Joint Secondary Transcriptomic Analysis of Non-Hodgkin’s B-Cell Lymphomas Predicts Reliance on Pathways Associated with the Extracellular Matrix and Robust Diagnostic Biomarkers

Abstract

Approximately 450,000 cases of Non-Hodgkin’s lymphoma are annually diagnosed worldwide, resulting in ~240,000 deaths. An augmented understanding of the common mechanisms of pathology among larger numbers of B-cell Non-Hodgkin’s Lymphoma (BCNHL) patients is sorely needed. We consequently performed a large joint secondary transcriptomic analysis of the available BCNHL RNA-sequencing projects from GEO, consisting of 322 relevant samples across ten distinct public studies, to find common underlying mechanisms and biomarkers across multiple BCNHL subtypes and patient subpopulations; limitations may include lack of diversity in certain ethnicities and age groups and limited clinical subtype diversity due to sample availability. We found ~10,400 significant differentially expressed genes (FDR-adjusted p-value < 0.05) and 33 significantly modulated pathways (Bonferroni-adjusted p-value < 0.05) when comparing BCNHL samples to non-diseased B-cell samples. Our findings included a significant class of proteoglycans not previously associated with lymphomas as well as significant modulation of genes that code for extracellular matrix-associated proteins. Our drug repurposing analysis predicted new candidates for repurposed drugs including ocriplasmin and collagenase. We also used a machine learning approach to identify robust BCNHL biomarkers that include YES1, FERMT2, and FAM98B, which have not previously been associated with BCNHL in the literature, but together provide ~99.9% combined specificity and sensitivity for differentiating lymphoma cells from healthy B-cells based on measurement of transcript expression levels in B-cells. This analysis supports past findings and validates existing knowledge while providing novel insights into the inner workings and mechanisms of transformed B-cell lymphomas that could give rise to improved diagnostics and/or therapeutics.

Transcriptomics secondary analysis of severe human infection with SARS-CoV-2 identifies gene expression changes and predicts three transcriptional biomarkers in leukocytes

“…A random forest classification method (using R randomforest version 4.6–14) was then used to classify the severity phenotype of each sample [51] . The hyperparameters for the random forest model were 10,000 decision trees per forest, gini index as impurity criterion, and the square root of the number of features (genes in this case) to use for each split in the decision tree, as described previously [52] , [53] .…”

Pathway2Targets: an open-source pathway-based approach to repurpose therapeutic drugs and prioritize human targets

“…Similarly, the enrichr pathway enrichment software only required a gene symbol, log 2 fold-change values, and FDR-adjusted p -values as input from the DEG list. The statistically significant differentially expressed genes (DEGs; FDR-corrected p -value < 0.05) were then subjected to signaling pathway analysis using the Signal Pathway Impact Analysis (SPIA) algorithm with 3,000 bootstrap replicates to generate a null distribution for each of over 2,000 public signaling pathways ( Tarca et al, 2009 ), as reported previously ( Scott, Jensen & Pickett, 2021 ; Gray et al, 2022 ; Moreno et al, 2022 ; Rapier-Sharman, Clancy & Pickett, 2022 ; Ferrarini et al, 2021 ; Scott et al, 2022 ; Gifford & Pickett, 2022 ). The lists of pathways were derived from publicly available versions of KEGG ( Aoki-Kinoshita & Kanehisa, 2007 ), Reactome ( Jassal et al, 2020 ), Pathway Interaction Database ( Schaefer et al, 2009 ), BioCarta, and Panther ( Mi et al, 2017 ).…”

“…This updated algorithm retrieves additional target information, clinical trial data, automatically fetches the reactome pathway diagrams for the signaling pathways with the highest number of targets, and accepts reactome pathway enrichments generated by the enrichr algorithm ( Xie et al, 2021a ). This additional data and prioritization method are used by the updated algorithm to generate ranked lists of targets and therapeutics that can be applicable to multiple use cases ( Scott, Jensen & Pickett, 2021 ; Gray et al, 2022 ; Moreno et al, 2022 ; Rapier-Sharman, Clancy & Pickett, 2022 ). The entities in these lists can then be evaluated as candidates for condition-specific repurposing efforts based solely on the unique signaling pathway “profile” for the disease/condition of interest.…”

Secondary Analysis of Human Bulk RNA-seq Dataset Suggest Potential Mechanisms and Transcriptional Biomarkers for Letrozole Resistance in Estrogen Positive (ER+) Breast Cancer

“…Separately, the Salmon read counts for each sample were combined into a tabular format and labeled as "resistant" or "sensitive" to treatment with Letrozole. This table was then used as input to the XGboost algorithm [39], which uses a tree-based method to train a model from 80% of the dataset and then quantifies its performance using the remaining 20% of the data [40,41], which minimizes model overfit. For the initial analysis, the gain metric was calculated from the read counts for all detected genes across all samples.…”

“…Given the gain metrics from the whole transcriptome, the number of genes/features being evaluated was reduced to the best two biomarkers from the original analysis since this is a number that is easily accommodated by qRT-PCR (or similar) molecular methods. This approach has been successfully applied previously with acceptable performance and accuracy [40,41,123]. The XGboost algorithm was selected since prior work has shown that tree-based classifiers are faster and more accurate than other machine learning-based methods such as support vector machine, neural networks, and Bayesian approaches [126].…”