOI] [PubMed] [Google Scholar]
- 34. Li M, Schweiger MW, Ryan DJ et al. Olfactory receptor 5B21 drives breast cancer metastasis. iScience. 2021; 24:103519. 10.1016/j.isci.2021.103519. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 35. Yu N, Zhang P, Wang L et al. RBBP7 is a prognostic biomarker in patients with esophageal squamous cell carcinoma. Oncol Lett. 2018; 16:7204–11. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 36. Chen H, Padia R, Li T et al. Signaling of MK2 sustains robust AP1 activity for triple negative breast cancer tumorigenesis through direct phosphorylation of JAB1. NPJ Breast Cancer. 2021; 7:91. 10.1038/s41523-021-00300-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.