Benchmarking Dimensionality Reduction Methods for Livestock Transcriptomic Data: A Comparative Analysis of Visualization, Clustering, and Classification Performance

Dimensionality reduction methods play a critical role in the analysis and visualization of high-dimensional transcriptomic datasets. In this study, the clustering and classification performances of Principal Component Analysis (PCA), Kernel Principal Component Analysis (Kernel PCA), t-distributed Stochastic Neighbor Embedding (t-SNE), and Uniform Manifold Approximation and Projection (UMAP) were compared using two independent livestock microarray datasets (GSE20552 and GSE24560) obtained from the GEO database. The performance of the methods was evaluated using Silhouette score, Davies–Bouldin index (DBI), Calinski–Harabasz index(CHI), accuracy, area under the receiver operating characteristic curve (AUC), and F1-score. Examination of the results revealed that t-SNE exhibited the highest clustering performance on the GSE20552 dataset, whereas the highest classification success was achieved by PCA (Accuracy=92.5%, AUC=0.9750, F1-score=0.9278). For the GSE24560 dataset, t-SNE was the most successful method in both clustering and classification analyses (Accuracy=80.59%, AUC=0.9165, F1-score=0.8057), ranking first in overall performance. Based on average rank values, the methods were ordered as t-SNE, PCA, Kernel PCA, and UMAP. The findings indicate that the choice of dimensionality reduction method significantly influences downstream analytical outcomes in livestock transcriptomic data. While t-SNE generally demonstrated the strongest performance, PCA was found to provide high classification accuracy in certain datasets. These results offer a comparative and application-oriented methodological framework for selecting appropriate dimensionality reduction techniques in livestock transcriptomic studies.