K-Medoids guided cell reduction and Node2Vec based complementary graph fusion for gene regulatory network inference from single-cell expression data

Abstract

Gene regulatory network (GRN) inference from single-cell gene expression data remains challenging because such datasets are often noisy, sparse, and computationally large. Moreover, different inference algorithms may recover different regulatory edges. To address these challenges, this study proposes a two-stage framework that combines representative cell selection with complementary graph-based fusion of existing GRN inference methods. The framework was also examined using an Alzheimer’s disease single-nucleus RNA-seq dataset as a case study, focusing on excitatory neurons from the entorhinal cortex and superior frontal gyrus. In the first stage, the gene expression matrix, with genes as rows and cells as columns, is reduced using K-medoids-based cell selection. Since K-medoids selects real observed cells rather than centroids, it helps preserve biologically meaningful expression profiles while reducing downstream computational burden. Benchmark GRN inference methods are then applied to the reduced expression matrix to obtain ranked candidate regulatory interactions. In the second stage, pairwise complementary fusion is performed by constructing a union graph from the predicted edges of different methods. For each gene pair, method-specific edge score, rank information, method support, local graph-topological features, and node2vec-based edge similarity are combined into an unsupervised fusion score. The fused rankings are evaluated using Area under the receiver operating characteristic curve(AUROC) and area under the precision-recall curve(AUPR). Among the tested combinations, AGRN and Normi produced the best fused results, suggesting complementary regulatory signals. The fusion achieved average absolute improvements of 0.14 in AUPR and 0.06 in AUROC, corresponding to relative gains of 34.8% and 10.37%, respectively.