To better learn miRNAs-related feature and get better prediction performance, a combination of MLP (Multilayer Perceptron) and CNN (convolutional network) is utilized, named CCNET. MLP employs fully connected layers in each layer, allowing the neural network to learn global feature representations. On the other hand, CNN utilizes local receptive fields and weight sharing mechanisms through convolutional and pooling layers, enabling automatic extraction of local features. MLP can be described as follows:

where w_j is the weight vector associated with the connections from the j-th neuron in the i-1th layer to the i-th layer, x is the input vector, and b_j is the bias term associated with the j-th neuron. f(·) is the activation function to introduce non-linearity. And CNN is described as follows:

where y[i] represents the i-th element of the output feature map, b is the bias term, x[j] represents the j-th element of the input sequence, w[i-j] represents the (i-j)-th element of the convolutional kernel.

In the paper, we chose miRNA-related genes as the features in colon cancer. The input features are processed in parallel through two separate branches in the neural network. One branch consists of a 3-layer MLP that focuses on learning global features, while the other branch includes a 4-layer CNN that emphasizes local features. Subsequently, a 2-layer Deep Neural Network (DNN) is employed to fuse and make decisions based on the features learned from both branches. The structure of CCNET as shown in Figure 1 and Table 1 .

In survival analysis, LASSO (Least Absolute Shrinkage and Selection Operator) is widely-used in identifying biomarkers, genes, phenotype features and other variables that important on survival time. Its principle is to add an L1 regularization term to limit the size of the regression coefficients, thereby achieving the screening of useless variables and improving the predictive and generalization abilities of the model. LASSO can not only effectively reduce the risk of overfitting, but also select informative variables with a significant impact on survival time. In the paper, we used LASSO to identify key genes in survival analysis of colon cancer.

Then Cox regression model is used to find the optimal model by transforming survival problem into a problem of modeling hazard ratios. The core assumption of the Cox regression model is the proportional hazards assumption, which states that the effects of covariates on hazard ratios are constant over time. This implies that the effects of covariates on hazard ratios are multiplicative and do not vary with time. Based on this assumption, the Cox regression model estimates the regression coefficients of covariates by maximizing the partial likelihood function, obtaining estimates of the effects of covariates on hazard ratios. The Cox proportional hazards model posits that covariates and risks have a multiplicative relationship, and the formula for calculating its risk function is as follows:

where h(t|x) represents the conditional hazard function at time t given the covariates x, h₀(t) is the baseline hazard function, and α₁, α₂,…, α_n are the regression coefficients corresponding to the covariates x₁, x₂,…, x_n, respectively.

CIBERSORT (Cell-type Identification By Estimating Relative Subsets Of RNA Transcripts) is a computational method used for quantifying and analyzing the presence and distribution of immune cells within tissue samples, specifically designed for immune cell infiltration assay. CIBERSORT is used by deconvolving the gene expression data of mixed cell populations into the contribution of individual cell types. This is achieved by utilizing a reference gene expression signature for each immune cell type. Then the gene expression profile of the mixed tissue sample is then compared to these reference signatures to estimate the proportions of different immune cell types.

CIBERSORT employs a support vector regression (SVR) model to calculate the cell type proportions, taking into account the potential noise and variability in the gene expression data. It uses a linear model to estimate the proportions of immune cells based on the expression levels of their signature genes, with a regularization step to improve accuracy and robustness.

One of the main advantages of CIBERSORT is its ability to provide quantitative assessment of immune cell infiltration using gene expression data. Additionally, CIBERSORT allows for simultaneous consideration of multiple immune cell subtypes, providing comprehensive information on immune cell infiltration within the tissue sample. Furthermore, CIBERSORT does not rely on prior knowledge of immune cell subtype proportions, making it applicable to various sample types and research designs. The workflow of this paper is shown as Figure 2.

According to biological experiments, 133 miRNAs have been identified as colon cancer-related miRNAs, which we used as positive samples for our model. To generate negative samples, we randomly selected 500 miRNAs with no known relationship to colon cancer. Using these samples, we developed our model and assessed its accuracy in identifying miRNA patterns associated with colon cancer using 10-fold cross-validation to test the AUC and AUPR. Figures 3A, B depict the AUC and AUPR curves of each test, along with an average curve for all tests. Furthermore, to demonstrate the superiority of our method, we compared it to other state-of-the-art methods such as Deep Neural Network (DNN), Support Vector Machine (SVM), and K-Nearest Neighbors (KNN), and found that our method achieved significantly higher AUC and AUPR scores, reaching 0.84 ( Figures 3C, D ).

We then used our validated method to input potential colon cancer-associated miRNAs to infer their association with colon cancer. These miRNAs included those related to enteritis, cholecystitis, various cancers, and other digestive diseases, which we believe have potential association with colon cancer due to their association with the disease itself.

Initially, we retrieved immune-related genes from InnateDB, and subsequently collected 133 miRNAs from HMDD, which were reported to be associated with colon cancer. Using miRTarBase as the source of information, we found that these 133 miRNAs had a total of 10,679 target genes, of which 836 were immune-related. Further analysis revealed that 60 of these miRNAs had immune related target genes ( Figure 4A ). We found that 47 of the 83 miRNAs had target genes that were functional in immune responses ( Figure 4B ).

Subsequently, we utilized our method to predict colon cancer-related miRNAs and identified 83 candidates. These miRNAs had a total of 5,954 target genes, of which 525 were immune-related. We found that 47 of the 83 miRNAs had target genes that were functional in immune responses.

As 525 of these miRNAs target immune genes, we investigated their role in colon cancer prognosis. Using clinical and gene expression data from TCGA, we employed LASSO and Cox regression to determine whether these genes were associated with patients’ survival time as shown in Figures 5A, B . After LASSO, we identified seven candidate genes: YBX1, ITGA3, SLC3A2, IL-15, APLN, TCF7, and CD47. Figures 5C–I illustrates the differences in survival curves under varying gene expressions. The P-values for these genes were 0.013, 0.062, 0.087, 0.082, 0.0061, 0.078, and 0.05, respectively (as shown in Figures 5C–I ). Three genes passed the significance threshold: YBX1, APLN, and CD47.

A unique phosphorylation site Ser-176 was found on YBX1 by (32). HEK293 cells and colon cancer HT29 cells that overexpressed the YBX1-S176A (serine to alanine) mutant showed a significant decrease in NF-κB activation ability compared to WT-YBX1-κB, confirming that Ser-176 phosphorylation activates NF-κB for YBX1-κB. Notably, the Ser-176 mutant and the Ser-165 locus each have unique and irreplaceable functions, making them potential targets for cancer treatment strategies. Furthermore, multiple studies have reported on the strong regulation between YBX1 and NF-κB in colon cancer (33, 34).

In a study by (35), the importance of apelin and apelin receptor (APJ) in regulating colon cancer cell motility was examined. The authors found that silencing the APLN gene had an effect on colon cancer cell migration. Additionally, (36) reported that exogenous apelin had anti-apoptotic effects on colon cancer cells.

CD47 binds to a receptor on immune cells called SIRPα, which initiates a signaling pathway that leads to the inhibition of phagocytosis. (37) found that CD47 is overexpressed in colon cancer cells, which can contribute to the immune evasion of these cells. Additionally, they found that the enzyme SHP2 plays a role in regulating CD47 expression and function in colon cancer cells. Through a series of experiments, the authors showed that inhibiting SHP2 activity in colon cancer cells reduced CD47 expression and increased phagocytosis by immune cells. This suggests that targeting the CD47/SIRPα axis may be a potential therapeutic strategy for colon cancer immunotherapy. (38) found that the expression levels of CD47 and SIRPA correlated with poor prognosis and advanced tumor stage. Furthermore, high expression of CD47 and SIRPA was associated with low infiltration of CD8+ T cells and M1 macrophages, which are important effector cells of antitumor immunity. They speculated that targeting the CD47-SIRPA checkpoint might be a potential therapeutic strategy to enhance antitumor immunity in colorectal cancer.

Furthermore, the expression of these genes was classified into high-risk and low-risk categories, and a survival analysis (as shown in Figure 5J ) demonstrated a statistically significant difference in survival time between these categories with a P-value of 0.017.

As shown in Figures 6A, B , through our analysis of cell infiltration levels of various immune genes, we observed a high abundance of CD4 memory resting cells, macrophages.M2, and Mast.cell.activated cells.

CD4 memory resting cells are a subtype of T lymphocytes that can mount a rapid immune response upon re-exposure to a previously encountered antigen (39). CD4 memory resting cells can play a dual role in both the immune response to cancer cells as well as the development and progression of colon cancer (40). These cells can recognize and attack cancer cells expressing specific antigens through direct cytotoxicity or by releasing cytokines that activate other immune cells and recruit them to the site of the tumor. However, CD4 memory resting cells can also promote the growth and progression of cancer by promoting angiogenesis and remodeling of the extracellular matrix. Studies have found that CD4 memory resting cells are positively associated with the expression of PD-L1 (41), an immune checkpoint molecule that inhibits the anti-tumor immune response, and can have a pro-tumor effect. The role of CD4 memory resting cells in colon cancer is complex and is likely influenced by various factors such as the cancer stage, antigenic profile, and the immune microenvironment(42). In early-stage colon cancer, CD4 memory resting cells may be associated with a better prognosis by contributing to the anti-tumor immune response. In contrast, in advanced-stage colon cancer, the pro-tumor effects of CD4 memory resting cells may outweigh their anti-tumor effects, resulting in a poorer prognosis (43).

Macrophages.M2 are a type of immune cell that plays a role in tissue repair and has anti-inflammatory properties (44). However, in the context of colon cancer, macrophages.M2 cells can promote tumor growth and progression through various mechanisms. They can secrete growth factors and cytokines that stimulate cancer cell proliferation and migration (45). Additionally, they can promote angiogenesis and remodeling of the extracellular matrix, both of which facilitate tumor growth and metastasis. Furthermore, macrophages.M2 cells can inhibit the anti-tumor immune response by producing cytokines and chemokines that recruit other immune cells to the tumor microenvironment, including regulatory T cells and myeloid-derived suppressor cells (46). These cells, in turn, can suppress the activity of cytotoxic T cells and natural killer cells. Additionally, macrophages.M2 cells can express immune checkpoint molecules, such as PD-L1 and CTLA-4 (47), that can inhibit the activity of cytotoxic T cells. The role of macrophages.M2 cells in colon cancer is complex and likely dependent on various factors, including the stage of the cancer, the specific cytokine and chemokine profile of the tumor microenvironment, and the composition of the immune microenvironment. For example, in early-stage colon cancer, the presence of macrophages. M2 cells in the tumor microenvironment may be associated with a better prognosis due to their contribution to tissue repair and anti-inflammatory responses. However, in advanced-stage colon cancer, the pro-tumor effects of macrophages.M2 cells may outweigh their anti-inflammatory effects, leading to a poorer prognosis (48).

Mast.cell.activated cells are a type of immune cell that is involved in allergic reactions and inflammation. The role of mast cell activation in colorectal cancer is not fully understood (49). Several studies have shown that mast cell activation can promote tumor growth and progression by promoting angiogenesis and extracellular matrix remodeling (50). In addition, mast cell activation can promote the recruitment of other immune cells into the tumor microenvironment, such as regulatory T cells and myeloid-derived suppressor cells, which can suppress antitumor immune responses (51). However, other studies have shown that mast cell activation can exert anti-tumor effects by stimulating the activity of cytotoxic T cells and natural killer cells, which recognize and attack cancer cells (52). In addition, mast cell activation can stimulate the production of cytokines and chemokines that recruit other immune cells, such as dendritic cells and macrophages, into the tumor microenvironment, thereby enhancing antitumor immunity reaction (53).

We also explored the function of these genes by KEGG and Gene Ontology. Figures 6C, D displays only the top 25 pathways and GO terms. Among them, the most significant pathway is ko04668, which contains 114 genes, including 66 colon cancer-related genes. This pathway, known as the cytokine-cytokine receptor interaction pathway, encompasses various cytokines and their receptors involved in regulating the immune response. In colon cancer patients, the dysregulation of this pathway can occur, contributing to cancer development and progression (54). Alterations in cytokine expression and signaling, resulting from the dysregulation of the ko04668 pathway, can impact the immune microenvironment of the tumor. Studies have found that cytokines and their receptors in this pathway are related to colon cancer progression and prognosis (55). For instance, interleukin-6 (IL-6) and its receptor (IL-6R) are overexpressed in colon cancer tissues, and their expression levels correlate positively with tumor stage and poor prognosis (56). Other cytokines and receptors in this pathway, such as IL-10, tumor necrosis factor (TNF), and transforming growth factor-beta (TGF-β), have also been linked to colon cancer progression and metastasis (57).

The most significant GO term is GO:0019221, which pertains to the cytokine-mediated signaling pathway. This term encompasses the series of molecular events that take place when cytokines bind to receptors on the surface of cells, resulting in the activation of intracellular signaling pathways and eventually changes in gene expression or cellular behavior. This GO term is similar to the ko04668 pathway, which is the most significant pathway we found. Cytokines are crucial regulators of immune responses and have been associated with the development and progression of colon cancer. The dysregulation of cytokine-mediated signaling pathways can contribute to the formation of an immunosuppressive tumor microenvironment, enabling cancer cells to avoid immune surveillance and promoting tumor growth and metastasis. Several cytokines, including interleukin-6 (IL-6), tumor necrosis factor-alpha (TNF-α), and transforming growth factor-beta (TGF-β), have been identified as important regulators of colon cancer development and progression (58). These cytokines can activate various signaling pathways (59), including the JAK/STAT pathway and the NF-κB pathway, both of which have been shown to be dysregulated in colon cancer (60, 61).