Multidimensional decoding of colorectal cancer heterogeneity: Artificial intelligence-enabled precision exploration of single-cell and spatial transcriptomics

Abstract

As a common malignant tumor, the heterogeneity of colorectal cancer plays an important role in tumor progression and treatment response. In recent years, the rapid development of single-cell transcriptomics and spatial transcriptomics technologies has provided new perspectives for resolving the heterogeneity of colorectal cancer. These techniques can reveal the complexity of cellular composition and their interactions in the tumor microenvironment, and thus facilitate a deeper understanding of tumor biology. However, in practical applications, researchers still face technical challenges such as data processing and result interpretation. The aim of this paper is to explore how to use artificial intelligence (AI) technology to enhance the research efficiency of single-cell and spatial transcriptomics, analyze the current research progress and its limitations, and explore how combining AI approaches can provide new ideas for decoding the heterogeneity of colorectal cancer, and ultimately provide theoretical basis and practical guidance for the clinical precision treatment.

Key Words: Artificial intelligence; Single-cell transcriptomics; Spatial transcriptomics; Colorectal cancer; Tumor heterogeneity

Core Tip: Colorectal cancer remains a major global health threat with rising incidence in younger populations and limited response to immunotherapy in most patients, largely due to its complex tumor microenvironment and high cellular heterogeneity. Recent advances in single-cell transcriptomics and spatial transcriptomics have opened new avenues for decoding this heterogeneity, offering unprecedented resolution into tumor biology and immune interactions. However, the massive and multidimensional nature of these datasets poses significant analytical challenges. This paper explores how the integration of artificial intelligence (AI), particularly machine learning and deep learning techniques, can enhance data interpretation in single-cell and spatial transcriptomics, improve the identification of novel biomarkers and tumor subtypes, and ultimately support personalized treatment strategies. By systematically reviewing current progress and proposing AI-driven solutions, this study aims to bridge the gap between complex omics data and clinically actionable insights in colorectal cancer precision medicine.

INTRODUCTION

Colorectal cancer (CRC) is one of the most common malignant tumors and poses a serious threat to human health. According to cancer statistics from 2022, the incidence and mortality rates of CRC remain among the top three for malignant tumors worldwide[1]. In recent years, influenced by various factors such as changes in dietary habits and increased psychological stress, the age of CRC onset has shown a marked downward trend. It has become the leading cause of cancer-related death among men under 50 and the second leading cause among women in the same age group[2,3]. This epidemiological shift poses more severe challenges for basic research and clinical treatment of CRC. With the development of precision medicine, in addition to traditional surgery, radiotherapy, and chemotherapy, targeted therapy and immunotherapy have gradually become key treatment options for patients with advanced CRC. In particular, immunotherapy, by relying on the immune system’s ability to monitor and eliminate tumors, activates anti-tumor immune responses, disrupts immune tolerance, and even establishes long-term immune memory. This enhances efficacy while sparing normal tissue damage to normal tissues caused by traditional radiotherapy and chemotherapy. Although the overall survival rate of CRC patients has significantly improved in recent years especially with the survival time of advanced-stage patients more than doubling studies have found that conventional chemotherapy combined with targeted therapy still causes considerable toxicity and side effects, making it difficult to further break through the survival bottleneck[4]. Against this background, the importance of immunotherapy has become increasingly prominent. However, compared with its widespread application in tumors such as lung cancer, the effectiveness of immunotherapy in CRC remains limited. At present, only about 15% of patients with defective mismatch repair (dMMR)/microsatellite instability (MSI)-high-type CRC show a favorable response to immunotherapy[3,5]. The underlying reason is that CRC exhibits a highly complex tumor microenvironment (TME) and significant cellular heterogeneity, which lead to fundamental differences in the mechanisms of tumorigenesis, progression, and immune evasion among different CRC subtypes[6,7]. This heterogeneity suggests that the molecular mechanisms by which normal cells transform into tumor cells may differ completely across subtypes, and this is precisely the key reason why current immunotherapies fail to achieve ideal efficacy in most CRC patients.

Fortunately, the rapid development of single-cell transcriptomics [single-cell RNA sequencing (scRNA-seq)] and spatial transcriptomics (ST) has provided unprecedented tools and perspectives for precisely dissecting the heterogeneity of CRC[8-10]. These technologies have enabled a leap from generating a mere “catalog of cell types” to constructing comprehensive “spatial functional maps”, significantly advancing our understanding of critical mechanisms such as tumor evolution and immune evasion. As such, they have become vital technical supports for formulating personalized therapeutic strategies[11-13]. However, the multidimensional and high-throughput data generated by scRNA-seq and ST are massive and complex, posing considerable challenges in terms of effective analysis, integration, and interpretation. The rapid advances in artificial intelligence (AI), particularly in machine learning and deep learning, offer promising solutions to these challenges[14,15]. For instance, in data analysis, machine learning enables a shift from traditional instruction-based models to data-driven learning frameworks, allowing for systematic predictions in areas such as gene function and therapeutic target identification[16-19]. This capability aligns precisely with the outcomes that our research aims to achieve.

Therefore, an in-depth exploration of how AI can be applied to the analysis of scRNA-seq and ST in CRC research is of great significance for advancing the comprehensive decoding of CRC heterogeneity. This paper focuses on the application of scRNA-seq and ST in CRC studies, systematically reviewing current achievements and existing limitations in the field. Furthermore, it proposes new approaches and strategies that integrate AI technologies, with the goal of providing a solid theoretical foundation and technical support for the precise treatment of CRC, ultimately benefiting patients.

THE BIOLOGICAL BASIS AND CLINICAL SIGNIFICANCE OF CRC HETEROGENEITY

Overview

CRC heterogeneity is not only observed between different patients but also prominently exists within individual tumors, manifesting across multiple biological levels. Since the introduction of the classical genetic mutation theory, studies on CRC have revealed that, in addition to common mutations in genes such as APC, TP53, and KRAS, other genes like ARID1A, SOX9, and FAM123B also undergo frequent mutations[20-22]. Based on these genomic alterations, classical CRC subtype classification systems have been established, including chromosomal instability and MSI[23,24]. With growing awareness of CRC heterogeneity, researchers have expanded their focus beyond genetics to encompass a broader spectrum of biological features, such as epigenetic regulation, metabolic reprogramming, angiogenesis, and the degree of immune activation[25-27]. In light of these findings, the consensus molecular subtype (CMS) classification system was introduced in 2015, dividing CRC into four subtypes with distinct biological characteristics. In recent years, the advancement of scRNA-seq has offered an even finer resolution for studying CRC. In 2022, building upon the CMS framework, researchers proposed two additional intrinsic subtypes iCMS2 and iCMS3 based on the diversity of tumor epithelial cells[28]. In addition to vertical refinement of classifications, transcriptomics-based horizontal classification systems have also emerged. For example, from a prognostic perspective, researchers have analyzed genes such as WNT, EGFR, and mitochondrial CYB to construct the CRC prognostic subtype system[29]. On the basis of these molecular subtypes, it is essential to explore the intrinsic and extrinsic heterogeneity of CRC cells for better clinical translation.

Intrinsic heterogeneity of CRC cells

The intrinsic heterogeneity of CRC cells primarily arises from their inherent biological properties. While classical genetic mutations are well-recognized contributors, epigenetic regulation also plays a pivotal role in shaping this heterogeneity. Among the various epigenetic mechanisms, DNA methylation stands out as one of the most influential in regulating gene expression. For instance, the CpG island methylator phenotype subtype frequently correlates with MSI in CRC patients[30-32], and in patients under the age of 55, MLH1 gene methylation is a major driver of dMMR CRC[33]. Advances in scRNA-seq have revealed that CRC stem-like cells (CCSCs) significantly contribute to the development of intrinsic heterogeneity, largely through asymmetric division. The formation of these CCSC subtypes is regulated by transcription factors such as ATF6 and FOXQ1[34]. Metabolic reprogramming also plays a substantial role. For example, altered glutamine metabolism has been implicated in promoting CRC metastasis to the ovaries[35]. Furthermore, cancer-associated fibroblasts (CAFs) and non-coding RNAs through mechanisms such as extracellular vesicle-mediated miR-200 transport and epithelial-mesenchymal transition (EMT) further enhance tumor heterogeneity[36]. Additionally, key signaling pathways including Wnt/β-catenin and phosphatidylinositol 3-kinase (PI3K)/protein kinase B (AKT)/ mammalian target of rapamycin exhibit substantial variability in expression and activation among CRC cells, contributing another layer to intrinsic heterogeneity. Taken together, this multifactorial heterogeneity underpins therapeutic resistance, metastatic potential, and inter-patient variability in clinical outcomes. Addressing this complexity is essential for improving treatment efficacy. Technologies such as scRNA-seq and ST provide powerful tools for uncovering the mechanisms of CRC heterogeneity and guiding personalized therapeutic strategies. Figure 1 intuitively demonstrates the intrinsic heterogeneity of CRC.

Figure 1 Overview of the integration of colorectal cancer heterogeneity with single-cell RNA sequencing and spatial transcriptomics. This schematic illustrates key drivers of cellular diversity within defective mismatch repair (dMMR) colorectal cancer (CRC) tumors. Epigenetic origin: Hypermethylation of the MLH1 gene promoter CpG island drives the transition from normal colonic epithelium to dMMR CRC. Cellular heterogeneity: Non-mutational epigenetic reprogramming enables both symmetric division (expanding the cancer cell pool) and asymmetric division, generating distinct subpopulations with divergent properties: Proliferation-related, chemotherapy tolerance-related, invasion-related. Key regulators in this process include ATF6 and FOXQ1. Activating invasion and metastasis: Invasion-prone subclusters exhibit metabolic reprogramming centered on glutamine/amino acid metabolism, generating α-ketoglutaric acid to fuel the tricarboxylic acid cycle. Activation of hypoxia-inducible factor 1-alpha and oxoglutarate dehydrogenase, ultimately promoting invasive and metastatic behavior. dMMR: Defective mismatch repair; CRC: Colorectal cancer; CCSC: Cancer stem cell-like cell; ASCT2: Alanine-serine-cysteine transporter 2; SN2: Sodium-coupled neutral amino acid transporter 2; GLS: Glutaminase; GLS2: Glutaminase 2; IL-4: Interleukin-4; Gln: Glutamine; Glu: Glucose; GLUD: Glutamate dehydrogenase; OGDH2: Oxoglutarate dehydrogenase; α-KG: Alpha-ketoglutarate; TCA cycle: Tricarboxylic acid cycle; HIF-1: Hypoxia-inducible factor 1-alpha.

Extrinsic heterogeneity of CRC cells

Extrinsic heterogeneity in CRC cells refers to intercellular differences caused by the TME or external factors, which are mainly reflected in the diversity of non-tumor cellular components, variations in physicochemical conditions, and differential responses to external interventions. First, Immune cells in CRC show functional diversity. For instance, neutrophils can adopt pro-tumor or anti-tumor phenotypes via NFKB1 or STAT4 activation, respectively, while CXCL14 loss in peritoneal metastasis impairs neutrophil-mediated immune surveillance[37]. In addition, ST studies reveal distinct B cell distributions across tumor sites, with cluster of differentiation (CD) 20⁺ B cell enrichment in right-sided CRC associated with favorable prognosis[38]. Additionally, interferon-producing cytotoxic T cells induce CD74 expression on neighboring cells, contributing to an “interferon-high” immunophenotype predictive of better immunotherapy response[39]. CAFs also exhibit significant heterogeneity. At least four spatial CAF subtypes (S1-S4) have been identified, with S4-CAFs enriched in Crohn’s-like reactions, which correlate with improved outcomes[40]. Matrix CAFs, through THBS2-CD47 signaling, promote invasion via the mitogen-activated protein kinase/extracellular regulated protein kinase 5 pathway and are linked to poor prognosis[41]. Additionally, myofibroblastic CAFs, associated with stem-like traits, contribute to liver metastasis[42].

Extrinsic heterogeneity further involves non-cellular factors. In the extracellular matrix (ECM), transcription factor 21 high tumor perivascular cells stiffen the perivascular niche and remodel collagen to facilitate liver metastasis[43]. Hypoxia also shapes metabolic heterogeneity; differences in vascularization create oxygen gradients, leading to distinct metabolic subtypes (C1-C3). Of these, the C1 subtype is characterized by severe hypoxia, CD8⁺ T cell exhaustion, and resistance to agents like oxaliplatin[44]. The intestinal flora adds another layer. Certain bacteria, such as Fusobacterium nucleatum and Bacteroides fragilis, colonize specific tumor regions and promote carcinogenesis through inflammation and metabolite production[45]. Conversely, butyrate-producing bacteria may suppress tumor progression via anti-inflammatory and immunomodulatory effects. Notably, extrinsic heterogeneity is dynamic and evolves in response to therapy. Treatment-induced remodeling of the TME alters both cellular composition and signaling context, complicating clinical management. Therefore, integrating approaches such as immunotherapy, anti-angiogenic therapy, and microbiota modulation guided by single-cell analysis will be key to overcoming CRC’s extrinsic complexity. Figure 2 intuitively demonstrates the extrinsic heterogeneity of CRC.

Figure 2 Overview of the extrinsic of colorectal cancer heterogeneity with single-cell RNA sequencing and spatial transcriptomics. This figure shows representative regulatory mechanisms in the cellular vs non-cellular components of colorectal cancer (CRC) extrinsic heterogeneity. First, different genes induce different types of neutrophil formation such as tumor-promoting, tumor-suppressing, and depleted types. Second, different subtypes of CRC-associated fibroblasts have different roles. S1-cancer-associated fibroblast (CAF) causes immunosuppression by affecting the extracellular matrix (ECM) under the regulation of actin alpha 2 (ACTA2) and transforming growth factor-β-1, and also causes immunosuppression by the subtype S3-CAF, which is achieved through the modulation of macrophages by CXCL14 and PDGFRB. S2-CAF reduces the number of macrophages through the up-regulation of interleukin-6, the ACTA2 expression to elicit an inflammatory response. S4-CAF, on the other hand, is associated with a favorable prognosis for inflammatory bowel disease in precancerous lesions of CRC. In the noncellular component, pro-vascular CRC cells and CRC cells distant from the vasculature produce different metabolites due to different metabolic patterns resulting in different CRC subtypes. CRC cells far from the vasculature tend to produce drug-resistant subtypes, and some metabolites of intestinal flora are also involved in the formation of this resistance. In addition, regulation of the ECM, such as transcription factor 21, shapes CRC subtypes that are susceptible to hematogenous metastasis by increasing the stiffness of the ECM, degrading the basement membrane, and collagen rearrangement. CRC: Colorectal cancer; ECM: Extracellular matrix; ACTA2: Actin alpha 2; TGFB1: Transforming growth factor-β-1; CAF: Cancer-associated fibroblasts; IL-6: Interleukin-6; TCF21: Transcription factor 21; BM: Basement membrane.

Clinical significance of heterogeneity in CRC

The intrinsic and extrinsic factors outlined above collectively contribute to the complex heterogeneity of CRC, which has profound clinical implications. This heterogeneity underlies treatment resistance the biological basis for therapy failure in many tumors[46-48]. This heterogeneity not only limits the efficacy of immunotherapy benefiting only approximately 15% of CRC patients[49], but also affects responses to chemotherapy and targeted therapy[50]. For instance, THBS2⁺ CAFs promote oxaliplatin resistance by secreting COL8A1, which activates PI3K-AKT signaling via ITGB1 binding, inducing EMT[51]. In the context of targeted therapies, CRC cells resistant to bevacizumab often show enhanced glycolysis and lactate production[52]. Importantly, such resistance can also emerge during treatment; EGFR or BRAF inhibitors may induce DNA damage, increase mutational burden, or trigger MSI, enabling tumor escape[53]. Thus, when discussing the clinical relevance of CRC heterogeneity, treatment resistance emerges as a central concern. On one hand, identifying different CRC subtypes can enable more precise prognostic stratification and prediction of treatment responses, facilitating the development of individualized therapeutic plans. On the other hand, with a deeper understanding of resistance mechanisms, future clinical approaches may adopt combinatorial treatment strategies for instance, co-administering drugs that target resistance pathways alongside standard therapies to reverse resistance and enhance efficacy. Although such strategies were once difficult to realize, the continuous advancement of technologies like scRNA-seq and ST provide unprecedented resolution to decode tumor heterogeneity, and thus, the goal of overcoming treatment resistance through precise and dynamic therapeutic design is increasingly within reach.

SCRNA-SEQ AND ST IN CRC HETEROGENEITY

Technical principles and development

ScRNA-seq, since its introduction in 2009[54] has rapidly evolved into a core tool for characterizing cellular heterogeneity. Technological advances particularly the development of high-throughput and barcoding platforms have significantly improved sensitivity and scalability[55-59], more recently, integration with multi-omics technologies has enabled simultaneous profiling of gene expression, mutations, and epigenetic states within individual cells[60,61]. ST complements scRNA-seq by retaining the spatial organization of gene expression within tissue architecture[62]. While early spatial methods like fluorescence in situ hybridization and liquid crystal monomers offered localized insights, they were limited by low throughput. The formalization of ST in 2016 by Lundeberg’s team marked a turning point, introducing standardized workflows but with relatively low resolution (approximately 100 μm). Recent advances have greatly improved spatial precision platforms such as Stereo-seq now reach near single-cell resolution (1 μm-10 μm) through innovations like bead arrays and DNA nanoballs[63].

ScRNA-seq and ST are increasingly integrated into clinical research due to their complementary strengths. While scRNA-seq excels at capturing cellular heterogeneity and dynamic gene expression at single-cell resolution, ST retains spatial context, allowing researchers to map gene activity within intact tissue architecture. Together, they enable high-resolution analysis of TME, immune landscapes, and cell–cell interactions factors critical to understanding treatment resistance and disease progression. More recently, both technologies have advanced toward multi-omics integration, incorporating proteomic and metabolomic layers (e.g., CO-detection by indexing, imaging mass cytometry, matrix-assisted laser desorption/ionization mass spectrometry imaging). This expansion offers a more comprehensive view of tumor biology. Although ST has lower throughput than scRNA-seq, its spatial precision from multicellular regions to subcellular resolution makes it particularly valuable for localizing therapy-resistant niches, stratifying tumor subtypes, and guiding spatially informed interventions. Collectively, these tools provide a powerful framework for advancing precision oncology and tailoring patient-specific therapeutic strategies.

Application of scRNA-seq and ST technologies in the study of CRC heterogeneity

The application of scRNA-seq in studying CRC heterogeneity primarily focuses on three key areas: Cell type identification, subpopulation analysis, and trajectory/pseudo time analysis all of which have been reflected to varying extents in earlier discussions. In cell type identification, scRNA-seq enables the efficient discrimination of tumor cell types by analyzing copy number variations and stemness gene expression. At single-cell resolution, it can simultaneously capture alterations at multiple molecular levels, including gene mutations and epigenetic modifications[64]. The refined molecular subtyping of CRC discussed earlier is a direct result of this capability. Subpopulation analysis delves deeper into the complexity of both tumor cells and their surrounding microenvironment. This approach allows researchers to investigate functional heterogeneity within immune cells, such as macrophages in the TME[65], while also characterizing key tumor-intrinsic features like metabolic reprogramming[66]. These insights are crucial for identifying novel biomarkers and informing clinical treatment strategies and prognostic assessments. In contrast, trajectory or pseudo time analysis offers a dynamic perspective by mapping the continuous progression of cell states along a temporal axis. This method allows for the systematic identification of upstream and downstream regulators of key molecular events, as well as the elucidation of co-regulatory networks. It is particularly valuable in processes such as cell differentiation, EMT, immune cell exhaustion, and the development of treatment resistance[67-69]. For example, by leveraging deep transfer learning frameworks that integrate large-scale cell line datasets, researchers can predict patient-specific therapeutic responses at the single-cell level[70]. This provides novel strategies for addressing clinical challenges such as immune evasion and drug resistance.

Compared to the cell-level insights offered by scRNA-seq, ST expands the analytical dimension by preserving the spatial context of gene expression within tissue architecture. This technology plays an irreplaceable role in several key aspects of tumor biology, including spatial compartmentalization of tumor tissues, functional localization of gene activity, mapping of metastatic trajectories, reconstruction of intercellular communication networks, and spatial localization of treatment-resistant regions. For instance, in CRC tissues, studies have revealed highly active cell-cell interactions between adjacent stromal and tumor regions, mediated notably by the C5AR1-RPS19 ligand-receptor pair[71]. Similar interaction patterns have been observed between fibroblasts and CRC cells[72]. Furthermore, ST analyses have uncovered that FAP⁺ tumor-specific fibroblasts and SPP1⁺ macrophages are frequently co-localized in CRC tissues, and their synergistic activity promotes the formation of a decalcified ECM, which hinders T cell infiltration. This mechanism contributes to poor responsiveness to programmed cell death ligand 1 immune checkpoint inhibitors[73]. In other words, ST provides a powerful means to pinpoint regions of resistance to chemotherapy, radiotherapy, or immunotherapy, offering spatially informed strategies to counteract such resistance. When combined with trajectory analysis from scRNA-seq, ST enables a transformation of our understanding of CRC heterogeneity from a two-dimensional cellular framework to a three-dimensional, spatially and temporally resolved dynamic structure. This integrative approach has proven particularly valuable in the study of metastatic CRC. Leveraging both technologies, researchers have identified FOXD1 as a novel marker of metastatic CSCs and demonstrated the central role of the CD74-macrophage migration inhibitory factor signaling axis in driving CRC liver metastasis[74]. Such findings underscore the complementary strengths of scRNA-seq and ST in resolving the multilayered complexity of CRC and guiding the development of more targeted, spatially-informed. In Table 1, we have listed the current representative clinical studies on CRC related to scRNA-seq and ST technologies.

Table 1 Clinical trials of colorectal cancer involving single-cell RNA sequencing and spatial transcriptomics.

NCT	Date	Name	Organization	Type	Status
NCT04789252	January 02, 2024	Heterogeneity of dendritic cells in colon and non-small cell lung cancer	University of Milano Bicocca	Observational	Recruiting
NCT04622423	October 01, 2024	Advanced therapies for liver metastases	IRCCS San Raffaele	Observational	Recruiting
NCT05056896	December 04, 2024	Aspirin intervention for the reduction of colorectal cancer risk -extension	Massachusetts General Hospital	Interventional	Active, not recruiting
NCT03984578	October 15, 2024	Window of opportunity study in colorectal cancer	National Cancer Centre, Singapore	Interventional	Recruiting
NCT05398380	October 02, 2024	Liver transplantation for non-resectable colorectal liver metastases: Translational research	Hospital Vall d’Hebron	Interventional	Recruiting
NCT06833866	February 19, 2025	Phase I trial of 5-fluorouracil-based therapy in combination with hydroxytyrosol in patients with advanced or metastatic colorectal cancer	The Methodist Hospital Research Institute	Interventional	Recruiting
NCT06886282	March 20, 2025	Spatial radiomics and transcriptomics to the discovery of the cross-link between colon cancer and chronic kidney disease	National Cancer Institute, Naples	Observational	Recruiting
NCT06762405	February 18, 2025	PRODIGE 90-(FFCD 2204) neoadjuvant dostarlimab with short course radiotherapy in a watch-and-wait strategy for microsatellite unstable or mismatch repair-deficient locally advanced rectal cancer patients	Centre Hospitalier Universitaire Dijon	Interventional	Recruiting

NCT: National clinical trial.

Bottlenecks in the study of CRC heterogeneity by scRNA-seq and ST technologies

Technical limitations: Despite the tremendous potential of scRNA-seq and ST in uncovering CRC heterogeneity, no technology is without limitations. These upper bounds stem not only from the techniques themselves but also from challenges in data analysis, clinical translation, and functional validation[75-78]. At the technical level, scRNA-seq exhibits limited sensitivity in detecting low-abundance transcripts, such as long non-coding RNAs, which may lead to the omission of critical regulatory molecules and, consequently, compromise the completeness and accuracy of the gene expression profile[79]. Secondly, biases and inefficiencies in cell capture remain a major bottleneck in current scRNA-seq platforms. For instance, the widely adopted 10 × Chromium system is optimized for cells with diameters ranging from approximately 5 μm to 40 μm. This means it fails to efficiently capture larger cells, such as megakaryocytes or multinucleated tumor cells, as well as smaller components, such as cellular debris or fragments smaller than 5 μm. Moreover, the system’s capture principle based on randomly pairing cells with barcoded beads in oil-in-water droplets introduces inherent stochasticity and failure rates. Under high-throughput conditions, multiplet artifacts, where multiple cells and beads are co-encapsulated, become more likely and can seriously distort downstream data interpretation. For ST, the primary technical constraint lies in its spatial resolution. Whether employing sequencing-based approaches (e.g., visium, slide-seq) or imaging-based platforms (e.g., multiplexed error-robust fluorescence in situ hybridization, sequential fluorescence in situ hybridization), current technologies must make trade-offs between resolution and tissue coverage. As a result, accurately reconstructing the spatial microenvironment at single-cell or subcellular resolution remains a major challenge, limiting the full potential of ST for fine-scale biological discovery. Encouragingly, ongoing innovations are pushing these boundaries. For example, spotiphy, a recently introduced technology, combines the strengths of imaging and sequencing to achieve whole-tissue-section, single-cell, spatial full-transcriptome profiling, significantly enhancing spatial resolution without sacrificing tissue coverage[80]. Additionally, the single-cell combinatorial fluidic indexing (scFI) technique offers a solution to the throughput limitations of droplet-based scRNA-seq. By loading multiple cells into each droplet and using combinatorial barcodes for identification, scFI enables high-throughput, efficient data capture, thereby greatly improving scalability and efficiency in large-scale single-cell studies[81].

Data analysis limitations: Despite the unprecedented high-resolution insights offered by scRNA-seq and ST technologies in elucidating CRC heterogeneity, the resulting high-dimensional, multimodal datasets pose significant analytical challenges. Limitations in algorithms, computational complexity, and the integration of multi-omics data have become major obstacles hindering the clinical translation of these technologies. Firstly, single-cell data are inherently high-dimensional, sparse, and noisy, increasing computational demands and processing time. More critically, standard dimensionality reduction or filtering steps risk discarding biologically meaningful signals. For instance, rare cell populations such as circulating tumor cells (frequency < 0.5%) can easily be obscured post-reduction, leading to biased biological interpretations[82]. Similarly, lowly expressed genes, including key transcription factors involved in CRC stemness and drug resistance may be mistakenly treated as background noise, especially under insufficient sequencing depth. To address this, advanced analytical tools such as BayesSpace have recently been developed, leveraging spatial Bayesian modeling to improve the detection of low-expression signals across spatial domains[83]. Secondly, the integration of multimodal omics data particularly the fusion of scRNA-seq with spatial location information remains technically challenging. scRNA-seq data lack inherent spatial coordinates, and current integration methods rely on computational inference to align them with ST data. However, this inferred spatial mapping is especially uncertain in complex tissues like CRC. A case in point: Whether FAP⁺ fibroblasts directly regulate oncogenes such as MYC in neighboring CRC cells through spatial proximity remains difficult to confirm with existing algorithms. Thirdly, batch effects and systematic errors introduced by sample handling and experimental procedures also present major concerns. Differences in sequencing platforms, tissue dissociation methods, sample batches, and even inherent tumor heterogeneity can introduce confounding variables. For example, intrinsic differences among CRC molecular subtypes (e.g., CMS1-CMS4) may obscure cell subtype-specific signals relevant to treatment, thus undermining the accuracy of clinical association analyses. In the cell annotation stage, inconsistencies in marker usage across studies further complicate reproducibility. For example, CAFs have been identified using varying markers such as interleukin-6 and actin alpha 2 (ACTA2), leading to different classification outcomes for the same cell types across studies. This lack of standardized annotation criteria hampers cross-study comparisons and increases the risk of misclassification. While automated annotation tools can aid in large-scale studies, they often require manual intervention when confidence scores are low or reference labels are absent. However, manual annotation introduces subjectivity, reducing reproducibility and potentially affecting downstream analyses. Such annotation biases may directly impact the reliability and generalizability of models used for clinical risk prediction[84]. In summary, computational bottlenecks have become one of the most critical barriers to unlocking the clinical potential of scRNA-seq and ST technologies. Compared with experimental challenges, these analytical issues may be more amenable to resolution in the current era of rapid advances in AI. Therefore, the next phase of research must focus on developing AI-powered analytical tools to enable efficient, accurate, and clinically meaningful translation from data to diagnosis and treatment.

AI ENABLED ANALYSIS OF CRC HETEROGENEITY

Data preprocessing

The integration of AI into scRNA-seq and ST analysis marks a major advancement in addressing the challenges of high-dimensional, sparse, and heterogeneous data. AI enhances analytical efficiency, improves reproducibility, and accelerates data processing key benefits in translational research. In scRNA-seq, feature selection is a critical step for effective dimensionality reduction and accurate identification of cellular subpopulations. Compared to conventional statistical approaches, deep learning-based feature selection models have demonstrated clear superiority in terms of cell type classification accuracy, reproducibility and diversity of selected features, and computational efficiency[85]. For example, the dynamic batching adversarial autoencoder has been used in CRC to build glycosyltransferase-associated risk signatures, aiding prognosis and immunotherapy planning[86]. In ST, models like spatial variational autoencoder (spaVAE) integrate spatial and molecular data, improving resolution and biological interpretation. In the field of ST, researchers have developed the spaVAE-a model that integrates spatial dependency structures with gene expression profiles for joint modeling. By capturing both spatial context and molecular variation, spaVAE significantly improves the representation of count-based data and enhances spatial resolution[87]. AI also supports upstream workflows, such as tissue alignment, enabling standardized analysis across patient samples and enhancing biomarker discovery. A recent example is a deep learning-optimized tissue section alignment strategy, which enables the standardized processing of tissue slices from multiple patients prior to ST analysis. This approach not only increases sample utilization efficiency but also facilitates the systematic identification of large-scale prognostic biomarkers, accelerating clinical translation.

Cell type annotation

Following data preprocessing, accurate cell type annotation is a critical step in scRNA-seq and ST analyses, particularly for dissecting CRC heterogeneity. Traditional annotation methods rely heavily on reference datasets and often struggle with incomplete labeling and limited adaptability. AI offers a robust solution by enhancing both the accuracy and completeness of annotations through supervised learning. Tools like scDeepInsight and stAI have significantly improved annotation performance by inferring missing labels and optimizing reference usage[88,89]. To address poor generalizability in traditional algorithms, deep learning models such as Cancer-Finder have been developed to accurately detect malignant cells in both scRNA-seq and ST data, enabling precise localization of tumor regions[90].

In addition, AI-driven cell type discovery can accelerate the identification of critical subpopulations, such as the recently characterized FAP⁺SPP1⁺ stromal subset (CAF-M1) and help refine the spatiotemporal distribution map of CRC cellular subtypes. These advances provide a stronger data foundation for precision-targeted therapies and clinical translation. Once cell type annotation is complete, the next essential task is to deeply interpret the annotated results and construct spatial maps, a stage where AI again demonstrates transformative potential this will be further elaborated in the following section.

Data analysis and graph interpretation

AI plays an even more powerful role in analyzing scRNA-seq and ST data. Deep learning efficiently handles high-dimensional data, while graph neural networks integrate spatial relationships, and transfer learning enhances generalizability across samples. AI-driven models such as single-cell graph convolutional network support cross-platform integration[91], and deconvolution tools like deconvoluting spatial transcriptomics data through graph-based convolutional networks and Vec2imag resolve mixed signals at each spatial location, enabling accurate reconstruction of cellular composition and tissue structure[92,93]. These capabilities allow the construction of detailed spatial maps of the TME, improving prognosis prediction and treatment response modeling[94]. Tools like SPACE further enable single-cell resolution analysis, segmenting tissue modules and modeling intercellular communication[95]. AI also advances spatial inference by integrating gene expression with histological features. For example, STASCAN can predict spatial cell distributions in unmeasured regions, enhancing resolution and interpretability, such as ResNet have achieved classification accuracies as high as 99.99%, while also effectively mitigating issues such as small sample sizes, batch effects, and technical variability inherent in single-cell data[96]. These technologies help model tumor progression, invasion, and therapy resistance in spatially resolved detail. AI-based prediction systems can map localized treatment responses, supporting target discovery and personalized therapy design. Ultimately, the development of CRC-specific multimodal databases will provide a foundation for individualized clinical decision-making.

Multi-omics data integration

In traditional data integration methods, linear models are limited in their ability to capture complex nonlinear relationships, and the manual integration process is time-consuming and inefficient in handling interactions among high-dimensional data. In contrast, AI methods (such as deep learning, graph neural networks, and multimodal learning) can automatically learn complex internal association structures through nonlinear modeling, thereby significantly improving the efficiency and accuracy of multi-omics data integration. This advantage is mainly reflected in the following aspects: First, multimodal data alignment models can fully integrate the feature spaces of transcriptomics, proteomics, and epigenomics to infer the dynamic relationships of the research objects. For example, the survival analysis learning with multi-omics neural networks model aggregates and simplifies the processing of gene expression and cancer biomarker data through a multi-omics neural network, significantly improving the accuracy of patient survival prediction[97], and by using AI to align the distributions of two or more different modalities of data, we can quickly and intuitively reveal the heterogeneity of the T cell exhaustion process among different CRC patients and identify key signaling pathways and regulatory genes involved in this process[69,98]. Second, deep learning-based models can effectively utilize copy number variation, gene expression, and point mutation data across omics to autonomously uncover the latent factors behind lower-dimensional data, further enabling refined CRC subtype classification[99]. AI-supported multimodal integration can aggregate large-scale cellular features to make tumor cell heterogeneity characterization more comprehensive. Based on spatial analysis of whole-slide images, combined with tumor proliferation and immune response characteristics, CRC samples have been successfully classified into four distinct subtypes, aiding precise subtyping and individualized treatment[100]. Third, the construction of cross-omics maps promotes causal inference and helps to reveal causal chains among genes, proteins, and metabolites. For example, machine learning algorithms have been used to analyze the role of angiogenesis-related genes in CRC, providing new insights into their biological function and clinical significance[101]. So, relying on AI’s powerful feature-learning capabilities, it is widely recognized that use multi-omics analysis to discover new therapeutic targets and their regulatory networks will provide strong support for clinical translation and precision treatment. Figure 3 summarizes the specific ways in which AI empowers scRNA-seq and ST research in CRC.

Figure 3 Artificial intelligence-powered analytical framework for single-cell RNA sequencing and spatial transcriptomics in colorectal cancer heterogeneity. Standard transcriptomics workflow: Illustrates sequential steps from single-cell separation through messenger RNA capture and labeling. Reverse transcription and amplification, library construction and sequencing, to data analysis. Artificial intelligence (AI)-enhanced analytics: AI transforms multi-modal data interpretation through: Advanced data preprocessing (noise reduction and lower dimension via AI models such as dynamic batching adversarial autoencoder and spatial variational autoencoder); Cell type annotation (supervision, supplementation, and optimization through deep learning models such as stAI and scDeepInsight); Data analysis (dataset integration via single-cell graph convolutional network and accurate resolution of spatial information through deconvoluting spatial transcriptomics data through graph-based convolutional networks); Ensemble image features for accurate prediction; Integrating multi-omics analysis for causal analysis, target prediction, and precision medicine. mRNA: Messenger RNA; DB-AAE: Dynamic batching adversarial autoencoder; spaVAE: Spatial variational autoencoder; AI: Artificial intelligence; scGCN: Single-cell graph convolutional network; DSTG: Deconvoluting spatial transcriptomics.

AI-enabled scRNA-seq and ST participation in CRC clinical decision making

Under the assistance of AI, clinical analysis models can generally be categorized into three types. The first is the multimodal predictive model. Empowered by AI, technologies such as scRNA-seq and ST can be integrated with expression profiles, spatial data, pathological features, and biomarkers to construct composite models that enable intelligent diagnosis at the cellular level. This includes identifying tumor stem cell populations or trace amounts of circulating tumor cells, thereby improving the coverage and detection rates of early screening for CRC. For example, researchers developed the ZAHV-AI system, which combines tumor biomarkers associated with extracellular vesicles and AI-based detection. It demonstrated excellent performance in detecting stage 0-1 CRC (area under the curve = 1.0)[102]. Furthermore, it enables accurate localization of tumor malignancy and spread. Taking positron emission tomography-computed tomography as an example, AI enhancements may further upgrade this technique, allowing for early identification of cells with metastatic potential or micrometastases comprising only a few cells[103]. The second type is the personalized treatment recommendation system. Based on deep learning, big data processing, and inference models, it can classify CRC patients into subtypes comprehensively and at an early stage across multiple dimensions, including the immune microenvironment, epithelial-mesenchymal ratio, and signaling pathway activities. This classification precision far exceeds the limitations of manual interpretation and lays a solid foundation for personalized therapeutic regimens[104,105]. Figure 4 illustrates the specific applications of AI-empowered scRNA-seq and ST technologies in the clinical translation of CRC. Moreover, by participating in and analyzing the treatment process in real time, AI can construct individualized risk scoring systems that predict postoperative recurrence, chemotherapy response, targeted therapy efficacy, and immunotherapy outcomes. These predictions assist in timely treatment adjustments. For instance, the artificial neural network model can predict the recurrence risk after local excision of rectal cancer in laboratory settings with an accuracy of up to 97.9%, potentially reducing the need for more invasive surgeries in 34.9% of early-stage CRC patients and thereby improving long-term outcomes[106]. Finally, there is the multi-center deep learning model built upon the aforementioned two frameworks. By aggregating patient data from different regions, such models can identify region-specific risk factors and disease characteristics, which are of great significance for regionalized CRC prevention strategies. Additionally, AI-powered patient navigation tools such as MyEleano have increased colonoscopy screening rates by 36%[107]. These self-learning and self-updating models will accelerate the translation of clinical research into practical applications, benefiting more CRC patients at an earlier stage.

Figure 4 Artificial intelligence-enabled single-cell RNA sequencing and spatial transcriptomics-involved colorectal cancer heterogeneity analysis and its guided clinical decision-making. This figure includes an artificial intelligence-based graphic neural convolutional network to drive clinical translation of colorectal cancer (CRC) intrinsic heterogeneity research through enhanced pathology image characterization and integrated data analysis. Clinical translation of CRC extrinsic heterogeneity research is driven by artificial intelligence-mediated neighborhood analysis revealing spatial information reorganization of intercellular networks, identification of immune-rejection regions highlighting stroma-cancer cell interactions, and characterization of tertiary lymphoid structures with extrinsic heterogeneity. These analyses support clinical predictive algorithms for prognostic risk scoring and treatment response prediction, enabling high-risk tumor metastasis localization and clinical precision therapy, as well as supporting early screening and prevention based on comprehensive clinical data. scRNA-seq: Single-cell RNA sequencing; CNN: Convolutional neural network; AI: Artificial intelligence; ST: Spatial transcriptomics.

LIMITATIONS

Admittedly, AI has brought us a transformative opportunity for technological advancement, and its potential in CRC heterogeneity research and precision medicine is substantial. However, it is equally important to recognize that AI still faces numerous limitations at this stage. First, in terms of model training, early AI development often suffered from insufficient data volume and homogeneous databases, leading to overfitting and poor generalizability. For example, underdeveloped regions are constrained by limited economic and educational resources, making it difficult to provide sufficient and high-quality samples to support robust AI training and optimization. Although this study has made every effort to include representative research especially regarding CRC heterogeneity some key mechanisms or significant findings may still have been omitted. Second, the application of AI faces the challenge of data heterogeneity in real-world settings. Differences in institutional policies, data collection principles, and ethical regulations across countries and regions result in inconsistent data distributions, which are often overlooked during AI model development, thereby limiting the generalizability of the models. Moreover, many of the conceptual integrations of AI with CRC heterogeneity proposed in this study remain at a theoretical level and have yet to be fully validated through empirical research. Third, although AI algorithms may surpass conventional standardized classification systems, such as tumor node metastasis staging and histopathological subtyping, in terms of predictive accuracy, they often sacrifice interpretability due to their complexity. This “black-box” nature makes it difficult to clarify the reasoning behind subtype classifications or TME scoring, and experimental processes are often irreproducible, which undermines clinical credibility. Therefore, the construction of multi-source, authentic, and standardized data platforms, along with unified validation criteria, is essential for improving data quality, security, and trustworthiness in clinical pathways. These measures are also crucial for facilitating the translation of AI outcomes from theoretical concepts into practical clinical applications.

CONCLUSION

This review takes CRC heterogeneity as a starting point, first analyzing the standardized workflows of scRNA-seq and ST, discussing their limitations and possible optimization strategies, and summarizing the achievements and potential breakthroughs these technologies have brought to CRC heterogeneity research. On this basis, it is widely recognized that further explore the application potential of AI and propose future directions for its development. In recent years, AI has emerged as a hot topic in academic research and has been widely applied across disciplines, with its influence expected to continue for the long term. In the fields of scRNA-seq and ST, AI thanks to its powerful computational and self-learning capabilities has significantly improved the efficiency of data preprocessing, analysis, cell type annotation, and multi-omics integration, effectively addressing the limitations of traditional methods and pushing the boundaries of current technologies. Although AI applications in these areas are still in an exploratory phase, their rapid iteration and advancement far exceed expectations, suggesting this phase will be short-lived. While revolutionary breakthroughs may not emerge immediately due to the trial-and-error nature of early exploration, quantitative change will eventually lead to qualitative transformation. We firmly believe that in the near future, AI technologies will reach a tipping point, offering life-changing opportunities for CRC patients and the broader cancer patient community.