---
title: "Key genes in colorectal adenoma-to-cancer progression identified by a ChebNet–Transformer graph mo"
id: "plos-one-12-exploring-the-key-genes-of-colorectal-adenoma-cancer-based-on-graph-transformer"
canonical_url: "https://medichelpline.com/clinical-feed/plos-one-12-exploring-the-key-genes-of-colorectal-adenoma-cancer-based-on-graph-transformer"
content_type: "clinical_feed_article"
specialty: "Oncology"
source_name: "PLOS ONE (Medicine)"
source_url: "https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0359785"
published_at: "2026-10-05T14:00:00.000Z"
evidence_level: "Journal Feed"
license: "CC-BY-NC-4.0 / Informational Use"
---
# Key genes in colorectal adenoma-to-cancer progression identified by a ChebNet–Transformer graph mo
## Provenance & Clinical Metadata
- **Canonical URL:** https://medichelpline.com/clinical-feed/plos-one-12-exploring-the-key-genes-of-colorectal-adenoma-cancer-based-on-graph-transformer
- **Specialty:** [Oncology](https://medichelpline.com/clinical-feed/oncology.md)
- **Primary Source:** PLOS ONE (Medicine)
- **Source URL:** [Original Journal Publication](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0359785)
- **Published At:** 2026-10-05T14:00:00.000Z
- **Evidence Rating:** Journal Feed
## Executive GIST (TL;DR)
- This study applied a graph neural network model named **ChebTs** (ChebNetII + Transformer + attention pooling) to identify key genes associated with the colorectal **adenoma-cancer** transition using microarray datasets from GEO (GSE41657, GSE31905). - The merged dataset contained 155 samples across normal, adenoma, and colorectal cancer groups after probe-level processing and batch correction. - Differential expression analysis was performed for normal vs adenoma, adenoma vs carcinoma, and normal vs carcinoma using limma and GEO2R with a P 0.4) and visualized in Cytoscape; PPI topology was embedded into the model. - Node features combined expression data with Laplacian positional encodings (PE) and random-walk structural encodings (SE) to enrich topology-aware representations prior to model input. - The ChebTs architecture uses ChebNetII (K-order Chebyshev polynomial convolutions) to capture local graph structure, followed by Transformer layers (L = 2, 4 attention heads) to model global, multi-hop interactions, and attention pooling to aggregate node-level importance for sample classification. - The model applied dropout, SiLU activation, residual connections, and a two-layer MLP to produce predicted probabilities; training used AdamW, warmup + cosine annealing, gradient clipping, and 5-fold cross-validation. - Gene-level interpretability was obtained using GNNExplainer and importance scores derived from learned node representations; virtual knockout experiments were used for validation. - Eight genes were prioritized with importance scores and reported associations with overall survival in CRC: **FN1**, **CLDN2**, **IL33**, **MMP1**, **STC2**, **IGFBP7**, **NOX4**, and **SFRP1**. - The authors conclude these molecules may serve as diagnostic biomarkers and participate in extracellular remodeling and other biological processes relevant to adenoma carcinogenesis. - Data sources used were GSE41657, GSE31905, and other GEO datasets listed; no new data were generated by the study.
## Clinical Analysis & Structured Key Points
Exploring the key genes of colorectal “adenoma-cancer” based on graph transformer | PLOS One Browse Subject Areas ? Click through the PLOS taxonomy to find articles in your field. For more information about PLOS Subject Areas, click here . Article Authors Metrics Comments Media Coverage Reader Comments Figures Figures Abstract Colorectal cancer (CRC) is one of the most lethal malignancies worldwide, and the precise identification of biomarkers from colonic adenoma to cancer is of great significance for preventing the development of adenocarcinoma. Given that existing methods inadequately capture the topological network relationships among genes, this study proposes a graph neural network model based on multifeature learning, named ChebTs, to investigate the correlation between key genes involved in the colorectal “adenoma-cancer” transition. The GSE41657 and GSE31905 datasets from the GEO database were stratified into normal, adenoma, and colorectal cancer groups. Feature encoding was introduced to enhance node features, followed by Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses of differentially expressed genes (DEGs). A protein-protein interaction (PPI) network was constructed using Cytoscape software, and the aggregated information was embedded into the model for training to generate a list of key genes with corresponding importance scores. An attention pooling mechanism aggregated node-level representations into graph level representations for sample classification, and a two layer fully connected network, following activation and regularization, produced predicted probabilities. Furthermore, we provided interpretable analyses at the gene level using GNNExplainer. The results were validated through virtual knockout techniques. The eight screened genes, Fibronectin 1 ( FN1 ), Claudin 2 ( CLDN2 ), Interleukin-33 ( IL-33 ), Matrix Metallopeptidase 1 ( MMP1 ), Stanniocalcin 2 ( STC2 ), Insulin-Like Growth Factor-Binding Protein 7 ( IGFBP7 ), NADPH Oxidase 4 ( NOX4 ), and Secreted Frizzled Related Protein 1 ( SFRP1 ), were all found to be associated with overall survival (OS) in CRC. In this study, eight molecules closely related to the development of colorectal adenocarcinoma were screened out, and they may be diagnostic biomarkers of colorectal cancer. These genes affect the prognosis of patients by participating in biological processes such as remodeling of extracellular mechanisms, and are of great significance for preventing the carcinogenesis of adenoma. Citation: Yuan Y, Liu Y, Fan J, Ren Z (2026) Exploring the key genes of colorectal “adenoma-cancer” based on graph transformer. PLoS One 21(10): e0359785. https://doi.org/10.1371/journal.pone.0359785 Editor: Kangkang Ji, Huazhong Agriculture University, CHINA Received: July 21, 2026; Accepted: September 17, 2026; Published: October 5, 2026 Copyright: © 2026 Yuan et al. This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability: No data was generated by this study. The following existing data sources were used: [GSE41657,GSE31905,GSE54236,GSE62452] from [GEO Database] available via [ https://www.ncbi.nlm.nih.gov/geo/ ]. Funding: The following three fund projects are all intended to support the first author YY. 1.Innovation and Entrepreneurship Fund Project of Gansu University of Chinese Medicine(2026CXZX-942). 2.Innovation Fund Project for College Teachers of Gansu Province(2026B-123). 3.Key Talent Project of Health Care in Gansu Province(2026SWJWRC008). The funder provided support in the form of salaries for authors, but did not have any additional role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript. The specific roles of these authors are articulated in the ‘author contributions’ section. Competing interests: No competing interests exist. Introduction Colorectal cancer (CRC) is one of the most prevalent and lethal malignancies worldwide, and its pathogenesis is characterized by a distinct stepwise progression, with the “adenoma-cancer” sequence being recognized as the classical paradigm colorectal tumor development [ 1 ]. The mutational process by which normal mucosa evolves into invasive carcinoma through adenomatous polyps is driven by a series of genomic alterations affecting cancer related pathways. However, this transformation process is also influenced by the gut microbiota. Shi et al. [ 2 ] discovered that diosgenin oligosaccharides from Dendrobium officinale could alleviate chronic colitis by regulating the inflammatory pathways and gut microbiota. Establishing gene regulatory networks through graph structures can better understand the above complex biological theories. Therefore, the precise identification of key genes involved in this process is essential not only for elucidating the molecular mechanisms underlying colorectal carcinogenesis but also for providing a theoretical basis for the discovery of early diagnostic biomarkers and the implementation of personalized therapeutic strategies. In recent years, the identification of key genes has relied primarily on two major approaches. The first approach is based on differential expression analysis and mutation spectrum profiling using high throughput sequencing data. By comparing omics data from normal mucosa, adenomas, and carcinomas, this strategy enables the screening of genes that exhibit sustained differential expression or mutation enrichment throughout the sequential transformation process. The second approach relies on functional validation in animal models, where gene editing techniques are employed to simulate specific mutation combinations and to observe their effects on malignant transformation of cells [ 3 ]. However, existing methods have certain limitations in identifying core genes. Traditional differential expression analysis fails to account for the complex network structures formed by gene-gene interactions, co-expression, or pathway associations. Tumor development, in essence, reflects an imbalance in network regulatory systems. Furthermore, high throughput omics data are characterized by high dimensionality, small sample sizes, high noise levels, and sparsity. Machine learning based screening methods applied to such data are prone to issues such as the “curse of dimensionality”, leading to model overfitting and poor generalization ability [ 4 ]. Graph Neural Networks (GNNs) have demonstrated considerable efficacy in processing graph structured data with non-Euclidean geometries. Their core principle involves learning low dimensional vector representations of target nodes through iterative aggregation of information from neighboring nodes. Through a multi-layer message passing mechanism, they capture topological dependencies among genes and incorporate network based prior knowledge into the learning process, thereby achieving information enhancement and noise reduction in high dimensional, sparse expression data [ 5 ]. EMOGI [ 6 ] is an interpretable model based on graph convolutional neural networks (GCNs) that classifies cancer driving genes by integrating features such as DNA methylation data, gene expression data, and PPI networks. However, its use of a single variant of GNN prevents it from capturing global information. MTGCN [ 7 ] introduces a ChebNet based graph convolutional network that identifies cancer driver genes through joint optimization of node classification and link prediction tasks, yet its lack of structural encoding limits its capacity to capture gene-gene dependencies. DGMP [ 8 ] combines directed graphs and MLPs to learn node features from gene regulatory networks and omics data for key gene prediction. However, its use of global average pooling results in the loss of structural information. To address these limitations, this study proposes a new GNN model, ChebNetII, based on Chebyshev interpolation and combined with a Transformer to identify key genes in colorectal “adenoma-cancer” progression. By embedding various types of encodings to enrich node information and integrating ChebNetII with the Transformer, this approach offers significant advantages in capturing long range dependencies among genes and perceiving the topological structure of the graph, thereby providing a new perspective for investigating biomarkers that drive the adenoma to cancer transformation process. Materials and methods Data collection The microarray datasets GSE41657 and GSE31905 used in the present study were sourced from the GEO database [ 9 ]. The original CEL files were converted into expression matrices at the probe level using the GEOquery package. Both datasets were generated using the GPL6480 platform (Agilent-014850 Whole Human Genome Microarray 4x44K G4112F), and a total of 155 samples were included after dataset merging. Specifically, GSE41657 consists of human colorectal tissue specimens obtained by colonoscopic resection, comprising 51 adenomas, 25 colorectal cancer samples, and 12 normal mucosa samples. GSE31905 was established to identify prognostic gene expression biomarkers for early stage primary colorectal cancer and provides expression profiles for 55 colorectal cancer samples, 7 normal samples, and 5 adenomas. Data processing DEGs analysis. Data preprocessing was performed using the sva and ComBat packages in R Studio (version 4.5.1), encompassing data cleaning, normalization, log transformation, and batch effect correction. When the probes were mapped to gene symbols, no obvious missing values were found. The merged dataset was corrected for batch effects using principal component analysis, to ensure high quality data for subsequent model construction. Differential expression analysis of the datasets was conducted using the limma and ggplot2 packages, in conjunction with the GEO2R tool. Three pairwise comparisons of DEGs were established, normal vs adenoma, adenoma vs carcinoma, and normal vs carcinoma. To ensure robustness of the results, the screening criteria were set at and P < 0.05, which were considered statistically significant. Construction of the PPI network. Differentially expressed genes were extracted using the rbioapr and dplyr packages in R Studio. PPI network data were obtained from the STRING database (version 12.0) [ 10 ], and only interactions with a combined score greater than 0.4 were retained [ 9 ]. The resulting network data were then imported into Cytoscape for visualization. Overview of the model We propose a graph neural network model, named ChebTs, that combines local topological encoding with a global attention mechanism to identify key genes in colorectal adenoma to cancer progression. Among them, ChebTs consists of ChebNetII, Transformer layers and attention pooling layers. Enhanced gene features were obtained by aggregating original expression data (node features and PPI network) with Laplacian positional encodings (PE) and random walk structural encodings (SE), allowing the model to learn from both expression profiles and network topology. The three inputs were independently projected into a unified space via linear projection layers, and dropout was applied to generate initial node representations (dropout = 0.3). To identify key genes, the ChebNetII convolutional layer (K = 5) was used to aggregate local neighborhood information. A normalized adjacency matrix was constructed using PPI edge indices and the maximum eigenvalue of the Laplacian matrix, and the Chebyshev polynomial recursion was subsequently applied. Following SiLU activation and residual connections, the updated node representations were fed into a Transformer layer (L = 2) with 4 head attention per layer, ultimately generating a list of key genes with corresponding importance scores. An attention pooling mechanism based on these gene importance scores was then used to aggregate node level representations into graph level representations for sample classification. Finally, a two layer fully connected network (MLP), following activation and regularization, produced predicted probabilities from the graph level representations. The overall model architecture is illustrated in Fig 1 . Download: PNG larger image TIFF original image Fig 1. Model architecture diagram. In this study, different genetic data were combined into an undirected network graph. After introducing Laplacian feature encoding, an enhanced feature set was formed for ChebTs input, and a list of key genes was obtained. Finally, a multi-layer perceptron classified the graph structure based on the gene importance scores. https://doi.org/10.1371/journal.pone.0359785.g001 Experimental environment and hyperparameter. All experiments in this study were conducted on an NVIDIA GeForce GTX 4080 GPU. The proposed model was implemented using the PyTorch framework, and a 5 fold cross validation strategy was adopted, with the dataset partitioned into training, validation, and test sets at a ratio of 6:2:2. The AdamW optimizer was employed with a learning rate of , hidden dimension of 128, and the model was trained for 200 epochs. To enhance model robustness and generalization performance, a learning rate scheduler incorporating warmup and cosine annealing strategies was applied. Additionally, gradient clipping was implemented to prevent gradient explosion resulting from message passing in the deep graph neural network architecture. Coding of features. Positional encodings are designed to provide information about the position of a given node within the graph. In this study, the normalized Laplacian matrix of the graph was computed to capture hierarchical positional information for each node, we define normalized magnetic Laplacian as formula Eq (1) . The positional encoding matrix is defined as . Structural encodings, in contrast, are intended to provide graph structure embeddings by enabling the transfer of richer information to each node through random walk based structural encodings, thereby enhancing the generalization capacity of the graph neural network. Consequently, when two nodes share similar subgraphs, their respective PE and SE are expected to be closely aligned [ 11 ]. (1) Where I denotes the identity matrix, D represents the degree matrix, and A is the adjacency matrix. ChebNetII model. GNNs have become one of the most powerful tools for a wide range of graph based learning tasks due to their ability to aggregate local information through message passing mechanisms. However, the illegal coefficients learned by ChebNet when approximating analytical filter functions can easily lead to overfitting. At the same time, most methods still rely on semi-supervised learning, and the scarcity of labels limits their applicability in real world systems. Although GCN can expand the receptive field through multiple layers of stacking, each layer of GCN can only aggregate 1-hop neighborhood information, thus unable to learn the discriminative information of high-pass filtering representation. The attention heads in GAT perform attention-driven feature aggregation on the graph structure based on node features and learnable parameters. Therefore, its spectral characteristics cannot be directly analyzed. ChebNetII, a GNN model based on Chebyshev interpolation. Unlike full spectrum convolution, Chebnet avoids the eigenvalue decomposition of the Laplacian operator through recursive polynomial computation. It implements the k-order Chebyshev polynomial expansion of the spectral graph filter, which has faster convergence speed and reduces the Runge phenomenon when approximating the filter function. Which can be defined in Eq (2) . Previous experimental studies have demonstrated that low-order polynomials are effective for function approximation, whereas higher order polynomials are required to capture more complex functional relationships [ 12 ]. (2) In this formula, the parameter K denotes the order of Chebyshev polynomials, represents the Chebyshev nodes of , and denotes the learnable parameters. Transformer layer. In practical scenarios, the number of GNN layers is critical to model performance, as the limited depth of these networks constrains their capacity to capture multi-hop neighbors and global contextual information. This limitation is particularly pronounced when processing graph structured data with large scale and complex topologies, where the absence of multi-hop information often leads to degraded task performance. In contrast to GNNs, Transformer layers are capable of effectively modeling global information and multi-hop interactions through the multi-head attention mechanism, and they exhibit greater robustness to the oversmoothing problem commonly encountered in deep GNNs. The attention scores are computed using the scaled dot product formulation [ 13 ]. The multi-head attention mechanism performs multiple independent attention heads in parallel, and the resulting outputs are concatenated and subsequently transformed through a linear projection to produce the final output. The formulas are as follows Eqs (3) and (4) . (3) (4) The self-attention module projects the input features into three subspaces, namely the query matrix Q , the key matrix K , and the value matrix V , where , , and , with denoting the hidden dimension. It is worth noting that the direct application of Transformers to graph data presents a significant challenge, as graph data contain both node features and structural topological information, with the latter being inherently difficult to encode as token sequences required by the Transformer architecture. To address this issue, existing approaches have primarily adopted two strategies [ 14 ]. The first strategy involves incorporating structural encoding modules directly into the Transformer, whereas the second strategy employs GNNs as auxiliary modules to preprocess graph structural information. The present study adopts the second strategy, utilizing ChebNetII as a graph structure preprocessor. Specifically, the input features are first passed through a ChebNetII layer, which aggregates local neighborhood information via K-order Chebyshev polynomial convolution, thereby encoding the graph structure into the node features. The enhanced features are then fed into the Transformer layer, enabling the Transformer to not only leverage its powerful global modeling capability but also indirectly perceive the graph topology through the input features. Furthermore, to strengthen the model’s ability to perceive graph structure, we introduce Laplacian positional encodings and random walk structural encodings as auxiliary information, which are injected into the network together with the original features. This approach effectively combines the inductive bias of GNNs in graph structure learning with the global interaction capacity of Transformers. Screening key genes To accurately identify key genes in the “adenoma-cancer” transition of colorectal cancer, this study leverages the message passing mechanism of graph neural networks and the global capture capability of attention modules to replace traditional methods that rely on static centrality metrics, thereby dynamically learning the extent to which genes contribute to classification decisions. Following the learning of node representations by ChebTs, each gene is mapped to a value between 0 and 1 to compute its importance score, and the genes are subsequently ranked in descending order to generate a candidate gene list, which is defined as formula Eq (5) . A higher score indicates more significant differential expression of that gene between the adenoma and carcinoma stages, suggesting its potential as a biomarker. (5) Where , W 1 denotes learnable weight, and represents the sigmoid activation function. Functional enrichment analysis To investigate the
## Related Clinical Research

- [Camupepimut (ELI-002 7P) in KRAS-mutant PDAC: Phase 2 AMPLIFY-7P results](https://medichelpline.com/clinical-feed/nature-1-lymph-node-targeted-mkras-specific-amphiphile-vaccine-in-pancreatic-cancer-the.md)
- [Chromosomal instability drives continual flux in cancer genomes: rates, distribution, and ploidy l](https://medichelpline.com/clinical-feed/biorxiv-9-chromosomal-instability-drives-continual-flux-in-cancer-genomes.md)
- [Cardiorespiratory fitness and sex modify colorectal cancer–suppressive effects of post-exercise se](https://medichelpline.com/clinical-feed/biorxiv-7-cardiorespiratory-fitness-and-sex-impact-the-colorectal-cancer-suppressive.md)
- [MSI‑High Colorectal Cancer in an Adolescent: Integrated Surgical, Molecular, and Adjuvant Manageme](https://medichelpline.com/clinical-feed/pubmed-42892709.md) (DOI: 10.12659/AJCR.953248)
- [CXCL14+ fibroblasts drive IBD-associated fibrosis and cancer through collagen IV–SPP1+ macrophage](https://medichelpline.com/clinical-feed/british-journal-of-cancer-0-cxcl14-fibroblasts-promote-development-of-ibd-associated-fibrosis-and-cancer-by.md)

## Navigation
- [← Back to Oncology Feed](https://medichelpline.com/clinical-feed/oncology.md)
- [← All Clinical Specialties](https://medichelpline.com/clinical-feed.md)
## Medical & Regulatory Disclaimer

> [!CAUTION]
> MedicHelpline content is structured for research, educational, and professional discovery purposes. It does not constitute individual medical advice, clinical diagnosis, or treatment recommendations.
> Always verify dosing, contraindications, and regulatory alerts against official product labeling and primary regulatory sources before clinical decision-making.