Integration of pathomics and single-cell omics in pancreatic ductal adenocarcinoma: Applications and clinical translation prospects

Abstract

Pancreatic ductal adenocarcinoma (PDAC), the predominant pathological subtype of pancreatic cancer, presents significant challenges in early diagnosis and treatment due to its high degree of heterogeneity. The emergence of single-cell omics and pathomics are providing powerful new tools and insights that are advancing PDAC research. Single-cell omics elucidates the molecular profiles of malignant epithelial cells, immune cells, and stromal cells within the PDAC tumor microenvironment, uncovering key pathways and cellular subpopulations that drive PDAC progression and drug resistance. In contrast, pathomics quantitatively extracts subtle morphological features from digitized whole-slide images, employing machine and deep learning to build diagnostic and prognostic prediction models. The multi-omics integration based on single-cell and pathology data provides deeper insights into tumor microenvironment. This integrated approach not only enables the prediction of molecular subtypes and immune status from routine hematoxylin and eosin-stained images, providing a low-cost and rapid auxiliary diagnostic tool for clinical practice, but also accurately identifies therapeutic targets, predicts drug responses, and screens potential beneficiaries for immunotherapy. This minireview aims to dissect PDAC from a multi-omics perspective, with the objectives of fostering greater integration and exploration across these fields and thereby deepening the molecular and spatial understanding of PDAC and laying the groundwork for future precision medicine approaches.

Key Words: Pancreatic ductal adenocarcinoma; Pathomics; Single-cell omics; Artificial intelligence; Multimodal integration

Core Tip: This minireview systematically summarizes the progress of single-cell omics and pathomics in pancreatic ductal adenocarcinoma (PDAC) research. Single-cell omics provides deep insights into the cellular heterogeneity, tumor microenvironment, and drug-resistance mechanisms of PDAC. Pathomics employs artificial intelligence to quantitatively analyze histopathological images, enabling automated diagnosis, molecular subtyping, and prognostic evaluation. The integration of these two approaches constructs a multidimensional “morphological-molecular-spatial” perspective, which significantly advances the precision management of PDAC. Future work should focus on standardizing multi-omics techniques, building interpretable models, and promoting clinical translation to address the current therapeutic challenges in PDAC.

INTRODUCTION

Pancreatic ductal adenocarcinoma (PDAC) is among the most aggressive malignant tumors, with a five-year survival rate as low as 13%[1]. Its pronounced heterogeneity, a complex tumor microenvironment (TME), and the inherent difficulty of early diagnosis collectively pose the central challenges in the clinical diagnosis and treatment of this disease[2]. From a biological perspective, PDAC is characterized by distinct histo-ecological features: A prominent desmoplastic stroma typically dominates the tumor architecture, forming a dense fibrotic barrier[3]. The considerable heterogeneity of cancer-associated fibroblasts (CAFs) and their dynamic crosstalk with tumor cells[4], along with myeloid cells[5] and exhausted T cells[6], collectively establish an immunosuppressive TME. This renders PDAC a prototypical “immune-cold” tumor, which shows general insensitivity to immune checkpoint inhibitors[7].

The diagnosis of PDAC remains based on conventional histopathology as the gold standard. However, the evaluation of routine hematoxylin and eosin (HE)-stained sections mainly depends on the subjective qualitative assessment by pathologists, which inherently restricts reproducible quantification, cross-center standardization, and in-depth interpretation of the associations between morphological and molecular features[8]. Particularly in the context of classic PDAC, which is characterized by sparse tumor cells, abundant stroma, and complex tissue architecture, purely descriptive morphological assessment is inadequate for the systematic and quantitative characterization of key spatial phenotypes that are strongly associated with metastasis, recurrence, and therapy resistance. These phenotypes include the spatial distribution patterns of immune cells, the tumorstroma interface niche, and perineural or perivascular invasion[9-11]. Addressing these limitations necessitates the advancement of quantitative spatial phenotyping approaches grounded in digital pathology and artificial intelligence (AI)[12].

In recent years, the innovation of high-throughput omics technologies has driven a paradigm shift in PDAC research. Single-cell/single-nucleus RNA sequencing (sc/snRNA-seq) enables high-resolution profiling of malignant epithelial cells, delineating their lineage evolution, molecular subtype plasticity, and dynamic crosstalk with the microenvironment[7,13]. This approach also facilitates the systematic identification of functionally heterogeneous CAFs subpopulations and maps their differentiation trajectories toward either protumor or tumor-restraining phenotypes[11,14]. Moreover, sc/snRNA-seq provides a comprehensive view of the immunosuppressive network and intercellular communication pathways within the TME, thereby revealing the cellular underpinnings of drug resistance, metastasis, and immune evasion in PDAC[15]. Spatial transcriptomic platforms, such as 10× Visium and CosMx Spatial Molecular Imaging, combined with multiplexed imaging techniques including CODEX and imaging mass cytometry, allow for the characterization of gene expression patterns in relation to precise cellular spatial contexts while maintaining the original tissue architecture. These methodologies further support the identification of spatially organized niches with significant prognostic and therapeutic implications[16]. However, these advanced technologies still encounter several practical limitations, such as cell-type bias and transcriptional distortion resulting from sample preprocessing steps like tissue dissociation, insufficient spatial resolution or restricted throughput in some platforms, along with high costs and intricate data analysis requirements. Consequently, their translation into large-scale, routine clinical application remains a considerable distance away[17]. Complementing these approaches, digital pathology combined with AI-driven pathomics enables the high-throughput extraction of morphological, textural, and spatial-topological features from whole-slide images (WSI). This framework captures fine-grained architectural variations in tissue that are difficult to quantify visually and leverages these features for diagnostic classification, molecular subtyping, therapyresponse prediction, and prognostic stratification[8]. For PDAC, HE-staining provides the practical benefits of ease acquisition and low cost. Moreover, the resulting images contain rich information on stromal architecture, glandular morphology, and immune infiltration patterns, rendering them inherently suitable for digital quantitative analysis[18]. Nevertheless, pathological images fundamentally reflect the phenotypic endpoint of cellular and tissue organization. In the absence of direct molecular anchoring and integration, the interpretability of models built from such images, as well as their ability to infer underlying biological mechanisms, remains inherently limited[19].

Therefore, multimodal integration of single-cell omics and pathomics is emerging as a pivotal direction for precision diagnosis and treatment in PDAC[16,20]. Molecular annotations derived from single-cell omics, which encompass cellular identities and pathway activation states, provide a mechanistic basis for decoding the cellular composition and biological processes underlying histopathological phenotypes. The spatial registration of these annotations onto tissue architecture enables the inference of molecular subtypes, immunological profiles, and predicted therapeutic responses directly from routine HE-stained sections. This integrative strategy supports the development of cost-effective, scalable, and clinically translatable decision-support systems[21]. Based on this rationale, this minireview systematically synthesizes existing evidence on the role of pathomics in PDAC diagnostic classification, molecular subtyping, treatment response evaluation, and prognostic stratification. It also consolidates findings from single-cell omics regarding cellular lineage evolution and TME crosstalk. A major emphasis is placed on discussing integration strategies, model interpretability, and translational pathways for combining these two modalities in clinical practice.

PATHOMICS WORKFLOW

Pathomics, an advancing interdisciplinary domain that bridges digital pathology and AI, is fundamentally reshaping the comprehension, diagnosis, and prognosis of diseases via high-throughput extraction and analytical processing of quantitative features from WSI. This section aims to delineate a standardized and reproducible workflow in pathomics, which encompasses four pivotal phases: Data acquisition and preprocessing, tissue region segmentation, feature extraction, data analysis, model development, and clinical validation (Figure 1).

Figure 1 Workflow for artificial intelligence-based prognostic analysis using histopathological whole-slide images. CNNs: Convolutional neural networks.

Data acquisition and preprocessing

WSI: The fundamental step in data acquisition is the digitization of glass slides into high-resolution WSI using dedicated slide scanners, typically achieving a resolution of 0.25-0.5 μm/pixel. This process fully preserves the morphological information of the original tissue section, enabling seamless navigation from a macroscopic overview down to subcellular structures. A WSI is a multi-level, computationally accessible image file (formats such as .svs, .ndpi, .mrxs) produced by highresolution panoramic scanning of conventional glass histopathology slides. In essence, it acts as a standardized digital repository of pathological morphology and exhibits three defining attributes: It supports smooth, continuous zooming across scales from tissue-level panoramas to subcellular features; it delivers ultra-high resolution; and it comprehensively retains all morphological information contained in the original glass slide.

The public ecosystem of WSI resources has matured into a multi-layered, complementary, and fully-structured system, offering a solid data foundation for both foundational research and algorithmic development. Large-scale integrated repositories, exemplified by The Cancer Genome Atlas and the Clinical Proteomic Tumor Analysis Consortium, provide tens of thousands of WSIs across a wide range of cancer types. These datasets are deeply integrated with genomic and proteomic profiles, thus establishing a critical infrastructure for pathomics and multi-omics integration research. In response to specific algorithmic demands, specialized datasets with carefully curated annotations have been developed. Notable examples include CAMELYON[22] for lymph node metastasis detection, the PANDA dataset[23] for prostate cancer grading, and the NCT-CRC-HE colorectal cancer image library[24] for tissue classification. Each of these serves as a valuable reference for algorithm training and performance benchmarking. Additionally, The Cancer Imaging Archive serves as a centralized platform that consolidates WSI resources from a wide array of studies. Meanwhile, open-source initiatives such as OpenSlide provide standardized tools for image access and test samples, effectively lowering the technical barriers to implementation. While most of these resources are subject to data-use agreements for non-commercial research purposes, they exhibit variability in scanner specifications, staining procedures, and annotation criteria. This heterogeneity requires researchers to conduct strict data harmonization and cross-institutional validation to maintain the robustness and reproducibility of the resulting scientific insights.

Image preprocessing and quality control: Image preprocessing serves as an essential step to minimize technical variability and ensure analytical reliability. It primarily involves color normalization and quality control. Color normalization is designed to reduce variations in stain hue and intensity caused by differences in staining batches and scanning equipment. The prevailing algorithms fall into two principal categories: Classical statistical methods and deep learning-based approaches. Classical techniques, exemplified by Reinhard and Macenko, rely on color-space statistics or physical models. In contrast, deep learning methods, such as CycleGAN and StainNet, improve normalization precision and efficiency through domain adaptation or streamlined network designs.

In the architectural evolution of deep learning-based color normalization for pathological images, the generative adversarial network (GAN) framework has established a pivotal foundation[25]. Notably, unpaired imagetranslation models such as CycleGAN enable domain adaptation across different staining styles via cycle-consistency loss. They conceptualize staining discrepancies arising from distinct laboratories or scanners as separate “domains” to be harmonized[26]. To improve performance, Tellez et al[27] introduced structural-preservation constraints, perceptual loss and structural similarity loss, along with multi-scale generators and domain-adaptation strategies within the base GAN framework. These enhancements better maintain nuclear morphology and bolster the model’s generalization capacity. Subsequently, lightweight task-specific networks, typified by StainNet, have been introduced. These architectures avoid the complexities of adversarial training and instead adopt designs that incorporate shallow invertible structures and instance normalization. This approach preserves high normalization accuracy while substantially accelerating inference, thereby aligning well with the demands of real-time clinical workflows. Self-supervised and unsupervised learning pathways, through stain-consistency learning which constructs self-supervisory signals from serial tissue sections and contrastive learning frameworks, reframe the normalization task as a representation-learning problem. This paradigm allows models to learn stain-invariant features directly from the data, significantly diminishing the dependency on large-scale annotated datasets.

Image quality assessment: Automated or manual screening is performed to identify regions exhibiting issues such as out-of-focus blur, tissue folding, section tearing, or over-/under-staining, followed by a decision on whether to exclude these areas from subsequent analysis. Common types of image quality defects and their countermeasures can be found in Table 1[28-33].

Table 1 Types of quality defects and detection methods.

Defect type	Causes	Impact on analysis	Detection and evaluation methods
Out-of-focus blur	Autofocus failure or uneven tissue section thickness	Loss of cell boundaries and nuclear details, distortion of texture features	Traditional methods: Feature-based. Quantify blurriness by calculating image edge sharpness (e.g., Sobel gradient magnitude, Laplacian variance). Frequency-domain analysis. Detect the attenuation of high-frequency components using Fourier transform or wavelet transform[28]; deep learning methods: Train a binary convolutional neural network (e.g., ResNet, Inception) to intelligently classify the sharpness of image patches[29]
Tissue folding/tissue tearing	Improper sectioning or mounting	Introduces artificial textures and edges, interfering with segmentation	Traditional methods: Edge and morphological analysis. Perform edge detection on low-resolution overview images to identify abnormal edges (e.g., excessively long, overly straight, or sharply curved). Morphological operations are then applied to locate elongated “ridges” (folds) or “gaps” (tears). Texture consistency analysis: Calculate texture features of local image patches (e.g., gray-level co-occurrence matrix contrast, energy) and identify regions that significantly deviate from normal tissue through statistical testing; deep learning methods (mainstream): Segmentation models (e.g., U-Net): Input image patches and directly output pixel-level masks of defect areas[30]. Classification or anomaly detection models: Perform classification or detect anomalous regions[31,32]
Hyper-staining/hypo-staining	Improper staining time or concentration	Loss of contrast, bias in feature extraction	Staining quality assessment methods: Color distribution analysis. Calculate intensity histograms in specific color channels of hematoxylin and eosin-stained images (e.g., the blue channel corresponding to hematoxylin) to detect abnormal distributions that are excessively concentrated near extremes[32]. Contrast measurement. Evaluate the clarity of staining layers by calculating local or global contrast metrics (e.g., root mean square contrast, Weber contrast)
Air bubbles/blade marks	Mounting or sectioning defects	Misidentified as tissue structures	Traditional methods: Morphological and texture analysis can be combined for preliminary screening (similar to fold detection). Deep learning methods (mainstream): Segmentation or classification models. Identify defect regions through segmentation or classification[31,33]. Anomaly detection models: Detect defects as unseen anomalous structures[32]
Uneven illumination	Scanner light source issues	Intensity gradients within the same tissue	Traditional methods: Background analysis. Fit a brightness plane over pure background areas and detect significant brightness gradients by evaluating the fitting residuals. Tissue region analysis. Calculate the spatial distribution of brightness within uniform tissue regions to detect systematic gradients; deep learning methods: Classification or segmentation models can be used to directly learn and identify regions with abnormal brightness distribution[31]

The central aim of tissue-region detection is to precisely delineate biologically relevant tissue areas from non-informative background regions, thereby removing potential interference from the latter in subsequent quantitative analyses. Currently, algorithmic development in this area has progressed along three primary trajectories. The first category encompasses conventional image-processing techniques grounded in thresholding and morphological operations. These methods typically involve color-space transformation and intensity-channel utilization, coupled with adaptive thresholding algorithms like Otsu for initial segmentation, and subsequent morphological refinement of boundaries. While computationally efficient and effective for standard HE-stained images, their performance is often constrained in low-contrast or specially stained specimens (e.g., immunohistochemistry). The second category consists of traditional machine-learning approaches that rely on hand-engineered features. By incorporating expert-designed color, texture, and other descriptors, and employing classifiers such as support vector machines, these methods achieve enhanced robustness compared to purely threshold-based segmentation in specific contexts. The third category is represented by deep-learning-based strategies, with the U-Net semantic segmentation network standing as the dominant paradigm. Its encoder-decoder architecture facilitates effective multi-scale feature fusion. Trained on a modest set of manually annotated mask data, the model produces pixel-precise binary segmentation masks and excels particularly in delineating tissue regions characterized by intricate borders and structural heterogeneity.

To address the memory constraints imposed by the ultra-high resolution of WSIs, a multi-scale pyramid strategy has emerged as an efficient computational solution. This hierarchical processing pipeline first conducts rapid coarse segmentation at a lower-resolution pyramid level (e.g., down-sampled to 5 × objective magnification) to identify candidate tissue regions. Subsequently, detailed segmentation is performed only within these candidate areas at higher resolutions (e.g., 20 × objective), dramatically reducing memory overhead. The final tissue mask is obtained by integrating results across scales. This approach achieves an effective trade-off between computational efficiency and segmentation precision, and has evolved into a standard methodology for large-scale WSI processing, thereby providing a robust spatial framework for downstream quantitative pathological analyses.

Tissue and region segmentation (region of interest definition)

Tissue and cell segmentation serves as a pivotal step in transforming macroscopic images into quantitative biological data. The goal is to precisely decompose a WSI into discrete analytical units that carry explicit biological significance. This process consists of three interconnected core phases. Tissue semantic segmentation leverages deep-learning architectures, notably U-Net and its derivatives, to perform pixel-wise classification of various tissue compartments, such as tumor parenchyma, stroma, necrotic foci, and immune-infiltrated regions. This lays the groundwork for region-specific analytical workflows[34]. Nuclear segmentation employs dedicated models like StarDist and HoVer-Net[35] to precisely segment individual nuclei while concurrently categorizing cell types (e.g., tumor cells, lymphocytes, fibroblasts), thereby supplying direct data for immune microenvironment characterization. Region-of-interest definition involves selecting an appropriate sampling strategy, such as whole-slide analysis, random multi-field sampling, or targeted region assessment, according to the study aims. It also necessitates standardized protocols to minimize selection bias.

Feature extraction

Feature types: The essence of pathomics feature extraction is the systematic quantification of tissue information across multiple organizational levels, primarily organized into three broad feature categories. Morphological features concentrate on the geometric attributes of individual objects, precisely capturing the microscopic form through measurements such as area, perimeter, and equivalent diameter of nuclei or glands, as well as shape descriptors like circularity, elliptical eccentricity, and fractal dimension. Derivation of these features is highly dependent on accurate pixel-level segmentation. Additionally, it’s important to recognize that two-dimensional tissue sections are projections of three-dimensional structures, which can introduce inherent measurement biases.

Texture features aim to characterize the spatial distribution patterns of pixel intensities within tissue regions, providing a mathematical representation of structural heterogeneity and spatial organization. Classical approaches include the gray-level co-occurrence matrix, which derives metrics such as contrast and homogeneity by quantifying the spatial relationships between pixel pairs; local binary patterns, which encode neighborhood intensity variations; and Gabor filter banks, which mimic the multi-scale, multi-orientation sensitivity of the visual system to capture textural information. Additionally, deep-learning methods can extract more abstract texture representations from convolutional neural network (CNN) activation maps. Appropriate parameter selection is required based on the analysis scale, whether focused on cellular or tissue-level structures.

Topological and spatial-relational features provide a higher-dimensional quantification of spatial interaction patterns within tissue architecture. Graph-theoretic approaches depict nuclear centroids as nodes in spatial networks, deriving metrics such as clustering coefficients and characteristic path lengths to describe the collective spatial organization of cellular ensembles[36]. Spatial-statistical measures, including Ripley’s K function, facilitate multi-scale analysis of spatial point-pattern distributions. Inter-regional relational features quantify spatial adjacency and interface characteristics across distinct histological compartments. Collectively, these features provide critical quantitative descriptors for investigating the spatial ecology of the TME[37].

Feature aggregation and standardization: The aim of feature aggregation is to condense extensive cell- or region-level features into sample-level descriptors. For example, area measurements from tens of thousands of nuclei can be aggregated into distributional statistics such as the mean, standard deviation, skewness, and kurtosis[38]. The selection of an aggregation strategy should be tightly coupled with biological rationale: The mean captures the central tendency, the median provides robustness to outliers, whereas histogram binning or complete distribution characterization retains richer information about population heterogeneity[12].

Subsequently, feature standardization is applied to eliminate scale discrepancies and ensure comparability among different features in subsequent analytical steps. The most widely adopted Z-score standardization transforms features to a distribution with a mean of zero and a unit variance, whereas min-max normalization linearly rescales features to the [0, 1] range. Crucially, standardization parameters (e.g., mean, SD) must be estimated solely from the training set and then consistently applied to the validation and test sets, a core principle for maintaining model generalizability.

Confronted with high-dimensional feature spaces, feature dimensionality reduction and selection become essential steps. Filter methods conduct rapid preliminary screening based on variance or correlation metrics. Wrapper approaches, such as recursive feature elimination, iteratively assess feature subsets according to model performance. Embedded methods integrate feature selection directly into the model training process, as exemplified by LASSO regression or tree-based algorithms[39]. Unsupervised dimensionality reduction techniques like principal component analysis can effectively reduce dimensionality, though often at the cost of diminishing the direct biological interpretability of the features[40]. Quality control procedures include visualizing feature distributions to identify anomalous patterns, projecting numerical results back onto tissue space using feature heatmaps for spatial validation, and cross-validating findings with pathological domain knowledge. These steps ensure that the extracted features possess not only statistical significance but also well-defined pathological-biological relevance[8].

Data analysis, modeling and validation

This phase constitutes the value-actualization step in pathomics, centered on establishing robust association models between high-dimensional omic features and clinical endpoints, including diagnostic categories, pathological grades, survival outcomes, and therapeutic responses, using rigorous statistical and machine-learning frameworks. The entire workflow must be built upon principled data governance, deliberate model selection, and rigorous validation to ensure both the robustness of the derived insights and their translational potential.

Data partitioning: Data partitioning constitutes the first line of defense against model overfitting and inflated performance estimates. All specimens derived from a given patient, including any associated WSIs or image tiles, should be assigned as a complete unit to either the training, validation, or independent test set. Strict separation must be maintained to guarantee that no samples from the same patient appear in more than one partition, thereby eliminating data leakage caused by intra-patient correlation[41]. The validation set is employed for hyperparameter optimization and model selection, whereas the independent test set is set aside for evaluating the model’s final performance on unseen data. The most stringent form of validation is derived from an external test set, comprising data collected from different healthcare institutions, using varied scanners and staining protocols, which constitutes the “gold standard” for assessing model generalizability and clinical applicability[42]. Furthermore, for imbalanced datasets (e.g., rare subtypes), stratified sampling should be applied during partitioning to preserve the proportional representation of each class.

Model construction: Traditional machine-learning pipeline. This approach employs manually engineered features, often following dimensionality reduction, as input. Logistic regression is commonly applied to examine associations between selected features and clinical endpoints; support vector machines exhibit robustness in high-dimensional, low-sample-size scenarios; and ensemble learners such as random forests, XGBoost, and LightGBM automatically handle nonlinear relationships and feature interactions, generally achieving higher predictive performance while providing interpretability through featureimportance rankings. This route offers computational efficiency, modest sample-size demands, and comparatively transparent model decision processes.

End-to-end deep learning. This paradigm directly consumes raw image patches or entire WSIs as input, leveraging CNNs, vision transformers, or multiple-instance learning frameworks to autonomously learn discriminative feature representations and generate predictions. This strategy can reveal intricate image patterns that transcend conventional human-defined priors, holding considerable promise. Its main challenges include the demand for large-scale, meticulously annotated datasets; high computational costs during training; and limited interpretability as a “black-box” model, which impedes clinical acceptance. Recently, the integration of attention mechanisms and explainable AI methodologies has been actively pursued to address this interpretability gap.

Model evaluation, validation, and interpretability: Model evaluation should be rigorously aligned with the task type. Classification tasks typically report accuracy, precision, recall, F1-score, and area under the curve (AUC)-receiver operating characteristic; survival analysis uses the C-index; regression tasks focus on metrics such as mean squared error. Reporting should include confidence intervals where appropriate. Beyond performance metrics, model interpretability is essential for clinical adoption. Approaches like SHapley Additive exPlanations and Local Interpretable Model-agnostic Explanations can quantify the contribution of individual features to specific predictions. Meanwhile, feature visualization, for instance, overlaying important features as heatmaps on the original histology images, enables pathologists to comprehend and trust the model’s reasoning and may even uncover novel morphological biomarkers. Ultimately, the highest-level validation of a model derives from prospective clinical trials that demonstrate its tangible improvement of clinical decisions or patient outcomes[43].

PATHOMICS IN CLINICAL DECISION-MAKING FOR PDAC

As one of the most aggressive solid tumors, PDAC poses a major clinical challenge in achieving precise diagnosis and treatment decision-making. Traditional pathology, which relies on manual microscopic examination, is inherently limited by subjectivity, low reproducibility, and difficulties in quantification. However, with the deep integration of whole-slide digital scanning and AI, pathomics has emerged as a transformative approach. By mining large-scale pathological image data, it extracts sub-visual morphological features beyond human perception, offering a revolutionary tool for precise diagnosis, tailored treatment guidance, and accurate prognosis prediction in PDAC. This paper systematically reviews the application of pathomics in the clinical decision-making workflow for PDAC. It delineates an integrated framework by examining its role in automated diagnostic subtyping, treatment response prediction, and individualized prognostic assessment, while also outlining future directions for the field.

Pathomics for precise diagnosis and subtyping of PDAC

Automated tissue segmentation: Pathological diagnosis of PDAC has traditionally relied on the subjective evaluation of HE-stained slides by pathologists, facing inherent limitations in quantification, standardization, and efficiency. Key challenges include inter- and intra-observer variability, ambiguous TME boundaries, and the inability to quantify continuous variables such as stromal ratio and lymphocyte density. To overcome these barriers, deep learning-based automated tissue segmentation has emerged, transforming pathological assessment from qualitative description to objective, quantitative analysis[8].

Semantic segmentation networks, such as U-Net and DeepLab variants, enable pixel-wise classification of images into specific tissue types[44]. Initial efforts primarily focused on binary segmentation, such as separating tumor from non-tumor regions, demonstrating the feasibility of AI for basic diagnostic tasks. However, the highly heterogeneous TME of PDAC contains diverse components including tumor cells, stroma rich in CAFs, immune cells, nerves, and blood vessels. Merely distinguishing tumor from non-tumor areas is inadequate for advanced research and clinical needs. Therefore, recent studies have shifted toward multi-class segmentation, with key challenges including morphological similarities between tissue types and severe class imbalance, such as vast stromal regions alongside minute structures like nerves and pancreatic islets[45].

To address these challenges, researchers have developed several innovative network architectures. For instance, Gao et al[46] proposed Selective Multi-Scale Attention Network, which preserves high-resolution details and enhances discriminative features, achieving accurate segmentation of five tissue types, malignant epithelium, blood vessels, nerves, pancreatic islets, and ducts, on an in-house PDAC dataset (mean Dice: 0.769). Chen et al[47] introduced channel-wise and spatial self-attention modules that effectively capture long-range dependencies in extensive tissue regions (e.g., stroma) within pathological images at a low computational cost, thereby mitigating the limited receptive field of conventional CNNs. Notably, these methods also demonstrate strong generalization performance on public benchmark datasets such as the Gland Segmentation Challenge dataset for colorectal histopathology[46,47].

Automated multi-tissue segmentation offers significant clinical value. It standardizes the quantitative analysis of the TME, forming the basis for deriving prognostic biomarkers such as the tumor-stroma ratio (TSR), an established predictor of lymph node metastasis and poor prognosis[48]. Additionally, it reduces the manual workload of pathologists, enhancing efficiency. Ultimately, this objective, reproducible tool provides standardized inputs for advanced prognostic and treatment response models, advancing pathology into a data-driven era.

Molecular subtyping via morphological signatures: The molecular subtypes of PDAC, especially the basal-like and classical subtypes, demonstrate significant clinical heterogeneity. The basal-like subtype is characterized by greater invasiveness, poorer prognosis[49], and potentially lower responsiveness to standard chemotherapy (e.g., FOLFIRINOX)[50]. Consequently, accurate preoperative subtyping is critical for personalized treatment planning. However, traditional molecular subtyping techniques, such as RNA sequencing, are often limited by high costs, lengthy turnaround times (typically several weeks). As a result, they are ill-suited to support timely clinical decision-making in the rapidly progressing context of PDAC.

To address this challenge, research has shifted toward leveraging routinely available and cost-effective HE-stained pathological slides as a readily accessible medium for predicting molecular subtypes. The key methodological advancement involves the adoption of weakly supervised and multiple-instance learning frameworks. These models operate without pixel- or region-level annotations, relying solely on patient-level molecular subtype labels to autonomously learn discriminative deep morphological patterns associated with each subtype. Through integrated attention mechanisms, they can automatically localize and focus on the most informative tissue regions. Such approaches have demonstrated strong predictive performance in subtype classification tasks. Ahmadvand et al[45] reported an AI classifier that achieved 83.03% accuracy in distinguishing PDAC molecular subtypes on an independent test set, with a positive predictive value of up to 1.0 for the basal-like subtype. Visualization via explainable AI techniques revealed that the model’s decision-making process aligns closely with established pathological principles. Differentiation was primarily driven by gland formation patterns. The classical subtype is characterized by attention to regions with well-formed glandular structures, whereas the basal-like subtype conversely focuses on areas that exhibit solid growth patterns with absent or poorly developed glands. These findings confirm that the AI model effectively captures key morphological phenotypes linked to molecular subtypes.

The most significant translational progress of this technology is demonstrated by its successful validation on preoperative biopsy samples. Even when applied to biopsy specimens with limited tissue volume, the optimized model retains robust subtyping capability[51]. This indicates that during the early diagnostic phase, the AI system can rapidly predict molecular subtypes from routinely prepared biopsy slides in a minimal amount of time. As a result, molecular subtyping transitions from an expensive, time-consuming specialized assay to a widely applicable and rapid ancillary pathological diagnostic procedure, offering an efficient and accessible pathway toward achieving true precision medicine.

Pathomics for treatment guidance and response prediction

Neoadjuvant therapy response prediction: Neoadjuvant therapy (NAT) has established a clear role in treating PDAC by improving surgical outcomes and survival. However, treatment responses are markedly heterogeneous. Consequently, nearly half of all patients are exposed to the toxicity of ineffective treatment and the risk of disease progression. This reality underscores the critical need for precise predictive tools to enhance the efficacy of NAT and guide individualized therapy. Current standard assessments, which rely primarily on post-treatment imaging and declines in carbohydrate antigen 19-9 (CA19-9), are limited by their inherent delay and suboptimal specificity[52,53].

To address current prediction bottlenecks, multimodal fusion offers a necessary path forward. Watson et al[54] showed that a deep learning model using only pre-treatment computed tomography images predicted the pathological response with an AUC of approximately 0.738. However, a hybrid model combining computed tomography features with a dynamic drop in CA19-9 (≥ 10%) improved the performance to an AUC of 0.785. This enhancement reveals a critical insight: Static imaging features reflect the intrinsic spatial heterogeneity of the tumor, whereas dynamic biochemical indicators capture the systemic response of the tumor to therapy, thereby providing complementary information across different temporal and biological dimensions. Li et al[55] employed digital pathology and spatial point pattern analysis to quantify the spatial correlation between CD8+ cytotoxic T cells and cancer cells. Their findings showed that in PDAC patients receiving NAT, CD8+ T cells exhibited significantly greater spatial clustering with cancer cells compared to those undergoing upfront surgery. This suggests that NAT may enhance effector T-cell infiltration into tumor tissue and direct tumor cell contact, potentially improving immune surveillance and cytotoxicity. Notably, this enhanced spatial correlation was independently associated with longer postoperative survival.

This successful paradigm demonstrates a clear direction for the advancement of pathomics. The predictive power of pathomics models fundamentally depends on the feature extraction capabilities of deep learning. At the pathological level, deep learning enables the extraction of digitized morphological patterns from pre-treatment HE-stained biopsy slides, including the complexity of tumor glandular architecture, the organization of stromal collagen, and the spatial distribution of immune cells. Research has confirmed that a Pathomics Score derived from pre-treatment pathological slides can effectively differentiate patients with a favorable pathological response (tumor regression grade 0-1) from those with a poor response (tumor regression grade 2-3) following NAT[56]. Looking forward, the most promising and clinically translatable predictive framework will be a deeply integrated system that fuses multi-source data. Such a system would synthesize quantitative deep-learning features from pre-treatment pathology images, dynamic liquid-biopsy biomarker profiles tracked during treatment [e.g., clearance kinetics of circulating tumor DNA (ctDNA), evolving molecular phenotypes of circulating tumor cells, specific microRNA signatures carried by extracellular vesicles], and patients’ baseline clinical information to build a multidimensional, dynamic, and intelligent prediction platform.

Risk evaluation for lymph node metastasis: Lymph node metastasis status represents one of the most critical prognostic indicators in PDAC and serves as a key determinant for pathological staging. Chen et al[51] established the TSR, a morphological parameter readily assessable from routine HE slides, as a strong, independent biomarker for predicting lymph node metastasis. The prognostic utility of TSR has been consistently validated across large-scale, multicenter studies. Notably, tumors with a low TSR (indicative of high stromal content) show a significantly elevated risk of lymph node metastasis. In the training cohort, TSR achieved an AUC of 0.749 (sensitivity 76.5%, specificity 71.6%) for predicting lymph node involvement, and its performance remained robust in an independent validation cohort. Clinically, the relevance of TSR extends beyond surgical resection specimens, demonstrating comparable predictive efficacy in pretreatment endoscopic ultrasound-guided fine needle aspiration biopsy samples (AUC = 0.747). Indeed, this indicates that TSR can reliably assess the risk of lymph node metastasis, even in limited preoperative tissue samples. The biological rationale for TSR-based prediction stems from the stroma-rich TME, which promotes tumor cell invasion and migration and may concurrently limit chemotherapy penetration.

Integrating TSR into combined models can further improve its predictive performance. While informative as a standalone biomarker, TSR achieves greater power when combined with other clinical or molecular dimensions. For example, pairing TSR with programmed death ligand 1 (PD-L1) expression status increases the AUC for lymph node metastasis prediction from 0.749 to 0.793. High PD-L1 expression typically reflects an immunosuppressive microenvironment, which together with the pro-invasive stroma indicated by a low TSR defines a tumor phenotype with heightened metastatic risk. Similarly, combining TSR with elevated serum CA19-9 levels (> 37 U/mL) also enhances predictive accuracy (AUC = 0.777)[51]. Such multimodal integration, linking local tissue architecture with systemic tumor burden markers, supports a more holistic clinical risk assessment.

Therefore, as a low-cost, easily accessible, and reproducible biomarker, TSR offers considerable potential for refining preoperative staging in PDAC. In patients presenting with both a low TSR and either PD-L1 positivity or elevated CA19-9 levels, a high suspicion of occult lymph node metastasis is warranted, thereby strongly indicating the need for NAT[51].

Immune microenvironment dynamic profiling: The significant heterogeneity, fibrotic stroma, and immunosuppressive microenvironment of PDAC pose major challenges for both treatment and evaluation. Pathomics provides a pivotal approach to address these complexities. AI-driven automated multi-tissue segmentation of WSIs enables precise quantification of immune cell spatial distribution, facilitating the classification of classic immune phenotypes, such as “immune-desert”, “immune-excluded”, and “immune-inflamed”, and allowing for objective immune phenotyping along with quantification of immunosuppression in PDAC[57]. The integration of multiplex immunofluorescence with cell-identification algorithms further supports spatial mapping and quantification of intratumoral immune subsets, including dendritic cells, CD4+ T cells, and diverse immunomodulatory cell types[58]. Spatial proximity analysis, such as measuring the nearest-neighbor distance between CD8+ T cells and tumor cells, offers a quantifiable metric for assessing immune activity at the tumor-immune interface, or “killing front”[59]. Additionally, automated detection and grading of tertiary lymphoid structures provide morphological insights into adaptive immune responses[60]. Collectively, these high-throughput and reproducible quantitative measures form a comprehensive digital framework for evaluating the immunosuppressive landscape of the PDAC microenvironment.

Pathomics provides a distinct advantage in tracking the evolving immune microenvironment of PDAC throughout neoadjuvant or immunotherapy. In predicting treatment efficacy, pathomic analysis of post-neoadjuvant surgical specimens shows that increased CD8+ T-cell density in the tumor core, mature tertiary lymphoid structures (with germinal centers), and a higher proportion of M1-macrophages are significantly correlated with prolonged disease-free survival[61]. Regarding resistance mechanisms, analysis of progressive post-treatment samples quantitatively captures immune “re-suppression”, marked by accumulations of regulatory T cells (Tregs) and myeloid-derived suppressor cells, as well as increases in specific CAF subsets [e.g., inflammatory CAFs (iCAFs)] that enhance stromal barrier function[61]. Dynamically monitored pathomic profiles can further guide combination strategies: If chemotherapy remodels stroma without boosting T-cell infiltration, adding immune checkpoint inhibitors may be warranted; if tertiary lymphoid structures form but remain functionally suppressed, pairing with immune agonists could be beneficial[62]. Importantly, integrating serial pathomic features with dynamic biomarkers such as ctDNA allows early prediction of long-term therapeutic response during treatment[8].

Pathomics for prognostic stratification and recurrence prediction

Development of a novel prognostic scoring system: Prognostic evaluation in PDAC must advance beyond the traditional tumor-node-metastasis (TNM) staging system to better capture its marked biological heterogeneity. Pathomics is enabling the construction of next-generation prognostic scoring systems through the quantitative analysis of complex morphological features within the TME. Initial studies emphasized biologically interpretable morphological ratios, including tumor-infiltrating lymphocytes and TSR[63]. The field has since evolved from relying on single quantitative indicators toward integrating multifaceted, data-driven composite scores.

Liu et al[64] confirmed the critical role of pathomics in PDAC prognosis and immune profiling using multicenter clinical data. Their study showed that pathomic features extracted from tumor and adjacent HE images significantly enhance the prediction of postoperative overall survival, with a LASSO-based composite model performing optimally. Notably, the derived Pathscore independently predicted prognosis beyond conventional TNM stage and CA19-9 levels. Mechanistically, high Pathscore was linked to an immunosuppressed microenvironment, characterized by reduced CD8+ T-cell infiltration, lower immune scores, and disrupted immune spatial architecture. Integrated transcriptomic, single-cell, and spatial analyses further associated the Pathscore with expression patterns of immune-related genes including ITGAX, ACTA2, and CD8A. This work offers a quantifiable, non-invasive prognostic tool for pancreatic cancer while uncovering immune-related pathways underlying pathomics-based prediction.

In a parallel effort, Liu et al[56] established a deep learning-based pathomics scoring system (Pathomics Score) using whole-slide HE images from 864 PDAC patients. Using weakly supervised and multi-instance learning methods, the model distilled 7 prognostic features from an initial set of 206, generating an independent pathomics score. Multivariable Cox regression demonstrated its strong and independent prognostic value, with hazard ratios of 3.90 in the training cohort, 3.49 in the validation cohort, and 4.82 in the NAT cohort (all P < 0.001). Furthermore, a predictive model incorporating this score achieved an AUC of 0.81 for 2-year survival prediction in the validation cohort, with a concordance index of 0.74, significantly outperforming a clinical model that used only traditional pathological factors. This demonstrates that the deep-level information encoded in tumor morphology can effectively capture the biological heterogeneity strongly associated with survival, a dimension not reflected by the TNM system. By merging data-driven insights with clinical knowledge, pathomics-based prognostic scoring systems create a quantitative digital profile of intrinsic tumor aggressiveness. This provides an objective tool for granular patient risk stratification and establishes a precise basis for tailoring personalized adjuvant therapy and follow-up strategies.

Evaluation of multidimensional prognostic models: With the rapid advancement of pathomics technology, prognostic prediction models for PDAC have evolved from single-dimensional frameworks into multidimensional, multi-layered systems. Current cutting-edge models now converge and compete along three principal axes: The composition of input data, the spatial scale of analysis, and the temporal dynamics of prediction.

Regarding data composition, there has been a clear trend has shifted from models relying solely on pathology toward integrated frameworks that fuse pathological and clinical features. Pure pathology models are constructed exclusively from information derived from pathological images. Their core strength lies in their directness: They quantify the intrinsic biological phenotype of tumor tissue without influence from clinical confounders such as patient age, performance status, or treatment choice, thus offering a more direct measure of the tumor’s inherent aggressiveness. However, patient outcomes ultimately arise from the interplay between tumor biology, host factors, and therapeutic interventions. Integrated clinico-pathological models address this gap by incorporating conventional clinicopathological variables, including TNM stage, tumor grade, and CA19-9 levels. For example, the combined model developed by Hu et al[65] achieved a C-index of 0.77 in the test set, outperforming both a clinical-only model (0.75) and a pathology-only model (0.73).

In spatial resolution, current model development is shifting from macro-scale tissue quantification toward micro-scale cellular spatial architecture analysis. Tissue-level approaches focus on area ratios or macroscopic distributions of tissue components, such as TSR, lymphocyte-to-stroma ratio, and lymphocyte-to-tumor ratio. Although computationally straightforward and clinically practical, these metrics lack insights into critical cell-to-cell interactions. By contrast, cellular-level spatial models represent a more advanced frontier. Using spatial pathology techniques, they assess the spatial proximity, clustering patterns, and interactions between specific cell types, such as immune cells, and cancer cells or stromal elements. For example, spatial enrichment of CD8+ T-cells in the tumor core has proven to be a stronger favorable prognostic marker than diffuse infiltration along the invasive margin[59]. This strategy enables a more fundamental characterization of the functional state of the tumor immune microenvironment.

Along the temporal axis of prediction, model development is moving from static, single-time-point assessments toward dynamic approaches that integrate longitudinal multi-time-series data. Currently, most existing models are static, relying on a single biopsy or surgical specimen collected at diagnosis or surgery to forecast long-term outcomes. While useful for baseline risk stratification, these models cannot account for tumor evolution during treatment. In contrast, dynamic monitoring models represent the logical next step, centered on the integration of longitudinal data sampled at multiple time points throughout therapy. This includes comparative pathology analysis before and after neoadjuvant treatment (capturing therapyinduced morphological shifts), serial liquid biopsies (e.g., tracking dynamic clearance of ctDNA mutant allele frequency), and sequential imaging. Although more complex to develop, their potential is significant: They enable updated risk assessment, early detection of recurrence, real-time evaluation of therapeutic response, and ultimately, adaptive clinical decision-making. Preliminary work analyzing spatial remodeling of the immune microenvironment before and after therapy to predict survival already offers a methodological blueprint for building such dynamic prognostic frameworks[55].

SINGLE-CELL OMICS APPROACHES IN PDAC RESEARCH

ScRNA-seq

ScRNA-seq is a high-throughput technique that enables transcriptome-wide analysis at single-cell resolution. Its general workflow encompasses single-cell isolation, RNA capture and reverse transcription, complementary DNA amplification, library construction, and high-throughput sequencing. Currently prevalent scRNA-seq platforms can be broadly divided into two categories: Droplet-based high-throughput systems (e.g., 10× Genomics Chromium) and full-length transcript-focused platforms (e.g., Smart-seq2). The former is optimal for profiling large cell numbers, whereas the latter offers higher sensitivity and is better suited for lower-throughput applications. Additionally, snRNA-seq is widely used in PDAC studies, particularly for frozen or difficult-to-dissociate tissue specimens. More recently, spatial transcriptomics technologies, including 10× Visium and GeoMx DSP, have emerged to overcome the spatial information loss associated with conventional scRNA-seq. These approaches allow gene expression analysis while maintaining tissue spatial context. Together, these integrated methodologies provide an unprecedented toolkit for systematically dissecting the cellular composition, functional states, and spatial architecture of the TME in PDAC.

Recent advances in scRNA-seq for PDAC

In PDAC research, scRNA-seq has emerged as a pivotal technology for elucidating key molecular mechanisms. The application of scRNA-seq has propelled PDAC research across several interconnected frontiers. These include the characterization of precursor lesions, the deconvolution of intra-tumoral heterogeneity, the high-definition mapping of the TME, the tracing of metastatic trajectories, and the elucidation of drug resistance mechanisms. Among these, the detailed characterization of cellular heterogeneity and TME composition serves as the cornerstone for understanding the subsequent, more complex processes of metastasis and therapy failure.

Decoding cellular heterogeneity and TME alterations in PDAC: ScRNA-seq has uncovered extensive heterogeneity within the ductal cell compartment of PDAC. By applying methods such as consensus non-negative matrix factorization and copy number variation analysis, ductal cells have been categorized into distinct molecular subtypes, including classical, basal-like, quasi-mesenchymal, progenitor-like, and normal epithelial-like subtypes. Among these, the basal-like subtype is strongly linked to poor prognosis and chemotherapy resistance. Specifically, Park et al[63] identified that a basal-like cell proportion of ≥ 22% serves as a prognostic threshold and correlates with an immunosuppressive TME. Furthermore, epithelial-mesenchymal transition appears to be subtype-specific, with classical and basal-like cells activating epithelial-mesenchymal transition-related gene programs through different transcriptional networks[66,67].

Integrated scRNA-seq and spatial transcriptomics offers a detailed dissection of the PDAC TME. CAFs have been further stratified into distinct subsets, including iCAF, myofibroblastic, antigen-presenting, and complement-secreting subtypes, each contributing differentially to tumor progression, immune modulation, and therapy resistance[68]. The resulting immune cell atlas delineates mechanisms driving the accumulation of immunosuppressive populations such as Tregs, M0/M2-polarized macrophages, and exhausted T cells. Du et al[69] demonstrated that basal-like ductal cells and M0 macrophages communicate through the CXCL14-CXCR4 axis, fostering an immunosuppressive niche. Smith et al[70] proposed that targeting intratumoral Tregs is key to overcoming immunotherapy resistance in PDAC. They suggest that precisely modulating Tregs, via markers such as CCR8 or TIGIT, to remodel the immunosuppressive microenvironment holds considerable therapeutic promise. Additionally, Saleem et al[71] documented pronounced T-cell exhaustion in PDAC and uncovered novel molecular markers linked to this dysfunctional state, identifying 16 core hub genes that regulate the exhaustion program. Collectively, these studies provide novel, single-cell-resolved insights into the PDAC TME.

Dissecting metastatic and differentiation mechanisms in PDAC: Park et al[63] performed scRNA-seq on paired samples from 21 treatment-naive PDAC patients and observed that tumor clonality increased progressively from non-metastatic primary lesions to liver metastases, indicative of dominant clone selection. This evolutionary path was subtype-specific: Basal-like tumors progressively acquired KRAS copy-number gains along with losses in SMAD2 and MAP2K4, while classical tumors mainly exhibited ETV1 copy-number gains. These findings suggest that different molecular subtypes of PDAC, KRAS- and ETV1-driven, promote metastatic progression through distinct transcriptional alterations, highlighting a subtypedependent evolutionary paradigm.

Wang et al[72] discovered that early-stage tumor-associated macrophages secrete tissue inhibitor of metalloproteinases 1, which binds to CD63 on the surface of pancreatic cancer cells, thereby activating the extracellular signal-regulated kinase signaling pathway and suppressing dual specific phosphatase 2 within the cancer cells. Tissue inhibitor of metalloproteinases 1-CD63 signaling drives immune evasion and metastasis in KRAS-mutant pancreatic cancer.

Song et al[73] reported that in PDAC, the cell-in-cell death process known as entosis does not lead to simple cell death but rather selects for a highly invasive cancer cell subpopulation characterized by elevated neuroepithelial transforming gene 1 expression. This entotic activity is enhanced in liver metastases compared to primary tumors and correlates directly with worse clinical outcomes. Thus, entosis and its key effector neuroepithelial transforming gene 1 emerge as a novel mechanism and potential therapeutic target that promotes aggressive progression and metastasis, particularly to the liver, in PDAC.

Unveiling mechanisms of therapeutic resistance: Treatment resistance represents a primary driver of poor clinical efficacy in PDAC. By delineating the unique transcriptional profiles of resistant cell populations and their dynamic interplay with the TME, scRNA-seq offers essential insights into the mechanisms of resistance to chemotherapy, targeted agents, and immunotherapy. For example, Principe et al[74] applied scRNA-seq to uncover a gemcitabine-resistant subpopulation of pancreatic cancer cells, distinguished by pronounced activation of calcium/calmodulin signaling. Follow-up studies demonstrated that calcium channel blockers can augment gemcitabine response, pointing to a feasible strategy for overcoming this form of resistance. Cui et al[75] showed that in patients receiving gemcitabine and paclitaxel, iCAFs upregulate metallothionein gene expression, a feature linked to multi-drug chemoresistance and potentially indicative of active resistance pathways.

Hypoxia is a prevalent and critical feature of PDAC, strongly associated with tumor invasion, metastasis, and therapeutic resistance. ScRNA-seq analyses have delineated adaptive cellular state changes within these hypoxic niches. In particular, hypoxia-inducible factor-1α (HIF-1α), the master transcriptional regulator of the hypoxic response, is markedly upregulated in pancreatic cancer and orchestrates a range of malignant phenotypes that fuel tumor progression. Raghavan et al[76] defined a key cellular state in PDAC, the “stress-like state”, induced by microenvironmental stresses, notably hypoxia and nutrient scarcity. This state is characterized by robust activation of the HIF-1α signaling axis and the glycolytic reprogramming it orchestrates. The stress-like state becomes markedly enriched after chemotherapy, where it promotes treatment resistance by bolstering processes like DNA damage repair. It also engages with specific subsets of CAFs to jointly sustain an immunosuppressive niche. Together, these single-cell findings reveal how the hypoxia-HIF-1α axis fosters PDAC adaptability and chemoresistance by establishing a plastic cellular program, providing a compelling rationale for therapeutic targeting of the TME.

Beyond this, single-cell and spatial omics are instrumental in unraveling the cellular and molecular basis of immune escape during PDAC treatment. For example, Pan et al[77] reported that elevated CD47 expression in pancreatic cancer is linked to poor prognosis, and targeting CD47 can reprogram the tumor immune microenvironment, specifically, in sensitive Panc02 models, it polarizes macrophages to a pro-inflammatory, anti-tumor state and activates CD8+ T cells. However, therapeutic efficacy was model-dependent: Anti-CD47 alone or combined with anti-PD-L1 only worked in select models (Panc02 or MPC-83), highlighting that the outcome is determined by the differential immune responsiveness of distinct TMEs. Thus, the therapeutic potential of CD47 blockade is strongly contingent on the preexisting immune context of the tumor. Table 2 provides a consolidated overview of the representative studies discussed above, representing the key contributions of scRNA-seq to our current understanding of PDAC biology.

Table 2 Key advances in pancreatic ductal adenocarcinoma research using single-cell RNA sequencing in recent years.

Research direction	Key findings	Key techniques	Ref.
Precursor lesions and early detection	High transcriptional similarity between pancreatic intraepithelial neoplasia and pancreatic ductal adenocarcinoma; NKX6-2 drives gastric-like differentiation in intraductal papillary mucinous neoplasm	Spatial transcriptomics	[7,13]
Tumor heterogeneity	Identification of distinct ductal cell subtypes (classical, basal-like, normal-like, cycling); basal-like subtype proportion ≥ 22% is associated with poor prognosis	Consensus non-negative matrix factorization, copy number variation analysis, pseudotime trajectory inference	[66-68]
Tumor microenvironment	Discovery of CAF subtypes (myofibroblastic CAFs, inflammatory CAFs, antigen-presenting CAFs); characterization of immunosuppressive roles of Tregs and M0/M2 macrophages	Cell-cell communication analysis (CellPhoneDB), spatial transcriptomic integration	[69-72]
Metastasis and evolution	Increased clonal homogeneity in liver metastases, accompanied by KRAS/ETV1 amplification and SMAD2/MAP2K4 deletion; identification of metastasis-related transcriptional modules	Single-cell copy number variation analysis, non-negative matrix factorization-based module detection	[66,73,74]
Therapeutic resistance	Upregulation of calcium signaling in gemcitabineresistant cells; HIF-1α-mediated angiogenesis in hypoxic niches; TAM polarization associated with immunotherapy resistance	Differential gene expression analysis, pathway enrichment, validation in animal models	[63,75,76]

CAF: Cancer-associated fibroblast; HIF-1α: Hypoxia-inducible factor-1α; TAM: Tumor-associated macrophage; MAP2K4: Mitogen-activated protein kinase kinase 4.

INTEGRATIVE PATHOMICS AND SINGLE-CELL OMICS ANALYSIS OF PDAC

Although pathomics has shown substantial promise for clinically relevant phenotyping in PDAC, morphology-based models alone often provide limited insight into the underlying cellular states and molecular programs that give rise to these histomorphologic phenotypes. Conversely, single-cell omics provides high-resolution molecular profiles but lacks spatial context. Therefore, integrating these two modalities offers a complementary framework that bridges morphological phenotypes with molecular programs, thereby deepening the interpretability and translational value of both approaches. In the context of PDAC, extensive desmoplasia and pronounced spatial heterogeneity impose unique constraints on both imaging-only and dissociation-based approaches, making multimodal integration particularly valuable.

Advantages and rationale for integrating pathomics with single-cell analyses

PDAC differs from most solid tumors in one fundamental respect: The stroma can comprise 70%-90% of the tumor tissue, and tumor cells and stromal populations (including fibroblasts and immune cells) are arranged in a highly heterogeneous spatial manner. Tumor-stroma crosstalk is a key driver of malignant progression. Given these histologic features, no single omics approach can simultaneously capture tissue-level spatial architecture and the cell-resolved molecular programs that underpin it. Pathomics enables rigorous quantification of spatial morphologic phenotypes in PDAC, such as glandular architectural distortion and the degree of desmoplasia, but it cannot explain which cell types and molecular states give rise to these patterns. Single-cell omics, by contrast, resolves cellular-state diversity and regulatory rewiring in PDAC; however, the dense desmoplastic stroma creates practical barriers that make it difficult to reliably map dissociated single-cell profiles back onto defined regions of the original tissue. This issue is far less prominent in tumors with limited stromal components, such as lymphoma or papillary thyroid carcinoma. Taken together, integrating pathomics with single-cell omics is a PDAC-specific and conceptually well-matched strategy to address a central bottleneck in the field. Figure 2 illustrates the workflow of the integrated multi-omics approach used in the study of PDAC.

Figure 2 Schematic workflow of an integrated multiomics approach in pancreatic ductal adenocarcinoma. This workflow starts by integrating three primary data types: Single-cell RNA sequencing, pathomics and clinical data. The combined multiomics dataset is analyzed with artificial intelligence to identify patterns and construct predictive models. These models are ultimately leveraged for clinical applications, including patient stratification, molecular subtyping, treatment response prediction, and prognosis evaluation. H&E: Hematoxylin and eosin.

Applications of integrative pathomics-single-cell analysis in PDAC

The integration of histopathology with single-cell omics has emerged as a powerful approach for dissecting the complexity of PDAC and TME. Chen et al[78] combined scRNA-seq with pathomics to identify a transforming growth factor-beta-induced protein+ Schwann-cell population that facilitates neural invasion: Through transforming growth factor-β signaling, these cells appear to support directional migration of tumor cells toward neural tissue, thereby enhancing invasiveness. Importantly, multimodal integration enabled spatial contextualization of this Schwann-cell subset within the TME, highlighting a specific cellular niche associated with perineural infiltration, while single-cell profiling provided the underlying molecular programs of the relevant cell types. Complementing these findings, Chen et al[79] integrated multiplex immunohistochemistry with scRNA-seq to delineate the immune landscape of PDAC and to interrogate receptor-ligand interactions between PD-1+ T cells and PD-L1+ tumor cells. They reported relatively low PD-L1 expression in PDAC and mapped the distribution of PD-1+ T cells across distinct T-cell states, underscoring the complexity of immune evasion within the TME. Beyond descriptive profiling, the study derived a prognostic model based on immune features and further highlighted the relevance of immune checkpoint molecules in tumor progression. By combining multiplex immunohistochemistry with single-cell genomics, this framework enables high-resolution analysis of immune subsets and their spatially informed interactions with tumor cells, offering actionable insights for therapeutic strategies. Kung et al[80] synthesized single-cell multi-omics, spatial transcriptomics, and machine learning modeling to propose a novel conceptual framework for the PDAC immune microenvironment. By integrating single-cell and spatial data, they defined the core features of the PDAC microenvironment: CAF-driven desmoplasia, profound intratumoral heterogeneity, and deep immunosuppression (the “cold” tumor phenotype). Through machine learning models, they identified a spatial tissue architecture associated with immune evasion, T cells sequestered in stromal compartments, physically separated from tumor cells. They further proposed actionable therapeutic strategies to convert “cold” tumors into “hot” phenotypes, including enhancing T cell priming, reversing exhaustion programs, and targeting myeloid-derived suppression. This group explicitly outlined combination therapeutic strategies based on integrative analysis, demonstrating how single-cell molecular maps can be linked to spatial tissue architecture and translated into clinically actionable insights, embodying a complete translational pipeline from mechanism to clinical application. Overall, integrative pathomics-single-cell analysis capitalizes on the strengths of conventional pathology for visualizing tissue architecture and of single-cell genomics for resolving cell-type-specific molecular states. This convergence not only deepens mechanistic understanding of tumor-immune interactions in PDAC but also opens new avenues for targeted intervention, particularly in the context of immune checkpoint blockade.

Future progress in PDAC is likely to benefit from tighter integration of computational pathology with single-cell genomics to support patient stratification and individualized therapeutic decision-making. Given the pronounced desmoplastic stroma and spatial heterogeneity of PDAC, integrative pipelines should be explicitly optimized for this disease context rather than adopted in a one-size-fits-all manner. A priority is the development of robust, automated methods to map dissociated single-cell profiles onto histopathologic regions in stroma-rich tissues. In addition, diagnostic frameworks that jointly model pathology-derived morphologic scores and single-cell molecular signatures may enable more refined classification. Such multimodal stratification could, in turn, inform personalized treatment strategies that couple pathology-guided risk grouping with single-cell-informed target prioritization, with the overarching goal of improving outcomes in this highly lethal malignancy.

CONCLUSION

PDAC diverges from most solid tumors in a fundamental aspect: The stroma can comprise 70%-90% of the tumor tissue, and tumor cells and stromal populations (including fibroblasts and immune cells) are arranged in a highly heterogeneous spatial manner[81]. Tumor-stroma crosstalk is a key driver of malignant progression. Given these histologic features, no single omics approach can simultaneously capture tissue-level spatial architecture and the cell-resolved molecular programs that underpin it. Pathomics enables rigorous quantification of spatial morphologic phenotypes in PDAC, such as glandular architectural distortion and the degree of desmoplasia, but it cannot explain which cell types and molecular states generate these patterns[9]. Single-cell omics, by contrast, resolves cellular-state diversity and regulatory rewiring in PDAC[13,14]; however, the dense desmoplastic stroma creates practical barriers that make it difficult to reliably map dissociated single-cell profiles back onto defined regions of the original tissue[82]. This limitation is considerably less conspicuous in tumors characterized by limited stromal components, such as lymphoma or papillary thyroid carcinoma[83]. Taken together, integrating pathomics with single-cell omics represents a PDAC-specific and conceptually well-matched strategy to address a central bottleneck in the field.

Accumulating evidence indicates that single-cell omics technologies can resolve PDAC molecular heterogeneity at cellular resolution, delineate tumor microenvironmental composition, and capture dynamic cellular-state transitions, thereby providing unprecedented insight into the PDAC cellular ecosystem[13,82]. In parallel, pathomics objectively characterizes tissue spatial architecture and phenotypes by extracting quantitative, high-dimensional morphologic features from routine HE-stained sections, frequently using deep learning models. These features enable systematic links between macroscopic histomorphology and clinical outcomes[84]. By bridging these complementary levels of information, their integration can narrow the gap between molecular mechanisms and morphologic phenotypes, offering a new multiscale framework for precision diagnosis, treatment decision-making, and prognostic assessment in PDAC. From a translational perspective, pathomics has shown clear utility in PDAC for automated diagnosis, molecular subtype prediction, treatment response assessment, and prognostic stratification (as discussed above). Importantly, pathomics workflows based on standard HE slides are cost-effective and scalable, making them particularly suitable for implementation in resource-constrained clinical settings. Single-cell omics, in turn, provides mechanistic insight into key processes such as therapeutic resistance, metastasis, and immune evasion, and can nominate actionable targets and rational combination strategies for targeted and immunotherapies.

Equally important are interpretability, ethical governance, and regulatory-grade evidence. The “black-box” nature of many AI-driven pathomics models limits clinical trust; explainable AI methods (e.g., SHapley Additive exPlanations, Local Interpretable Model Agnostic Explanations, and attention-based visualization) can improve transparency and facilitate discovery of morphologically meaningful biomarkers[85]. To avoid algorithmic bias and inequitable performance, training and validation cohorts should be demographically diverse and geographically representative[86]. Privacy-preserving frameworks such as federated learning can further enable multi-institutional model development without direct data sharing[87]. From a regulatory perspective, clinical implementation typically requires evidence of analytical validity, clinical validity, and clinical utility[88], motivating prospective multicenter studies to test whether AI-augmented workflows improve diagnostic accuracy, treatment selection, and patient outcomes[89,90]. Even after approval, successful adoption depends on seamless workflow integration into digital pathology infrastructure, interoperability with laboratory information systems and electronic health records, user-centered reporting interfaces (including uncertainty and explainability outputs)[91], and adequate clinician training to support real-world use[92]. Multimodal decision-support platforms that can synthesize pathomics with genomic, proteomic, or radiomic data will be increasingly important for actionable recommendations in clinical practice[93].

It is worth emphasizing that integrative pathomics-single-cell approaches also have intrinsic constraints. Most such frameworks are fundamentally correlational, providing indirect links between morphology and molecular states rather than direct in situ measurements. Consequently, they cannot resolve cell-cell interactions or spatial signaling circuits with the same fidelity as spatial transcriptomics or highly multiplexed imaging. Using the multiplexed imaging platform CODEX and a “cellular neighborhood” framework, Xia et al[94] identified an immune-evasion pattern in PDAC driven by microenvironmental topology. They described a T-cell-enriched neighborhood with abundant CD8+ and CD4+ T cells that nonetheless associated with worse outcomes. Spatial proximity analyses further showed that CD8+ T cells within this neighborhood preferentially clustered with other T cells, including CD4+ T cells, rather than engaging tumor cells. These observations argue for a mode of immune escape beyond classic immune deserts or exclusion barriers: PDAC may spatially sequester infiltrating effector T cells into discrete stromal units that are physically separated from malignant glands, thereby limiting effective tumor contact and blunting T-cell function at the tissue-organization level. This provides a spatially grounded explanation for immunotherapy resistance in PDAC. Kung et al[80] combined single-cell and spatial multi-omics with machine-learning-based modeling to delineate defining features of the PDAC microenvironment, including a CAF-driven desmoplastic stroma, substantial intra- and inter-tumoral heterogeneity, and a profoundly immunosuppressed (“cold”) state. These insights translate into actionable therapeutic concepts centered on converting cold tumors into inflamed (“hot”) ecosystems, by enhancing T-cell priming and activation, reversing exhaustion programs, and mitigating myeloid-mediated immunosuppression. The study also pointed to integrative vulnerabilities across the tumor-stroma-immune axis, including metabolic crosstalk within the TME and oncogenic KRAS-linked signaling dependencies.

Methodologically, integration is highly dependent on the mapping accuracy from imaging to molecular labels achieved by AI models, and remains vulnerable to limitations in interpretability, stability, and generalization ability, as well as batch effects introduced by tissue handling and staining. Furthermore, rare cell populations or functional niches lacking distinctive morphologic correlates may not be reliably identified by image-based surrogates[95]. Future development should prioritize standardization of multi-omics acquisition and integration protocols in PDAC, improved spatial omics-histology alignment algorithms tailored to stroma-rich tissue, and multimodal predictive models that integrate pathomics, single-cell signatures, and clinical variables while remaining interpretable and generalizable. Longitudinal frameworks that combine pathomics with complementary modalities (e.g., liquid biopsy and radiomics) may enable dynamic monitoring of tumor evolution and provide actionable decision support. Ultimately, prospective multicenter studies will be required to establish real-world clinical utility and cost-effectiveness, facilitating translation into clinical guidelines and routine practice.

In summary, integrative single-cell omics with pathomics is increasingly clarifying the biological basis of intratumoral heterogeneity and microenvironmental crosstalk in PDAC. Continued methodological advances, together with rigorous clinical validation, should enable this multimodal framework to support earlier and more accurate diagnosis, individualized therapeutic stratification, and longitudinal monitoring, with the goal of improving outcomes in this highly lethal malignancy.

ACKNOWLEDGEMENTS

We are grateful to the School of Medicine of Wuhan University for providing a conducive research environment.