Efficient differential latent network analysis: applications to colon cancer Authors: Yewon Han, Lee Sael Published: 2026 ABSTRACT BackgroundColon cancer (CC) presents significant molecular heterogeneity, complicating our understanding of its initiation and progression. Identifying CC-associated genes and their interactions is crucial for improving diagnostics and therapeutics. However, current gene analysis methods struggle to holistically capture complex gene interactions due to their inherent complexity.MethodWe propose a novel, simple, and scalable method called Efficient Differential Latent Network Analysis (EDLNA) for detecting alterations in gene interactions using latent co-expression patterns. Our approach applies non-negative matrix factorization to construct separate latent gene networks for normal and cancer samples. We then identify differential interactions, followed by protein-protein interaction network and transcription factor (TF) analyses to detect functional modules and regulatory relationships.ResultsWe evaluated EDLNA against conventional methods using both simulated and colon cancer gene expression data (GSE44076, GSE50760). In simulation studies, EDLNA was significantly faster among differential network analysis techniques while maintaining comparable accuracy. When applied to colon cancer data, our method outperformed differential gene expression analysis in identifying biologically relevant gene clusters and stage-specific traits.ConclusionsEDLNA provides an efficient and scalable framework for identifying stage-specific gene interactions that bridge molecular mechanisms with clinical phenotypes. These findings underscore its potential for discovering novel biomarkers and advancing targeted therapies in colon cancer. Background Colon cancer (CC) ranks as the third most prevalent cancer and the second leading cause of cancer-related mortality worldwide [1]. CC is characterized by significant heterogeneity, encompassing a complex network of genes and molecular pathways. This complexity makes it difficult to comprehensively understand the molecular mechanisms underlying CC initiation and progression. Consequently, advancing methodologies to identify CC-associated genes and unravel their interactions is crucial for deepening our insight into CC pathogenesis, thereby facilitating the development of more effective diagnostic tools and therapeutic strategies. Moreover, as CC prognosis deteriorates with disease progression, early detection is important [2]. The two primary approaches for identifying cancer-associated genes are paired differential gene expression analysis and paired differential network analysis. Paired differential gene expression (DEG) analysis focuses on identifying individual genes exhibiting significant expression level changes between matched normal and cancer tissues. While DEG analysis offers a straightforward approach for biomarker detection by directly identifying genes with significantly altered expression, it is limited in capturing the complex interplay between genes [3]. In contrast, paired differential network analysis (DNA) methods detect alterations in gene interactions across matched normal and cancer tissues. DNA methods [4–7] often uncover coordinated changes that might be overlooked by DEG analysis, which relies on individual gene expression variations. These methods provide insights into functional modules and regulatory relationships. However, DNA methods frequently face scalability issues due to their reliance on complex network analysis [8]. In this study, we propose Efficient Differential Latent Network Analysis (EDLNA), a simple, scalable, but effective approach for analyzing differential networks derived from latent co-expression patterns among genes. Our method proceeds as follows: First, we filter for genes exhibiting substantial alterations between the normal and cancer states. Our EDLNA then employs non-negative matrix factorization (NMF) to identify latent factors in gene expression. These factors are subsequently utilized to construct separate latent association networks for paired normal and cancer gene expression samples. We then select gene pairs that demonstrate significant alterations in their interactions. The selected genes undergo further analysis through protein-protein interaction (PPI) network analysis to detect functional modules. In addition, we perform transcription factor analysis on these genes to elucidate their regulatory relationships. Methods Our proposed EDLNA consists primarily of two steps: the paired non-negative matrix factorization (NMF) step and the differential latent interaction detection step. In the following, we detail the two steps. Paired non-negative matrix factorization The paired NMF step facilitates the identification of latent gene expression patterns unbiased by patient-specific traits by ensuring consistency in the sample-latent factor matrix across paired conditions. Initially, NMF is applied to the normal-tissue gene expression matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X}_{N} \in \mathbb {R}^{m\times n}$$\end{document}, where m denotes the number of genes and n represents the number of samples. The NMF decomposes \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X}_N$$\end{document} into a gene-latent factor matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{W}_{N} \in \mathbb {R}^{m\times r}$$\end{document} and a sample-latent factor matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{H} \in \mathbb {R}^{r\times n}$$\end{document}, such that the residual sum of squares (RSS) is minimized and all entries are non-negative. The variable r represents the rank or size of the latent factors, which is a hyperparameter. Subsequently, NMF is performed on the tumor-tissue gene expression matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X}_{T} \in \mathbb {R}^{m\times n}$$\end{document} while maintaining the sample-latent matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{H}$$\end{document} constant. Only the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{W}_T$$\end{document} matrix is updated, minimizing the RSS between \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X}_T$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{W}_T\textbf{H}$$\end{document}. This approach ensures that the sample-specific features captured in the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{H}$$\end{document} matrix remain consistent across both normal and tumor samples, while allowing tumor-specific gene expression patterns to be captured in the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{W}_T$$\end{document} matrix. The RSS is defined as follows [9]:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} RSS = \left\| \textbf{X} - \textbf{WH} \right\| ^2,\qquad \text {subject to} \quad W \ge 0,\; H \ge 0. \end{aligned}$$\end{document} The iterative update rules for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{H}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{W}_N$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{W}_T$$\end{document} that minimize the RSS given rank r and normal and tumor gene expression matrices \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X}_N$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X}_T$$\end{document} are as follows:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}&\textbf{H} \leftarrow \textbf{H} \odot \frac{(\textbf{W}_N^T \textbf{X}_N)}{(\textbf{W}_N^T\textbf{W}_N\textbf{H})},\nonumber \\&\textbf{W}_N \leftarrow \textbf{W}_N \odot \frac{(\textbf{X}_N \textbf{H}^T)}{(\textbf{W}_N \textbf{H}\textbf{H}^T)},\nonumber \\&\textbf{W}_T \leftarrow \textbf{W}_T \odot \frac{(\textbf{X}_T \textbf{H}^T)}{(\textbf{W}_T \textbf{H}\textbf{H}^T)} \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\odot $$\end{document} denotes element-wise multiplication. To determine the optimal size of latent factors r, the elbow method is employed. This method plots the RSS against r and selects the “elbow” point where increasing r yields diminishing returns in error reduction. Differential latent interaction detection The differential latent interaction detection step identifies variations in gene-gene associations between paired conditions by computing correlation coefficients from the corresponding rows of the gene-latent factor matrices \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{W}_N$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{W}_T$$\end{document}, which are obtained from the paired NMF step. To construct the latent association matrices \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{P}_N$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{P}_T$$\end{document}, Pearson correlation coefficients were calculated using rows from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{W}_N$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{W}_T$$\end{document}. The resulting matrices \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{P}_N$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{P}_T$$\end{document} matrices are then used to derive a difference matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{D}$$\end{document} by calculating the absolute differences between the cell values. Differential latent edges and genes were identified based on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{D}$$\end{document}. To identify differential latent edges in the simulation study, we applied a threshold t corresponding to the top 0.1% of edges, and selected those exceeding this threshold while excluding duplicates. In the empirical study, because the total number of edges was much larger than in the simulation, visualizing all selected edges and examining functional modules was impractical. Therefore, we applied a stricter threshold (0.0001% of edges), counted the occurrences of genes at both ends of the selected edges, and used only the top 40 most frequently occurring genes for visualization and functional module analysis. Experiments and results Dataset Simulation Data: The simulation dataset used in this study was obtained from a previous study by Galindez et al [10]. The dataset was generated using the SimulatorGRN tool [11], which simulates the gene expression of knockdowns. The regulatory network and normal gene expression data of Saccharomyces cerevisiae contained 150 genes. The dataset includes simulations of 500 random knockdowns. Microarray Data and Preprocessing: The raw microarray data (CEL files) of colon cancer GSE44076 was retrieved from the NIH Gene Expression Omnibus (GEO) [12]. This dataset comprises paired normal adjacent mucosa and tumor samples from 98 individuals, along with 50 healthy colon mucosae, all obtained using Affymetrix Human Genome U219 Arrays. Importantly, all tumor samples in this dataset correspond to stage II colorectal cancer. Preprocessing was performed using quantile normalization using the RMA method from the ‘Affy’ package (version 1.80.0) in R, followed by removal of control probes. To map probe IDs to gene symbols, we retained the probe with the highest standard deviation among duplicates as the representative for each gene, discarding the remaining probes. This filtering step ensures the retention of probes that exhibit significantly different expression levels between the three types of sample (healthy, adjacent normal, and tumor colon cells), based on the probe annotation files from the GPL570 platform. This process resulted in a final set of 20,035 genes. Subsequently, the data were segregated into two matrices: one containing 98 normal samples and another with 98 tumor samples. RNA-seq Data and Preprocessing: The RNA-seq data of colorectal cancer GSE50760 was retrieved from the GEO [13]. This dataset was generated from a cohort of 18 colorectal cancer (CRC) patients. A total of 54 samples were collected, comprising normal colon tissue, matched primary CRC, and synchronous liver metastasis. Importantly, all patients in this cohort had stage IV colorectal cancer with synchronous liver metastasis. In this study, we used only the paired samples of normal colon and primary CRC. Sequencing libraries were prepared using the TruSeq RNA Sample Preparation kit v2(Illumina), which included mRNA enrichment, fragmentation, cDNA synthesis, adapter ligation, and PCR amplification. Paired-end sequencing(2 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times $$\end{document} 100 bp) was performed on the Illumina HiSeq-2000 platform. Compared methods Our proposed EDLNA was compared with a differential gene expression (DEG) method and two paired differential network analysis (DNA) methods. We used the code provided by the original authors with the default parameters and implemented our own code for the remaining methods. Differential Gene Expression (DEG): DEG analysis was performed using the limma package (version 3.60.4). Genes were ranked in descending order based on the absolute value of their log fold change. The top-N genes were then compared with those identified by our method. Pearson Correlation Coefficients (PCC): The PCC method serves as a baseline DNA approach. It computes the Pearson correlation coefficients between gene expressions across samples, to generate association matrices for each condition. These matrices are then compared to identify the differential links. This PCC approach can be considered equivalent to our method without the paired NMF step. chNet: The chNet is a DNA method that integrates changes in gene expression and gene interactions using an optimization framework with hierarchical constraints. In this framework, differential edges are allowed only when at least one gene is differentially expressed [14]. BoostDiff: The BoostDiff is an ensemble-based DNA method that uses adaptively boosted differential regression trees with a novel variance-based splitting criterion. It is designed to detect context-specific regulatory network changes between two conditions [10]. Simulation studies The simulation data includes ground truth, with three truth types: Direct, Influence, and Association. We use Association as the gold standard, since it encompasses all relationships, including indirect interactions between genes. To evaluate whether EDLNA can effectively detect hub genes, we defined transcription factor (regulator) genes corresponding to the edges as the label set. The average number of genes in this label set was approximately 43. Based on this, we calculated the precision, recall, and F1-score for the compared methods. Figure 1 presents these results.Fig. 1Precision, recall, and F1-scores of EDLNA, Boostdiff, chNet, DEG, PCC. Comparisons were made between EDLNA and each method based on the F1-score, and p-values are shown only when they are less than 0.05 Precision, recall, and F1-scores of EDLNA, Boostdiff, chNet, DEG, PCC. Comparisons were made between EDLNA and each method based on the F1-score, and p-values are shown only when they are less than 0.05 Precision, recall, and F1-score were defined as follows: Precision was the proportion of correctly identified differential genes among all genes predicted as differential by a given method. Recall was the proportion of correctly identified differential genes among all genes in the ground-truth set. The F1-score, calculated as the harmonic mean of precision and recall, served as a balanced measure of overall performance. The number of edges selected by EDLNA averaged 11, corresponding to an average of 17 unique genes (excluding duplicates). For consistency, the top 11 edges were also selected for BoostDiff, PCC, and chNet. For DEG, the top 17 genes were selected, matching the selection in EDLNA. In terms of F1-score and precision, EDLNA achieved slightly higher values than BoostDiff, whereas BoostDiff showed marginally higher recall. However, these differences were not statistically significant. Among the other three methods, EDLNA had the highest F1-score, with statistical significance. Overall, these results demonstrate that EDLNA can effectively and efficiently identify hub genes in differential gene networks through a simple computational process. Empirical studies in the colon gene expression The colon cancer datasets GSE44076 and GSE50760 comprised 20,035 and 23,492 genes, respectively. These genes were analyzed using EDLNA, PCC, and DEG. The chNet and BoostDiff methods were excluded due to their high computational complexity, which prevented them from scaling effectively to the colon cancer datasets. Gene-set mapping and PPI expansion To evaluate the biological relevance of the genes identified by the three methods, we selected the top-40 genes from EDLNA, PCC, and DEG, then expanded each set using a two-hop search. Since our method is particularly effective at detecting hub genes, hop expansion was expected to be especially beneficial. Hop expansion was performed using experimentally and literature-validated protein–protein interactions (PPIs) provided by the STRING database. Specifically, 1-hop expansion refers to adding proteins that are directly connected to the input protein by a single edge, whereas 2-hop expansion refers to adding proteins that are two interaction steps away from the input protein. Because EDLNA is effective in identifying hub genes within a gene network, hop expansion was applied to capture broader functional modules surrounding these hubs. However, as the expansion distance increases, the direct functional relevance to the original hub gene decreases, and the likelihood of including biologically less relevant interactions increases. To preserve biological validity, the 1-hop expansion was limited to a maximum of 10 additional proteins, and the 2-hop expansion was restricted to a maximum of 5 additional proteins. This expansion strategy enabled the identification of regulatory clusters by incorporating indirectly associated genes, thereby producing richer and more functionally meaningful modules. Figures 2 and 3 shows the results for without expansion and Figs. 4 and 5 show two-hop expansion of DEG, PCC, and our EDLNA. The resulting PPI topologies differed markedly across DEG, PCC, and our EDLNA. With no-hop expansion, almost no connections were observed among the top-40 genes selected by EDLNA or PCC However, after applying the two-hop expansion, EDLNA revealed a single, densely connected cluster, revealing a regulatory module that was not apparent in the original gene set. In contrast, PCC produced relatively weakly connected clusters after expansion. This demonstrates that the NMF process enables EDLNA to detect hub genes more effectively. DEG, on the other hand, produced moderately connected clusters even without expansion, primarily reflecting direct PPIs among highly expressed genes. These results demonstrate that our EDLNA is particularly effective in discovering functional modules when used with network expansion.Fig. 2STRING PPI network of the top 40 genes selected from GSE44076 by the three methods: a DEG, b PCC, and c EDLNA. Edge thickness represents the strength of supporting evidenceFig. 3STRING PPI network of the top 40 genes selected from GSE50760 by the three methods: a DEG, b PCC, and c EDLNA. Edge thickness represents the strength of supporting evidenceFig. 4STRING PPI network of the top 40 genes selected from GSE44076 by three methods — a DEG, b PCC, and c EDLNA — expanded with their 2-hop neighbors. Edge thickness represents the strength of supporting evidenceFig. 5STRING PPI network of the top 40 genes selected from GSE50760 by three methods — a DEG, b PCC, and c EDLNA — expanded with their 2-hop neighbors. Edge thickness represents the strength of supporting evidence STRING PPI network of the top 40 genes selected from GSE44076 by the three methods: a DEG, b PCC, and c EDLNA. Edge thickness represents the strength of supporting evidence STRING PPI network of the top 40 genes selected from GSE50760 by the three methods: a DEG, b PCC, and c EDLNA. Edge thickness represents the strength of supporting evidence STRING PPI network of the top 40 genes selected from GSE44076 by three methods — a DEG, b PCC, and c EDLNA — expanded with their 2-hop neighbors. Edge thickness represents the strength of supporting evidence STRING PPI network of the top 40 genes selected from GSE50760 by three methods — a DEG, b PCC, and c EDLNA — expanded with their 2-hop neighbors. Edge thickness represents the strength of supporting evidence Finding regulatory clusters Under the assumption that upstream genes with significant connections to colon cancer play regulatory roles in the PPI clusters, we examined the gene sets within the largest clusters of each method (Figs.4 and 5). Table 1 lists the genes found in these largest clusters for each method.Table 1List of genes included in the largest clustersClusterGene List of GSE44076DEG (28 genes)KRT23, CHGA, ZG16, TMIGD1, OTOP2, BEST4, GUCA2B, GUCA2A, MS4A12, CLCA1, AQP8, CLCA4, CEACAM7, KRT20, CD177, CA7, SLC9A3, SLC4A4, CFTR, SLC26A2, CA1, CA3, CA4, ABCG2, SLC51B, ANPEP, CA2, GUCY2CPCC (18 genes)KCTD17, SPOP, KLHL41, KLHL21, KLHL3, KLHL2, KBTBD8, KLHL25, KBTBD6, COMMD6, STK24, MTMR9, LASP1, DCTN6, CNKSR3, ARF6, COMMD7-2, NOL4LEDLNA (15 genes)LMO1, LDB2, LMO2, LMO4, SSBP4, SSBP3, ISL2, LMX1B, SSBP2, LDB1, LHX4, IDI2, DDX4, CDK7, NDST2ClusterGene List of GSE50760DEG (43 genes)DAPL1, RPS27, RPS10-NUDT3, RPL37, RPL31, RPS15A, RPL36A, RPL28, RPL24, RPL36AL, RPL39, RPS21, RPL35, RPL29, RPL30, RPL7, RPL7A, RPL37A, RPL32, RPL8, RPL41, RPL36, RPS6, RPS18, RPS2, EEF1A1, TPT1, TMSB10, CCDC124 PELO, NACA2, NACA, LGALS3, LGALS4, RPL36A-HNRNPH2, GAPDH, B2M, KRT8, IFITM3, CKB, FABP1, MT1G, MT2APCC (19 genes)OR2D2, OR2D3, OR2AG1, OR6A2, OR10A5, BRI3 (cluster 1) CSF1R, PLCB3, LPAR6, LPAR4, PLCB2, CAMK2A (cluster2) PATE2, PATE3, PATE1, PINLYP, SMIM34B, PATE4, LY6L (cluster3)EDLNA (15 genes)DKKL1, SPP1, LRP5, DKK1, SFRP5, WNT8A, WNT8B, NOTUM, WNT10A, WNT9A, FZD8, WNT6, WNT2B, DKK2, DKK4 Table 1List of genes included in the largest clustersClusterGene List of GSE44076DEG (28 genes)KRT23, CHGA, ZG16, TMIGD1, OTOP2, BEST4, GUCA2B, GUCA2A, MS4A12, CLCA1, AQP8, CLCA4, CEACAM7, KRT20, CD177, CA7, SLC9A3, SLC4A4, CFTR, SLC26A2, CA1, CA3, CA4, ABCG2, SLC51B, ANPEP, CA2, GUCY2CPCC (18 genes)KCTD17, SPOP, KLHL41, KLHL21, KLHL3, KLHL2, KBTBD8, KLHL25, KBTBD6, COMMD6, STK24, MTMR9, LASP1, DCTN6, CNKSR3, ARF6, COMMD7-2, NOL4LEDLNA (15 genes)LMO1, LDB2, LMO2, LMO4, SSBP4, SSBP3, ISL2, LMX1B, SSBP2, LDB1, LHX4, IDI2, DDX4, CDK7, NDST2ClusterGene List of GSE50760DEG (43 genes)DAPL1, RPS27, RPS10-NUDT3, RPL37, RPL31, RPS15A, RPL36A, RPL28, RPL24, RPL36AL, RPL39, RPS21, RPL35, RPL29, RPL30, RPL7, RPL7A, RPL37A, RPL32, RPL8, RPL41, RPL36, RPS6, RPS18, RPS2, EEF1A1, TPT1, TMSB10, CCDC124 PELO, NACA2, NACA, LGALS3, LGALS4, RPL36A-HNRNPH2, GAPDH, B2M, KRT8, IFITM3, CKB, FABP1, MT1G, MT2APCC (19 genes)OR2D2, OR2D3, OR2AG1, OR6A2, OR10A5, BRI3 (cluster 1) CSF1R, PLCB3, LPAR6, LPAR4, PLCB2, CAMK2A (cluster2) PATE2, PATE3, PATE1, PINLYP, SMIM34B, PATE4, LY6L (cluster3)EDLNA (15 genes)DKKL1, SPP1, LRP5, DKK1, SFRP5, WNT8A, WNT8B, NOTUM, WNT10A, WNT9A, FZD8, WNT6, WNT2B, DKK2, DKK4 List of genes included in the largest clusters DAPL1, RPS27, RPS10-NUDT3, RPL37, RPL31, RPS15A, RPL36A, RPL28, RPL24, RPL36AL, RPL39, RPS21, RPL35, RPL29, RPL30, RPL7, RPL7A, RPL37A, RPL32, RPL8, RPL41, RPL36, RPS6, RPS18, RPS2, EEF1A1, TPT1, TMSB10, CCDC124 PELO, NACA2, NACA, LGALS3, LGALS4, RPL36A-HNRNPH2, GAPDH, B2M, KRT8, IFITM3, CKB, FABP1, MT1G, MT2A OR2D2, OR2D3, OR2AG1, OR6A2, OR10A5, BRI3 (cluster 1) CSF1R, PLCB3, LPAR6, LPAR4, PLCB2, CAMK2A (cluster2) PATE2, PATE3, PATE1, PINLYP, SMIM34B, PATE4, LY6L (cluster3) In the case of GSE50760, however, a single dominant cluster could not be observed for PCC. Instead, three smaller clusters were extracted separately and analyzed as representative modules. Using EDLNA, we identified a functional module consisting of 15 genes that was not detected by other methods. Gene Ontology (GO) analysis of this module was performed using the STRING database. The strength score, defined as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$log_{10}(observed/expected)$$\end{document}, quantifies the degree of over-representation for a functional term by comparing the observed number of annotated proteins in the network with the expected count in a random network of equal size. Strength values greater than 2 were considered significant, indicating a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\ge $$\end{document}100-fold enrichment, and thus representing a highly overrepresented functional term. The results are summarized in Table 2.Table 2Gene ontology analysis of functional moduleGO-term (GSE44076)Gene ontologyStrengthGO:0048382 (Biological process)Mesendoderm development2.42GO:0021522 (Biological process)Spinal cord motor neuron differentiation2.12GO:0030274 (Molecular Function)LIM domain binding2.57GO-term (GSE50760)Gene ontologyStrengthGO:0061303 (Biological process)Cornea development in camera-type eye2.55GO:0061072 (Biological process)Iris morphogenesis2.47GO:0060070 (Biological process)Canonical Wnt signaling pathway2.12GO:0039706 (Molecular Function)Co-receptor binding2.61GO:0005109 (Molecular Function)Frizzled binding2.32GO:0048019 (Molecular Function)Receptor antagonist activity2.24GO:0042813 (Molecular Function)Wnt receptor activity2.24GO:0017147 (Molecular Function)Wnt-protein binding2.12GO:1990851 (Cellular Component)Wnt-Frizzled-LRP5/6 complex2.82 Table 2Gene ontology analysis of functional moduleGO-term (GSE44076)Gene ontologyStrengthGO:0048382 (Biological process)Mesendoderm development2.42GO:0021522 (Biological process)Spinal cord motor neuron differentiation2.12GO:0030274 (Molecular Function)LIM domain binding2.57GO-term (GSE50760)Gene ontologyStrengthGO:0061303 (Biological process)Cornea development in camera-type eye2.55GO:0061072 (Biological process)Iris morphogenesis2.47GO:0060070 (Biological process)Canonical Wnt signaling pathway2.12GO:0039706 (Molecular Function)Co-receptor binding2.61GO:0005109 (Molecular Function)Frizzled binding2.32GO:0048019 (Molecular Function)Receptor antagonist activity2.24GO:0042813 (Molecular Function)Wnt receptor activity2.24GO:0017147 (Molecular Function)Wnt-protein binding2.12GO:1990851 (Cellular Component)Wnt-Frizzled-LRP5/6 complex2.82 Gene ontology analysis of functional module In the GSE44076 dataset, the functional module was enriched for developmental processes such as mesendoderm development and spinal cord motor neuron differentiation, as well as LIM domain binding activity, suggesting reactivation of developmental programs in colon cancer progression. Consistent with these findings, SSBP2 and SSBP3 are known to stabilize the LDB1–LMO4 transcriptional complex by preventing LDB1 degradation. This axis has been implicated in increased proliferation and invasiveness in CRC, and in metastasis/angiogenesis in other tumor contexts, supporting its potential oncogenic relevance in early-stage disease [15]. Importantly, this axis was not revealed by differential expression analysis, as the individual genes did not display marked expression differences. Instead, it emerged through significant changes in gene–gene correlations between tumor and normal tissues, which were successfully detected by EDLNA. In the GSE50760 dataset, the functional module was enriched for multiple processes related to Wnt signaling, including the canonical Wnt pathway, Wnt–Frizzled–LRP5/6 complex formation, and receptor–ligand interactions, indicating strong engagement of Wnt receptor–ligand interactions and downstream signaling in advanced disease. Although the two cohorts revealed distinct modules, both converged on a common association between LDB1 and Wnt signaling. Taken together, analyses of GSE44076 (stage II) and GSE50760 (stage IV) consistently capture stage-appropriate biology—developmental reprogramming in stage II and heightened Wnt signaling in stage IV—linked through LDB1. Independent studies show that LDB1 overexpression associates with poor survival and correlates positively with Wnt targets (CTNNB1, AXIN2, MYC, CCND1), and enhances proliferation/invasiveness in CRC [16]. Notably, the interaction between LDB1 and SSBP2, first highlighted in the GSE44076 dataset, provides a direct structural and functional connection to Wnt signaling. Because the LDB1–SSBP2 complex constitutes a critical element of the Wnt enhanceosome and mediates Wnt/\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta $$\end{document}-catenin–dependent transcription, its disruption represents a promising therapeutic strategy for suppressing aberrant Wnt signaling in colorectal cancer [17]. Overall, these results highlight the biological importance of the LDB1–SSBP2 interaction in connecting developmental reprogramming with Wnt activation, and showcase EDLNA’s ability to prioritize differential correlation–defined rewiring. Identifying significant TFs We then identified transcription factors (TFs) known to regulate the expression of genes in each cluster. For this analysis, we employed the Enrichr [18] web tool along with the ChEA 2022 [19] dataset. Since the number of enriched TFs was relatively large, we applied a more stringent criterion by selecting only those with an adjusted p-value below 0.05 (Table 3). This investigation helps confirm whether the selected gene clusters not only demonstrate protein-protein interactions but also share common upstream regulatory mechanisms, thereby providing stronger evidence for their biological relevance.Table 3Transcription factors listMethodTranscription factors of GSE44076 (adjusted p-value \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{<}$$\end{document} 0.05)EDLNA (5/17)STAT3, KDM2B, JARID2, TEAD4, RNF2, SUZ12, KLF4, EGR1, UTX, SETDB1, TCF7, RING1B, CTCF, CBX2, REST, PHC1, EZH2DEGFOXM1PCC-MethodTranscription factors of GSE50760 (adjusted p-value < 0.05)EDLNA (5/11)EZH2, SUZ12, MTF2, BMI1, JARID2, TRIM28, TP53, SMC3, ZFP281, TCF3, EEDDEG (11/39)MYC, XRN2, TTF2, E2F1, GABP, TAL1, ZFX, NELFA, ETS1, VDR, YY1, OCT4, ASH2L, SPI1, POU5F1, CREB1, ZFP42, FLI1, SOX9, HOXC9, MYBL2, CREM, SOX2, STAT1, FOXO3, FOXA1, ELK1, KLF4, ASXL1, HNF1A, EKLF, TBP, MYCN, CNOT3, FUS, TRIM28, AR, CEBPB, FOXP3PCC-TFs are listed in the order of lowest adjusted p-values. Known TFs to colon cancer is underlined Table 3Transcription factors listMethodTranscription factors of GSE44076 (adjusted p-value \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{<}$$\end{document} 0.05)EDLNA (5/17)STAT3, KDM2B, JARID2, TEAD4, RNF2, SUZ12, KLF4, EGR1, UTX, SETDB1, TCF7, RING1B, CTCF, CBX2, REST, PHC1, EZH2DEGFOXM1PCC-MethodTranscription factors of GSE50760 (adjusted p-value < 0.05)EDLNA (5/11)EZH2, SUZ12, MTF2, BMI1, JARID2, TRIM28, TP53, SMC3, ZFP281, TCF3, EEDDEG (11/39)MYC, XRN2, TTF2, E2F1, GABP, TAL1, ZFX, NELFA, ETS1, VDR, YY1, OCT4, ASH2L, SPI1, POU5F1, CREB1, ZFP42, FLI1, SOX9, HOXC9, MYBL2, CREM, SOX2, STAT1, FOXO3, FOXA1, ELK1, KLF4, ASXL1, HNF1A, EKLF, TBP, MYCN, CNOT3, FUS, TRIM28, AR, CEBPB, FOXP3PCC-TFs are listed in the order of lowest adjusted p-values. Known TFs to colon cancer is underlined Transcription factors list TFs are listed in the order of lowest adjusted p-values. Known TFs to colon cancer is underlined The EDLNA identified 17 transcription factors (TFs) in the GSE44076 dataset, DEG detected one TF and PCC failed to identify any TFs. In the GSE50760 dataset, the EDLNA identified 11 TFs, DEG detected 39 TFs, and PCC did not identify any significant TFs. In GSE44076, 5 of the 17 TFs identified by our method and in GSE50760, 5 of the 11 TFs were included in the MalaCards [20] list of genes confirmed to be associated with colorectal cancer. Compared with DEG and PCC, our method identified a substantially larger number of colon cancer–related TFs, demonstrating its ability to detect biologically relevant regulators. Notably, the MalaCards database catalogs 1,351 genes associated with colorectal cancer, including 142 elite (high-evidence) genes and 1,209 non-elite or text-mined gene associations. Pathological stage distribution of TFs found by EDLNA We further validated the significance of the TFs identified by EDLNA using the cBioPortal platform [21] on the Colon Adenocarcinoma(TCGA, GDC) dataset, which comprises 464 colon adenocarcinoma samples. We compared the pathological stage distribution of colon cancer between samples harboring mutations in the TFs detected by our method and the remaining cases (Fig. 6). In GSE44076, which is entirely composed of stage II tumors, TF-mutated samples (94 of 464 samples) showed a higher proportion of stage II compared with the rest of the dataset. Similarly, in GSE50760, which is entirely composed of stage IV tumors, TF mutations (271 of 464 samples) were enriched in stage IV. Consistent with the inclusion of liver metastatic samples, these cases also exhibited a lower proportion of N0 in the N-stage distribution. Overall, the TFs identified by our method not only capture biologically relevant regulators but also accurately reflect the clinical features of the respective datasets.Fig. 6Stage distribution of TF‑mutated samples in (a) GSE44076 and (b)(c) GSE50760 Stage distribution of TF‑mutated samples in (a) GSE44076 and (b)(c) GSE50760 Discussion and conclusion Our proposed method, Efficient Differential Latent Network Analysis (EDLNA), demonstrates a significant advancement in detecting alterations in gene interactions related to colon cancer (CC). By utilizing latent co-expression patterns, we successfully identified functional gene clusters and their regulatory transcription factors (TFs) more effectively than conventional methods such as differential gene expression (DEG) and paired differential network analysis (DNA) with improvements in performance, speed, and scalability. Methodologically, denoising via low-rank factorization followed by differential correlation testing reduces noise and composition-driven artifacts, thereby prioritizing coherent axes rather than broad housekeeping signals. In addition, our analysis revealed that significant differences in cancer stage distribution were observed between patients with mutations in these TFs and those without them. Specifically, the high proportion of samples with mutations in the identified TFs suggests that our method successfully captures stage-specific characteristics. This provides insights into the molecular underpinnings of stage-2 colon cancer, which is essential for improving early detection and targeted therapies. Likewise, when focusing on stage IV (GSE50760), samples harboring mutations in the identified TFs were disproportionately enriched in advanced metastatic cases, underscoring that our method not only detects biologically meaningful regulators but also faithfully reflects clinically relevant features across different stages of CC progression. Taken together, these findings highlight the ability of our approach to bridge molecular mechanisms with clinical phenotypes, thereby providing a foundation for the identification of novel biomarkers and therapeutic targets. Moreover, by detecting functional modules and their associated TFs, our method provides a more comprehensive understanding of gene interactions in CC by detecting functional modules and associated TFs. The identification of both known and novel TFs offers potential avenues for further investigation into CC pathogenesis and treatment strategies. The improved performance and reduced computational complexity of EDLNA compared with existing techniques make it a promising tool for advancing colon cancer diagnostics and therapeutics.