Pan‐cancer analysis shapes the understanding of cancer biology and medicine

Abstract

Advances in multi‐omics datasets and analytical methods have revolutionized cancer research, offering a comprehensive, pan‐cancer perspective. Pan‐cancer studies identify shared mechanisms and unique traits across different cancer types, which are reshaping diagnostic and treatment strategies. However, continued innovation is required to refine these approaches and deepen our understanding of cancer biology and medicine. This review summarized key findings from pan‐cancer research and  explored their potential to drive future advancements in oncology.

List of Abbreviations

AI

artificial intelligence

ASC

antibody‐secreting cell

AtM

atypical memory

CCF

cancer cell fraction

ctDNA

circulating tumor DNA

CAF

cancer‐associated fibroblast

EF

Extrafollicular

EMT

epithelial‐mesenchymal transition

gLOH

genomic loss of heterozygosity

GC

Germinal Center

GNN

graph neural network

HRR

homologous recombination repair

ICI

immune checkpoint inhibitor

ITH

intratumoral heterogeneity

MSI

microsatellite instability

MP

Meta‐program

meCAF

metabolic cancer‐associated fibroblast

NLP

Natural Language Processing

PTM

post‐translational modification

scRNA‐seq

single‐cell RNA sequencing

SNP

single nucleotide polymorphism

TCGA

The Cancer Genome Atlas

TME

tumor microenvironment

TIC

tumor‐immune cell

TMB

total mutation burden

1. BACKGROUND

Cancer is a complex, chronic disease and a leading cause of death worldwide [1, 2, 3]. It is traditionally classified into different types based on tissues of origin, with significant divergence among these cancer types in various aspects. Cancer types show considerable variation in terms of incidence and mortality (Figure 1). Furthermore, the molecular mechanisms underlying these different cancer types differ substantially. Despite these differences, there are shared characteristics across different cancers, such as common pathways, etiology, and therapeutic targets. André et al. [4] suggested that the naming of cancer should be named based on molecular characteristics rather than traditional classification based on tissues of origin. This suggestion highlights the increasing importance of pan‐cancer analyses in advancing cancer research.

FIGURE 1.

Comparison of incidence and mortality rate of different cancer types. Significant differences were observed among various cancer types. Lung cancer, breast cancer, gastric cancer and liver cancer exhibit highest incidence and mortality rates [5, 6, 7].

Pan‐cancer analysis, an integrative approach to studying multiple cancer types, has revolutionized our understanding of cancer biology and medicine. By identifying key biological processes that are dysregulated across different cancer lineages, this approach uncovers both differences and similarities that are essential for the development of new therapeutic targets. The development of The Cancer Genome Atlas (TCGA) molecularly characterized over 20,000 primary cancer samples across 33 cancer types [8], enabling comprehensive cancer analysis using multi‐omics data [9]. Today, pan‐cancer analysis provides a more systematic understanding of cancer mechanisms at the genomic, epigenomic, transcriptomic, proteomic, and clinical levels [10] (Figure 2).

FIGURE 2.

Applications of multi‐omics analysis in pan‐cancer studies. Integrating genomics, epigenomics, transcriptomics, proteomics, single‐cell/spatial omics and multi‐modal data provides diverse perspectives on cancer biology. These multi‐omics approaches collectively contribute to optimizing cancer medicine, enabling more precise diagnostics, personalized treatments, and deeper biological insights (created with biorender.com).

Advances in artificial intelligence (AI) have significantly enhanced the efficiency of data processing and pattern recognition in pan‐cancer analysis. Prelaj et al. [11] demonstrated the ability of AI‐based methods to discover biomarkers across a wide range of cancers. Martínez‐Jiménez et al. [12] highlighted the potential of pan‐cancer genome‐wide analysis in identifying unique features of advanced tumors. Geffen et al. [13] integrated proteomics and post‐translational modifications (PTMs) data to uncover a pan‐cancer pattern of changes in protein acetylation and phosphorylation, emphasizing the role of PTMs in cancer biology. These studies provide valuable insights into cancer biology and medicine.

In this review, we summarized key findings from pan‐cancer studies and discussed how pan‐cancer analysis will shape the future of cancer research. Our review provides a more comprehensive understanding of the current role of pan‐cancer analysis in tumor biology and its clinical applications.

2. THE DEVELOPMENT AND LANDMARKS OF PAN‐CANCER STUDIES

The development of pan‐cancer analysis can be divided into two period, 2013‐2019 and 2020‐2024 (Figure 3). Prior to 2020, the field was focused on genomics, transcriptomics, and epigenomics, which laid the foundation for shared molecular insights. In recent years, however, the field has seen notable shifts in research focus and methodology.

FIGURE 3.

The development and landmarks of pan‐cancer analysis. Key milestones from the establishment of TCGA database to the latest proteogenomic analysis were highlighted, demonstrating the evolution from genomic analysis to single‐cell sequencing and artificial intelligence technologies [14, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55 (created with biorender.com). Abbreviations: Ref, reference.

Early studies primarily focused on identifying biomarkers and genomic features. The establishment of TCGA project provided valuable data resources for large‐scale pan‐cancer studies, marking significant advancements in the field [14]. Zack et al. [15] conducted a pan‐cancer analysis of somatic copy number alterations, revealing common characteristics of cancer genomes. Hamilton et al. [16] identified a pan‐cancer oncogenic microRNA superfamily anchored by a central core seed motif. Subsequent studies expanded to analyze various omics data and explore cancer biology in greater depth. For instance, Cabanski et al. [17] performed pan‐cancer transcriptome analysis, uncovering conserved functions of long non‐coding RNAs. Aran et al. [18] systematically analyzed the impact of tumor purity across cancers. Ma et al. [19] conducted pan‐cancer genomic and transcriptomic analyses of 1,699 pediatric leukemias and solid tumors, providing a comprehensive understanding of molecular characteristics in childhood cancers. Saghafinia et al. [20] analyzed pan‐cancer DNA methylation pattern, highlighting the role of epigenetic alterations in cancer. Danaher et al. [21] defined pan‐cancer adaptive immune resistance, offering new perspectives for immunotherapy research. Priestley et al. [22] carried out pan‐cancer whole‐genome analysis of metastatic solid tumors, deepening the understanding of cancer metastasis mechanisms. Robichaux et al. [23] analyzed the role of ERBB2 mutations across cancers, providing new targets for precision medicine. Tan et al. [24] and Liu et al. [25] studied microRNA‐associated gene activation and DNA methylation markers, respectively, offering new tools for cancer diagnosis and prognosis evaluation. These studies span multiple omics levels, including genomics, transcriptomics, epigenomics, and proteomics, revealing key aspects of cancer such as molecular characteristics, driver genes, signaling pathways, and the immune microenvironment.

However, these studies also have limitations, such as sample bias and insufficient functional validation. Research conducted after 2019 has built upon these foundations, expanding sample diversity, deepening functional studies, enhancing clinical translation, and integrating more dimensional data (such as single‐cell omics and spatial omics) to create a more comprehensive molecular atlas of cancer. Since 2020, pan‐cancer research has evolved in several key directions. Firstly, the application of single‐cell sequencing technologies has enabled detailed exploration of tumor microenvironment (TME) and heterogeneity. For example, single‐cell analysis of Qian et al. [26] revealed a heterogeneity blueprint of TME. Secondly, the integration of multi‐omics data has deepened our understanding of complex biological processes. Advancements in computational methods have indeed accelerated the development of clinical prognostic prediction models. The study of Arora et al. [27] utilized outcome‐weighted integrative clustering to identify clinically relevant genomic subtypes across cancers, providing novel tools for prognosis prediction. Biomarker‐driven precision medicine has become a primary focus, with drug development based on pan‐cancer analyses increasingly emphasizing individualized treatment strategies. The 2021‐study of Kaushik et al. [28] on GPR139 and GPR142 pan‐cancer analysis and drug design demonstrated how pan‐cancer data can guide the development of personalized therapies.

A key trend in recent years has been the growing focus on immunotherapy and the immune microenvironment. Researchers are increasingly exploring T cell‐targeted therapies and immune resistance mechanisms, providing new perspectives for optimizing immunotherapy [29]. Additionally, the application of AI and machine learning technologies is enhancing our ability to comprehensively analyze complex datasets, identify novel therapeutic targets, and predict treatment responses.

By providing robust datasets, comprehensive analytical frameworks, and crucial biological insights, early pan‐cancer studies have become an integral part of the ongoing development in cancer research and treatment. Pan‐cancer research is shifting from descriptive analysis to mechanistic interpretation and clinical application. These advancements are expected to lead to breakthrough progress in cancer diagnosis, treatment, and prevention, driving the realization of individualized and precise cancer care.

3. THE DEVELOPMENT OF ANALYTICAL METHODS IN IMPROVING PAN‐CANCER ANALYSIS

3.1. Improving data integration

Pan‐cancer analysis typically involves the integration of large‐scale datasets (Table 1). Traditional methods often fall short when handling these large‐scale multi‐omics and multi‐modal datasets, whereas AI‐based methods excel in data integration and are therefore more effective. Specifically, multimodal deep learning has proven its strengths in pan‐cancer analysis by integrating multidimensional biomedical data for tasks, such as drug interaction prediction, tumor classification, and cancer prognosis prediction, showing strong performance [30, 31, 32]. Its scalability also indicates considerable potential for advanced tasks [33].

TABLE 1.

Existing pan‐cancer datasets.

Database name	Data Type	Website	Accessibility a
Pan‐Cancer Metabolism Data Explorer	metabolome	http://www.sanderlab.org/pancanmet	Active
SynMICdb	multi‐omics	http://synmicdb.dkfz.de/	Active
TSNAdb	genome	https://pgx.zju.edu.cn/tsnadb/	Active
FusionGDB	multi‐omics	https://compbio.uth.edu/FusionGDB2	Active
CancerSplicingQTL	multi‐omics	http://www.cancersplicingqtl‐hust.com/	Active
Pancan‐meQTL	genome	http://bioinfo.life.hust.edu.cn/Pancan‐meQTL	Active
LOGpc	genome	http://bioinfo.henu.edu.cn/DatabaseList.jsp	Active
LncMAP	transcriptome	https://ngdc.cncb.ac.cn/databasecommons/database/id/5020	Active
MutEx	multi‐omics	http://www.innovebioinfo.com/Databases/Mutationdb	Inaccessible
Cancer‐mbQTl	genome	http://canmbqtl.whu.edu.cn/	Active
IncExplore	transcriptome	http://lncexplore.bmi.nycu.edu.tw/	Inaccessible
Cancerouspdomains	multi‐omics	http://www.cancerouspdomains.ir/	Inaccessible
Pancan‐ncQTM	multi‐omics	http://bigdata.njmu.edu.cn/Pancan‐ncQTM	Inaccessible
SPACE	multi‐omics	http://fun‐science.club/SPACE/	Inaccessible
FOX2Cancer	proteome	http://hainmu‐biobigdata.com/FOX2Cancer	Active
ClGene	genome	http://soft.bioinfo‐minzhao.org/cigene	Active
MethMarkerDB	genome	https://methmarkerdb.hzau.edu.cn/	Active
TICCom	transcriptome	http://bio‐bigdata.hrbmu.edu.cn/TICCom/	Inaccessible
TCIA	imaging	https://www.cancerimagingarchive.net/	Active
TCGA	multi‐omics	https://www.cell.com/pb‐assets/consortium/pancanceratlas/pancani3/index.html	Active
CPTAC	proteome	https://proteomics.cancer.gov/programs/cptac	Active
GEO	multi‐omics	https://www.ncbi.nlm.nih.gov/gds/	Active
PCAWG‐TCGA	multi‐omics	https://icgc.bionimbus.org/	Active
TIMER	multi‐omics	http://timer.cistrome.org/	Active
GEPIA	transcriptome	http://gepia.cancer‐pku.cn/	Active
GTEx	multi‐omics	https://gtexportal.org/home/	Active
ICGC	multi‐omics	https://docs.icgc‐argo.org/docs/data‐access/icgc‐25k‐data	Inaccessible
EGA	multi‐omics	https://ega‐archive.org/	Active
OrganoidDB	transcriptome	http://www.inbirg.com/organoid_db/	Active
linkedomics	multi‐omics	https://www.linkedomics.org/login.php	Active
UALCAN	transcriptome /proteome	https://ualcan.path.uab.edu/index.html	Active
OrganoidBase	organoid multi‐omics	https://organoid.crownbio.com/OrganoidModels/Index	Active
SORC	spatial/single‐cell transcriptome	http://bio‐bigdata.hrbmu.edu.cn/SORC/	Active

^a The information presented in this table reflects the research landscape as of the latest search conducted on February 11, 2025. Due to the dynamic nature of online resources, subsequent updates may affect content availability.

The integration of multimodal data in pan‐cancer research remains a relatively niche area, warranting further exploration. Most existing models primarily focused on image and text data, limiting the diversity of data formats utilized. Hoang et al. [34] and Pizurica et al. [35] trained deep learning models on pan‐cancer datasets to generate transcriptomics from histopathology images, demonstrating superior performance and applicability in clinical scenarios. This includes improved prediction of cancer treatment response and recurrence risks using histopathology images alone. The accumulation of multi‐modal pan‐cancer datasets highlights the need for future research in this area to realize the potential of cancer precision medicine.

3.2. Identifying commonalities and specificities

Many pan‐cancer studies applied statistical methods, including frequency statistics and hypothesis testing, providing intuitive insights into the distribution of certain biological events in pan‐cancer. However, in‐depth analysis requires the extensive development of sophisticated analytical algorithms.

Machine learning and deep learning have demonstrated strong performance in predictive tasks, enabling the inference of a wide range of gene mutations, molecular tumor subtypes, gene expression signatures, and traditional pathological biomarkers directly from conventional histology (Figure 4, Table 2) [36]. These technologies play important roles in investigating the prognostic ability and the relationship between biological substances and pan‐cancer, thus facilitating the discovery of potential biomarkers [37, 38, 39, 40]. Al‐Fatlawi et al. [41] predicted features with strong pan‐cancer commonality and ranked them according to applicability, improving the versatility of the predicted biomarkers. Another key area of interest is the improvement of model architectures and network designs [42]. Lewis et al. [43] addressed the challenge of machine learning accuracy by integrating genome‐scale metabolic modeling, providing important advances for identifying prognostic metabolite biomarkers and predicting radio‐sensitivities in individual patients. Qiu et al. [44] developed an interpretable deep learning method to train pan‐cancer gene expression data, revealing both common and specific pathways among different cancer types. Designing appropriate deep learning methods to learn shared and distinct information across different cancer types and applying interpretable analysis methods to decode the information is an ongoing trend for pan‐cancer analysis.

FIGURE 4.

Roles of different analysis methods for pan‐cancer studies. Statistical methods revealed distributions and landscape of molecular signatures among pan‐cancer, while machine learning and deep learning methods are good at performing biomarker discovery, prediction and tumor subtyping. Notably, deep learning techniques can manipulate complex multi‐modal data, enhancing the depth of analysis (created with biorender.com).

TABLE 2.

AI‐based models for pan‐cancer analysis.

Model name	Application	Performances	Source
MRNGCN	Identification of Cancer Driver Genes	AUC = 0.9192, AUPRC = 0.8446	https://doi.org/10.1186/s12859‐023‐05140‐3
MILTS	PD‐L1 expression prediction	AUC = 0.83 on fresh‐frozen and AUC = 0.74 on formalin‐fixed specimens	https://doi.org/10.1038/s41467‐024‐46764‐0
PC‐RMTL	Pan‐cancer classification	Accuracy = 96.07%	https://doi.org/10.1038/s41598‐021‐03554‐8
Stem.Sig	Characterization of pan‐cancer stemness for predicting immunotherapeutic response	AUC = 0.71 in validation cohort	https://doi.org/10.1186/s13073‐022‐01050‐w
ATLAS	Tumor Spectrum and Site Classification	Accuracy = 91.4% for cancer site‐of‐origin, accuracy = 97.1% for cancer lineage	https://doi.org/10.1038/s42003‐024‐05981‐5
PUREE	Tumor purity prediction based on gene expression profile	Spanning median purity ranges from 45% to 76%	https://doi.org/10.1038/s42003‐023‐04764‐8
DNNraw	Cancer type prediction	Overall accuracy = 0.9640	https://doi.org/10.1038/s41598‐023‐47805‐2
MetaWise	Metastatic tumors classification	Test set accuracy = 88.5%	https://doi.org/10.1038/s41598‐023‐35842‐w
	Tumor type classification and microsatellite status prediction	Accuracy = 86%	https://doi.org/10.1038/s41551‐023‐01120‐3
Cancer‐Finder	Identify malignant cells in single‐cell data	Accuracy = 95.16% in internal validation dataset and accuracy = 98.30% in external test dataset	https://doi.org/10.1038/s41467‐024‐46413‐6
CAMIL	Prediction of molecular biomarkers from pathology slides	AUC = 0.82 in internal validation set, AUC = 0.96 in external validation set	https://doi.org/10.1038/s41467‐024‐45589‐1
Precily	Inferring tumor drug sensitivity	AUC‐PR = 0.85	https://doi.org/10.1038/s41467‐022‐33291‐z

Abbreviations: AUC, area under the curve; AUPRC, area under the precision‐recall curve; AUC‐PR, area under the precision‐recall curve.

3.3. Improving individual cancer analysis

From a computational view, pan‐cancer analysis can be considered a broader domain compared to the study of individual cancer types. In many individual cancer analyses, particularly for rare cancer types, data insufficiency poses significant challenges, especially in tasks like tumor classification. The development of self‐supervised learning and foundation models has demonstrated strong generalization capabilities for analyzing individual cancers when the model is trained on large‐scale pan‐cancer datasets, and fine‐tuned for specific cancer types. Virchow, a typical foundation model, is pretrained on the histopathology images from multiple cancer types and has achieved superior performances in clinical‐grade diagnosis of specific cancer types, particularly rare cancer types [56]. Similar examples include recently developed foundation models for histopathology imaging [57, 58]. These models facilitate the high‐level learning of data distributions across different cancer types, providing generalizability in characterizing the complexity of specific cancer types.

While pan‐cancer deep learning methods facilitate robust feature extraction from multi‐modal, multi‐omics datasets, uncovering the biological meanings of these learned features remains challenging but is important for cancer biology studies. Recently, Howard et al. [59] developed HistoXGAN, a generative adversarial network trained on 8,120 whole‐slide images across 29 cancer types. This model enables the reconstruction of histopathology images from feature vectors extracted by the existing models on histopathology images. Further evaluation of HistoXGAN has suggested its biological interpretability, including the retention of gene expression in the reconstructed images. The continued development and application of such methods will be essential to uncover the potential cancer mechanisms underlying these models.

4. PAN‐CANCER ANALYSIS IN UNRAVELING CANCER BIOLOGY

Recent years have seen remarkable advancements in uncovering the mechanisms underlying cancer biology. As a crucial domain, pan‐cancer analysis has provided novel insights into cancer biology from various perspectives, including etiology, pathogenesis, heterogeneity, TME, and cancer evolution (Figure 5).

FIGURE 5.

Key cancer biology problems studied in pan‐cancer studies. The cancer etiology, pathogenic mechanism, tumor heterogeneity, cancer evolution, and tumor microenvironment are mostly studied (created with biorender.com).

4.1. Pan‐cancer analysis adds insights into cancer etiology

The etiology of cancer is influenced by both genetic and environmental factors. Environmental factors, such as radiation, viruses, and chemical substances, are exogenous influences. Genetic factors, however, play a more critical role as they determine an individual's genetic susceptibility to cancer. As a result, the etiology of cancer exhibits substantial heterogeneity across populations and among individuals. Identifying cancer risk factors and elucidating the underlying mechanisms of cancer etiology are essential for effective cancer prevention.

The role of genetic variation in cancer etiology is complex and multifaceted. Pan‐cancer genomic studies have identified shared heritability across cancers, as well as gene mutations associated with cancer risk and treatment efficacy [60, 61]. These studies have also identified cancer driver gene mutations, which enhanced the understanding of common carcinogenic mechanisms. Among these genetic variations, single nucleotide polymorphisms (SNPs), chromosomal structural variations, chromosomal number variations, copy number variations, and regional events are all implicated in cancer induction [62]. With the construction of the pan‐cancer genome‐wide map, common patterns of somatic mutations, cancer type‐specific mutation events, and their oncogenic mechanisms have been revealed [63]. Pan‐cancer whole genome analysis has also uncovered typical cancer driver mutations and early events such as chromothripsis, along with associations between abnormal telomere maintenance and germline variants that influence somatic mutation patterns [45]. Additionally, frequent and diverse somatic retrotransposition events have been discovered in cancer, which are linked to genome instability and tumor suppressor genes [64]. Furthermore, the analysis of mutational signatures has provided novel insights into cancer etiology, including not only endogenous factors but also exogenous factors. By leveraging large‐scale genomic datasets, such as TCGA and ICGC, researchers have revealed both shared and cancer‐specific signatures linked to factors like DNA repair deficiencies, environmental exposures, and endogenous enzymatic activity [65, 66]. These comprehensive findings provide insight into cancer etiology at the molecular genetic level and underscore the complexity and diversity of genetic alterations that drive cancer development.

Driver genes are those whose mutations promote the development and progression of cancer by providing a growth advantage to tumor cells. There are a total of 568 cancer driver genes, and mutations in these genes enable malignant cells to rapidly and uncontrollably proliferate, evade the immune system and other defense mechanisms, spread to other tissues, and alter the TME [67]. Among these, TP53, PIK3CA, KMT2C, and ten other driver genes exhibit the most malignant functions. Additionally, with the aid of proteogenomic data, significant cis‐ and distal trans‐effects can be quantified across RNA, protein, and phosphoprotein levels, revealing the impact of cancer driver factors [68], thus providing a valuable resource for future research. The DriverPower tool also facilitates the identification of cancer‐specific driver and passenger mutations in whole‐genome and exome sequencing data [69]. To fully understand cancer etiology, focusing solely on classic genetic drivers is insufficient. Some passenger mutations, which are often considered “non‐driving” mutations, can significantly impact the genome and may influence tumor growth just as much as driver mutations. Interestingly, certain passenger mutations may even play a role in suppressing cancer development [70].

Microbial communities within primary tumor lesions interact with tumor's commensal flora, and may play a role in cancer initiation, progression, metastasis, and treatment. Tumor‐associated microbes can also influence the quality of life of patients by affecting tumor microsatellite instability (MSI) and its molecular characteristics. For instance, microbes such as Dialister and Casatella have been linked to improved patient survival rates [71]. Moreover, an analysis of 6,876 tumor samples across 21 cancer types revealed that microbial dysregulation within tumors is associated with abnormal DNA methylation, which can impact cancer cell proliferation [72].

4.2. Pan‐cancer analysis can reveal tumor heterogeneity

Tumor heterogeneity is a fundamental characteristic of cancer, manifesting as variations within and between tumors in terms of their genetic, epigenetic, transcriptomic, and phenotypic features. Pan‐cancer analysis provides a powerful approach to comprehensively investigate tumor heterogeneity by comparing multiple cancer types across various datasets.

Different gene subtypes of cancer cells have been identified in cell lines, xenograft tumors, and patients‐derived tumors [73], highlighting significant variability of tumor cells even within the same individual. This heterogeneity impacts treatment efficacy and patient prognosis, although cancer cells from different tumors can share conserved functional states. Several pan‐cancer studies identified these shared characteristics among different tumors. For instance, Barkley et al. [74] conducted pan‐cancer single‐cell RNA sequencing (scRNA‐seq) analysis on 15 cancer types, and identified a series of gene modules that define cancer cell states, such as “Stress”, “Hypoxia”, “Interferon response”, and “Epithelial mesenchymal transition”. Another study analyzed single‐cell and spatial omics, identified 41 meta‐programs (MPs) in malignant cells and mapped them to 11 hallmarks of intratumoral heterogeneity (ITH) [74], revealing the transcriptional patterns associated with ITH. Notably, non‐malignant epithelial cells also exhibited similar MPs, suggesting that many malignant ITH programs may predate tumor formation.

Gene modules represent the primary functions of cancer cells within the TME. Cancer cell states cannot always be clearly defined by cell types alone [75, 76], and some studies use gene modules to characterize cancer cell states. Cancer cells often exhibit specific gene modules with high expression,  helping to reveal tumor heterogeneity. The analysis of scRNA‐seq data identified gene modules in various cancer cell lines, uncovering repetitive patterns of heterogeneity common to both tumors and specific cell lines [77]. Furthermore, a pan‐cancer spatial transcriptome map revealed spatially related and frequently active pathways across six different tumor types [78]. Additionally, combining scRNA‐seq with single‐cell sequencing assay for transposase‐accessible chromatin (scATAC‐seq) allows for simultaneous characterization of both transcriptomic and epigenetic heterogeneity of cancer cells, providing dynamic insights into the underlying mechanism [79]. Moreover, the development of single‐cell genome sequencing has enabled the analysis of intra‐tumor heterogeneity at even small samples with limited cells [80], highlighting the need for analyzing tumor heterogeneity at pan‐cancer single‐cell genome level. Overall, these studies demonstrate that single‐cell and spatial omics analysis offer profound insights into the molecular heterogeneity of cancer cells. However, further in‐depth research is needed to integrate multi‐omics pan‐cancer data at both single‐cell and spatial levels to achieve a comprehensive understanding of tumor heterogeneity.

4.3. Pan‐cancer analysis in characterizing TME

Tumors are closely linked to their surrounding microenvironment and are in constant interaction with it. They influence the microenvironment by releasing signaling molecules, promoting angiogenesis, and inducing immune tolerance. TME plays a crucial role in the formation of tumor heterogeneity. The immune cells and stromal cells within the TME can significantly affect the growth and development of cancer cells.

Various studies have explored the common features and characteristics of the TME. Wang et al. [81] developed the DeClust model, which allows estimation of cancer type‐specific microenvironment signals from a large‐scale oncogenomic data, aiding in the development of personalized treatment plans for patients. However, traditional genomic data, such as bulk sequencing, struggle to capture cell‐to‐cell interactions. In contrast, pan‐cancer single cell and spatial omics analyses can deeply characterize the interaction network between cancer cells and other cells, revealing both shared and cancer‐specific cell types. Furthermore, there is a need to construct comprehensive database to characterize gene expression within the TME across various cancer types [82].

Stromal cells, a key component of the TME, play crucial roles in tumor initiation, progression, invasion, and metastasis. Qian et al. [26] identified 68 stromal cell subsets in colorectal, lung and ovarian cancers, with 46 subsets shared across cancer types and 22 unique to specific cancers. Resident cells exhibit significant tissue specificity, while tumor‐infiltrating cell types are largely conserved across different cancers. Among stromal cells, fibroblasts are a major component of TME. Cancer‐associated fibroblasts (CAFs) are highly plastic and perform diverse roles in supporting cancer cell survival [83, 84]. CAFs can be classified into 6 major subtypes based on their functional status, and their activation trajectories share similar characteristics across various cancers, particularly in epithelial‐mesenchymal transition (EMT) scores [85]. In a pan‐cancer study, Ma et al. [46] identified four CAF subtypes, including a new subtype of metabolic CAFs (meCAFs), which may mediate the reverse Warburg effect. These findings highlight the plasticity and complexity of CAFs. Further investigations are needed to explore the precise therapeutic potential of these cells.

T cells heterogeneity within TME is highly complex, and pan‐cancer studies using single‐cell multi‐omics have significantly advanced our understanding of this complexity. These studies have identified numerous T cell subsets, providing detailed insights into their differentiation pathways, transcriptional regulation, and the factors shaping tumor‐infiltrating T cells [86, 87]. Notably, two main trajectories of T cell exhaustion have been uncovered: one through effector memory T cells and the other through tissue‐resident T cells. The exhaustion process is generally associated with transcription factors TOX and PRDM1. The expression of certain exhaustion‐related transcription factors presents certain cancer‐type bias [88]. The abundance and state of dysfunctional T cells exhibit the most significant changes, while regulatory T cells remain relatively stable In addition to T cells. Studies have also shed light on B cell subpopulations within the TME, revealing two independent developmental pathways for antibody‐secreting cells (ASCs): one through the germinal center (GC) and another through an alternative extrafollicular (EF) pathway [89]. While further in‐depth studies are required, these findings offer detailed and holistic insights into TME.

Overall, identifying these tumor‐specific cell populations and characterizing their functions can not only deepen our understanding of the heterogeneity of TME, but also provide a scientific foundation for downstream applications, such as tumor typing, cancer detection, and the development of targeted therapeutics.

4.4. Pan‐cancer analysis can characterize cancer evolution

Pan‐cancer data analysis has significantly advanced our understanding of cancer evolutionary trajectories and identified early events in cancer evolution. A Bayesian mutation landscape approach, developed and applied to pan‐cancer genome data, revealed four evolutionary patterns and their associations with clinical aggressiveness and functional mechanisms [90]. By analyzing 2,658 pan‐cancer genomes, this study characterized the evolutionary trajectories of cancer and early driver events across different cancer types [47]. Evolutionary dynamics were modeled based on cancer cell fractions (CCFs), uncovering four evolutionary dynamic signatures representing neutral evolution, clonal expansion, and fixation in cancer [91]. Further analysis revealed that the switching between adaptive and innate immune cell populations is accompanied by clonal expansion of driver genes, which in turn modulate the tumor immune microenvironment. The evolutionary studies at the pan‐cancer level provided insights into cancer's evolutionary landscapes and the functional mechanisms underlying cancer evolution and progression. Moreover, the recent development of spatial omics enabled the analysis of cancer evolution in spatial resolution. Integrating Visium spatial transcriptomics (ST) and single‐nucleus transcriptomics on 78 cases across six cancer types, researchers identified spatial subclones and explored spatial tumor evolution through interactions with the local microenvironment in 2D and 3D space [92]. Pan‐cancer analyses thus provide a deeper understanding of common patterns in cancer evolutionary dynamics.

5. PAN‐CANCER ANALYSIS IN THE DECADE OF PRECISION MEDICINE

5.1. Cancer detection and screening

Cancer detection and screening are critical areas in both medical research and clinical practice. Early detection and accurate diagnosis are essential for effective cancer treatment, and achieving pan‐cancer early screening and detection could significantly reduce clinical costs. Studies suggested that blood testing combined with imaging methods, such as positron emission tomography‐computed tomography (PET‐CT) [93], as well as plasma cell free DNA (cfDNA) [94], are promising approaches for multi‐cancer early screening and detection. Additionally, the establishment of Pan‐cancer early detection by Multi‐omics biomarkers study (PROMISE) has highlighted the potential of integrating cfDNA methylation, ctDNA mutations, and protein biomarkers from blood tests for multi‐cancer detection [95].

Proteomic studies provide valuable insights into the functional status of cancer tissues. By integrating proteomic data with genomic, transcriptomic and clinical data, researchers are able to construct a comprehensive cancer biomarker landscape that facilitates cancer detection. A large‐scale proteomic study identified genetic instruments for 3,991 plasma proteins, uncovering 39 circulating protein biomarkers with robust causal evidence for seven site‐specific cancers, including 11 proteins with cross‐cancer effects [96]. Alvez et al. [97] proposed a strategy for exploring protein profiles in blood, enabling simultaneous cancer identification using only a few microliters of blood. This analytical platform, combined with simple sample collection methods, facilitated cost‐effective pan‐cancer population screening by detecting multiple cancer types in a single assay. Elmas et al. [98] developed a new algorithm, OPPTI, to identify overexpressed kinase proteins across 10 cancer types using global mass spectrometry proteomics data from 1,071 cases. Gonçalves et al. [99] created a comprehensive proteomic map of 949 human cancer cell lines spanning over 40 cancer types. They combined proteomic data with multi‐omics, drug response, and CRISPR‐Cas9 gene essentiality datasets to mine biomarkers of pan‐cancer vulnerability. With the advancement of AI, deep learning is increasingly utilized in pan‐cancer proteomic analysis. Wang et al. [100] trained deep learning models on pan‐cancer proteogenomic data and histopathology image, which effectively distinguished tumor from normal tissues and identified the tissue of origin, mirroring human pathologists’ assessments. Moreover, integrating transcriptomics and proteomics unveiled pathway‐level signatures associated with predictive histology features. Pan‐cancer proteomics analysis revealed potential protein biomarkers for cancer detection.

Among pan‐cancer biomarkers, cancer driver genes play a central role, as their mutations disrupt key cellular functions. Graph Convolutional Networks and Graph Attention Networks can analyze the relationship of multiple variables of interest and aggregate multi‐dimensional information to identify pan‐cancer driver genes, improving identification robustness and accuracy [101, 102, 103, 104, 105]. Furthermore, machine learning methods can investigate the activity of pan‐cancer driver genes, providing a detailed landscape of pan‐cancer biology [106, 107]. These pan‐cancer biomarkers can serve as potential drug targets and therapeutic targets [108]. In addition, observing cancer drug sensitivity can also help identify potential therapeutic targets for cancer [109, 110].

5.2. Cancer classification and subtyping

Pan‐cancer analyses have identified molecular subtypes from a comprehensive perspective of cancer. For example, Zhang et al. [111] identified four expression‐based metastatic subtypes that transcend the tumor spectrum. Combes et al. [112] revealed the dominant pattern of immune composition in cancer and delineated cancer classifications. Microbiome analysis has also proven useful in distinguishing between different types of cancer. One study collected 1,106 tissue samples from patients with no recurrence and metastasis and 912 tissue samples from 11 cancer types with recurrence and metastasis, finding significant differences in microbial composition across tissues [113]. Another study focusing on fungi in gastrointestinal and lung cancers suggested that Candida was highly abundant in tumor tissues and associated with reduced cancer survival [114]. In addition, fungal studies across multiple cancer types have shown that Malassezia species present in skin, breast and pancreatic cancer can significantly reduce patient survival [115].

Traditional cancer classification and staging methods, such as the TNM system and tumor molecular subtyping in breast cancer (e.g., Luminal A, Luminal B, HER2+, basal‐like and low‐Claudin subtypes), no longer fully meet the needs of clinical treatment [116]. As single‐cell omics and spatial omics expand across various cancer types, the heterogeneity of the co‐evolution of various cell subpopulations in the TME has been revealed. Summarizing these patterns, refining molecular classifications, and guiding detection strategies are essential steps toward achieving precision medicine and standardizing clinical paradigms.

Cancer molecular typing is not limited to cancer cells. For instance, based on the differences in tumor‐infiltrating T cells across cancer types, researchers have categorized cancers into groups with a high proportion of terminally exhausted CD8⁺ T cells and those with a high proportion of tissue‐resident memory CD8⁺ T cells [86]. scATOMIC, a modular annotation tool for malignant and non‐malignant cells that defines 19 common cancer references, can accurately classify breast cancer into clinically relevant subtypes and predict the significance of tumor origin in metastatic cancer [117]. Cancer‐Finder can identify malignant cells or regions in single‐cell or spatial transcriptome data with high accuracy [118]. The tool can also identify the tumor‐normal interface by using ten genes co‐located at thus boundary, providing a strategy for tumor detection. Pan Cancer Seeker is also a similar cancer cell identification [119] and is capable of detecting distinguishing tumor types and subtypes with greater precision based on single cell or spatial omics data. While single‐cell analysis enabled high‐resolution identification of cancer subtypes, the cost of single‐cell sequencing limited the clinical application of the discoveries. Some other methods can be used instead, including whole‐genome sequencing, clinical imaging, etc. In fact, cancer subtyping systems are mostly developed on whole‐genome sequencing data or multi‐omics data for some specific cancer types [120], and are also studied for pan‐cancer subtyping [121]. Developing related AI models for translation from bulk level omics data or histopathology image data to single‐cell level omics data would promote the clinical applications of pan‐cancer subtyping.

The rapid development of machine learning and deep learning has significantly improved the efficiency of cancer diagnosis. Deep learning‐based radiomics signatures show promise in identifying cancer subtypes, which is crucial for precision cancer treatment [122]. Integrating histology and proteomics can implement effective classification of pan‐cancer [123]. Microscopically, these models extract significant differences between cancer types and subtypes by learning early methylation patterns, miRNA signatures, and other cancer‐specific features [124, 125]. Cancer classification plays a crucial role in cancer diagnosis and aligns well with the functional mechanisms of machine learning and deep learning [126, 127]. With the emergence of new techniques such as Graph Neural Networks(GNNs) and multimodality models, cancer classifiers have become more accurate and robust [128]. Meanwhile, advancements in network architecture have further improved pan‐cancer classification efforts [129, 130, 131]. Additionally, deep learning has demonstrated strong performance in differentiating tumor origin, aiding in the identification of malignant tumors and predicting tumor origins in pleural fluid and ascites [132].

5.3. Cancer treatment

Due to tumor heterogeneity and the complexity of TME, may existing cancer drugs or therapies are insufficient to meet clinical needs. The precise treatment of cancer remains an enduring challenge. However, recent biological discovery from pan‐cancer perspective have significantly advanced our understanding of cancer biology. By associating these discoveries with treatment data, researchers have been able to reveal and validate potential new treatment strategies.

Certain cell subtypes within TME are closely associated with drug resistance and may serve as therapeutic targets. One study found that extracellular matrix associated CAFs are enriched at the tumor boundary where they regulated depleted CD8⁺ T cells through immune checkpoint ligand receptors [133]. This interaction forms a barrier that expels immune cells and promotes tumor immune escape. Additionally, TaNKs and iCAF were significantly correlated with resistance to immunotherapy [46, 134]. Furthermore, some studies have leveraged single‐cell transcriptome atlases and pan‐cancer cell line datasets to predict drug response and identify effective therapeutic drugs for drug‐resistant cell populations [73, 135]. Overall, pan‐caner single‐cell and spatial omics research provides valuable insights into cancer immune response, though the mechanisms underlying drug resistance remain poorly understood.

The microbiome plays a critical role in shaping responses to cancer treatments, particularly immune checkpoint inhibitors (ICIs). A recent study has revealed that fungal profiles can serve as multi‐kingdom microbial markers with strong predictive performance for ICI responses [136]. Additionally, the gut microbiota has been closely linked to the efficacy of immune checkpoint blockade therapies in human cancers [137]. Moreover, integrated analyses of metagenomics, genomics, transcriptomics, and clinical data have shown that microbial diversity correlates with tumor‐infiltrating neutrophils, with Fusobacterium specifically associated with resistance to immune checkpoint blockade [138].

Integrating pan‐cancer multi‐modal data facilitated the development of innovative tools that aid cancer treatment. Ruppin's group has made significant advancements in precision oncology by developing computational frameworks that predict cancer patients’ responses based on multimodal data. SELECT utilized tumor transcriptome data and gene interactions to predict treatment responses, showing broad applicability across multiple clinical trials [48]. PERCEPTION integrated bulk and single‐cell transcriptomics data from large‐scale cell line drug screens to construct predictive models based on patients’ tumor single‐cell transcriptomes [49]. The LORIS scoring system, which integrated common clinical, pathological, and genomic features, excels in predicting responses to immune checkpoint blockade therapy [50]. ENLIGHT identified clinically relevant genetic interactions from tumor transcriptomes to predict responses to various treatments across various cancer types [51]. ENLIGHT‐DeepPT predicted patient responses to targeted and immunotherapies by analyzing routine hematoxylin and eosin‐stained slides, eliminating the need for additional molecular testing [34]. Collectively, these methods constitute a comprehensive computational platform for precision oncology, spanning various modalities and demonstrating the immense potential of big data and AI in advancing personalized cancer treatment strategies.

3D organoids are 3D cell cultures that mimic the native cell microenvironment, enabling the simulation organ histogenesis and physiopathology. Organoids have been successfully developed for various cancers, providing a powerful platform for drug screening, personalized therapy, and cancer pathogenesis research. Genomic and single‐cell sequencing analyses have demonstrated that tumor organoids preserve the heterogeneity of the original tumors, retaining their genetic and phenotypic characteristics [139]. As a result, patient‐derived tumor organoids are valuable preclinical models for both academic research and personalized medicine [140]. Larsen et al. [141] developed a pan‐cancer patient‐derived tumor organoids, which demonstrated model consistency, and optimized culture conditions. The growing collection of pan‐cancer organoid datasets (Table 1) highlights the need for further analysis and validation to identify potential cancer treatments.

6. CHALLENGES AND FUTURE PROSPECTS

Pan‐cancer analysis has provided extensive insights into cancer biology and medicine. Unlike studies focused on specific cancer, pan‐cancer analysis offers a more integrative and holistic view of cancer mechanisms. By mapping the pan‐cancer multi‐omics landscapes, researchers are able to systematically uncover the underlying biological processes that drive cancer developments. Moreover, identifying both the molecular commonalities and specificities across different cancer types opens up potential diagnostic and therapeutic opportunities.

Considering the tumor heterogeneity among different cancer types, most of the current studies focuse on individual cancer types. However, recognizing the shared features across various cancers is equally important. To better address this, future studies on single cancer type should also incorporate pan‐cancer analysis to validate the specific discoveries or to uncover commonalities. This approach will require the development of comprehensive analytical frameworks, methods and databases to facilitate the seamless translation of single‐cancer findings into pan‐cancer insights.

Integrative analysis of large‐scale datasets is important for pan‐cancer analysis. Although many algorithms have been developed for analyzing pan‐cancer datasets, the full potential of these algorithms in identifying both commonalities and unique features across cancer types remains largely unexplored. A promising direction would be the development of biologically interpretable deep learning models that take into account the underlying biological mechanisms in pan‐cancer datasets [44, 142]. Moreover, advanced AI methods hold significant potential for integrating multi‐modal multi‐omics pan‐cancer datasets [143], yet few studies have been developed [144]. Future studies should focus on developing biologically interpretable deep learning models capable of effectively integrating and analyzing pan‐cancer multi‐modal datasets, ultimately advancing the biological understanding of cancer and enhancing clinical applications.

From an ecological perspective, cancer can be considered as an ecosystem, with each cancer type representing a distinct species. The tumor heterogeneity observed across different cancer types is one of the primary reasons for variations in survival rates and the curability of these cancers. Uncovering the evolutionary mechanisms that drive the diversity within cancer ecosystem will provide invaluable understandings into cancer biology. However, extensive collection of longitudinal datasets and development of spatio‐temporal analysis are required, which are also suggested in the commentary of Chang et al. [145]. Notably, cross‐organ regulations in cancer should be further studied [146]. As multi‐omics datasets, particularly spatial omics, continue to expand, future pan‐cancer analyses will offer holistic insights into the dynamic evolution of cancer, including the interactions within each cancer types and organ‐to‐organ interactions in cancer. These efforts would also be important for providing novel cancer treatment strategies.

Translating pan‐cancer discoveries into clinical applications is a complex challenge. However, pan‐cancer analysis has already led to the identification of unified diagnostic biomarkers and common therapeutic targets that could revolutionize cancer diagnosis and treatment. Additionally, pan‐cancer molecular subtyping has emerged as a promising tool for more precise cancer classification. While many of these findings are still in the research phase, advancements in datasets, algorithms, and emerging technologies such as quantum computing could accelerate the identification of optimized molecular subtypes and therapeutic targets. Moreover, dedicated departments for treatment planning and risk monitoring of patients are suggested to be established, which should leverage multi‐omics pan‐cancer data and discoveries to design patient treatment plans. This approach transitions from the traditional department‐specific treatment model to a centralized, interdisciplinary framework would facilitate personalized treatment plans and systematic care tailored to specific patient profiles.

In conclusion, pan‐cancer analyses have provided profound insights into cancer biology and medicine, and will likely shape the future of cancer treatment. Moving forward, it is essential to expand the scope of cancer types and omics data included in pan‐cancer research. By incorporating a broader range of cancer types and diverse omics datasets, we can achieve a more comprehensive understanding of pan‐cancer commonalities and specificities from the view of biological mechanisms predictive biomarkers, and therapeutic targets. Expanding the analysis to encompass a wider spectrum of cancer species will enhance our ability to identify common molecular pathways and alterations across different cancers, paving the way for the discovery of novel therapeutic targets and personalized treatment strategies.

AUTHOR CONTRIBUTIONS

Xiaoping Cen and Huanming Yang conceived and conceptualized the concept of this writing. Xiaoping Cen, Yuanyuan Lan, Jiansheng Zou, and Ruilin Chen wrote the original draft. Can Hu, Yahan Tong, Chen Zhang, Wei Dong, Qingqing Gao, and Man Ma generated the figures and tables and revised the manuscript. Jingyue Chen, Yuanmei Wang, Run Zhou, Weiwei He, Tianyu Lu take part in the revision of figures the manuscript writing. Fred Dubee and Dragomirka Jovic modified the language and logic of the whole manuscript. Youyong Lu, Yu Xue, Xiangdong Cheng, Yixue Li, and Huanming Yang supervised the manuscript writing and revisions. All the authors performed revision on the manuscript and agreed the final version.

CONFLICT OF INTEREST STATEMENT

The authors declare that they have no conflicts of interest to disclose regarding the publication of this article.

ETHICS APPROVAL AND CONSENT TO PARTICIPATE

Not applicable.

Cen X, Lan Y, Zou J, Chen R, Hu C, Tong Y, et al. Pan‐cancer analysis shapes the understanding of cancer biology and medicine. Cancer Commun. 2025;45:728–746. 10.1002/cac2.70008

DATA AVAILABILITY STATEMENT

Not applicable.