2.1. Screening of differentially expressed mutation-prone genes in GC

We developed a deep learning prioritization and ranking Graph Neural Networks (GNNs) model, specifically GCNConv (Graph Convolutional Networks), to identify GC pathway-specific differentially expressed mutation-prone genes. At first, RNAseq gene expression data were retrieved from the GEPIA2 database (http://gepia2.cancer-pku.cn/) [27], and GC-specific mutation-prone genes were retrieved from the cBioPortal database (https://www.cbioportal.org/) [28]. The common genes between dysregulated genes and mutation-prone genes were identified using a Venn diagram from the InteractiVenn tool (https://www.interactivenn.net/) [29]. The protein-protein interaction network was constructed using the common genes from the STRING database (https://string-db.org/) [30], which contains both experimentally validated and predicted protein-protein interactions. Furthermore, GC pathway-specific genes were collected from the KEGG pathway database (https://www.kegg.jp/) [31] that contains pathways of molecular mechanisms for disease development, including cancer pathways [32]. These commonly mutation-prone differentially expressed genes (DEGs), genes from protein-protein interaction networks, and GC pathway-specific genes were utilized as input to develop a robust computational framework to identify pathway-specific, highly dysregulated, mutation-prone genes responsible for GC pathogenesis. Secondly, the NetworkX library of Python was implemented for graph construction and topological analysis (degree, betweenness, closeness) [33]. Thirdly, the PyTorch library for tensor operations and neural network architecture, and torch_geometric for graph neural network implementation, were utilized. Finally, the StandardScaler of the scikit-learn library was used for data preprocessing [34]. Specifically, we employed a Graph Convolutional Network (GCN) implemented through the GCNConv layers, which enabled the propagation of node features across the network topology [35,36]. Node features were normalized using StandardScaler to ensure consistent scaling across heterogeneous molecular data types [34]. The network architecture consisted of two GCNConv layers followed by ReLU (Rectified Linear Unit) activation functions to introduce non-linearity into the neural network, which is crucial because most of the real-world problems are non-linear, and neural networks need non-linear activation functions to learn complex patterns [37]. Subsequently, dropout regularization (rate = 0.3) was applied between layers to prevent overfitting. Model performance was rigorously assessed through train-validation-test splitting (60:20:20 ratio) using the train_test_split module from scikit-learn, where training and validation datasets were utilized for learning and internal validation purposes, whereas the testing set was utilized and masked for external validation. Moreover, a nested 5-fold cross-validation (CV) strategy was applied to maintain both hyperparameter tuning and unbiased test estimate with 2 loops: outer loop for model evaluation and inner loop for hyperparameter tuning. In this framework, the inner loop was exclusively used to optimize model hyperparameters without exposing the outer test folds, whereas the outer loop evaluated performance on completely unseen data partitions to provide an unbiased estimate of generalization performance. To mitigate information leakage, we implemented a gene-level partitioning strategy in which all features and network relationships associated with a given gene are confined to a single data split, preventing cross-set propagation of topological or annotation-derived information. Finally, the performance metrics of the model were evaluated using training and validation loss over 200 epochs.

To develop the model, the GNN is first trained to predict a propensity score based on the relevance of 3 gene categories.

1. (1) Mutation-prone Differentially Expressed Genes (FDR-adjusted p-values < 0.05).
2. (2) Genes from the protein-protein interaction network (confidence score: 0.4 (medium to balanced network).
3. (3) Gastric cancer (GC) pathway-specific genes.

This represents the integrated likelihood that a gene contributes to gastric cancer pathogenesis due to mutations, expression dysregulation, and network-level functional importance. A continuous propensity score is operationalized by combining normalized gene expression magnitude (Log2FC) with multiple graph-topological centrality measures (degree, betweenness, and closeness) and, when available, pathway membership information. These biological components reflect the premise that genes that are both dysregulated and centrally positioned in protein interaction networks, along with being present in the GC pathway, are more likely to act as key disease drivers. The resulting composite score serves as the dependent variable in a regression framework and ranks the top genes. Therefore, the primary regression performance metrics were mean squared error (MSE), root mean squared error (RMSE), and coefficient of determination (R²).

Afterward, the classification metrics were used as a secondary interpretive analysis (AUC-ROC reporting) rather than the primary modeling objective. For this purpose, the continuous propensity scores were binarized using a biologically interpretable threshold (≥0.5 after sigmoid normalization), representing high-confidence versus low-confidence pathogenic candidates. This transformation allowed us to assess the model’s ability to discriminate high-risk oncogenic genes from background genes, resulting in identifying the top 20 hub genes based on both regression and classification metrics. Due to the complexity of several layers of input data in the model, we did not consider the Youden index or cost-sensitive approaches in this case; rather, we chose a simplified classification boundary threshold (≥0.5 after sigmoid normalization) to remove background genes from more valuable genes. The regression captured fine-grained biological risk ranking, while ROC and precision-recall analyses provided the discrimination performance to remove background genes for prioritization tasks. All reported metrics were computed on independent training, validation, and test partitions to ensure robustness.

2.2. Identification of high-frequency mutation-prone genes

We have screened high-frequency mutation-prone genes by analyzing 147 GC patient samples from the cBioPortal database due to their impact on the dysregulation of proteins. Based on the deep learning neural network model and frequency analysis, we have found that tumor suppressor proteins are the most frequently mutated proteins. We have further examined their involvement in the GC pathway from the KEGG database (https://www.genome.jp/kegg/), which is a comprehensive resource for understanding high-level biological functions and utilities, including molecular-level information such as genes, proteins, and biochemical pathways [31]. Another crucial reason for choosing tumor suppressor proteins for single-nucleotide polymorphism analysis is that they play critical roles in inhibiting GC, and mutations in these genes are likely to have been implicated in the development and progression of GC [15,16].

2.3. Identification of tumor suppressor domains in proteins

The tumor suppressor domains of proteins were identified using the NCBI Conserved Domain Search Tool (https://www.ncbi.nlm.nih.gov/Structure/cdd/wrpsb.cgi) [38] and validated through the UniProt database (https://www.uniprot.org/) [39]. The NCBI Conserved Domain Search Tool was used to search for conserved domains within the target protein sequences. This allows searching against various domain databases, including the NCBI-curated Conserved Domain Database (CDD), to identify significant matches to known functional domains. The UniProt database was then used to gather additional information on the identified tumor suppressor domains responsible for tumor suppressor activity.

2.4. Retrieval of clinical variations of tumor suppressor proteins

Clinically significant variations were retrieved from the ClinVar database of NCBI (https://www.ncbi.nlm.nih.gov/clinvar/) to investigate their association with cancer [40]. ClinVar consists of curated and annotated clinically significant variants, such as frameshift, missense, nonsense, splice site, ncRNA, and UTR, to prioritize potentially pathogenic variations and provide evidence of a link between these variants and disease phenotypes. Only pathogenic variations were selected, and likely pathogenic clinical variations were considered only for missense SNP variation analysis, as sophisticated methodologies have been established to analyze missense variations compared to other variations, and these variants are often directly linked to functional consequences [41].

2.5. Identifying damaging, deleterious, and disease-causing missense SNPs

To identify potentially damaging, deleterious, disease-causing, and oncogenic missense SNPs, we employed a multi-faceted bioinformatics approach utilizing seven established prediction tools. We leverage the capabilities of PolyPhen2 (https://genetics.bwh.harvard.edu/pph2/) [42], SIFT (https://sift.bii.a-star.edu.sg/) [43], SNPs&GO (https://snps.biofold.org/snps-and-go/) [44], PhD-SNP (https://snps.biofold.org/phd-snp/) [45], SNPs3D (http://snps3d.org/) [46], GVGD (http://agvgd.hci.utah.edu/) [47], and Cscape (http://cscape.biocompute.org.uk/) [48]. Each tool was selected for its unique strengths to meet the ACMG variant classification guidelines [49], which are standard practice in SNP studies [50,51]. PolyPhen2 assessed the potential impact of amino acid substitutions on protein function using structural and comparative considerations, whereas SIFT predicted the effects of amino acid changes based on sequence homology and physical properties. SNPs&GO correlated SNPs with gene ontology to predict disease relevance, PhD-SNP evaluated the deleterious effects of mutations, SNPs3D provided structural insights into damaging SNPs, GVGD classified the likelihood of pathogenicity, and finally, Cscape evaluated the oncogenic potential of mutations. Incorporating the scores of seven computational tools improved forecast precision and ensured the stringency and accuracy of results.

2.6. Determining the impact of missense SNPs on protein structural stability

To determine the impact of missense SNPs on the structural stability of tumor suppressor proteins, we utilized the I-Mutant 2.0 tool [https://folding.biofold.org/i-mutant/i-mutant2.0.html] [52], which is an online support vector machine-based tool that predicts the effects of single-site mutations on protein stability from the sequence or structure of the protein. We also employed the MUPro tool [https://mupro.proteomics.ics.uci.edu/] [53] to corroborate our findings, as it is another well-established method for predicting the impact of mutations on protein stability.

2.7. Examining the effects of missense SNPs on protein 3D Structure

The structural effects of missense SNPs were examined using two complementary bioinformatics tools: Missense3D (https://missense3d.bc.ic.ac.uk/) [54] and HOPE server (https://www3.cmbi.umcn.nl/hope/) [55]. Missense3D, a web server developed by the Structural Bioinformatics Group at Imperial College London, predicts the structural changes introduced by an amino acid substitution by analyzing both experimental protein structures (PDB) and homology-predicted structures. HOPE is another web server hosted by the CMBI group at the University of Nijmegen that collects and combines structural information from a variety of web services and databases to produce a comprehensive report on the effects of a point mutation. The combined use of these two tools allowed for a thorough investigation of how missense SNPs might impact protein 3D structure, with the goal of better understanding the potential functional consequences of these genetic SNPs.

2.8. Analysis of protein evolutionary conservation

To assess the evolutionary conservation of the proteins, we employed the ConSurf server (https://consurfdb.tau.ac.il/) [56]. This analysis was undertaken to identify functionally important residues within these proteins, as highly conserved residues are often crucial for protein structure and function. The ConSurf server utilizes a phylogenetic approach, combining multiple sequence alignments with evolutionary modeling to predict the evolutionary conservation of each amino acid position within a protein structure. The conservation scores, which range from 1 (variable) to 9 (highly conserved), were then mapped onto the protein structures to identify the evolutionarily conserved regions.

2.9. Comparison of protein energy and minimized energy in wild-type and mutant proteins

To investigate the impact of mutations on tumor suppressor proteins, we further employed UCSF Chimera to introduce mutations into the proteins of interest [57]. Subsequently, the energy calculation and minimized energy calculation of these mutated proteins were performed using the Swiss-Pdb Viewer [58]. The selection of Swiss-Pdb Viewer was based on its intuitive interface and capabilities for energy minimization calculations and analysis of structural features such as H-bonds, angles, and interatomic distances. This process provided quantitative data on the energetic consequences of the introduced mutations, thereby elucidating their potential effects on protein stability and function. To assess the effects of minimized energies of the mutated proteins, they were compared to the energies of wild-type proteins.

2.10. Prediction of post‑translational modification sites

The impact of missense SNPs on post-translational modification (PTM) sites was elucidated by employing GPS-MSP 1.0 (https://msp.biocuckoo.org/) [59], GPS-6.0 (https://gps.biocuckoo.cn/) [60], and GPS-Uber (https://gpsuber.biocuckoo.cn/) [61] to predict methylation, phosphorylation, and ubiquitination sites, respectively. These tools were chosen due to their high accuracy and specificity in identifying PTM sites. GPS-MSP 1.0 integrates deep neural networks, convolutional neural networks, and penalty logistic regression to predict methylation sites. On the other hand, GPS-6.0, a group-based prediction system, employs transfer learning and machine learning algorithms to predict phosphorylation sites. Finally, GPS-Uber is a ubiquitin-protein ligase enzyme-substrate relationship prediction tool to predict ubiquitination sites. By applying these tools, we aimed to identify potential PTM sites in missense SNPs and elucidate their effects on protein function and stability.

2.11. Validation of GC-specific variations

Before further analysis, clinical variations were validated by establishing a link between SNPs and GC pathogenesis using the cBioPortal database (https://www.cbioportal.org/) [28]. The reason behind choosing the cBioPortal database is that it is a reliable and widely used specialized database focusing only on cancer genomics data for exploring the complex relationships between genomic alterations and cancer outcomes, and retrieving and analyzing genomic data from a large cohort of GC patients. Therefore, this analysis enables the identification of statistically significant correlations between genetic variations and GC pathogenesis [62].

2.12. Comparison of energetic patterns between wild-type and mutant proteins

Specific missense SNPs that were correlated with GC pathogenesis were further analyzed using titration curve analysis to elucidate mutation-induced changes in protein electrostatic properties. We utilized Schrodinger’s Biologics Package to generate titration curves, as this approach provides high-resolution insights into protein ionization behaviors across physiological conditions [63]. The analysis was conducted by calculating the net charge of proteins at pH intervals from 2 to 12. This methodology enabled quantitative comparison of wild-type and mutant protein ionization behaviors, revealing shifts in pKa values, protonation states, and isoelectric points that impact protein stability and function, ultimately providing insights into mutation-specific perturbations in protein electrostatic properties.

Furthermore, energetic pattern analysis was implemented to quantify mutation-induced biophysical perturbations affecting protein stability and function. We employed Schrodinger’s Biologics Package, which utilized sophisticated force field implementations with accurate energy minimization algorithms for a multi-parameter energetic analysis combining hydropathy calculations, potential energy distributions, and interaction energy computations [63]. To reveal crucial biophysical perturbations of mutation in hydrophobicity patterns, the energetic patterns exhibit changes in thermodynamic stability and molecular interactions, providing mechanistic insights into mutation-induced protein dysfunction.

2.13. Atomic-level analysis of proteins utilizing unsupervised machine learning

The atomic-level evaluation was conducted using unsupervised machine learning to detect subtle conformational changes and disrupted interaction networks in mutant proteins. This applied to atomic-level structural features to explore potential conformational patterns and provide comparative diversity between wild-type and mutant structures at the atomic level. We implemented scikit-learn’s k-means clustering algorithms in Python 3, combined with dimensionality reduction techniques, as this approach enables unbiased pattern recognition and scalable analysis of protein structural features [64]. The methodology involved preprocessing atomic coordinates through principal component analysis, followed by k-means clustering. Atomic features were encoded using a combination of spatial coordinates, atomic properties, and atomic interaction patterns to enable the identification of conserved atomic clusters critical for protein function, detection of mutation-induced perturbations in atomic interaction networks, and quantification of structural changes, providing a systematic framework for understanding protein dysfunction mechanisms.

2.14. Examination of protein expression and patient survival

To investigate the possible impact of missense SNPs on protein expression, we analyzed the expression profiles of proteins in GC using the GEPIA2 web server (http://gepia2.cancer-pku.cn) [27]. This is a comprehensive and interactive web application for gene expression analysis based on The Cancer Genome Atlas (TCGA) and Genotype-Tissue Expression (GTEx) data. We utilized the differential expression analysis module (Expression DIY) of GEPIA2 to compare the expression levels of the selected proteins between tumor and normal gastric tissues to assess whether missense SNPs in these genes might lead to altered protein expression patterns, which could potentially impact protein function, contribute to disease development, and compromise patient survival. The patient survival differences were analyzed using Kaplan-Meier plot analysis (https://kmplot.com/), which is suitable for addressing the correlation between genes or proteins and survival in more than 35 thousand samples from 21 tumor types [65]. The clinical covariates analysis of survival in different stages of cancer was analyzed using TIMER 3.0 (https://compbio.cn/timer3/) [66].

2.15. Analysis of the network and hallmarks of tumor suppressor proteins

We conducted network and pathway analysis of proteins utilizing the STRING database (https://string-db.org/) [30] and cancer hallmark database (https://cancerhallmarks.com) [67] respectively. STRING reconstructed the protein-protein interaction networks of tumor suppressor proteins, which integrates multiple sources of experimental and predicted protein interaction data, providing a comprehensive view of protein networks. On the other hand, the cancer hallmark database employed the cancer hallmarks enrichment plot to identify key cancer-relevant hallmarks associated with tumor suppressor proteins.

2.16. Analysis of gene regulatory networks of tumor suppressor proteins

A gene-regulatory network has been constructed to determine the key elements that might interfere with the tumor suppressor proteins of GC. NetworkAnalyst 3.0 (https://www.networkanalyst.ca/) was used to identify gene-regulatory networks (GRNs) and the intricate relationships between gene-microRNAs (miRNAs) and gene-transcription factors (TFs) [68]. To build and evaluate the GRNs, we specifically used TTRUST, a database of TF-target interactions, and miRTarBase v9.0, an extensive library of miRNA-target interactions [69]. We chose NetworkAnalyst 3.0 because it can integrate and analyze large biological networks and contains large databases like miRTarBase v9.0 and TTRUST.

2.17. Altered gene expression assessment of oncogenes

To investigate the impact of mutations on gene expression patterns in GC pathway oncogenes, we performed differential gene expression analysis between tumors with and without specific mutations. The analysis was conducted using the Gene Mutation module of TIMER 2.0 (http://timer.comp-genomics.org/) [70]. TIMER 2.0 was selected because it provides a robust platform for analyzing associations between gene mutations and expression patterns across multiple cancer types using TCGA data. The Gene Mutation module analyzed and visualized the effect of gene mutations on gene expression by comparing expression levels between mutated and wild-type tumors. For each gene of interest, we examined the statistical significance of expression differences between tumors harboring mutations and those without mutations to identify potential correlations between specific gene mutations and altered expression patterns

3.1. Identification of the most significant dysregulated mutation-prone genes

Understanding the molecular basis of GC requires identifying genes that exhibit both dysregulated expression and high mutation propensity, as these represent critical drivers of oncogenesis. Our graph-based neural network model, developed using differentially expressed genes, mutation-prone genes, common genes between DEGs and mutation-prone genes, protein-protein interaction network, and GC pathway genes, revealed a substantial overlap between GC-specific mutation-prone genes and differentially expressed genes (DEGs) (S1–S5 Tables, respectively). As illustrated in Fig 2A, we identified 10,981 mutation-prone genes and 3,741 differentially expressed genes, with 1,886 genes (50.4% of DEGs) showing both characteristics. The network visualization highlights the protein-protein interaction landscape of these genes, with red nodes representing the most important GC pathway-specific genes that serve as key hubs within the network (Fig 2B). Fig 2C depicts our GCN architecture with two hidden layers. The GCN model takes an input graph, processes it through two hidden layers with ReLU activation, and produces an output graph. Each layer consists of multiple nodes with weighted connections to highlight relationships within the graph, indicated by the color-coding of nodes and edges. Our graph neural network model effectively identified differentially expressed mutation-prone genes with exceptional accuracy across all datasets. The model demonstrated robust predictive capabilities after 5-fold nested cross-validation (CV) with low mean square error values (MSE: 0.00482 ± 0.00023 to 0.07108 ± 0.00437) and strong goodness-of-fit metrics (R²: 0.85098 ± 0.01903 to 0.85899 ± 0.01987) across training, validation, and testing datasets (Table 1). The error distribution histograms for training (Fig 3A), validation (Fig 3B), and testing (Fig 3C) datasets showed tightly clustered prediction errors around zero, confirming the model’s precision. The learning curve (Fig 3D) demonstrated rapid convergence within 50 epochs, stabilizing training and validation losses, indicating proper model fitting without overfitting. High discriminative power was confirmed through impressive AUC-ROC scores (Fig 3E) of 0.93287 ± 0.00373, 0.93095 ± 0.01758, and 0.93309 ± 0.00725 for training, validation, and testing datasets, respectively, significantly outperforming random classification (diagonal line). Similarly, the precision-recall curves (Fig 3F) yielded strong AUC-PR values (0.93879 ± 0.01815 to 0.94772 ± 0.00214), further validating the model’s reliability. Through this computational approach, we identified the top 20 GC pathway-specific dysregulated mutation-prone genes, with TP53 exhibiting the highest mutation frequency (47.6%), followed by ERBB2 (8.8%), CDH1 (8.2%), and APC (6.8%) (Table 2). Notably, 17 of these genes were confirmed to be present in established GC pathways, with importance scores exceeding 0.99. The network analysis revealed that TP53 exhibited the highest degree (0.239), betweenness centrality (0.081), and closeness centrality (0.521), indicating its crucial role as a hub gene in GC progression. Importance scores are derived based on co-expression, experimentally determined interaction, database annotation, automated text mining, and a combined score from the STRING database. On the other hand, the predicted scores are the final calculated scores from the model. For further analysis, we have selected TP53, CDH1, and APC based on their most important pathway-specific high-frequency mutation states and tumor suppressor capabilities.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Performance Metrics of training, testing, and validation datasets of differentially expressed mutation-prone genes with ± SD (standard deviations).

https://doi.org/10.1371/journal.pone.0358440.t001

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Graph neural network-based prediction of GC pathway-specific dysregulated mutation-prone genes and mutation frequency from the cBioPortal database.

https://doi.org/10.1371/journal.pone.0358440.t002

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Identification of differentially expressed mutation-prone genes using the graph neural network model and evaluation of model error.

(A) Identification of common DEGs and mutation-prone genes. (B) Protein-protein interaction network (Red indicates the most important GC pathway-specific genes). (C) GCNconv model with two hidden layers and ReLU activation.

https://doi.org/10.1371/journal.pone.0358440.g002

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Performance evaluation of the graph neural network model.

(A), (B), (C) Error distribution of training, validation, and testing datasets. (D) Training and validation data loss over 200 epochs. (E) ROC curve and (F) PR curve. Here, AUC = area under the curve, and AP = average precision. AUC and AP values range from 0 to 1, where 1 indicates a perfect classifier.

https://doi.org/10.1371/journal.pone.0358440.g003

3.2. Tumor suppressor proteins of the GC pathway

Our analysis of the GC pathway from the KEGG database revealed that TP53, APC, and CDH1 are key tumor suppressor proteins implicated in the GC pathway in three distinct mechanisms (Fig 4A). However, functional analysis revealed that both APC and CDH1 inhibit beta-catenin, thereby preventing the uncontrolled proliferation of cancer cells. In contrast, TP53 is activated in response to DNA damage to inhibit uncontrolled proliferation, increase survival, and mitigate genomic instability [31]. The functional consequences of these three crucial proteins that play key roles in suppressing GC have been presented in Fig 4B.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Tumor suppressor proteins mediated the regulation of the GC pathway.

(A) GC pathway from the KEGG database. (B) Functional consequences of APC, CDH1, and TP53 in suppressing GC development [31].

https://doi.org/10.1371/journal.pone.0358440.g004

3.3. Mapping tumor suppressor domains of proteins

Mapping tumor suppressor domains is essential to precisely identify and characterize the specific functional regions within key tumor suppressor proteins. This enables a deeper understanding of their roles in cancer prevention and facilitating targeted therapeutic development. We investigated the tumor suppressor domains of TP53, APC, and CDH1 and revealed specific domains essential for their tumor-suppressive functions. Specifically, the TP53 protein exhibits a tumor suppressor domain spanning amino acid positions 109–288, while APC encompasses a domain from positions 127–207, and CDH1 features a domain from positions 155–262. The tumor suppressor domains have been represented in Fig 5.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Tumor Suppressor Domains of Proteins.

(A) TP53, (B) CDH1, and (C) APC.

https://doi.org/10.1371/journal.pone.0358440.g005

3.4. Retrieval of clinical variations

The ClinVar database consists of frameshift, missense, nonsense, splice site, ncRNA, and UTR (untranslated region) variations. The pathogenic variations were initially retrieved from the database for TP53 (936), CDH1 (559), and APC (2370). Besides, for missense SNPs, both pathogenic and likely pathogenic variations were considered for further analysis. General missense SNPs from the dbSNP database are considered as common variants with uncertain clinical significance, whereas missense SNPs from ClinVar are curated and annotated variants with known or predicted clinical relevance and phenotypic disease associations. Therefore, from a total of 11,393 identified missense SNPs across the three selected proteins: APC, CDH1, and P53, only 216 missense SNPs of TP53, 19 of APC, and 22 of CDH1 were classified as clinically significant pathogenic or potentially pathogenic based on the NCBI ClinVar database that could have functional consequences and directly contribute to GC pathogenesis (Accessed at 24 February, 2025). The list of retrieved missense SNPs was incorporated in the form of the latest chromosome 38 version (GRCh38) in Fig 6.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Pathogenic or potentially pathogenic clinically significant missense SNPs that could have a link with GC pathogenesis.

Green Circle: Tumor Suppressor Proteins and Red Box: Missense SNPs.

https://doi.org/10.1371/journal.pone.0358440.g006

3.5. Damaging, deleterious, and disease-causing missense SNPs in tumor suppressor proteins

This computational analysis was performed to systematically identify and characterize deleterious missense SNPs in tumor suppressor proteins, thereby elucidating their potential role in cancer development and progression. Through comprehensive analysis, we identified 54 missense SNPs in the TP53 gene that were consistently classified as “high-risk,” as they were unanimously predicted to be “pathogenic” or “damaging” by all the employed tools. Similarly, 9 missense SNPs in APC and 11 missense SNPs in CDH1 were also defined as “high-risk” based on their consensus classification as “pathogenic” or “damaging” across all assessment approaches. The full list of these high-risk missense SNPs has been provided in Table 3.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. List of potential GC pathway-disrupting missense SNPs.

https://doi.org/10.1371/journal.pone.0358440.t003

3.6. Impact of missense SNPs on protein structural stability

The assessment of protein structural stability is required to elucidate the structural implications of missense SNPs on tumor suppressor proteins, thereby providing insight into their potential oncogenic effects. This assessment revealed that the majority of the identified high-risk missense SNPs exerted a detrimental impact on the structural integrity of the corresponding tumor suppressor proteins. Specifically, all of the 54 missense SNPs in TP53, 9 in APC, and 7 in CDH1 were found to significantly decrease the overall structural stability of these critical proteins, as predicted by the I-Mutant 2.0 and MUpro computational tools (Table 3).

3.7. Effects of missense SNPs on protein 3D Structure

Computational analysis through Missense3D and HOPE servers is required to predict the structure-disrupting effects of missense SNPs on protein 3D structure. Using computational analyses through Missense3D and the HOPE servers, we evaluated the effects of missense SNPs on the three-dimensional structures of tumor suppressor proteins TP53, APC, and CDH1. Our findings revealed 54 structural damages in TP53, three in APC, and two in CDH1, indicating significant alterations in protein conformation attributable to these SNPs. Furthermore, the HOPE scores for the point mutations corroborated the potential deleterious impacts on protein 3D structure (Table 3).

3.8. Protein evolutionary conservation sites

Protein evolutionary conservation analysis is required to identify evolutionarily conserved sites within tumor suppressor domains of TP53, APC, and CDH1 proteins. This highlights critical regions where mutations are likely to have significant functional consequences and contribute to tumorigenesis. The tumor suppressor proteins TP53, APC, and CDH1 revealed a high degree of evolutionary conservation throughout the tumor suppressor domains. The analyzed 54 TP53 SNPs exhibited a ConSurf conservation score of 7–9, indicating highly conserved, surface-exposed residues. Similarly, APC R141S and CDH1 D254Y showed a ConSurf score of 9, indicating their critical role within the tumor suppressor domain. This high conservation suggests that alterations at these sites may have profound effects on protein function and tumorigenesis. In contrast, CDH1 G239R displayed a lower score of 6, also suggesting good conservancy but potentially less stringent selective pressure at this site than APC R141S and CDH1 D254Y. Protein conservancy is shown in Fig 7.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 7. Evolutionary conservation in the region of the representative domain of proteins.

(A) APC, (B) CDH1, and (C) TP53. Conservation scores 6-9 (Purple) are considered highly conserved amino acid positions.

https://doi.org/10.1371/journal.pone.0358440.g007

3.9. Estimation of protein energy

The energy estimation and minimization analysis revealed insights into the immediate structural instability of mutant proteins compared to wild-type proteins. The mutations in the APC protein at position R141S and the CDH1 protein at position D254Y exhibited high energy values, indicating that the overall protein structure was unfavorable. This suggests that these mutations may have a significant impact on the structural integrity and stability of proteins. In contrast, the TP53 protein had a high number of mutated sites with 42 high-energy values, including 31 abnormally high energy values, often indicating that the protein is not in its ideal or stable state and is therefore at high risk. Specific mutations in TP53, such as D281A, D281G, D281H, R273P, R273G, G266E, G266R, and L265P, displayed particularly high energy minimization values, further suggesting the potential instability of these mutant forms of the protein. Protein Energy estimation before and after minimization, together with the wild-type form of proteins, was represented in S6 Table.

3.10. Post‑translational modification sites

The post-translational modification site analysis using GPS-MSP 1.0 and GPS-6.0 is essential to identify potential regulatory modifications that may influence the functional activity and stability of tumor suppressor proteins to elucidate mechanisms underlying their dysregulation in cancer. This analysis, using the GPS-MSP 1.0 and GPS-6.0 tools, revealed several post-translational modification (PTM) sites in the tumor suppressor genes TP53, APC, and CDH1. Specifically, the GPS-MSP 1.0 tool predicted nine methylation sites in TP53, one methylation site in the tumor suppressor domain of APC, and no methylation sites in the tumor suppressor domain of CDH1. Furthermore, the GPS-6.0 tool identified four phosphorylation sites in TP53, but no phosphorylation sites in APC or CDH1. Notably, no protein ubiquitin ligase sites were detected in the tumor suppressor domains of these genes. The PTM sites are represented in Fig 8.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 8. Post-translational modification sites of proteins.

(A) TP53, (B) CDH1, and (C) APC. TP53 contains both Methylation and Phosphorylation sites, CDH1 does not contain any PTM sites, and AMP contains 1 Methylation site in the tumor suppressor domain regions.

https://doi.org/10.1371/journal.pone.0358440.g008

3.11. Validation of clinically significant missense SNPs from GC patients

Validation of clinically significant missense SNPs is required to confirm the significance and relevance of identified missense SNPs in GC patients and establish their potential association with the GC and tumor suppressor genes TP53, APC, and CDH1. The validation of clinically significant missense SNPs was conducted using cBioPortal, analyzing data from 5,623 samples of GC patients across 22 studies. The results indicated that 35 dangerous missense SNPs in the TP53 gene were identified as clinically significant and directly associated with GC, while no clinically significant SNPs were found in the APC gene. Additionally, one clinically significant, dangerous missense SNP was detected in the CDH1 gene, as illustrated in Fig 9. Validation of missense SNPs using cBioPortal confirmed the clinical relevance of the identified oncogenic missense SNPs and their direct association with GC. Apart from missense SNPs, other pathogenic clinical variations, such as frameshift, nonsense, splice site, ncRNA, and UTR, were also validated from the cBioPortal database to identify mutations that can alter the balance of the GC pathway through the disruption of tumor suppressor proteins. Our analysis found that only missense (36) and nonsense (60) pathogenic variations were linked to GC pathogenesis. The list of all validated pathogenic clinical variations of TP53, CDH1, and APC is listed in S7, S8, and S9 Tables, respectively.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 9. Mutations disrupting GC suppression proteins.

(A) Missense SNPs, and (B) Nonsense SNPs. Light Green Triangle: Tumor suppressor Protein, Dark Red Circle: experimentally validated pathogenic SNPs in GC, and Light Red Circle: missense SNPs that may disrupt GC prevention.

https://doi.org/10.1371/journal.pone.0358440.g009

3.12. Comparison of energetic patterns between wild-type and mutant proteins

Titration curve analysis between wild-type and mutant proteins enables direct comparison of their ionization behaviors, revealing shifts in pKa values and protonation states that could indicate potential impact on protein stability, function, and intermolecular interactions. S2 Fig shows the analysis of the titration curves for TP53, revealing highly similar protonation behavior between wild-type (pI = 8.6) and T125R mutant (pI = 8.9) variants. Both proteins exhibit a characteristic sigmoidal curve with a maximum net charge of approximately +30 at pH 2, gradually decreasing to −20 at pH 12. The subtle difference in isoelectric points suggests the minimal impact of the T125R mutation on the protein’s overall charge distribution, with near-identical titration behavior observed across the entire pH range. The CDH1 wild-type (pI = 4.3) and D254Y mutant (pI = 4.3) proteins demonstrated identical isoelectric points but showed distinct titration patterns. The curves initiate at approximately +15 net charge at pH 2, converge through the isoelectric point, and progressively decrease to approximately −30 at pH 12. Despite sharing the same pI, subtle variations in the curves between pH 6–10 suggest localized effects of the D254Y mutation on charge distribution, while maintaining the protein’s overall acidic character. APC titration curves revealed minimal differences between wild-type (pI = 5.5) and R141S mutant (pI = 5.3) forms. Both variants displayed similar titration behaviors, starting at +20 net charge at pH 2, crossing the isoelectric point near pH 5, and declining to −15 at pH 12. The marginal 0.2 unit decrease in pI for the R141S mutant indicates a slight enhancement in overall protein acidity, though the nearly superimposable nature of the curves suggests conservation of the protein’s general charge characteristics despite the mutation.

Moreover, comparison of hydropathy, potential energy, and interaction energy between wild-type and mutant proteins reveals crucial biophysical perturbations that could affect protein conformational stability, folding dynamics, binding affinity, functional interactions, protein-ligand interactions, and solvation by quantifying changes in molecular forces and thermodynamic properties. Fig 10 displays the hydropathy, potential energy, and interaction energy plots for TP53, CDH1, and APC. Hydropathy profiles revealed distinct biophysical alterations across the three protein systems. The T125R mutant of TP53 maintained a near-identical hydropathic character to the wild-type, with localized fluctuations observed in amino acid residues 23, 26, and 29 where the hydropathy index of the wild-type was higher than the mutant T125R. In contrast, the hydropathy index of the wild type was lower than the mutant T125R in residue 32, suggesting a minimal deviation from the hydrophobic core organization. The CDH1 D254Y mutant exhibited less negative deviation than the wild type in residues 179, 180, and 211 (Green). The APC R141S mutant demonstrated minor deviation in residues 10 and 13. In the case of potential energy distributions that highlight thermodynamic instability in mutant variants, the T125R mutant of TP53 showed marginal deviations in residues 17–48, 138–143, 146, 180–194. D254Y mutant of CDH1 displayed a striking (−391.263 to +711.939 kcal/mol) spike near residue 178–183, exceeding wild-type values by 1103.192 kcal/mol, suggesting potential electrostatic repulsion in the extracellular cadherin repeat. Another potential deviation (+669.680 kcal/mol) was observed in mutant residue 211. R141S mutant of APC roughly deviated from wild type in residues 1–20 and 83–105. Subsequently, for molecular interaction energy patterns, TP53’s mutant T125R showed conserved interaction energies except at residues around 20–45, 135–145, and 177–211. In most cases, only the mutant T125R crosses the positive interaction energy boundaries, which could alter p53-MDM2 binding kinetics. On the other hand, the CDH1’s wild type manifested a −400 kcal/mol interaction energy peak at residue 179, and CDH1’s mutant D254Y manifested a + 300 kcal/mol interaction energy peak at residue 211. APC’s mutant R141S retained wild-type-like interaction energies between −300 to +100 kcal/mol except in the region (residues 2–19).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 10. Analysis and comparison of hydropathy, potential energy, and Interaction energy.

(A) TP53, (B) CDH1, and (C) APC. Both the potential energy and interaction energy shifts were observed in TP53 at amino acid region 20-44, and CDH1 at amino acid region 178-179.

https://doi.org/10.1371/journal.pone.0358440.g010

3.13. Atomic evaluation of proteins utilizing Unsupervised machine learning

Atomic-level evaluation and comparison of wild-type and missense mutant proteins through unsupervised machine learning may enable the detection of subtle intrinsic conformational changes, pattern recognition, and scalability. The atomic content of TP53, CDH1, and APC of both wild and mutant types has been utilized for an unsupervised machine learning algorithm (S10A-F Table). Native clusters of wild-type TP53 containing the atomic content of particular amino acids should be maintained to retain tumor suppressor activity (Table 4). In TP53, the first cluster group showed variations between wild type and mutant type (T125R) in cluster 1, indicating the most prominent atoms that were required to maintain the tumor suppressor activity of TP53. Till the analysis of cluster group 3, this scenario continued, which confirmed that T125R disrupted the atomic associations of particular amino acids responsible for tumor suppressor activity modulation. The structural and functional disrupting impact of missense SNPs T125R, D254Y, and R141S in TP53, CDH1, and APC provides hypothesis-generating outcomes from the disruption of Cluster 1, Cluster 3, and Cluster 4, respectively. For clusters of TP53 and APC, both contain 13 crucial amino acids in the tumor suppressor domain region that were extremely affected. On the other hand, the cluster of CDH1 contains 50 amino acids, the atomic content of which could be responsible for the distortion of the atomic content. The disruption of the atomic pattern was clearly visualized in wild-type TP53 and mutant-type TP53 (Figs 11A and 11B); however, the subtle variation was only observed in the case of CDH1 (Figs 11C and D) and APC (Figs 11E and F) after analysis of cluster groups 5 and 8, respectively.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Native clusters containing atomic content of particular amino acids are retained in the wild-type tumor suppressor protein, but these patterns are messed up in the mutant type, possibly disrupting structural and functional consequences. The amino acids in clusters 1, 3, and 4 for TP53, CDH1, and APC were retained in wild type but were messed up in mutant type (T125R of TP53), (D254Y of CDH1), and (R141S of APC), respectively.

https://doi.org/10.1371/journal.pone.0358440.t004

Download:

• PNG
  larger image
• TIFF
  original image

Fig 11. Cluster analysis using the unsupervised k-means clustering algorithm through dimensionality reduction.

(A), (C), (E) Wild-type versions of TP53, CDH1, and APC, respectively, and (B), (D), (F) Mutant versions of TP53 (T125R), CDH1 (D254Y), and APC (R141S), respectively. C1 to C8 denotes the number of clusters where the initial atomic distribution in the clusters was the same; however, atomic distribution changes throughout the increasing number of clusters due to mutation.

https://doi.org/10.1371/journal.pone.0358440.g011

3.14. Link among missense SNPs, protein expression, and patient survival

Evaluating protein expression and its association with patient survival in GC revealed that tumor suppressor genes TP53, CDH1, and APC were significantly upregulated in GC-developing tissues (Figs 12A, B, and C). Subsequently, survival analysis was assessed using the log-rank test, indicating a significant difference in patient outcomes associated with these genes, with FDR-adjusted p-values of 3.6e-08, 7e-05, 0.0081 for TP53, CDH1, and APC, respectively, with an FDR (False Discovery Rate) cutoff set to 1% for all. Moreover, the hazard ratios of TP53, CDH1, and APC were 1.68, 1.43, and 0.78, respectively (Figs 12D, E, and F). The survival was evaluated using the difference calculations between the low expression cohort (months) 57.63, 44.07, 24.03, and the high expression cohort (months) 24.5, 24.4, 57.8 for TP53, CDH1, and APC, respectively. Moreover, across TP53, APC, and CDH1, hazard ratios progressively increase with advancing cancer stage, indicating that although early stages show relatively protective effects, the prognostic risk consistently escalates in higher stages, reflecting stage-dependent worsening survival impact. The S11 Table presents significant covariate analysis of survival in different stages of cancer.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 12. Protein expression and patient survival differences.

Plots show significant expression of proteins (A) TP53, (B) CDH1, (C) APC, and significant survival differences in proteins (D) TP53, (E) CDH1, and (F) APC.

https://doi.org/10.1371/journal.pone.0358440.g012

3.15. Networks and cancer hallmarks of tumor suppressor proteins

We observed a strong correlation between TP53, APC, and CDH1 proteins and the GC pathway by analyzing protein-protein interactions using the STRING server. This analysis revealed CSNK1A1 as a common interacting protein between TP53 and APC, while CTNNA1 is linked to APC and CDH1. The identification of CDH1, APC, and CTNNA1 as a distinct cluster of proteins further suggested that APC and CDH1 may exert their tumor suppressor functions via distinct mechanisms compared to TP53, which we have already evaluated at the beginning of our study (Fig 4). Fig 13 represents the whole network and protein cluster. Therefore, this network and the identified protein cluster are highly susceptible to disruption by the identified missense SNPs. Further cancer hallmarks enrichment analysis indicated significant involvement of TP53, CDH1, and APC in multiple cancer hallmarks, including sustaining proliferative signaling, evading growth suppressors, genome instability, resisting cell death, tissue invasion and metastasis, and tumor-promoting inflammation (FDR-adjusted p-values <0.05). Only TP53 showed involvement in additional hallmarks such as evading immune destruction and sustained angiogenesis (Table 5 and S1 Fig).

Download:

• PNG
  larger image
• TIFF
  original image

Table 5. Proteins involved in Cancer Hallmarks.

https://doi.org/10.1371/journal.pone.0358440.t005

Download:

• PNG
  larger image
• TIFF
  original image

Fig 13. Protein-Protein interaction networks.

(A) Whole network with the first shell of interactors, and (B) the protein cluster that may function together. Red Box: Tumor Suppressor Proteins and Green Box: Interconnected Nodes.

https://doi.org/10.1371/journal.pone.0358440.g013

3.16. MicroRNAs and transcription factor-mediated gene regulatory networks

We elucidated the intricate gene regulatory networks involving tumor suppressor proteins TP53, APC, and CDH1, highlighting their potential interference in GC development. Our findings reveal that TP53, APC, and CDH1 interact with various transcription factors and microRNAs, suggesting a complex interplay that may regulate critical cellular processes such as apoptosis, cell cycle control, and epithelial integrity. EZH2 was a common TF among TP53, APC, and CDH1. On the other hand, HMGA1, DNMT1, HDAC1, STAT3, SIRT1, ESR1, NFKB1, RELA were common between TP53 and CDH1. Only YY1 was common between TP53 and APC. Additionally, only miRNA hsa-miR-129-5p was common between TP53, APC, and CDH1. The gene regulatory network has been presented in Fig 14.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 14. Gene regulatory networks of tumor suppressor proteins.

(A) Gene-TFs network and (B) Gene-miRNAs network.

https://doi.org/10.1371/journal.pone.0358440.g014

3.17. Altered gene expression of oncogenes

Identifying gene expression alterations in oncogenes is crucial for understanding the molecular mechanisms underlying gastric carcinogenesis and developing targeted therapeutic approaches. Our analysis revealed significant changes in the expression profiles of key oncogenes within the GC pathway following mutations in tumor suppressor genes TP53, APC, and CDH1. Specifically, the oncogenes TERT, ERBB2, KRAS, NRAS, CCNE1, CCNE2, MET, and FGFR2 exhibited abnormally elevated expression levels in tumor samples harboring these mutations compared to wild-type controls (Fig 15). The FDR-adjusted p-value < 0.05 was considered significant. In the case of TP53 mutation, the expression of CCNE1 (log2FC: 0.44) and TERT (log2FC: 0.49) was more significantly altered (upregulated). The mutation in APC altered the gene expression of CCNE1 (log2FC: 0.35), CCNE2 (log2FC: 0.22), ERBB2 (log2FC: 0.28), FGFR2 (log2FC: 0.16), and TERT (log2FC: 0.34). Finally, the mutation in CDH1 dysregulated (upregulated) the expression of FGFR2 (log2FC: 0.15).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 15. Changes in gene expression of oncogenes in the GC pathway (TERT, ERBB2, KRAS, NRAS, CCNE1, CCNE2, MET, and FGFR2) due to the mutational impact on TP53, APC, and CDH1.

(A) TP53, (B) APC, and (C) CDH1. GC falls under ESCA: Esophageal carcinoma, and STAD: Stomach Adenocarcinoma.

https://doi.org/10.1371/journal.pone.0358440.g015