Comparison of Methods for Feature Selection in Clustering of High-Dimensional RNA-Sequencing Data to Identify Cancer Subtypes

Tumor Subtype Classification Tool for HPV-associated Head and Neck Cancers

Importance

Molecular subtypes of HPV-associated Head and Neck Squamous Cell Carcinoma (HNSCC), named IMU (immune strong) and KRT (highly keratinized), are well-recognized and have been shown to have distinct mechanisms of carcinogenesis, clinical outcomes, and potentially differing optimal treatment strategies. Currently, no standardized method exists to subtype a new HPV+ HNSCC tumor. Our paper introduces a machine learning-based classifier and webtool to reliably subtype HPV+ HNSCC tumors using the IMU/KRT paradigm and highlights the importance of subtype in HPV+ HNSCC.

Objective

To develop a robust, accurate machine learning-based classification tool that standardizes the process of subtyping HPV+ HNSCC, and to investigate the clinical, demographic, and molecular features associated with subtype in a meta-analysis of four patient cohorts.

Data Sources

We conducted RNA-seq on 67 HNSCC FFPE blocks from University of Michigan hospital. Combining this with three publicly available datasets, we utilized a total of 229 HPV+ HNSCC RNA-seq samples. All participants were HPV+ according to RNA expression. An ensemble machine learning approach with five algorithms and three different input training gene sets were developed, with final subtype determined by majority vote. Several additional steps were taken to ensure rigor and reproducibility throughout.

Study Selection

The classifier was trained and tested using 84 subtype-labeled HPV+ RNA-seq samples from two cohorts: University of Michigan (UM; n=18) and TCGA-HNC (n=66). The classifier robustness was validated with two independent cohorts: 83 samples from the HPV Virome Consortium and 62 additional samples from UM. We revealed 24 of 39 tested clinicodemographic and molecular variables significantly associated with subtype.

Results

The classifier achieved 100% accuracy in the test set. Validation on two additional cohorts demonstrated successful separation by known features of the subtypes. Investigating the relationship between subtype and 39 molecular and clinicodemographic variables revealed IMU is associated with epithelial-mesenchymal transition (p=2.25×10-4), various immune cell types, and lower radiation resistance (p=0.0050), while KRT is more highly keratinized (p=2.53×10-8), and more likely female than IMU (p=0.0082).

Conclusions and Relevance

This study provides a reliable classifier for subtyping HPV+ HNSCC tumors as either IMU or KRT based on bulk RNA-seq data, and additionally, improves our understanding of the HPV+ HNSCC subtypes.

Classification of Long Non-Coding RNAs s Between Early and Late Stage of Liver Cancers From Non-coding RNA Profiles Using Machine-Learning Approach

Long non-coding RNAs (lncRNAs), which are RNA sequences greater than 200 nucleotides in length, play a crucial role in regulating gene expression and biological processes associated with cancer development and progression. Liver cancer is a major cause of cancer-related mortality worldwide, notably in Thailand. Although machine learning has been extensively used in analyzing RNA-sequencing data for advanced knowledge, the identification of potential lncRNA biomarkers for cancer, particularly focusing on lncRNAs as molecular biomarkers in liver cancer, remains comparatively limited. In this study, our objective was to identify candidate lncRNAs in liver cancer. We employed an expression data set of lncRNAs from patients with liver cancer, which comprised 40 699 lncRNAs sourced from The CancerLivER database. Various feature selection methods and machine-learning approaches were used to identify these candidate lncRNAs. The results showed that the random forest algorithm could predict lncRNAs using features extracted from the database, which achieved an area under the curve (AUC) of 0.840 for classifying lncRNAs between early (stage 1) and late stages (stages 2, 3, and 4) of liver cancer. Five of 23 significant lncRNAs (WAC-AS1, MAPKAPK5-AS1, ARRDC1-AS1, AC133528.2, and RP11-1094M14.11) were differentially expressed between early and late stage of liver cancer. Based on the Gene Expression Profiling Interactive Analysis (GEPIA) database, higher expression of WAC-AS1, MAPKAPK5-AS1, and ARRDC1-AS1 was associated with shorter overall survival. In conclusion, the classification model could predict the early and late stages of liver cancer using the signature expression of lncRNA genes. The identified lncRNAs might be used as early diagnostic and prognostic biomarkers for patients with liver cancer.

Comparison of cancer subtype identification methods combined with feature selection methods in omics data analysis

BACKGROUND

Cancer subtype identification is important for the early diagnosis of cancer and the provision of adequate treatment. Prior to identifying the subtype of cancer in a patient, feature selection is also crucial for reducing the dimensionality of the data by detecting genes that contain important information about the cancer subtype. Numerous cancer subtyping methods have been developed, and their performance has been compared. However, combinations of feature selection and subtype identification methods have rarely been considered. This study aimed to identify the best combination of variable selection and subtype identification methods in single omics data analysis.

RESULTS

Combinations of six filter-based methods and six unsupervised subtype identification methods were investigated using The Cancer Genome Atlas (TCGA) datasets for four cancers. The number of features selected varied, and several evaluation metrics were used. Although no single combination was found to have a distinctively good performance, Consensus Clustering (CC) and Neighborhood-Based Multi-omics Clustering (NEMO) used with variance-based feature selection had a tendency to show lower p-values, and nonnegative matrix factorization (NMF) stably showed good performance in many cases unless the Dip test was used for feature selection. In terms of accuracy, the combination of NMF and similarity network fusion (SNF) with Monte Carlo Feature Selection (MCFS) and Minimum-Redundancy Maximum Relevance (mRMR) showed good overall performance. NMF always showed among the worst performances without feature selection in all datasets, but performed much better when used with various feature selection methods. iClusterBayes (ICB) had decent performance when used without feature selection.

CONCLUSIONS

Rather than a single method clearly emerging as optimal, the best methodology was different depending on the data used, the number of features selected, and the evaluation method. A guideline for choosing the best combination method under various situations is provided.

A new gene-scoring method for uncovering novel glaucoma-related genes using non-negative matrix factorization based on RNA-seq data

Early diagnosis and treatment of glaucoma are challenging. The discovery of glaucoma biomarkers based on gene expression data could potentially provide new insights for early diagnosis, monitoring, and treatment options of glaucoma. Non-negative Matrix Factorization (NMF) has been widely used in numerous transcriptome data analyses in order to identify subtypes and biomarkers of different diseases; however, its application in glaucoma biomarker discovery has not been previously reported. Our study applied NMF to extract latent representations of RNA-seq data from BXD mouse strains and sorted the genes based on a novel gene scoring method. The enrichment ratio of the glaucoma-reference genes, extracted from multiple relevant resources, was compared using both the classical differentially expressed gene (DEG) analysis and NMF methods. The complete pipeline was validated using an independent RNA-seq dataset. Findings showed our NMF method significantly improved the enrichment detection of glaucoma genes. The application of NMF with the scoring method showed great promise in the identification of marker genes for glaucoma.

Majority voting based on different feature ranking techniques from gene expression

In bioinformatics studies, many modeling tasks are characterized by high dimensionality, leading to the widespread use of feature selection techniques to reduce dimensionality. There are a multitude of feature selection techniques that have been proposed in the literature, each relying on a single measurement method to select candidate features. This has an impact on the classification performance. To address this issue, we propose a majority voting method that uses five different feature ranking techniques: entropy score, Pearson’s correlation coefficient, Spearman correlation coefficient, Kendall correlation coefficient, and t-test. By using a majority voting approach, only the features that appear in all five ranking methods are selected. This selection process has three key advantages over traditional techniques. Firstly, it is independent of any particular feature ranking method. Secondly, the feature space dimension is significantly reduced compared to other ranking methods. Finally, the performance is improved as the most discriminatory and informative features are selected via the majority voting process. The performance of the proposed method was evaluated using an SVM, and the results were assessed using accuracy, sensitivity, specificity, and AUC on various biomedical datasets. The results demonstrate the superior effectiveness of the proposed method compared to state-of-the-art methods in the literature.

Entropy sorting of single-cell RNA sequencing data reveals the inner cell mass in the human pre-implantation embryo

A major challenge in single-cell gene expression analysis is to discern meaningful cellular heterogeneity from technical or biological noise. To address this challenge, we present entropy sorting (ES), a mathematical framework that distinguishes genes indicative of cell identity. ES achieves this in an unsupervised manner by quantifying if observed correlations between features are more likely to have occurred due to random chance versus a dependent relationship, without the need for any user-defined significance threshold. On synthetic data, we demonstrate the removal of noisy signals to reveal a higher resolution of gene expression patterns than commonly used feature selection methods. We then apply ES to human pre-implantation embryo single-cell RNA sequencing (scRNA-seq) data. Previous studies failed to unambiguously identify early inner cell mass (ICM), suggesting that the human embryo may diverge from the mouse paradigm. In contrast, ES resolves the ICM and reveals sequential lineage bifurcations as in the classical model. ES thus provides a powerful approach for maximizing information extraction from high-dimensional datasets such as scRNA-seq data.

Investigation of the Role of PUFA Metabolism in Breast Cancer Using a Rank-Based Random Forest Algorithm

Simple Summary

Polyunsaturated fatty acids (PUFAs) and their derivatives, oxylipins, are a constant focus of cancer research due to the relationship between cancer and processes of energy metabolism and inflammation, where a PUFA system is an active player. Only recently have methods been developed that allow for studying such complex systems. Using the Rank-based Random Forest (RF) model, we show that PUFA metabolism genes are critical for the pathogenesis of breast cancer (BC); BC subtypes differ in PUFA metabolism gene expression. The enrichment of BC subtypes with various genes associated with oxylipin signaling pathways indicates a different contribution of these compounds to the biology of subtypes.

Abstract

Polyunsaturated fatty acid (PUFA) metabolism is currently a focus in cancer research due to PUFAs functioning as structural components of the membrane matrix, as fuel sources for energy production, and as sources of secondary messengers, so called oxylipins, important players of inflammatory processes. Although breast cancer (BC) is the leading cause of cancer death among women worldwide, no systematic study of PUFA metabolism as a system of interrelated processes in this disease has been carried out. Here, we implemented a Boruta-based feature selection algorithm to determine the list of most important PUFA metabolism genes altered in breast cancer tissues compared with in normal tissues. A rank-based Random Forest (RF) model was built on the selected gene list (33 genes) and applied to predict the cancer phenotype to ascertain the PUFA genes involved in cancerogenesis. It showed high-performance of dichotomic classification (balanced accuracy of 0.94, ROC AUC 0.99) We also retrieved a list of the important PUFA genes (46 genes) that differed between molecular subtypes at the level of breast cancer molecular subtypes. The balanced accuracy of the classification model built on the specified genes was 0.82, while the ROC AUC for the sensitivity analysis was 0.85. Specific patterns of PUFA metabolic changes were obtained for each molecular subtype of breast cancer. These results show evidence that (1) PUFA metabolism genes are critical for the pathogenesis of breast cancer; (2) BC subtypes differ in PUFA metabolism genes expression; and (3) the lists of genes selected in the models are enriched with genes involved in the metabolism of signaling lipids.

Novel feature selection methods for construction of accurate epigenetic clocks

Epigenetic clocks allow us to accurately predict the age and future health of individuals based on the methylation status of specific CpG sites in the genome and are a powerful tool to measure the effectiveness of longevity interventions. There is a growing need for methods to efficiently construct epigenetic clocks. The most common approach is to create clocks using elastic net regression modelling of all measured CpG sites, without first identifying specific features or CpGs of interest. The addition of feature selection approaches provides the opportunity to optimise the identification of predictive CpG sites. Here, we apply novel feature selection methods and combinatorial approaches including newly adapted neural networks, genetic algorithms, and ‘chained’ combinations. Human whole blood methylation data of ~470,000 CpGs was used to develop clocks that predict age with R2 correlation scores of greater than 0.73, the most predictive of which uses 35 CpG sites for a R2 correlation score of 0.87. The five most frequent sites across all clocks were modelled to build a clock with a R2 correlation score of 0.83. These two clocks are validated on two external datasets where they maintain excellent predictive accuracy. When compared with three published epigenetic clocks (Hannum, Horvath, Weidner) also applied to these validation datasets, our clocks outperformed all three models. We identified gene regulatory regions associated with selected CpGs as possible targets for future aging studies. Thus, our feature selection algorithms build accurate, generalizable clocks with a low number of CpG sites, providing important tools for the field.

Epigenetic clocks accurately predict a person’s age by measuring the levels of a chemical mark (methylation) at specific sites of the DNA. More of these clocks are being built all the time, and there is a need for tools to best construct these clocks, and particularly to pick the specific DNA sites to include. We propose several novel machine-learning tools for the optimised selection of these DNA sites, known as feature selection approaches. We applied our approaches to a large human blood dataset to develop several clocks that accurately predict age using 35 or less DNA sites with more accuracy than previously published clocks when applied to other datasets for validation. Some of the DNA sites identified may be associated with interesting genes to explore further for their role in aging. These approaches should enable the building of more accurate, generalizable age prediction clocks from a low number of DNA sites.