The Prognostic and Biological Value of PGF‐Based H&E Pathomics in Hepatocellular Carcinoma

ABSTRACT

Purpose

Placental growth factor (PGF) is associated with the progression of hepatocellular carcinoma (HCC), but current research on this relationship remains limited. This study aims to establish a pathomics model for predicting PGF expression levels in H&E‐stained HCC sections, and to explore its prognostic relevance and underlying molecular mechanisms.

Methods

Retrospective analysis utilised H&E images and clinical data from TCGA and an external cohort. Prognostic significance of PGF was assessed via survival analysis. Image segmentation employed the OTSU algorithm, followed by PyRadiomics‐based feature extraction. Key features were selected using mRMR and RFE algorithms, with a gradient boosting machine (GBM) model constructed for PGF prediction. Model performance was validated through ROC and Precision‐Recall (PR) curves, calibration analysis along with Brier score, and decision curve analysis. Prognostic stratification, Cox regression, and subgroup analyses were conducted for high/low pathomics score (PS: a continuous score derived from a machine learning model based on H&E image features to predict PGF expression) groups. Bioinformatics approaches identified differentially expressed genes (DEGs) and immune infiltration patterns.

Results

PGF expression was identified as an independent prognostic factor for poor survival in HCC (HR = 1.922, 95% CI: 1.217–3.036, p = 0.005). A pathomics model integrating seven PGF‐associated features demonstrated strong predictive accuracy, achieving an AUC of 0.811 (95% CI: 0.749–0.873) in the training set, 0.747 (95% CI: 0.639–0.855) in the internal validation set, and 0.740 (95% CI: 0.632–0.849) in the external test set. Patients classified into the high‐pathomics score (PS) subgroup had significantly poorer survival (HR = 1.667, 95% CI: 1.024–2.713, p = 0.040). Functional analysis of DEGs in high‐PS tumours revealed enrichment in ribosome‐ and coagulation‐related pathways, upregulation of the inflammatory gene HBEGF, and increased infiltration of γδT cells. Moreover, TP53 mutations were frequently observed in this subgroup, with a mutation rate exceeding 20%.

Conclusion

PGF may serve as an independent prognostic biomarker in HCC. The developed pathomics model enables non‐invasive PGF expression prediction through H&E image analysis. Mechanistically, PGF‐associated molecular alterations involve inflammatory signalling, immune microenvironment remodelling, and frequent TP53 mutations, providing insights into HCC pathogenesis.

Lay Summary

We developed a novel machine learning‐based pathomics model that accurately predicts placental growth factor (PGF) expression in hepatocellular carcinoma using routine H&E‐stained tissue slides, and confirmed PGF as an independent prognostic biomarker for poor survival in liver cancer patients with high pathomics scores linked to significantly worse outcomes. This non‐invasive model further uncovered that high PGF expression is associated with perturbed inflammatory signalling, altered immune cell infiltration, and frequent TP53 mutations in liver cancer, providing key mechanistic insights into tumour progression and a practical, low‐cost tool for clinicians to stratify high‐risk patients and guide personalised treatment strategies.

Key Points

• What is already known: Placental growth factor (PGF) correlates with poor prognosis in hepatocellular carcinoma (HCC), yet its assessment relies on costly molecular assays.

• What this study adds: A machine‐learning pathomics model using routine H&E slides accurately predicts PGF expression, identifies high‐risk patients with worse survival (HR = 1.667), and links high scores to TP53 mutations and inflammatory/immune alterations.

• Potential impact: This non‐invasive, low‐cost tool enables prognostic stratification and personalized treatment guidance without additional tissue consumption.

1. Introduction

Hepatocellular carcinoma (HCC) is the sixth most common malignant tumour and the fourth leading cause of cancer‐related deaths worldwide [1]. HCC poses an enormous health burden globally in terms of morbidity and mortality, particularly in China [2]. Although surgical resection and liver transplantation are effective treatments for patients with HCC, numerous clinical studies indicate that the 5‐year recurrence rate of postoperative HCC is as high as 70% [3, 4]. Over the decades, management of HCC has followed several clinical staging systems that rely on tumour morphologic characteristics (e.g., size, number and vascular invasion) and clinical variables [5, 6, 7, 8, 9, 10, 11]. However, these algorithms struggle to fully capture the full picture of tumour invasiveness and cannot achieve precise prognostic stratification. Given the poor prognosis of HCC and the relative scarcity of prognostic assessment tools, there is an urgent need for additional biomarkers to evaluate the prognosis of HCC patients.

PGF has been implicated in previous studies as being associated with patient prognosis. As a member of the vascular endothelial growth factor family, PGF not only promotes angiogenesis but also exerts pro‐inflammatory effects [12], and inflammation is an essential pathway in the progression of hepatocellular carcinoma. It was found that the expression of PGF was significantly elevated in HCC tumour tissues compared to normal tissues [13, 14]. Meanwhile, PGF is closely related to the pathophysiological process of liver diseases, primarily involved in regulating the activation of epithelial‐mesenchymal transition, and plays an important role in the development of other tumours [15, 16]. Additionally, a competing risk model based on PGF has demonstrated strong predictive value for HCC prognosis, and PGF may serve as a new target for anti‐tumour immunity in HCC [17].

Given the potential value of PGF in the development and prognosis of hepatocellular carcinoma, it is crucial to accurately detect its expression level. However, current conventional assays have some limitations, such as high price, high time cost, difficulty in specimen collection, and results affected by operator and antibody. H&E‐stained sections are essential for clinical diagnosis and the most accessible image data, which have the advantages of being widely used, comprehensive information, stable, and inexpensive. However, there are limitations for pathologists to predict biomarker expression by observing H&E sections alone.

The development of pathomics models represents a major leap forward in the integration of histopathology and molecular biology. By utilising various data mining algorithms based on machine learning methods, high‐dimensional data can be transformed into predictive or prognostic information [18, 19, 20]. Recent seminal findings have shown the wide application of machine learning methods in pathomics, demonstrating high value in metastasis detection in lymph nodes, Ki67 scoring in breast cancer, Gleason grading in prostate cancer, and tumour‐infiltrating lymphocyte scoring in melanoma [Artificial intelligence as the next step towards precision pathology]. Previous studies have shown that machine learning can accurately recognise information in H&E images that is often not captured by pathologists through the human eye [21, 22] For example, Shamai et al. [23] successfully predicted the expression content of various molecular biomarkers, such as oestrogen receptor, in breast cancer H&E sections with high accuracy using a machine learning approach. The pathomics model constructed by Zhaoyong Yan et al. [24] successfully predicted TNF receptor superfamily member 4 (TNFRSF4) expression, and the model had good predictive performance for molecular expression and revealed a significant association between high histopathological scores and poorer OS.

This study demonstrates that molecular prediction in H&E sections can be achieved using machine learning, overcoming the limitations of traditional methods. Hence, we aim to use machine learning to predict the expression level of PGF in H&E sections of HCC patients, enabling us to evaluate their prognosis. This approach will serve as a complementary tool for prognostic evaluation in HCC patients. Additionally, we will integrate bioinformatics analysis to explore the potential molecular mechanisms underlying Pathomics.

2. Methodology

2.1. Patient Cohort

The flow of this study is illustrated in (Figure 1). The public database cohort used in this study was sourced from the publicly available TCGA dataset, and all data were de‐identified and approved by the ethics committee for public research purposes. Conversely, the hospital cohort was obtained from the General Hospital of Ningxia Medical University and received approval from its Research Ethics Committee (KYLL‐2025‐0520).

FIGURE 1.

Flowchart of this study to construct a clinical molecular prediction model based on a machine learning approach.

2.1.1. Public Database Cohort

First, 377 patients with HCC were selected from the TCGA database (https://portal.gdc.cancer.gov/). The inclusion criteria were (a) patients with available H&E‐stained section images from surgical resections. (b) Samples with primary solid tumour RNA‐seq. (c) samples with a primary diagnosis and initial treatment of hepatocellular carcinoma. The exclusion criteria were (a) samples lacking survival status, survival time, or having a survival time of less than 1 month. (b) samples with missing clinical data, or substandard quality of pathological images. Ultimately, this study included a sample of 267 HCC cases in the TCGA cohort.

2.1.2. Hospital Cohort

In addition, we included 80 HCC patients who underwent surgical treatment at the General Hospital of Ningxia Medical University from January 2016 to December 2018. Samples with hepatocellular carcinoma at both initial diagnosis and initial treatment were included, and samples with missing pathologic images or substandard pathologic image quality were excluded, resulting in 79 patients being included in the external validation cohort.

2.2. Survival Analysis

We downloaded the HCC‐related RNAseq data from the TCGA website, and analysed the intergroup differences between the HCC tumour group and the normal group by Toil pipeline [25], and the model outputs the probability of predicting the gene expression level of pathomics score after the modelling was completed, and compared the pathomics score in the PGF high and low by Wilcoxon test. groupings with or without differences between the subgroups. Subsequently, by plotting the Kaplan–Meier survival curves, we demonstrated the survival differences between the high and low PGF expression groups as well as the survival differences between the high PS and low PS groups. In addition, we investigated whether high PGF expression as well as high PS expression were independent influences on overall survival (OS) using univariate and multivariate Cox regression analysis. Finally, we performed exploratory subgroup analyses by univariate COX regression to analyse the effects of PGF (high expression group vs. low expression group) and PS (high expression group vs. low expression group) on patients' prognosis in different subgroups of each covariate; the likelihood ratio test was used to analyse the interactions between PGF as well as PS and other covariates.

2.3. Segmentation and Processing of Pathology Images

Firstly, image segmentation and processing were applied to the pathology images of the TCGA cohort. Given that the entire pathology image contains a significant amount of extraneous information, we used the OTSU algorithm (https://opencv.org/) to divide the image into two parts: the irrelevant background and the relevant tissue region for the study [26]. After removing the irrelevant background, the 40× image was segmented into multiple subimages of 1024 × 1024 pixels (Figure 2A), and the 20× image was segmented into multiple subimages of 512 × 512 pixels, which were then up‐sampled to 1024 × 1024 pixels. A final review was conducted by a pathologist to exclude subimages with poor image quality, such as contamination, blurriness, or blank areas exceeding 50%. Subsequently, 10 subimages from each segmented pathology image were randomly selected for further analysis (Figure 2B) to reduce computational cost while ensuring sufficient data [27, 28]. The pathology images from the external validation cohort were photographed and uploaded by specialised equipment and stored as 40× magnification images for subsequent model validation.

FIGURE 2.

Segmentation of H&E image with sub‐images.

2.4. Image Feature Extraction

We utilised PyRadiomics, a Python‐based tool (https://pyradiomics.readthedocs.io/en/latest/), to extract relevant features, including first‐order, second‐order, and higher‐order features, from the subimages. Feature values were then normalised across all samples using z‐score standardisation to reduce scale variation. Following feature extraction from the 10 sub‐images of each patient, we first computed the Intraclass Correlation Coefficient (ICC) for each pathomics feature to assess its stability across sub‐images. Only features demonstrating good reproducibility (ICC > 0.75) were retained. The mean value of these retained stable features was then calculated and used as the final pathomics feature representation for each sample in all subsequent analyses.

2.5. Feature Screening and Model Construction

The TCGA cohort was randomly divided into training and validation sets in a 7:3 ratio. To prevent information leakage, all feature selection steps were performed exclusively on the training set. We employed the Maximum Relevance Minimum Redundancy (mRMR) algorithm and the Recursive Feature Elimination (RFE) algorithm to identify the optimal subset of features. Both approaches represent widely adopted and established feature selection strategies in current pathomics/radiomics research [29, 30]. Initially, the mRMR algorithm was used to filter out a set of relevant features, followed by the RFE algorithm to further refine the selection until the best subset of features was obtained. The RFE algorithm continued to identify relevant features vital for generating accurate models. Finally, we constructed a model to predict PGF expression using the Gradient Boosting Machine (GBM). To ensure robustness and generalisability, the GBM model was trained on the training set with 10‐fold cross‐validation (10‐fold CV) implemented via the caret framework, which facilitated automated hyperparameter tuning. The intrinsic mechanisms of GBM—including iterative boosting, shrinkage, and tree‐depth control—further helped mitigate overfitting. The final model was built by assigning importance scores to the selected pathomics features based on this optimised procedure.

2.6. Evaluation and Validation of the Model

To enhance the interpretability of the predictive model, we performed SHapley Additive exPlanations (SHAP) analysis using the shap Python package. SHAP values were calculated for each prediction in the training set to quantify the marginal contribution of each pathomics feature to the model output.

For model evaluation, we assessed the predictive performance of the model by plotting ROC curves (Receiver Operating Characteristic curves) and calculating the respective AUC values (Area Under the ROC Curve) for the training, validation, and test sets. To provide a complementary perspective on classification performance, especially relevant for imbalanced outcomes, we also generated Precision‐Recall (PR) curves and reported the average precision (AP). The calibration of the predictive models was assessed by plotting calibration curves, conducting Hosmer‐Lemeshow goodness‐of‐fit tests, and calculating the Brier score to quantify the overall mean squared error of the probability predictions. Decision curve analysis (DCA) was also performed to illustrate the clinical benefits of the predictive models.

During the model validation phase, all pathological image data were integrated into the constructed pathomics model, which output the probability of predicting gene expression levels as the pathomics score. The Wilcoxon test was employed to compare differences in pathomics scores between the high‐ and low‐PGF expression subgroups within the training, validation, and test sets, and the results were visualised accordingly. The pathomics score was then merged with clinical data, and the survminer package was used to determine the cutoff value for the pathomics score, thereby categorising it into Low/High binary variables (PS).

2.7. Histopathological Mechanism Analysis

To explore the molecular mechanisms underlying pathomics, we first performed differential gene expression analysis between high‐ and low‐PS groups using the limma R package. Subsequently, Gene Set Enrichment Analysis (GSEA) using the KEGG and Hallmark gene sets was conducted to evaluate pathway enrichment. To account for multiple testing, enrichment results were adjusted using the false discovery rate (FDR) method. The primary exploratory screening utilised a nominal p‐value threshold (p < 0.05) to identify biologically plausible signals, while a comprehensive table reporting both nominal and FDR‐adjusted p‐values for all examined pathways is provided as Supporting Information. Subsequently, we analysed the differential expression of 200 inflammation‐related genes between the high and low PS subgroups using the Wilcoxon rank‐sum test, focusing on genes with a p‐value of less than 0.05 for visualisation [31, 32]. We uploaded the gene expression matrices of the HCC samples from the TCGA cohort to the ImmuCellAI database to calculate immune cell infiltration for each sample. We then analysed the differences in immune cell infiltration in high/low PS groups using the Wilcoxon rank‐sum test. Finally, we downloaded mutation data for TCGA‐LIHC patients from the TCGA database and intersected it with 267 samples from the TCGA cohort to analyse the mutation data. We visualised the top 15 genes with the highest mutation frequency, assessing the relationship between the mutation profiles of TCGA‐LIHC patients and the PS grouping model predictions.

2.8. Statistical Methods

Continuous variables were compared between groups using the Wilcoxon test. The calibration of the histopathological prediction model was assessed by plotting ROC and calibration curves and performing the Hosmer‐Lemeshow goodness‐of‐fit test. Kaplan–Meier survival curves illustrated variations in survival rates across different groups, and the log‐rank test evaluated the significance of survival rate differences between groups. Model performance was assessed through likelihood ratio test analysis. All analyses were conducted using R software version 4.1.0, with hypothesis testing being two‐sided and a p‐value of less than 0.05 indicating a significant difference.

3. Results

3.1. Correlation Analysis of PGF With HCC

In both the TCGA cohort and the external validation cohort, we categorised patients into high‐ and low‐expression groups based on the level of PGF expression and created baseline information tables for each cohort (Appendix Tables S1 and S2). The results showed no statistically significant differences in the distribution of clinical variables between the groups and were comparable. Analysis of variance showed that the PGF expression level in HCC tumour tissues was significantly higher than that in normal tissues (p < 0.0001) (Figure 3A). Kaplan–Meier curves showed that there was a significant difference in overall survival between patients with high and low PGF expression in the TCGA cohort (p = 0.001), the median survival time of patients in the low PGF expression group being 84.4 months, compared to only 46.2 months for those in the high PGF expression group (Figure 3B). In univariate analysis (Figure 3C), high PGF expression emerged as a risk factor for overall survival (OS) in patients (HR = 2.123, 95% CI: 1.368–3.293, p < 0.001). When included in multivariate analysis (Figure 3D), high PGF expression remained a distinct risk factor for OS (HR = 1.922, 95% CI: 1.217–3.036, p = 0.005). In the interaction test analysis (Figure 3E), there was an interaction between the expression content of PGF and the prognosis of patients in the three subgroups of whether or not they were treated with medication, whether or not they had undergone ablative embolisation, and the level of blood AFP (p < 0.05). The results of the subgroup analysis (Figure 3F) showed that in subgroups undergoing drug therapy, ablative embolisation therapy, and with AFP levels > 400, elevated PGF did not correlate with patient prognosis. However, in subgroups not defined by these conditions, PGF expression levels could accurately reflect patient prognosis. While drug therapy and ablation embolisation are commonly used clinical treatments that can affect patient survival time to varying degrees [33], they also influence PGF expression levels, thereby complicating the assessment of HCC patient prognosis based on PGF expression. In the subgroup with AFP > 400, PGF could not effectively predict the prognosis of HCC patients. This may be due to AFP's influence on PGF expression levels, though no detailed reports exist on the relationship between these factors in HCC patients domestically or internationally. This observation presents a novel perspective for exploring the impact of molecular levels on the survival prognosis of HCC patients.

FIGURE 3.

Correlation analysis between PGF and HCC. (A) Comparison of PGF expression in tumour tissues and normal tissues in TCGA pairs of columns; (B) Kaplan–Meier survival curves demonstrating the changes in survival between high and low PGF expression groups; (C, D) One‐way & multifactorial Cox regression analyses exploring whether high expression of PGF is an independent influencing factor for OS. (E, F) Thery are interaction tests and subgroup analysis plots to analyse the interaction between the PGF high expression group and the low expression group in different subgroups of each covariate and the effect on patients' prognosis in each subgroup.

3.2. Pathology Image Feature Extraction and Model Construction

A total of 267*10 sub‐images of 1024 × 1024 pixels were obtained after segmenting and filtering the pathology images of the TCGA cohort. After image normalisation, 1488 features were extracted and finally the average value was taken to represent the histopathological features of this sample [34, 35, 36]. The TCGA cohort was randomly divided into training and validation sets with a 7:3 ratio, comprising 188 cases in the training set and 79 cases in the validation set. A baseline profile table was created, indicating that the analysis of variance for each variable showed no significant differences between the groups (p > 0.05) (Appendix Table S3). Subsequently, in the training set, we used the mRMR method to screen the top 30 image features with strong predictive efficacy. These features were further refined using the RFE method, resulting in a total of 7 features significantly correlated with the prediction of the PGF molecule for subsequent analysis (Figure 4A). The seven selected pathomics features capture multi‐dimensional heterogeneity in tumour architecture, cellular composition, and stromal patterns. log‐sigma‐4‐0‐mm‐3D‐glrlm‐RunEntropy reflects high uncertainty in grey‐level runs, indicative of disorganised and complex tissue patterns. log‐sigma‐1‐0‐mm‐3D‐gldm‐LargeDependenceHighGrayLevelEmphasis highlights large, high‐intensity structures, often associated with vascular‐rich or densely cellular regions. square‐glcm‐InverseVariance points to local texture uniformity, typical of homogeneous tumour areas. Conversely, square‐glszm‐SmallAreaHighGrayLevelEmphasis identifies small, high‐intensity foci, suggestive of micro‐tumour nests or microvascular aggregates. wavelet‐LL‐glszm‐LargeAreaLowGrayLevelEmphasis corresponds to extensive low‐intensity zones, commonly seen in necrotic or fibrotic stroma. wavelet‐HH‐glcm‐ClusterShade emphasises skewness in high‐frequency textures, frequently found at infiltrative tumour margins. Lastly, exponential‐firstorder‐90Percentile quantifies the proportion of high‐intensity pixels, linked to areas of enhanced staining or elevated cellular density. To quantify and visualise each feature's contribution to the model's prediction, we performed SHAP (SHapley Additive exPlanations) analysis. The SHAP summary plot (Figure 4B) illustrates both the direction and magnitude of impact for each feature, confirming their collective role in predicting PGF expression. Finally, we calculated the importance scores of these 7 features in the model by the GBM algorithm (Figure 4C), and constructed the pathomics model in the training set by the relevant features and their corresponding weights.

FIGURE 4.

Construction of a machine learning‐based prediction model for PGF expression level. (A) screening of pathological features by machine learning; (B) SHAP analysis of the seven selected pathomics features, showing their mean absolute impact on model predictions. (C) constructing a prediction model by the seven pathological features screened.

3.3. Evaluation of the Model

The results of our model efficacy evaluation showed that the HCC pathomics model had high predictive accuracy in the training set (AUC = 0.811, 95% CI: 0.749–0.873), the internal validation set (AUC = 0.747, 95% CI: 0.639–0.855), and the external test set (AUC = 0.740, 95% CI: 0.632–0.849) (Figure 5A–C). We then plotted calibration curves for the training, validation, and test datasets (Figure 5D–F), and the calibration curves and the Hosmer‐Lemeshow goodness‐of‐fit test showed that the histopathological prediction model's predictive probability of whether PGF was highly expressed was in good agreement with the true value (p > 0.05); the DCA curves (Figure 5G–I) demonstrated that the model had clinical utility. To provide a more comprehensive evaluation of the model's predictive performance, we calculated the Brier Score and generated Precision‐Recall (PR) curves. The Brier Score, quantifying the overall calibration error of probability predictions, was 0.193, 0.210, and 0.244 for the training, validation, and external test cohorts, respectively. These results are consistent with the good fit observed in the calibration curves. Furthermore, the area under the PR curve (AUPRC, also reported as Average Precision, AP)—a metric particularly informative for imbalanced classification—achieved values of 0.807, 0.803, and 0.739 across the three cohorts (Figure 5J–L). The robust AUPRC values, especially in the training and validation sets, complement the AUC findings and underscore the model's strong ability to identify PGF‐high expression cases across different decision thresholds.

FIGURE 5.

Evaluation of the prediction model. (A–C) are the ROC curves for the molecular prediction scores of the training, validation, and test cohorts, respectively; (D–F) are the calibration curves for the training, validation, and test cohorts, respectively; and (G–I) are the DCA curves for the training, validation, and test cohorts, respectively; (J–L) are the PR curves for the training, validation, and test cohorts, respectively.

3.4. Prognostic Value of the Model

We divided the training, validation, and test sets into high expression and low expression groups based on PGF expression levels. By comparing the distribution of PS values across these three datasets, we observed that the PS of the PGF high expression group was significantly higher than those of the PGF low expression group (Figure 6A–C). The TCGA cohort was divided into high and low PS groups based on their PS. A baseline information table for each clinical variable was created (Appendix Table S4). The KM curves plotted (Figure 6D) showed that the median survival time for the high PS group was 56.17 months, while it was 82.87 months for the low PS group. High PS was significantly associated with lower OS in HCC patients (p = 0.006), highlighting the positive predictive value of predictive models in assessing patient prognosis.

FIGURE 6.

Validation of the prediction model. (A–C) represent the plots of variability in PS scores grouped by high and low PGF expression across the training cohort, validation cohort, and test cohort, respectively. (D) illustrates the KM curve plotted in the TCGA dataset, categorised by high and low PS.

Further univariate and multivariate Cox regression analyses (Figure 7A,B) confirmed the prognostic value of the pathomics model. In univariate analysis, a high PS was significantly associated with worse OS (HR = 1.786, 95% CI: 1.171–2.723, p = 0.007). After adjusting for clinicopathological variables in multivariate analysis, a high PS remained an independent risk factor for poor OS (HR = 1.667, 95% CI: 1.024–2.713, p = 0.040). Other significant covariates in the final multivariate model included advanced pathologic stage, specific ablation/embolisation status, unknown AFP level, hepatic inflammation, and residual tumour (all p < 0.05, detailed in Figure 7B and Appendix Table S5). Interaction test analysis showed (Figure 7C) that there was no significant interaction between PS and the majority of subgroups, indicating that PS scores are reliable predictors of patient prognosis across different subgroups for each covariate.

FIGURE 7.

Validation of the prediction model. (A, B) are univariate & multivariate Cox regression analyses related to PS to explore whether PS score is an independent influence on OS; (C) is interaction tests and subgroup analysis plots to analyse whether there is an interaction between high and low PS expression and each covariate as well as to analyse the effect of high and low PS expression on patients' prognosis in different subgroups of each covariate.

3.5. Biological Interpretability

Gene Set Enrichment Analysis revealed distinct molecular profiles between high‐ and low‐PS subgroups (Figure 8A,B). In the KEGG gene set, pathways including RIBOSOME and COMPLEMENT_AND_COAGULATION_CASCADES were significantly enriched (nominal p < 0.05; FDR‐adjusted p‐values are detailed in Supporting Information). Similarly, the Hallmark gene set analysis showed significant enrichment of pathways such as COAGULATION and INFLAMMATORY_RESPONSE (nominal p < 0.05). In the differential analysis of inflammation‐related genes between the high and low PS subgroups, we found (Figure 8C) that the expression of genes such as HBEGF was significantly elevated in the high PS group. In the immune cell infiltration analysis (Figure 8D), γδT cells were significantly different in the high and low PS subgroups (p < 0.01). Finally, we visualised the somatic mutation landscape across subgroups (Figure 8E,F). This analysis was descriptive, aiming to provide a panoramic view of mutational patterns rather than to test specific hypotheses. Within this context, TP53 mutations were frequently observed (29% in high‐PS vs. 25% in low‐PS samples). It is important to note that this comparison is observational, and no formal statistical test (e.g., Fisher's exact test) was performed to assess enrichment, as the goal was to describe, not infer causality from the mutational background.

FIGURE 8.

Histopathological mechanism analysis. (A, B) are plots of enrichment analysis of differential genes between high and low PS subgroups as a result of model prediction; (C) is a plot of differential analysis of PS and inflammation‐related genes as a result of model prediction; (D) is a plot of differential analysis of PS and immune cell abundance as a result of model prediction; and (E) is a plot of mutation analysis of genes between high and low PS subgroups as a result of model prediction.

4. Discussion

This study established a pathomics model based on H&E‐stained whole‐slide images to predict placental growth factor (PGF) expression in hepatocellular carcinoma (HCC). The model, constructed from seven selected features using a gradient boosting machine (GBM), demonstrated robust predictive performance across training, internal validation, and independent external test cohorts. Patients categorised into the high pathomics score (PS) group exhibited significantly worse overall survival, confirming the prognostic relevance of PGF inferred from histopathology. Bioinformatics analyses further revealed that high‐PS tumours were enriched in coagulation‐ and ribosome‐related pathways, displayed upregulated inflammatory signalling (e.g., HBEGF), altered immune infiltration (notably reduced γδ T cells), and frequent TP53 mutations. Our findings align with and extend previous studies implicating PGF in HCC progression. PGF, a member of the VEGF family, is overexpressed in HCC and correlates with poor prognosis, angiogenesis, and metastasis [13, 14, 37]. Consistent with Miao et al. [14], we confirmed that high PGF expression serves as an independent risk factor for reduced overall survival. Beyond its angiogenic role, emerging evidence indicates that PGF modulates immune responses and stromal remodelling in the tumour microenvironment (TME) [15, 24]. Our pathomics approach bridges histomorphological patterns with molecular phenotypes, similar to recent studies using radiomics or pathomics to predict molecular features from imaging [23, 24]. Notably, our model's ability to non‐invasively estimate PGF expression from H&E slides offers a practical complement to traditional biomarker assays, which are often limited by cost, variability, and tissue requirements.

A major strength of our study lies in the biological interpretability afforded by the pathomics model. The seven selected H&E features provide direct histopathological insights into the PGF‐associated tumour microenvironment. Features highlighting high‐grey‐level, large, or small focal structures may correspond to vascular proliferation and dense cellular regions, aligning with PGF's pro‐angiogenic role. Conversely, features indicating low‐grey‐level, large homogeneous areas (e.g., LargeAreaLowGrayLevelEmphasis, InverseVariance) likely reflect stromal remodelling or necrosis, common in aggressive tumours. The high textural complexity (RunEntropy, ClusterShade) further suggests architectural disorganisation, a hallmark of tumour progression.

The pathway differences between high‐PS and low‐PS tumours revealed by gene‐set enrichment analysis (GSEA) construct a coherent biological narrative linking PGF to multiple hallmarks of cancer.

Coagulation and inflammation pathways: We observed significant enrichment of complement/coagulation cascades and inflammatory‐response pathways. Coagulation factors can promote metastasis by enhancing cancer‐cell adhesion and endothelial activation [38], while the coagulation cascade plays a key role in shaping the tumour‐immune microenvironment (TME) [39]. As a key gene related to coagulation and fibrinolysis in HCC, PGF may influence HCC development through TME remodelling [24, 40]. Mechanistically, PGF activates PI3K/AKT, MAPK, and NF‐κB signalling via VEGFR‐1 and its co‐receptors, thereby promoting angiogenesis, endothelial activation, and inflammatory cytokine expression, which is closely coupled with complement/coagulation cascades and vascular inflammation [41]. Moreover, PGF can recruit and polarise monocyte‐macrophages, increase tumour‐associated macrophage (TAM) infiltration, and enhance their immunosuppressive phenotype [15, 42]. PGF has also been shown to promote dense stromal reaction and hypoxic microenvironment formation [43], consistent with the observed enrichment of the hypoxia pathway. In the differential analysis of inflammation‐related genes, we found that a considerable number of inflammatory genes were expressed at higher levels in the high‐PS group, with HBEGF being particularly prominent. Increased local blood supply is a characteristic of HCC progression. In a basic study of peripheral arterial disease, inhibition of relevant genes led to increased local blood supply in mouse hindlimbs, accompanied by upregulation of HBEGF and PGF expression, suggesting that HBEGF may alter local perfusion through angiogenesis or microenvironmental remodelling [44]. Together with previous reports and our present results, HBEGF and PGF may act synergistically in tumour progression or certain benign vascular disorders.

Ribosome and metabolic pathways: In contrast, low‐PS tumours showed preferential enrichment of pathways related to protein translation and metabolism, such as ribosome and oxidative phosphorylation. The ribosome pathway is closely associated with protein synthesis and cellular proliferation. Upregulation of PGF can activate the mTOR pathway, which participates in multiple processes of ribosome biogenesis and controls protein synthesis by promoting ribosome production [45, 46, 47]. Hyperactivation of ribosome biogenesis is recognised as a key event driving hepatocellular transformation and HCC development [48], indicating a fundamental metabolic reprogramming difference between high‐PS and low‐PS tumours.

Immune modulation and oncogenic network: The immunosuppressive microenvironment observed in high‐PS tumours—characterised by reduced γδ T‐cell infiltration and frequent TP53 mutations—positions PGF within a broader oncogenic network that influences both metastatic propensity and immunotherapy response. γδ T cells are a subset of T cells involved in anti‐tumour immunity, but their function is often impaired in HCC patients. Studies have shown that γδ T‐cell infiltration correlates with HCC progression and patient prognosis, and their infiltration decreases while anti‐tumour efficacy diminishes during HCC development [49]. PGF may regulate the activation and function of γδ T cells through the MAPK‐RAP1A signalling pathway, thereby promoting immune evasion and tumour progression [50]. This highlights the potential of targeting PGF and its downstream signalling as a strategy to enhance anti‐tumour immunity in HCC. Furthermore, TP53 mutations are the most common genetic alterations in HCC, predominantly missense mutations, which are involved in regulating DNA repair, cell‐cycle arrest, and apoptosis [51]. The high mutation rate of TP53 in our cohort (29% in the high‐PGF group vs. 25% in the low‐PGF group) underscores the importance of this gene in HCC pathogenesis and points to the need for targeted therapies that can restore TP53 function or bypass its tumour‐suppressive effects.

These findings align with key advances in metastasis immunotherapy: (1) Engineered T‐cell therapies emphasise the need to overcome T‐cell exhaustion, a state that may be exacerbated by the PGF‐driven inflammatory milieu [52]; (2) Driver mutations such as TP53 are recognised modulators of tumour immunogenicity and checkpoint‐inhibitor response [53]; (3) Immune‐stromal scoring models highlight the prognostic value of TME‐related gene signatures in predicting metastasis and therapeutic outcome, an axis consistent with our enriched coagulation and inflammatory pathways [54]; (4) Epigenetic regulation via non‐coding RNAs modulates immune evasion and epithelial‐mesenchymal transition (EMT), raising the question of whether PGF interacts with such regulatory layers [55]; (5) Single‐cell dissection of the TME has identified core immune‐escape genes linked to T‐cell function and patient prognosis, suggesting that integrating PGF‐pathomics with high‐resolution immune maps could improve patient stratification [56]. Collectively, these insights situate PGF as a central node within an “inflammation–coagulation–immune suppression” axis that promotes metastatic progression.

Our PGF‐pathomics model holds tangible potential for clinical translation. The non‐invasive prediction of PGF expression levels using routinely acquired H&E‐stained images offers a promising tool for advancing personalised treatment strategies in HCC. In practice, the model serves as a quantifiable auxiliary tool: after digitising routine H&E‐stained sections, it automatically computes a pathomics score for each patient, thereby indirectly predicting the tumour's PGF expression level. By integrating this score with other multi‐omics data (e.g., genomics, proteomics) and conventional indicators (e.g., clinical stage, AFP levels), clinicians can obtain a more holistic understanding of tumour behaviour and tailor therapeutic interventions accordingly.

Concretely, patients identified with high PGF expression/high pathomics scores—characterised by higher recurrence risk and a more active inflammatory/immune microenvironment—may be more suitable for anti‐angiogenic therapies (e.g., bevacizumab) or immune‐based combination regimens. Conversely, those with low PGF expression/low pathomics scores may be considered for more conventional surveillance and management strategies. Thus, the model could aid in postoperative risk stratification, treatment selection, and individualised follow‐up planning. Moreover, the association between PGF expression and specific inflammatory genes and immune infiltration patterns opens new avenues for enhancing immunotherapeutic strategies. Targeting PGF‐mediated pathways could potentially increase the efficacy of existing treatments and improve clinical outcomes.

This study has several limitations that should be acknowledged. First, the retrospective design and the single‐institution external cohort (2016–2018) may introduce selection bias and limit the model's generalisability to contemporary cohorts treated with newer immunotherapies. Second, despite robust external validation, inherent differences between the TCGA and our local cohort—in ethnicity, regional characteristics, and tissue processing/scanning protocols—could lead to domain shift and affect feature stability. Third, our mechanistic insights, derived from bioinformatics analyses, lack direct functional validation (e.g., in vitro or in vivo) to establish causality for PGF's role, and the absence of standardised immunohistochemistry data precluded validation at the protein level.

To address these limitations, future prospective, multi‐center studies with standardised imaging protocols, integrated multi‐omics data, and functional experiments are warranted to confirm the clinical utility and biological foundations of the model.

5. Conclusion

In conclusion, our study establishes PGF expression as a robust prognostic biomarker in HCC and further develops a pathomics‐based model capable of predicting PGF expression levels directly from H&E‐stained pathological slides. This model enables non‐invasive stratification of patients into distinct prognostic subgroups, thereby facilitating early identification of high‐risk individuals who may benefit from more aggressive or tailored therapeutic interventions. By leveraging widely available histopathological images, our approach offers a clinically practical tool for improving prognostic assessment and personalising treatment strategies in HCC. As research in pathomics continues to advance, we anticipate that such integrative analytical frameworks will play an essential role in precision oncology, ultimately contributing to improved outcomes for HCC patients.

Author Contributions

Long Chen and Xusheng Zhang contributed equally. Long Chen, Xusheng Zhang, and Bendong Chen participated in the conceptualisation and design of this study. Long Chen and WeiHu Ma organised the database and statistical analysis. Long Chen, Bendong Chen, Lin Ding, Shicai Liang, Xuebo Wang, and KeJun Liu divided the work and participated in the drawing. Long Chen wrote the first manuscript. Bendong Chen and KeJun Liu participated in the revision of the manuscript. All authors read and agreed to the final manuscript and authorship arrangement.

Funding

This work was supported by The Natural Science Foundation of Ningxia (Grant 2023AAC02073), The Science and Technology Support Project of the Science and Technology Bureau of Yinchuan City (Grant 2023SF13) and The Central Guidance for Local Science and Technology Development Special Project (Grant 2024FRD05060).

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Table S1: Baseline clinical characteristics of HCC patients in the TCGA cohort stratified by PGF expression level (high vs. low). This table compares clinicopathological variables between PGF‐high and PGF‐low subgroups in the TCGA cohort (n = 267), showing no significant differences for most covariates, thereby supporting the comparability of the two groups.

Table S2: Baseline clinical characteristics of HCC patients in the external validation cohort (General Hospital of Ningxia Medical University) stratified by PGF expression level. This table presents the distribution of age, gender, AFP level, and pathologic stage between PGF‐high and PGF‐low groups in the independent external cohort (n = 79), confirming consistency of patient characteristics across subgroups.

Table S3: Baseline clinical characteristics of the TCGA cohort divided into training and internal validation sets (7:3 ratio). This table demonstrates that no statistically significant differences were observed for any clinical variable between the training set (n = 188) and the internal validation set (n = 79), supporting the appropriateness of random split for pathomics model development.

Table S4: Baseline clinical characteristics of the TCGA cohort stratified by pathomics score (PS) groups (high vs. low). This table summarises the clinicopathological features of patients with high PS versus low PS, where PS is a continuous score derived from the gradient boosting machine model to predict PGF expression.

Table S5: Univariate and multivariate Cox regression analyses for overall survival in the TCGA cohort, incorporating PS and clinicopathological variables. This table provides detailed hazard ratios (HRs), 95% confidence intervals (CIs), and p‐values for all variables included in the Cox models. After adjusting for confounders, high PS remained an independent risk factor for poor overall survival (HR = 1.667, 95% CI: 1.024–2.713, p = 0.040).

Table S6: An Excel spreadsheet containing the full list of differentially expressed genes (DEGs) between high‐ and low‐pathomics score groups in the TCGA cohort, including gene symbols, log fold change, average expression, t‐statistic, nominal p‐values, and false discovery rate (FDR)‐adjusted p‐values.

Data S1: Paper‐related data parameters. A comprehensive technical appendix (Microsoft Word document) detailing all methodological parameters, computational steps, and analytical specifications used in this study, including data curation criteria, image processing pipeline, radiomics feature extraction settings, feature selection procedures, model construction details, and bioinformatics analysis parameters. This document is provided to enhance reproducibility and transparency.

Chen L., Zhang X., Liu K., et al., “The Prognostic and Biological Value of PGF‐Based H&E Pathomics in Hepatocellular Carcinoma,” Liver International 46, no. 5 (2026): e70655, 10.1111/liv.70655.

Data Availability Statement

The data supporting the findings of this study are available at [TCGA]. These data originate from the following public resources: https://portal.gdc.cancer.gov/projects/TCGA‐LIHC.