Abstract Despite of recent advances in machine learning methods with huge training data, classification of high-dimensional, low-sample-size (HDLSS) data remains challenging because limited sample sizes increase the risk of overfitting and hinder reliable parameter estimation. The nearest shrunken centroid classifier, also known as Prediction Analysis of Microarrays (PAM), combines centroid-based discrimination with feature selection through thresholding. However, its homoscedasticity assumption may limit classification performance when feature variances differ across classes, as commonly observed in gene expression data. We propose a heteroscedastic extension of the nearest shrunken centroid framework that incorporates class-specific variances into both the test statistic and discriminant function. The proposed method retains the simplicity and interpretability of the original classifier while accommodating variance heterogeneity. By integrating the framework with soft, hard, and order thresholding strategies, we develop three new classifiers. We evaluate their performance through comprehensive simulation studies and applications to multiple multiclass cancer gene expression datasets. The results show that accounting for heteroscedasticity substantially improves classification accuracy, particularly with hard and order thresholding, while maintaining competitive feature sparsity. These findings highlight the importance of modeling variance heterogeneity in high-dimensional classification and provide practical tools for improved cancer subtype prediction.
Abstract Discovering differentially expressed genes (DEGs) is an essential process for diagnosis and treatment. In this study, we evaluate the capacity of machine learning algorithms to identify genes that are simulated as truly differentially expressed. We first obtain nine real datasets from The Cancer Genome Atlas Program. We conduct Monte Carlo simulation by exploring the characteristics of the real datasets (i.e. the number of genes). Since these datasets have highly imbalanced class distributions, datasets with the same characteristics are also generated with balanced class ratios. In this way, 18 scenarios are obtained with a total number of 19,947 genes of which 10 genes are simulated significantly different between the two groups. After the pre-processing steps including filtering, normalization, transformation and univariate analysis, gene selection is conducted with five machine learning algorithms. These are biosigner algorithm, GMDH-type neural network algorithm, Determan’s optimal gene selection algorithm with random forest, support vector machines and elastic net generalized linear models. Five-fold cross validation is used for improving gene selection performances. Recall, precision and F-measure are taken into account for the evaluation of gene selection performances. All algorithms show higher average performance in balanced scenarios than in unbalanced scenarios. Elastic net generally achieves favorable average performance in both balanced and unbalanced scenarios.
Abstract The association between sarcopenia and renal function has been reported in observational studies; however, the directionality and potential causal nature of these associations remain uncertain. We assessed whether genetically predicted sarcopenia-related traits are associated with renal function and vice versa using bidirectional Mendelian randomization (MR). We conducted a bidirectional two-sample MR analysis using publicly available European-ancestry GWAS summary statistics for appendicular lean mass (ALM), hand-grip strength (left and right), and walking pace, and for renal function (cystatin C-based estimated glomerular filtration rate [eGFRcystatin C] and urinary albumin excretion [UAE]). Causal estimates were primarily obtained using inverse-variance weighted (IVW) models, complemented by sensitivity analyses (MR-Egger intercept, weighted median/mode, MR-PRESSO, Radial MR, and leave-one-out). In forward MR, genetically predicted walking pace was positively associated with eGFRcystatin C. Genetically predicted ALM and grip strength (right and left) were inversely associated with UAE. In reverse MR, genetically predicted UAE was inversely associated with ALM and right-hand grip strength. Estimates were broadly consistent across sensitivity analyses, and outlier-robust analyses (MR-PRESSO/Radial MR) yielded similar results. These findings provide genetic evidence consistent with bidirectional relationships between sarcopenia-related traits and renal function (particularly UAE), under standard MR assumptions. Given potential limitations (e.g., heterogeneity, pleiotropy, and possible sample overlap), the results should be interpreted cautiously and complemented by other lines of evidence.
Polygenic risk scores (PRS) are increasingly recognized as pivotal tools for quantifying disease risk through the aggregation of multiple genetic variants. As sample sizes in genome-wide association studies (GWAS) continue to expand and PRS become more powerful, they are set to play a key role in translational research and personalized medicine. Understanding the interplay of PRS with environmental factors is critical for interpreting and applying PRS in diverse contexts. This interplay manifests in two forms: PRS-by-environment interaction (PRS × E) and gene-environment correlation (rGE). However, despite the growing application and importance of PRS, there are limited guidelines for performing PRS × E interaction analyses while controlling for rGE, which can lead to inconsistencies across studies and misinterpretation of results. Here we provide a review of different methods for performing PRSxE interaction in various epidemiological study designs, propose recommendations for best-practice, and discuss future challenges.
p53 is a tumor suppressor protein that maintains genome integrity. Single amino acid variations (SAVs), especially hotspot mutations (such as R175 and R248), are closely associated with oncogenic transformation and weaken DNA-binding and transcriptional regulatory capabilities. To identify deleterious SAVs for understanding disease mechanisms and key cellular processes, including apoptosis, DNA repair, and cell cycle regulation. This study used quantitative biochemical and microbial descriptors of p53 peptide sequences, along with a CNN and a 2-layer bidirectional long short-term memory (Bi-LSTM) network architecture with attention mechanisms and ESM-2-based embeddings, to classify sequences as wild-type or mutant. A systematic analysis of the p53 mutation dataset was performed using molecular weight, instability index, hydrophobicity, motif enrichment, and amino acid substitution patterns to identify mutation-centered peptides. Using stratified 5-fold cross-validation, the proposed model was tested, achieving an accuracy of 0.98, an area under the ROC curve (AUROC) of 0.98, and a precision-recall AUC of 0.99. Furthermore, SHAP-based interpretation identified the key amino acid residues and biochemical factors that contribute to the model's predictive performance. The proposed approach provides insights into the structural, sequential, and biochemical effects of variants, an interpretable, robust framework for evaluating the functional consequences of p53 hotspot and other variants, and a computationally efficient tool for prioritizing high-risk p53 variants.
