Colorectal cancer (CRC) follows a stepwise progression in which normal mucosa can develop into invasive carcinoma via adenomatous polyps. Precise identification of biomarkers that mark the adenoma-cancer transition is important for early diagnosis and prevention of malignant transformation. Traditional differential expression analyses and mutation profiling often ignore complex gene–gene network structures, while experimental validation in models can be time-consuming. The authors argue that graph-structured models better capture gene regulatory networks and can improve biomarker discovery for colorectal tumorigenesis.
The study used microarray datasets from the GEO database. Two primary datasets—GSE41657 and GSE31905—were processed at the probe level. Both were generated on the Agilent GPL6480 platform. After merging and preprocessing, a total of 155 samples were included, comprising normal mucosa, adenoma, and colorectal cancer specimens. Specifically, GSE41657 contained 51 adenomas, 25 colorectal cancer samples, and 12 normal mucosa samples; GSE31905 provided additional cancer, normal, and adenoma samples as described in the source.
Data preprocessing included cleaning, normalization, log transformation, and batch effect correction using R packages (sva, ComBat) with principal component analysis to verify correction. Differential expression analysis was performed using limma, ggplot2, and GEO2R. Three pairwise DEG comparisons were constructed: normal vs adenoma, adenoma vs carcinoma, and normal vs carcinoma. The screening criteria required P < 0.05 for statistical significance. Differentially expressed genes were extracted using rbioapr and dplyr for downstream network construction.
PPI data were retrieved from the STRING database (version 12.0) and filtered to retain interactions with a combined score > 0.4. The resulting network was imported into Cytoscape for visualization. The PPI topology served as structural prior knowledge that was embedded into later graph model inputs.
To leverage both local graph structure and global dependencies, the authors proposed ChebTs, a graph neural network combining ChebNetII (a Chebyshev polynomial–based convolution) with Transformer layers and attention pooling. ChebTs integrates three input streams—gene expression (node features), Laplacian positional encoding (PE), and random-walk structural encoding (SE)—which are independently linearly projected, regularized with dropout (dropout = 0.3), and then fused as initial node representations.
Positional encodings (PE) were computed from the normalized Laplacian matrix to capture hierarchical positional information for each node in the graph. Structural encodings (SE) were derived from random-walk–based embeddings to convey local subgraph similarities and richer topological context. PE and SE were concatenated with expression-derived node features to provide the model with both attribute and structural signals.
ChebNetII implements K-order Chebyshev polynomial expansions of spectral graph filters to approximate convolution without explicit eigen decomposition. This recursion-based approach reduces approximation artifacts and supports expanded receptive fields via higher-order polynomials (the study used K = 5). ChebNetII aggregates local neighborhood information while applying SiLU activation and residual connections to produce updated node representations.
Following ChebNetII, the model applies Transformer layers (L = 2) with multi-head self-attention (4 heads per layer) to capture long-range, multi-hop interactions and global context that single GNN layers may miss. Because Transformers require sequence-like inputs, the pipeline uses ChebNetII as a graph-structure preprocessor so that the Transformer receives node features already enriched with topological information. This hybrid design aims to combine the inductive bias of GNNs with the global modeling strengths of Transformers.
Experiments ran on an NVIDIA GeForce GTX 4080 GPU and were implemented in PyTorch. A 5-fold cross-validation scheme split data into training, validation, and test sets at a 6:2:2 ratio. The AdamW optimizer was used with a hidden dimension of 128 and training for 200 epochs. A learning rate scheduler with warmup and cosine annealing and gradient clipping were applied to improve stability and generalization. Details such as the exact learning rate value were reported in the source but are not reproduced here.
After learning node representations, ChebTs produced per-gene importance scores by mapping node embeddings to values between 0 and 1 (via learned weights and a sigmoid activation). Genes were ranked by importance to generate candidate lists for the adenoma-to-cancer transition. An attention pooling mechanism aggregated node-level representations into graph-level representations for sample classification, and a two-layer fully connected network produced class probabilities.
The study performed Gene Ontology (GO) and KEGG enrichment analyses on DEGs (methods described earlier). For interpretability, GNNExplainer was used to provide gene-level explanations of model predictions. Virtual knockout experiments (in silico removal of gene nodes) were employed as a validation strategy to assess the impact of candidate genes on model outputs and implied biological relevance.
Eight genes were prioritized as closely related to colorectal adenocarcinoma development and were reported to associate with overall survival in CRC: FN1 (Fibronectin 1), CLDN2 (Claudin 2), IL33 (Interleukin-33), MMP1 (Matrix Metallopeptidase 1), STC2 (Stanniocalcin 2), IGFBP7 (Insulin-Like Growth Factor-Binding Protein 7), NOX4 (NADPH Oxidase 4), and SFRP1 (Secreted Frizzled Related Protein 1). The authors propose these molecules as potential diagnostic biomarkers that may influence prognosis by participating in processes such as extracellular matrix remodeling. The source states these genes were found to be associated with overall survival in CRC; no additional effect sizes or statistical details beyond that association are reported in the provided text.
By integrating topology-aware encodings with Chebyshev-based graph convolution and Transformer-based global modeling, the ChebTs framework aims to overcome limitations of traditional DEG analyses and prior GNN approaches. Applying this model to GEO microarray data yielded eight candidate genes implicated in the colorectal adenoma-to-cancer transition and linked to patient survival. These findings suggest that hybrid graph–transformer architectures can aid biomarker discovery in CRC, though the provided text does not report external experimental validation beyond in silico analyses.