Machine Learning-Based Pathomics Signature in Predicting MSH2 Expression and Prognosis in Gastric Cancer

Abstract

INTRODUCTION:

Gastric cancer (GC) is one of the most prevalent and lethal gastrointestinal malignancies. MutS homolog 2 (MSH2), a DNA mismatch repair protein, has emerged as a promising prognostic biomarker. However, traditional histopathological evaluation is limited by restricted fields compared with whole-slide imaging. This study aimed to investigate whether machine learning-derived digital pathomics features could predict MSH2 expression and clinical outcomes in GC.

METHODS:

Hematoxylin and eosin-stained whole-slide images from 234 patients were analyzed to extract quantitative pathological features. A pathomics score (PS) was developed to estimate MSH2 expression. The association between PS and overall survival (OS) was assessed using univariate and multivariate Cox regression. Survival differences between high-PS and low-PS groups were evaluated using Kaplan-Meier analysis. Functional enrichment and immune infiltration analyses were performed to explore potential biological mechanisms.

RESULTS:

Digital image analysis identified pathomics features associated with MSH2 expression. The PS served as a surrogate marker for MSH2 and effectively stratified patients into prognostic subgroups with significant different OS. High PS was associated with features suggestive of a stronger antitumor immune response, whereas low PS was linked to an immunosuppressive microenvironment.

DISCUSSION:

The machine learning-derived pathomics signature shows potential in predicting MSH2 expression. It can serve as a complementary research tool and provide clinically meaningful prognostic information for GC.

INTRODUCTION

Gastric cancer (GC) remains one of the most common and fatal malignancies of the gastrointestinal system (1). Gastric adenocarcinoma (GAC), which accounts for more than 95% of GC cases, continues to represent a major global health burden (1,2). Early-stage GAC (stage I) can be treated using endoscopic submucosal dissection or radical resection, whereas advanced GAC (stages II–III) typically requires multimodal therapy combining surgery with adjuvant treatment (3). However, traditional prognostic indicators are no longer sufficient in the era of precision medicine. Risk assessment based solely on the tumor node metastasis (TNM) staging system is often inadequate for guiding individualized therapeutic decision-making (4). Therefore, the identification of novel molecular biomarkers capable of refining prognostic stratification in GAC is urgently needed.

MutS homolog 2 (MSH2) is a key DNA mismatch repair (MMR) protein encoded by the tumor suppressor gene MSH2 on chromosome 2. During DNA replication, MSH2 forms heterodimers with MutS homolog 6 (MSH6) or MutS homolog 3 to detect base pair mismatches, while the MutL homolog 1 (MLH1)/PSM3 complex facilitates excision and repair (5,6). DNA damage is a fundamental driver of tumorigenesis (7), and loss of MSH2 expression or function results in microsatellite instability (MSI) and genomic instability—hallmark features observed across multiple cancer types (7–9). Notably, MSI-positive GC has been associated with more favorable survival outcomes compared with microsatellite-stable tumors (10,11). Furthermore, MSI-high tumors typically exhibit increased tumor mutational burden and upregulation of immune checkpoint pathways, rendering them more responsive to immunotherapy (12,13). Accordingly, MSH2 has gained attention as both a prognostic and predictive biomarker across various malignancies (14–16).

Current methods for finding MSH2 expression levels are limited by high costs, complex procedures, and challenges in sample collection. In routine practice, hematoxylin and eosin (H&E)-stained slides are widely used in determining TNM staging and histological subtypes. At present, the conventional methods to detect and quantify MSH2 expression are mainly divided into immunohistochemistry (for protein-level detection) and molecular genetic techniques (for gene-level or nucleic acid-level detection). Both methods are widely used in clinical and scientific research, but both have obvious shortcomings. It only reflects the amount of protein but not its function. Western blot is difficult to process a large number of samples at the same time and is not suitable for large-scale clinical detection because of its complicated operation process and high requirements for sample quality. Molecular genetic techniques such as polymerase chain reaction are prone to miss detection of rare mutation sites, with high false positive rate, poor stability, high probe cost, and relatively complex technical operation and time-consuming. The main drawback of next-generation sequencing is that it is expensive, which is not conducive to widespread clinical application. In summary, the novelty of this study is that relatively accurate results can be obtained to guide clinical practice without the need for expensive and complex machines, long detection times, and high labor costs. Machine learning-based pathomics signatures have recently gained considerable attention. Pathomics involves the extraction of high-dimensional, quantitative features from representative pathological images, enabling characterization of tumor morphology and its association with clinical outcomes (17,18). Based on this context, in this study, we hypothesized that digital pathomics signatures derived from H&E-stained images could serve as noninvasive surrogates for MSH2 expression in GC.

This study developed a gradient elevator (Gradient Boosting Machine [GBM])-based model as a supplementary research tool to the existing standard detection, integrating multiple pathological omics features and generating a pathological omics score that can predict the expression level of MSH2. The prognostic value of this score for overall survival (OS) was evaluated, and the potential mechanism was further explored through integrated bioinformatics analysis.

METHODS

Participants

We retrieved detailed information of a total of 443 patients with gastric adenocarcinoma (STAD) from the Cancer Genome Atlas (TCGA) database (https://tcga-data.nci.nih.gov/tcga). Patients with nonprimary or nonfirst diagnosed stomach adenocarcinoma (n = 3); patients with survival time less than 1 month or missing data (n = 49); patients with missing information on age, tumor stage, and grade (n = 42); and patients without ribonucleic acid (RNA)-seq (n = 54) were excluded. After excluding the cases with no pathological images or unqualified pathological images (n = 61), 295 cases met the inclusion criteria, and the pathological omics model was constructed (Figure 1). The inclusion criteria were as follows: (i) diagnosis of primary STAD, (ii) availability of RNA-sequencing data, (iii) complete clinical data, and (iv) an expected survival time ＞1 month. From this group, 234 patients with accessible histopathological slide images were further screened. These 234 participants were subsequently divided into training and validation groups in a 7:3 ratio.

Figure 1.

Flow diagram. TCGA, the Cancer Genome Atlas.

Histopathological samples and region of interest selection

Histopathological whole-slide images were obtained from the TCGA-STAD data set. Formalin-fixed, paraffin-embedding GC samples were sectioned at a thickness of 5 μm and stained with H&E. Images were scanned at $\times$20 or $\times$40 magnification. Tissue segmentation was performed using the Otsu's thresholding algorithm (https://opencv.org/) to remove background. Slides scanned at ×20 magnification were divided into 512 $\times$ 512-pixel subimages, whereas ×40 slides were split into 1,024 $\times$ 1,024-pixel subimages. All images underwent quality review by 3 independent pathologists. After quality control, 10 subimages were randomly selected from each slide for analysis. Subimage normalization and feature extraction were conducted using the PyRadiomics package (https://pyradiomics.readthedocs.io/en/latest/), yielding 93 original features and 465 wavelet-transformed features. For each participant, the mean values of each feature across the selected subimages were calculated and used as the final quantitative pathological feature set for downstream analysis.

Construction and evaluation of the GBM model

Feature selection was performed using the maximal relevance minimum redundancy (mRMR) algorithm followed by recursive feature elimination (RFE). Thirty features were identified using mRMR and further reduced to 10 through RFE. These features were used to construct a GBM model to predict MSH2 expression, generating both prediction probabilities and a corresponding PS for each participant.

Model performance was evaluated using receiver operating characteristic curves, with higher area under the curve values indicating better predictive accuracy. Calibration was assessed with calibration curves and the Hosmer-Lemeshow goodness-of-fit test, while the Brier score (range: 0–1) quantified the reliability of probabilistic predictions, with lower values reflecting higher accuracy. Decision curve analysis was conducted to assess the net clinical benefit of the GBM model.

Association between PS and prognosis

The “survminer” package (Version 0.4.9) was used to determine the optimal cutoff value of PS, enabling stratification of participants into high-PS and low-PS groups. Potential associations between PS with OS were assessed using univariate and multivariate Cox proportional hazards regression models. Survival differences between the high-PS and low-PS groups were evaluated using the Kaplan-Meier method.

Functional enrichment analysis

Gene set enrichment analysis was conducted to identify gene sets and pathways associated with PS. RNA-sequencing data were divided into high-PS and low-PS groups. Differentially expressed genes (DEGs) were identified using the criteria of |log₂ fold change| >2 and an adjusted P-value < 0.05. These DEGs were subjected to enrichment analysis using Kyoto encyclopedia of genes and genomes (KEGG) gene sets and Hallmark gene sets.

Genes associated with DNA damage and repair processes were visualized using a heatmap. Weighted gene coexpression network analysis (Version 1.73) was conducted to assess the relationship between gene modules and clinical outcomes. The module demonstrating the strongest correlation with prognosis was selected for hub gene identification. Gene ontology and KEGG pathway analyses were subsequently performed for the identified hub genes.

Immune infiltration analysis

Immune infiltration associated with PS was assessed using single-sample gene set enrichment analysis implemented through the gene set variation analysis R package (Version 1.52.3) (19). The relative abundance of 23 immune cell types was calculated for each sample. Differences in immune cell infiltration between the high-PS and low-PS groups were compared, with statistical significance thresholds set as follows: (*P < 0.05; **P < 0.01; ***P < 0.001).

High pathological score group and immune checkpoint genes

The immune checkpoint gene expression profile of the high/low PS group was further analyzed, and a subsection was added in the Results section of this study to enhance the explanation at the mechanistic level. The immune checkpoint gene set was referred to the existing literature (20), and the immune-related gene set of this study was extended appropriately. The expression of various immunomodulatory genes was detected. These include PDCD1LG2 (PD-L2), TLR4, ENTPD1 (CD39), SELP, CX3CL1, HAVCR2 (TIM-3), CD80, CD28, SLAMF7, and IL2RA (CD25) and a variety of human leukocyte antigen (HLA) class II molecules (e.g., HLA-DQA1, HLA-DQB1, HLA-DQB2, and human leukocyte antigen-DR alpha). These genes are closely related to processes such as antigen presentation, T-cell costimulation, dendritic cell activation, and inflammatory response.

Statistical analyses

Continuous variables were compared using 2-tailed Student t-tests, while categorical variables were analyzed using the Pearson χ² test or Fisher exact test. Univariate and multivariate Cox proportional hazards regression models were used to estimate hazard ratios (HRs) and corresponding 95% CIs. Kaplan-Meier survival analysis and the log-rank test were used to evaluate differences in survival distributions between groups. All statistical analyses were conducted using R software (version 3.6.3). A 2-sided P-value < 0.05 was considered statistically significant.

RESULTS

MSH2 expression in GC and normal tissues

Comparison of MSH2 mRNA expression between tumor and normal tissue samples indicated a significant upregulation in GC tissues (P < 0.001). The median expression difference was 1.552 with an interquartile range (IQR) of 1.414–1.627 (Figure 2a).

Figure 2.

Prognostic value of MSH2 in GC. (a) MSH2 expression between normal tissues and pan-cancerous samples. (b) Kaplan-Meier analysis of GC patients with different MSH2 expression levels. (c) Forest map of univariate Cox regression analysis. (d) Forest map of multivariate Cox regression analysis. (e) Univariate Cox regression analysis in different subgroups.

Association of MSH2 upregulation with improved OS in GC

Based on the inclusion criteria, 295 patients with GC from the TCGA-STAD data set were included in the analysis. Participants were divided into low-expression (n = 193) and high-expression (n = 102) groups using a cutoff value of 1.9452. The cutoff value of MSH2 was automatically calculated by the R package (survminer). This function is based on the maximum rank statistic, which is the optimal grouping by selecting the cut point from all possible thresholds that yields the most significant survival difference (minimum P value) (21–23). The cutoff value of MSH2 in this study was 1.9452. Baseline clinical characteristics are summarized in Table 1. No significant differences were observed between the 2 groups about clinicopathological features (P > 0.05).

Table 1.

Characteristics of patients with GC in the low-MSH2 and high-MSH2 groups

Variables	Total (N = 295)	Low-MSH2 (N = 193)		High-MSH2 (N = 102)		P value
		N	%	N	%
Age						0.308
≤65	132	91	47.15	41	40.20
>65	163	102	52.85	61	59.80
Sex						0.906
Female	110	71	36.79	39	38.24
Male	185	122	63.21	63	61.76
Pathologic stage						0.84
I–II	135	87	45.08	48	47.06
III–IV	160	106	54.92	54	52.94
Histologic type						0.502
NOS	96	59	30.57	37	36.27
Others	61	43	22.28	18	17.65
Intestinal type	138	91	47.15	47	46.08
Histologic grade						0.837
G1-G2	109	70	36.27	39	38.24
G3	186	123	63.73	63	61.76
Residual tumor						0.242
R0	250	160	82.90	90	88.24
R1/R2	23	15	7.77	8	7.84
RX/unknown	22	18	9.33	4	3.92
Tumor location						0.261
Distal	113	73	37.82	40	39.22
Proximal	42	33	17.10	9	8.82
Body	109	67	34.72	42	41.18
Gastroesophageal junction	31	20	10.36	11	10.78
Radiotherapy						0.276
No	256	171	88.60	85	83.33
Yes	39	22	11.40	17	16.67
Chemotherapy						1
No	158	103	53.37	55	53.92
Yes	137	90	46.63	47	46.08

MSH2, MutS homolog 2; NOS, not otherwise specific; Others, signet-ring-cell type and diffuse type.

Kaplan-Meier analysis demonstrated that upregulation of MSH2 was significantly associated with improved OS (P = 0.011) (Figure 2b). The median OS in the MSH2-high group was significantly longer compared with the MSH2-low group (56.2 vs 26.4 months).

Univariate and multivariate Cox proportional hazards regression analyses were conducted to further identify prognostic risk factors for OS. MSH2 expression, age, pathological stage, radiotherapy, and residual tumor status were significantly associated with OS (P < 0.05) (Figure 2c). Multivariate analysis identified high MSH2 expression as an independent protective factor (HR = 0.558, 95% CI: 0.395–0.907, P = 0.015) and receipt of radiotherapy (HR = 0.415, 95% CI: 0.195–0.886, P = 0.023). By contrast, age older than 65 years (HR = 1.725, 95% CI: 1.161–2.564, P = 0.007), stage III–IV disease (HR = 1.79, 95% CI: 1.173–2.731, P = 0.007), residual tumor classified as R1/R2 (HR = 2.622, 95% CI: 1.455–4.726, P = 0.001), and residual tumor classified as RX (HR = 3.507, 95% CI: 1.865–6.595, P < 0.001) were identified as risk factors for OS in patients with GC (Figure 2d).

Subgroup analyses were subsequently performed to assess the prognostic value of MSH2 expression across clinical and pathological variables, including age, sex, pathological stage, histological type, histological grade, residual tumor status, tumor location, radiotherapy, and chemotherapy. No significant interactions were observed (P > 0.05), indicating a consistent protective effect of high MSH2 expression on OS across all subgroups. In each subgroup, high MSH2 expression was associated with improved OS (HR < 1) (Figure 2e).

Development and evaluation of a GBM model for predicting MSH2 expression

A total of 234 eligible participants with complete clinical data, RNA sequencing data, and pathological images were included in the analysis. Participants were randomly divided into a training group (n = 165) and a validation group (n = 69) at a 7:3 ratio. Detailed clinicopathological characteristics are presented in Table 2. Among the 234 cases, 62% were male and 38% were female. Upregulation of MSH2 was observed in more than half of the patients (62%, 145/234).

Table 2.

Characteristics of patients with GC in the training and validation cohorts

Variables	Total (N = 234)	Training cohort (N = 165)		Validation cohort (N = 69)		P value
		N	%	N	%
Median OS (mo)	17.33 (11.55, 27.25)	18.73 (12.47, 28.53)		13.2 (9.3, 22.3)		0.014
MSH2						1
Low	145	102	61.82	43	62.32
High	89	63	38.18	26	37.68
Age						0.009
≤65	100	80	48.48	20	28.99
>65	134	85	51.52	49	71.01
Sex						0.646
Female	88	60	36.36	28	40.58
Male	146	105	63.64	41	59.42
Pathologic stage						1
I–II	102	72	43.64	30	43.48
III–IV	132	93	56.36	39	56.52
Histologic type						0.482
NOS	69	51	30.91	18	26.09
Others	47	35	21.21	12	17.39
Intestinal type	118	79	47.88	39	56.52
Histologic grade						0.584
G1-G2	87	59	35.76	28	40.58
G3	147	106	64.24	41	59.42
Residual tumor						0.085
R0	197	142	86.06	55	79.71
R1/R2	20	15	9.09	5	7.25
RX/unknown	17	8	4.85	9	13.04
Tumor location						0.102
Distal	99	72	43.64	27	39.13
Proximal	29	24	14.55	5	7.25
Body	88	60	36.36	28	40.58
Gastroesophageal junction	18	9	5.45	9	13.04
Radiotherapy						1
No	197	139	84.24	58	84.06
Yes	37	26	15.76	11	15.94
Chemotherapy						0.399
No	124	84	50.91	40	57.97
Yes	110	81	49.09	29	42.03

GC, gastric cancer; MSH2, MutS homolog 2; NOS, not otherwise specific; OS, overall survival; Others, signet-ring-cell type and diffuse type.

Most of the participants (84%) had R0 residual tumor status. No significant differences in baseline clinical characteristics were observed between the 2 groups (P > 0.05), except for age distribution. The median OS was 18.73 months (95% CI: 12.47–28.53) in the training group and 13.2 months (95% CI: 9.3–22.3) in the validation group.

The framework for constructing the pathomics signature is depicted in Figures 3a–3c. In the training group, a 10-feature pathomics signature was generated using the mRMR and RFE (Figure 3d). The GBM algorithm was subsequently applied to model the selected pathological features for predicting MSH2 expression. The importance of each feature in the GBM model is shown in Figure 3e.

Figure 3.

The workflow of histopathology image processing. HH, high-high; HL, high-low; LH, Low-High; LL, low-Low; MCC, maximal correlation coefficient; WSI, whole slide image.

Model performance was evaluated in the internal validation group. The area under the curve values were 0.821 (95% CI: 0.755–0.887) in the training group and 0.724 (95% CI: 0.604–0.844) in the validation group, reflecting favorable predictive capability (Figures 4a and 4b). Calibration curves and the Hosmer-Lemeshow goodness-of-fit test demonstrated strong concordance between predicted and observed values (Figures 4c and 4d). The Brier scores for the training and validation groups were 0.713 and 0.205, respectively, confirming predictive accuracy. Decision curve analysis further indicated that the GBM model provided net clinical benefit (Figures 4e and 4f).

Figure 4.

Performance evaluation of the GBM model. (a) and (b) ROC curves in the training cohort and the validation cohort. (c) and (d) Calibration curves of the training and validation cohorts. (e) and (f). DCA curves of the training and validation cohorts. AUC, area under the curve; DCA, decision curve analysis; FPR, false positive rate; HL, high-low; ROC, receiver operating characteristics; TPR, true positive rate.

Association between PS and MSH2 expression

The GBM model directly outputted a PS to estimate the probability of MSH2 expression. In the training group, the distribution of PS differed significantly between the low-MSH2 and high-MSH2 groups (P < 0.001) (Figure 5a), with higher PS values observed in the high-MSH2 group. A similar association was confirmed in the validation group (P < 0.01) (Figure 5b). Overall, PS demonstrated a consistent positive correlation with MSH2 expression levels.

Figure 5.

The differences of pathologics score between different levels of MSH2 (a) in the training cohort and (b) in the validation cohort. MSH2, MutS homolog 2.

Association between PS and prognosis

An optimal cutoff value of 0.4833 was applied to stratify participants into high-PS (n = 71, 30.3%) and low-PS (n = 163, 69.7%) groups. No significant differences (P > 0.05) in clinical variables were observed between the 2 groups, except for age and histological type (Table 3). The distribution of PS across survival time is presented in Figure 6a, demonstrating that higher PS values were associated with reduced mortality risk. The 6-year OS rate in the high-PS group was significantly higher compared with the low-PS group (log-rank P = 0.012). The median OS was also longer in the high-PS group compared with the low-PS group (56.2 vs 26.07 months). These findings indicate that higher PS was associated with improved OS and prognosis in patients with GC.

Table 3.

Characteristics of patients with GC in the low-PS and high-PS groups

Variables	Total (N = 234)	Low-MSH2 (N = 163)		High-MSH2 (N = 71)		P value
		N	%	N	%
Age						0.011
≤65	100	79	48.47%	21	29.58%
>65	134	84	51.53%	50	70.42%
Sex						0.597
Female	88	59	36.20%	29	40.85%
Male	146	104	63.80%	42	59.15%
Pathologic stage						0.309
I–II	102	67	41.10%	35	49.30%
III–IV	132	96	58.90%	36	50.70%
Histologic type						0.005
NOS	69	58	35.58%	11	15.49%
Others	47	27	16.56%	20	28.17%
Intestinal type	118	78	47.85%	40	56.34%
Histologic grade						0.792
G1-G2	87	62	38.04%	25	35.21%
G3	147	101	61.96%	46	64.79%
Residual tumor						0.617
R0	197	139	85.28%	58	81.69%
R1/R2	20	12	7.36%	8	11.27%
RX/unknown	17	12	7.36%	5	7.04%
Tumor location						0.185
Distal	99	62	38.04%	37	52.11%
Proximal	29	22	13.50%	7	9.86%
Body	88	64	39.26%	24	33.80%
Gastroesophageal junction	18	15	9.20%	3	4.23%
Radiotherapy						0.62
No	197	139	85.28%	58	81.69%
Yes	37	24	14.72%	13	18.31%
Chemotherapy						0.803
No	124	85	52.15%	39	54.93%
Yes	110	78	47.85%	32	45.07%

GC, gastric cancer; MSH2, MutS homolog 2; NOS, not otherwise specific; Others, signet-ring cell type and diffuse type; PS, pathologics score.

Figure 6.

Association of the pathologics score with prognosis in GC. (a) Kaplan-Meier analysis of GC patients with high-PS and low-PS. (b) Forest map of univariate Cox regression analysis. (c) Forest map of multivariate Cox regression analysis. (d) Univariate Cox regression analysis in different subgroups. GC, gastric cancer; NOS, not otherwise specific. PS, pathomics score.

In univariate Cox regression analysis, PS, age, pathological stage, radiotherapy, and residual tumor status were significantly associated with OS (P < 0.05) (Figure 6b). Multivariate analysis further confirmed that high PS (HR = 0.434, 95% CI: 0.254–0.744, P = 0.002) and receipt of radiotherapy (HR = 0.343, 95% CI: 0.151–0.78, P = 0.011) were independent protective predictors of OS. By contrast, age older than 65 years (HR = 1.957, 95% CI: 1.238–3.094, P = 0.004), stage III–IV disease (HR = 1.713, 95% CI: 1.067–2.752, P = 0.026), residual tumor classified as R1/R2 (HR = 2.817, 95% CI: 1.498–5.296, P = 0.001), and RX (HR = 3.716, 95% CI: 1.805–7.649, P < 0.001) were identified as risk factors for OS in GC (Figure 6c).

No significant interaction effects were found between PS and other clinical variables. In the subgroup with RX residual tumor, high PS seemed to represent a risk factor for OS, although the association did not reach statistical significance (P = 0.34). All other subgroup analyses consistently supported high PS as a protective factor for OS (Figure 6d).

Functional enrichment analyses

DEGs were identified between the high-PS and low-PS groups and analyzed through KEGG and Hallmark gene set enrichment analyses. The top 50 significantly enriched pathways are presented in Figure 7. In the KEGG analysis, several metabolic pathways including the pentose phosphate pathway, unsaturated fatty acid biosynthesis, glycolysis, gluconeogenesis, one-carbon cycle, and citrate cycle as well as signaling cascade pathways such as vascular endothelial growth factor, wingless/integrated, insulin, P53, TGF-β, and mTOR, were significantly enriched in the low-PS group. In the Hallmark gene sets, the low-PS group demonstrated marked upregulation of DNA damage-related and repair-related pathways.

Figure 7.

Functional enrichment analyses. (a) Kyoto encyclopedia of genes and genomes pathway gene set and (b) Hallmark gene set. GSVA, gene set variation analysis; PS, pathomics score.

Association of PS with DNA damage and repair

DEGs related to DNA damage and repair were compared between the high-PS and low-PS groups. Genes with statistically significant differences (P < 0.001) were visualized in a heatmap (Figure 8a). Expressions of DCLRE1C, poly(ADP-ribose) glycohydrolase, and aprataxin were significantly upregulated in the high-PS group.

Figure 8.

Association of PS with DNA damage and repair. (a) Heatmap of differentially expressed genes between high-PS and low-PS groups. (b) and (c) The determination of the soft threshold in weighted gene coexpression network analysis. (d) Diagram of module-trait relationship. (e) Gene ontology analysis of 118 hub genes in magenta module. (f) KEGG analysis of 118 hub genes in magenta module. ATP, adenosine triphosphate; BP, biological process; CC, cellular component; KEGG, Kyoto encyclopedia of genes and genomes; MF, molecular function.

Weighted gene coexpression network analysis was conducted to identify DNA damage-related and repair-related gene sets associated with prognosis. A soft threshold of 4 was applied to establish a scale-free network (Figures 8b and 8c). Six modules were identified based on average hierarchical clustering and the applied soft threshold, with 219 genes assigned across these modules. Module-trait relationship analysis (Figure 8d) indicated that the magenta module demonstrated a significant correlation with prognosis. Consequently, 118 hub genes within this module were selected for further investigation.

Gene ontology enrichment analysis (Figure 8e) indicated that hub genes in the biological process category were mainly associated with DNA catalytic activity and damaged DNA binding. Genes in the cellular component category were predominantly enriched in the replication fork and chromosome telomeric region, while those in the molecular function category were related to double-strand break binding and DNA recombination. KEGG pathway analysis further identified the top 5 significantly enriched pathways: the Fanconi anemia pathway, homologous recombination, base excision repair, nucleotide excision repair, and DNA replication (Figure 8f).

Association between PS and immune infiltration

Tumor immune infiltration is a critical determinant of immunotherapy response and long-term prognosis. The proportions of 23 immune cell subtypes in both high-PS and low-PS groups are presented in Figure 9. In the high-PS group, significantly higher infiltration was observed for CD8⁺ T cells (P < 0.001), exhausted T cells (P < 0.001), γδ T cells (P < 0.01), and cytotoxic T cells (P < 0.01). Conversely, Th17 cells (P < 0.001) and dendritic cells (P < 0.01) were significantly increased in the low-PS group.

Figure 9.

Immune infiltration of 23 immune cells in high-PS and low-PS groups. DC, dendritic cell; MAIT, mucosal-associated invariant T cell; NK, natural killer cell; NKT, natural killer T cell; PS, pathomics score.

The high PS group showed more active immune response potential

We further analyzed the expression of immune checkpoint genes in the high/low PS group and added a subsection in the Results section of this study to enhance the explanation at the mechanistic level. The immune checkpoint gene set was referred to the existing literature (20), and the immune-related gene set of this study was extended appropriately. The results showed that a variety of immunomodulatory genes were significantly upregulated in the high-PS group. These include PDCD1LG2 (PD-L2), TLR4, ENTPD1 (CD39), SELP, CX3CL1, HAVCR2 (TIM-3), CD80, CD28, SLAMF7, and IL2RA (CD25) and a variety of HLA class II molecules (e.g., HLA-DQA1, HLA-DQB1, HLA-DQB2, and human leukocyte antigen-DR alpha). These genes are closely related to the process of antigen presentation, T-cell costimulation, dendritic cell activation, and inflammatory response show the expression characteristics of immune activation. This feature is consistent with the trend observed in the immune infiltration analysis, that is, the high-PS group has more active immune response potential. The above supplementary analysis further strengthens the explanation of PS in immunobiological mechanism and makes the research conclusion more robust.

DISCUSSION

Accurate prognosis prediction is essential for risk stratification and treatment planning in patients with GC in clinical practice. Advances in artificial intelligence have substantially influenced traditional pathology-based prognostic approaches (24,25). The detection of MMR protein by immunohistochemistry has been a routine and widely used method. Mismatch repair gene is the core to maintain genomic stability, and its dysfunction will drive tumorigenesis. It is also a key marker for tumor classification and genetic screening. Unfortunately, the current MSH2 detection is usually expensive and based on local tumor tissue, which is time-consuming and costly. It not only cannot represent the overall situation of the tumor but also is difficult to achieve dynamic observation. This study aims to construct a pathological omics prediction model for MSH2 as a supplementary research tool to the existing standard tests. The objective, batch, and accurate prediction of MSH2 by pathological omics can more quickly screen the population with potential benefits for precision treatment, provide a basis for guiding personalized precision treatment, and save medical costs. However, studies applying computational pathology to identify prognostic markers in GC remain limited. In this study, quantitative analysis of digital pathological images extracted from H&E-stained slides was performed to estimate MSH2 expression. A PS was developed and applied as an alternative marker of MSH2 for prognostic evaluation. The PS successfully stratified participants into high-PS and low-PS groups with significant differences in OS. In addition, high PS was associated with features of an enhanced antitumor immune response, whereas low PS was linked to an immunosuppressive tumor microenvironment.

Although the TNM staging system remains the cornerstone for prognostic evaluation in GC (4,26), substantial survival variability exists among patients within the same stage. Molecular classifications based on next-generation sequencing can provide additional prognostic insights, aiding risk stratification and clinical decision-making (10,27). However, high costs and complex analytical procedures limit their widespread application. Tumor stromal immune components, including cytotoxic CD8⁺ T cells and tumor-associated macrophages, have been reported as prognostic indicators (28,29). Radiomics approaches have been explored for survival prediction (25,30), but the indirect relationship between imaging features and tumor microenvironment may constrain predictive accuracy. Thus, there is an ongoing need for reliable, readily assessable molecular biomarkers for prognostic evaluation in GC.

The MMR system is essential for maintaining genomic stability by preventing errors during DNA replication, genetic recombination, and damage repair (31). This highly conserved system relies on 4 essential genes—MLH1, post meiotic segregation increased 2 (PMS2), MSH2, and MSH6—that form heterodimeric complexes to detect and repair mismatches. Specifically, MLH1 and MSH2 are critical mediators of the MMR pathway through dimerization with PMS2 and MSH6, respectively. Loss of expression or functional alterations in MSH2 compromise DNA repair capacity and have been implicated in the development of several malignancies, including hepatocellular carcinoma (32), colorectal cancer (33), prostate cancer (34), breast cancer (35), and non-small cell lung cancer (36). The reason why MSH2 was chosen as a detection or research target, rather than combined MMR and MSI status, is that this study aims to find a more convenient detection method to guide patients' medication and prognosis assessment, and to complement traditional methods, thereby achieving a faster and more economical treatment plan. Therefore, the use of a single-key molecule, that is, MSH2, can achieve more efficient results. Moreover, MSH2 is the core key protein of the MMR pathway, and its mutation/deletion is one of the main causes of MMR dysfunction. Taking MSH2 as the research object can focus more on its specific function and mechanism of action. By contrast, composite MMR and MSI status involves multiple genes, and when studying disease association, it is difficult to determine which specific gene changes are responsible.

Deficiency in MMR genes often leads to MSI with mismatch repair deficiency and MSI-high status demonstrating a high degree of concordance (90%–95%) across various tumor types (37). The relationship between MSH2 deficiency and MSI status has been well documented (38). Although MSI status is commonly assessed through immunohistochemical analysis of MMR proteins (MSH6, MLH1, MSH2, and PMS2), it remains uncertain if MSH2 alone can serve as a surrogate marker for MSI (39). Previous studies have indicated that positive MSH2 expression in nodal metastases indicates less aggressive GC and favorable survival outcomes (40). Consistent with these reports, both univariate and multivariate Cox regression analyses in this study identified high MSH2 expression as an independent protective factor for OS. Collectively, these findings support the prognostic value of MSH2 upregulation and highlight its potential as a clinically relevant biomarker in GC.

Advances in histological feature recognition technologies have highlighted the potential of pathomics in cancer assessment and management (25,41,42). Conventional pathological examination performed by experienced pathologists at various magnifications remains a standard approach for evaluating tumor cell characteristics. However, traditional optical microscopy offers a limited field of view compared with whole-slide imaging, which enables comprehensive visualization of cellular morphology and tissue architecture (42). Pathomics has the potential to complement conventional pathological evaluation as an efficient and scalable tool (43). In this study, a machine learning algorithm was applied to identify optimal digital pathological features, and a GBM model was constructed to predict MSH2 expression levels. The model was subsequently validated in an independent group, demonstrating favorable discriminative and predictive capabilities.

A PS derived from the GBM model incorporating 10 pathomics features from H&E-stained slides was proposed as a surrogate measure of MSH2 expression. The PS demonstrated significant ability to distinguish patients with high and low MSH2 expression and enabled accurate estimation of MSH2 expression in individual cases. Moreover, the PS was associated with prognosis, with higher PS values corresponding to longer median OS. Nevertheless, whether the predictive values of PS reflect intrinsic tumor characteristics or extrinsic influences from the tumor microenvironment remains unclear. Integrative analyses combining pathomics with genomic and transcriptomic data have been proposed as feasible approaches to elucidate the underlying mechanisms linking PS with survival outcomes (44,45).

Functional enrichment analysis indicated that DEGs between high-PS and low-PS groups were primarily enriched in metabolic pathways, including the pentose phosphate pathway. These findings underscore the potential value of incorporating comprehensive multiomics approaches, including metabonomics, pathomics, and genomics, into future investigations. Immune infiltration analysis indicated that the high-PS group exhibited relatively greater infiltration of effector T cells, whereas the low-PS group showed enrichment of immunosuppressive cell types, such as Th17 cells. These results suggest that enhanced immune activity in the high-PS group may contribute to improved prognosis. Further research is required to determine and raise the possibility that PS could serve as a predictor of immunotherapy response.

High MSH2, a feature that should lead to worse prognosis, showed better survival in our data. The reason might be the response to commonly used chemotherapy regimens (e.g., fluorouracil plus platinum) in GAC is related to MMR status. Tumors with high expression of MSH2 (complete MMR) have low tumor mutational burden but intact DNA repair function. DNA damage induced by chemotherapy drugs can be “overloaded” and more likely to cause tumor cell apoptosis. If the patients in the high-MSH2 group included in the study had higher chemotherapy compliance, adequate dose, or no drug resistance cases, a survival advantage may occur. The high-MSH2 group was more likely to receive radical surgery (such as D2 lymph node dissection) or a higher proportion of postoperative adjuvant therapy, while the low-MSH2 group had a higher proportion of advanced cases and received only palliative treatment. The survival difference was not caused by MSH2 itself. At the same time, in addition to mismatch repair, MSH2 may be involved in other pathways such as regulating apoptosis signals of gastric mucosal cells and inhibiting angiogenesis in GAC. Its high expression may inhibit tumor progression through these MMR-independent pathways, rather than simply reflecting MMR function. Finally, the prognosis of GAC is significantly affected by the tumor microenvironment. High-MSH2 group may be accompanied by more CD8+ T-cell infiltration, less M2 macrophages, or differences in the composition of intestinal flora, which offset the immunogenicity disadvantage caused by low tumor mutational burden and improve survival. These may be the reasons why “high MSH2” in this study showed better survival in the data of this study.

We further analyzed the expression of immune checkpoint genes in the high/low PS group and added a subsection in the Results section of the manuscript to enhance the explanation at the mechanistic level. The immune-related gene set was expanded appropriately in combination with this study. The results showed that a variety of immunomodulatory genes were significantly upregulated in the high-PS group. These include PDCD1LG2 (PD-L2), TLR4, ENTPD1 (CD39), SELP, CX3CL1, HAVCR2 (TIM-3), CD80, CD28, SLAMF7, IL2RA (CD25), and a variety of HLA class II molecules (e.g., HLA-DQA1, HLA-DQB1, HLA-DQB2, and human leukocyte antigen-DR alpha). These genes are closely related to the process of antigen presentation, T-cell costimulation, dendritic cell activation, and inflammatory response, and show the expression characteristics of immune activation. This feature is consistent with the trend observed in the immune infiltration analysis, that is, the high-PS group has more active immune response potential. The above supplementary analysis further strengthens the explanation of PS in the immunobiological mechanism and makes the research conclusion more robust.

It is noteworthy that the Kaplan-Meier survival curves of high and low MSH2 expression groups converged at about 72 months of follow-up, which may be due to changes in sample characteristics during follow-up, changes in MSH2-related tumor biology, and differential effects of therapeutic interventions. (i) The sample error is enlarged in the late follow-up period, which affects the validity of the curve. The tail stability of the Kaplan-Meier curve depends on the sufficient sample size, and 72 months is a long-term follow-up, at which time the number of remaining follow-up cases in the high-MSH2 and low-MSH2 expression groups may be greatly reduced. (ii) The biological characteristics of tumors change dynamically with the course of the disease. As a mismatch repair gene, the effect of MSH2 expression on tumors is not unchanged but may change at the functional or regulatory level with the progression of the disease. (iii) The effects of different therapeutic interventions tend to be balanced in long-term follow-up. Tumor patients often receive adjuvant therapies such as chemotherapy and targeted therapy after surgery, and the differences in the effects of these treatments on different MSH2 expression groups may gradually disappear over time. (iv) The tumor microenvironment and the long-term changes in the immune status of the body and the difference in survival, at the same time, the inflammatory factors, stromal cells in the tumor microenvironment may also be reconstructed over time and finally promote the survival curve convergence.

Several limitations should be acknowledged. First, further validation of these findings in prospective, multicenter cohorts is required. In the future, we will add validation cohorts for preoperative or postoperative treatments, such as neoadjuvant chemotherapy, immune checkpoint inhibitor, or chemotherapy, among other conditions. Second, the biological mechanisms underlying the association between PS and prognosis require additional exploration using multiomics analyses. Third, variability in H&E staining protocols across institutions may have introduced bias into the extracted features and should be considered when interpreting the results.

This study has several limitations that should be considered when interpreting the results. First, there was a significant difference in the median survival time between the training set and the validation set, which was mainly due to the imbalance of baseline characteristics between the 2 groups. Specifically, the proportion of early-stage tumor cases in the training set was higher, and the proportion of patients who received radical surgery combined with standard adjuvant chemotherapy was higher. Although known baseline confounding variables were adjusted by propensity score matching and multivariate Cox regression models, and the stability of the results was verified by sensitivity analyses, there are still potential unmeasured confounders (e.g., patient adherence to treatment, severity of comorbidities, and timing of genetic testing). These factors may have an indirect effect on the association of MSH2 expression with prognosis, reducing the internal validity of the study.

Second, significant differences in baseline survival characteristics may weaken the external generalization of the findings. The survival baseline of the training data set is better, and the prognostic model based on this data set may overfit the characteristics of the specific population. When facing a more heterogeneous patient population in the real clinical scenario, the predictive performance of the model may be reduced, and it is difficult to accurately generalize to patients with GAC from different centers, different stages, or treatment regimens. In addition, this difference may also increase the risk of nonreproducibility of the results, and it may be difficult for other research teams to replicate the conclusion that high expression of MSH2 is associated with better prognosis in this study if they perform validation based on populations with different baseline survival characteristics.

Finally, the sample size of this study was relatively limited, especially the sample size of some subgroups was further reduced after propensity score matching , which may lead to insufficient statistical power to capture the potential subgroup effects of the association between MSH2 expression and prognosis (such as differences under different Lauren classifications and chemotherapy regimens). In the future, multicenter, large-sample prospective cohort studies are needed to strictly control the inclusion criteria to ensure that the baseline characteristics of the training set and validation set are balanced and consistent. At the same time, functional validation experiments are combined to further clarify the prognostic value and potential molecular mechanism of MSH2 in GAC, so as to provide more reliable evidence support for clinical transformation and application.

The machine learning-derived pathomics signature demonstrated potential as a novel supplementary research tool for estimating MSH2 expression levels in GC. As a supplementary marker of MSH2, the PS was significantly associated with OS and effectively stratified patients into prognostic subgroups. Furthermore, PS was linked to distinct patterns of tumor immune infiltration, reflecting differential antitumor immune responses. Collectively, these findings suggest that PS may provide valuable insights for prognostic evaluation and could also serve as a supplementary research tool to guide the reference basis for clinical decision-making of GC.

CONFLICTS OF INTEREST

Guarantor of the article: Xiao-Ming Wang, MD and Chao Zhang, MD.

Specific author contributions: Z.-R.Z.: funding acquisition, writing—original draft. Y.W.: data curation, formal analysis. W.-W.Y.: data curation, writing—original draft. H.-R.L.: conceptualization, data curation. Z.-W.C.: software, writing—review and editing. T.H.: conceptualization, software. C.Z.: conceptualization, writing—review & editing. X.-M.W.: conceptualization, funding acquisition, writing—review & editing. All authors read and approved the final draft.

Financial support: This study was funded by the Shanghai Pujiang Program (Grant No.: 23PJ140900), Scientific Research Project (Natural Science) of Anhui Province (Xiao-Ming Wang, Grant No.: 2023AH040254), Key Research Project (Natural Science) of Wannan Medical College (Zheng-Rong Zhang, Grant No.: WK2023ZZD19), Anhui Province University Teaching Quality and Teaching Reform Project General Project (Zheng-Rong Zhang, Grant No.: 2023jyxm1224), and Teaching Quality and Teaching Reform Project of Wannan Medical College General Project (Zheng-Rong Zhang, Grant No.: 2022kcszsfc14). The funding bodies had no role in the design of the study and collection, analysis, and interpretation of data and in writing the manuscript.

Potential competing interests: None to report.

Ethics approval and consent to participate: This study was conducted with approval from the Ethics Committee of The First Affiliated Hospital of Wannan Medical College (Approval Number: 2025-219). This study was conducted in accordance with the declaration of Helsinki. Informed consent was waived because of the retrospective nature of the study, which involved the analysis of existing clinical data with no direct patient contact or intervention.

Availability of data and materials: The data sets of this article were generated from the TCGA database.

Study Highlights.

WHAT IS KNOWN

• ✓ Gastric cancer (GC) is one of the most common and deadly gastrointestinal malignancies.

• ✓ MutS homolog 2 (MSH2), a key DNA mismatch repair protein, acts as a valuable prognostic biomarker in GC.

• ✓ Traditional histopathological assessment is limited by small viewing fields, unlike whole-slide imaging.

WHAT IS NEW HERE

• ✓ A constructed pathomics score (PS) effectively serves as a surrogate marker for MSH2 and stratifies patients into distinct prognostic groups based on overall survival.

• ✓ High pathomics score correlates with enhanced antitumor immune activity, while low score is associated with an immunosuppressive tumor microenvironment.

• ✓ This machine learning-derived pathomics signature provides a complementary tool and clinically useful prognostic information for gastric cancer.

ABBREVIATIONS:

AUC

area under the curve

GAC

gastric adenocarcinoma

GBM

gradient boosting machine

GC

gastric cancer

GO

Gene Ontology

H&E

hematoxylin and eosin

KEGG

Kyoto Encyclopedia of Genes and Genomes

MMR

mismatch repair

mRMR

maximal relevance minimum redundancy

MSH2

MutS homolog 2

MSI

microsatellite instability

OS

overall survival

PS

pathomics score

RFE

recursive feature elimination

ROC

receiver operating characteristic

STAD

stomach adenocarcinoma

TCGA

The Cancer Genome Atlas

TNM

tumor node metastasis

WSI

whole-slide imaging