Visual Multi-Omics Profiling Maps Molecular Programs Behind Renal Cell Carcinoma Heterogeneity

Clear cell renal cell carcinoma exhibits extensive morphological and genetic intra-tumoral heterogeneity, presenting substantial obstacles to accurate diagnosis and consistent therapeutic response. Tumors frequently contain an intricate mosaic of low-grade and high-grade histological regions coexisting within the same surgical specimen, with aggressive grade 4 components such as rhabdoid differentiation predicting particularly adverse patient outcomes. Despite the long-standing clinical reliance on the four-tiered World Health Organization and International Society of Urological Pathology grading system based primarily on nucleolar prominence, the precise molecular alterations underlying these distinct cell morphologies have remained elusive. In a study published in EMBO Molecular Medicine, a research team led by Hella A. Bolck, Peter Horvath, Holger Moch, and colleagues addressed this challenge by developing Deep Visual Multi-Omics, a technology framework that couples artificial intelligence-driven digital pathology and morphology-guided single-cell isolation with ultra-sensitive transcriptomic and proteomic profiling to interrogate defined cell populations directly from formalin-fixed paraffin-embedded tissue.

To capture the molecular consequences of histopathological variation, the investigators focused on five clear cell renal cell carcinoma cases exhibiting pronounced intra-tumoral micro-heterogeneity, confirmed through expert pathological examination to harbor spatially co-occurring grade 1/2, grade 3, and grade 4 rhabdoid regions. The researchers first established the baseline mutational profiles of these cases by performing whole-genome sequencing on macrodissected tumor areas pooled across all histological grades. The sequencing analysis confirmed consistent tumor mutational burdens across cases and revealed classic structural alterations typical of renal cell carcinoma, including chromosome 3p loss, 5q gain, and 14q loss, as well as recurring somatic mutations in the VHL tumor suppressor gene across all five patients, accompanied by alterations in SETD2, PBRM1, and BAP1. This confirmed that the cohort maintained canonical genetic features of renal cell carcinoma, providing a representative platform to explore how non-genetic molecular programs might differ between distinct histological grades within the same overall mutational background.

The Deep Visual Multi-Omics workflow relies on high-resolution whole-slide imaging and machine learning models implemented in the BIAS software suite to identify and isolate specific cell types. Hematoxylin and eosin-stained diagnostic sections were digitized into whole-slide images with multi-slice focus selection, followed by nucleus segmentation using the nucleAIzer deep neural network architecture. By extracting contextual features from individual nuclei, cytoplasmic regions, and surrounding cellular neighborhoods, the team applied multilayer perceptron models to pre-classify healthy kidney tissue from malignant regions while filtering out artifacts, stroma, and immune cells. For fine grading, the investigators trained convolutional neural networks to detect nucleoli characteristic of grade 1/2 versus grade 3/4 cells and specifically trained a model to detect rhabdoid cytomorphology. In tenfold cross-validation on annotated training sets, the image classification achieved sensitivity values ranging from 95.1 percent to 99.7 percent and specificities above 98 percent across classes, while retaining high accuracy for normal tubuli and grade 1/2 cells in independent test cases, with modest reductions in grade 3 and grade 4 distinctions reflecting the natural biological continuum between advanced tumor grades.

Following automated cell classification and high-confidence manual verification on the primary diagnostic slides, the researchers utilized laser capture microdissection on a Leica LMD6 system to isolate targeted cellular subsets. Nuclear contours were expanded by a six-micrometer margin to include cytoplasmic material, yielding enriched populations of low-grade, high-grade, rhabdoid, and adjacent normal tubular cells. For each morphological class per patient, two pools of 1,200 cells were excised without requiring tissue registration across adjacent sections. One pool was processed using an ultra-low-input total RNA sequencing protocol employing random priming, generating expression profiles covering 20,528 genes across nineteen high-quality samples. The second isolated pool was analyzed by data-independent acquisition liquid chromatography-mass spectrometry on an Evosep One coupled to a timsTOF Pro mass spectrometer, successfully quantifying 4,382 proteins across fourteen samples that passed strict analytical quality controls.

Principal component analyses and hierarchical clustering of the resulting transcriptomic and proteomic data demonstrated clear segregation between normal renal tubuli and malignant clear cell renal cell carcinoma cells. Within the tumor compartments, an ordered separation emerged along major principal components that aligned with increasing histological grade, a pattern that proved consistent across patients rather than driven by individual outliers. Differential expression analyses relative to normal renal tubuli revealed a progressive surge in both downregulated and upregulated molecules as tumors transitioned from grade 1/2 to grade 3 and grade 4 rhabdoid states. The gradual molecular shift was particularly evident in the proteomic landscape, demonstrating that histopathologically defined cell states correspond to systematic and detectable shifts in protein abundance.

Over-representation and pathway enrichment analyses of the differential expression data identified coordinated reprogramming across multiple essential metabolic pathways. Downregulated gene and protein sets across all tumor grades were heavily enriched for classical renal metabolic functions, reflecting progressive dedifferentiation. Specifically, cells demonstrated marked suppression of oxidative phosphorylation and tricarboxylic acid cycle enzymes, alongside reductions in normal renal processes including glucuronidation, transmembrane transport, and fatty acid beta-oxidation enzymes such as ACAT1, ECHS1, and EHHADH. Concurrently, upregulated programs across grades captured widespread activation of hypoxia-inducible factor 1 signaling and glycolytic cascades, evidenced by significant increases in VEGFA, GLUT1 (SLC2A1), ENO2, and PKM. These findings verified that the small, image-selected cell pools preserved canonical hallmarks of renal cell carcinoma biology while mapping their magnitude across advancing histological stages.

Although direct concordance between individual mRNA transcripts and their corresponding protein products was modest, identifying 40, 134, and 245 significantly overlapping genes in grade 1/2, grade 3, and grade 4 rhabdoid cells, respectively, network integration revealed deeper pathway alignment. Building on high-confidence protein-protein interaction networks from public repositories using the NOODAI platform and current-flow betweenness centrality metrics, the researchers analyzed modules unique to aggressive rhabdoid cells. In these cells, downregulated central nodes included factors essential for differentiated epithelial identity and cell adhesion, such as CDH1 (E-cadherin), ACE2, and ALDH3A2. In contrast, upregulated central nodes were enriched for regulators of the epithelial-to-mesenchymal transition, cell motility, and matrix interactions, including vimentin (VIM) and annexin A2 (ANXA2), indicating extensive cytoskeletal and architectural rewiring associated with discohesive growth.

Decomposition of the rhabdoid interaction networks highlighted elevated proliferation driven by the Forkhead box transcription factor FOXM1 as a distinctive biological feature. The FOXM1 network hub, which demonstrated substantially lower prominence in low-grade and intermediate-grade networks, coordinated an extensive cell cycle subnetwork comprising DNA replication factors such as CLSPN and CDC6, alongside critical cell cycle regulators including CCNA2, CDK3, CDC25A, CDC25C, PLK1, BUB1, and the anti-apoptotic factor BIRC5. Clinical survival evaluation within the TCGA-KIRC dataset confirmed that elevated FOXM1 expression correlated with significantly reduced overall survival, providing translational support for FOXM1-driven proliferative programs as a characteristic feature and candidate vulnerability in aggressive renal cell carcinoma subpopulations.

Extracellular matrix remodeling and cell-matrix interactions emerged as another defining feature of rhabdoid cells, characterized by prominent network centrality of integrin subunits including ITGA4, ITGA10, ITGB2, and ITGB7, alongside matrix-modulating factors such as FN1, PLAUR, and VEGFC. Concurrently, rhabdoid cells displayed intrinsic upregulation of immune-modulatory molecules, including type I interferon response targets, chemokines such as CCL4, CCL20, CXCL1, and CXCL13, the checkpoint ligand PD-L1 (CD274), and the multifunctional ectoenzyme CD38. Independent validation by immunohistochemistry across an expanded cohort of twelve heterogeneous clear cell renal cell carcinomas confirmed the selective loss of membrane E-cadherin alongside elevated expression of CD38 and ITGB2 in rhabdoid areas. Evaluation of the local immune microenvironment revealed dense CD8-positive T-cell infiltration in eleven of twelve rhabdoid specimens, with nine exhibiting co-expression of the T-cell exhaustion marker TOX, suggesting an active yet dysfunctional immune context influenced by tumor-intrinsic factors.

To evaluate the broader applicability of these observations, the researchers analyzed public clinical cohorts, identifying consistent upregulation of FOXM1, CD38, and ITGB2 alongside reduced CDH1 expression in rhabdoid tumors from the CheckMate 025 clinical trial and in grade 4 tumors from the TCGA-KIRC dataset. The authors noted several key limitations of the study, including the modest core cohort size necessitated by the rarity of surgical specimens containing clearly demarcated, spatially matched low-grade, high-grade, and rhabdoid components within the same tumor. In addition, the exploratory nature of the profiling lacked direct functional perturbation experiments, and the limited number of annotated rhabdoid cases with recorded immunotherapy exposure precluded definitive conclusions regarding therapeutic predictive value. Nevertheless, the study establishes Deep Visual Multi-Omics as an effective framework for resolving intra-tumoral heterogeneity in clinical tissues, directly linking spatial histopathology to the molecular machinery driving high-risk cancer phenotypes.

Subject of Research: Morphology-guided single-cell multi-omics profiling of clear cell renal cell carcinoma

Article Title: Deep visual multi-omics profiling links morphology and molecular programs in clear cell renal cell carcinoma

Article References: Bolck, H. A., Migh, E., Kriston, A., Zajac, N., Kreutzer, S., Totu, T., Leary, P., Kovacs, F., Rutishauser, D., Pfammatter, S., Grossmann, J., Lakshminarayanan, H., Basu, D., Litchfield, C., Buljan, M., Rupp, N. J., Horvath, P., & Moch, H. (2026). Deep visual multi-omics profiling links morphology and molecular programs in clear cell renal cell carcinoma. EMBO Molecular Medicine, 18(9), 3850-3877. https://doi.org/10.1038/s44321-026-00484-8

Image Credits: AI Generated

DOI: 10.1038/s44321-026-00484-8

Keywords: clear cell renal cell carcinoma, digital pathology, multi-omics, transcriptomics, proteomics, single-cell isolation, laser capture microdissection, rhabdoid differentiation, FOXM1, immune microenvironment, intra-tumor heterogeneity, histopathological grading, whole-genome sequencing, PD-L1, CD38, ITGB2