OMNIS: a spatially informed multi-omics deep-learning framework for tumor recurrence prediction and primary–metastatic tumor differentiation title page
By embedding three-dimensional genome organization into deep-learning models, OMNIS nominates biologically coherent, context-specific drivers of progression and may guide future biomarker development and personalized therapy in precision oncology.
Abstract
Background Cancer recurrence and distant metastasis are major causes of cancer-related death, yet existing biomarkers and single-omics models have limited accuracy and interpretability across tumor types. Methods We developed OMNIS (OMics Network Integration and Spatial representation), a convolutional deep-learning framework that embeds multi-omics profiles into a five-channel genomic image ordered by Hi-C–derived chromosomal proximity. Somatic mutation, copy-number alteration, DNA methylation and gene-expression data from 1,578 TCGA tumors across 33 cancer types were used to train classifiers for recurrence risk and for primary-versus-metastatic status. Performance was assessed by 10-fold cross-validation using AUROC, AUPR and threshold-based metrics. Integrated gradients yielded per-gene attribution scores; top-ranked genes were evaluated for prognostic value in two independent non-small cell lung cancer cohorts (GSE31210, n = 226; GSE135222, n = 27) using survival analyses. Results OMNIS achieved high discrimination for recurrence (AUROC/AUPR 0.970/0.937) and metastasis (0.980/0.883), with accuracies of 0.873–0.911 and negative predictive values ≥0.970 across tasks. Spatial genomic embedding accelerated convergence and outperformed non-spatial baselines. Attribution highlighted seven recurrence-associated genes (including IBA57, DNTTIP1, SLC20A2 and TMEM201) and ten metastasis-associated genes (including PLXNA1, POLR3D, TTLL4, SREBF2, TYMP and ZBTB7C). In external cohorts, expression of these genes showed independent, stage-dependent associations with progression-free and overall survival. Conclusion OMNIS is a spatially informed multi-omics framework that couples accurate prediction with gene-level interpretability. By embedding three-dimensional genome organization into deep-learning models, OMNIS nominates biologically coherent, context-specific drivers of progression and may guide future biomarker development and personalized therapy in precision oncology.
Deep learning models are increasingly used to analyze medical images, but their “black box” nature makes it hard to understand the underlying biology and slows down the development of targeted treatments. To tackle this, we built a multi-step approach that combines deep learning analysis of breast magnetic resonance imaging (MRI) with several types of molecular data, including gene activity, protein levels, and genetic information, along with laboratory experiments. Our MRI-based deep learning model accurately predicted whether breast cancer had spread to lymph nodes, and it performed consistently across 3 separate groups of patients. Causal inference using double least absolute shrinkage and selection operator (LASSO) and causal forest double machine learning established a significant effect of NBPF4 expression on the imaging-defined high-risk phenotype, independent of genomic confounders. When we looked at which genes were linked to the imaging-defined high-risk pattern, one gene called NBPF4 stood out because it was supported by all 4 kinds of evidence: imaging features, gene expression, protein data, and genetic association studies. Follow-up experiments in cells and animals showed that boosting NBPF4 activity made tumor cells grow faster, move more, form new lymphatic vessels, and spread to lymph nodes. Mechanistically, NBPF4 worked by activating the mitogen-activated protein kinase (MAPK) signaling pathway and triggering a process known as epithelial mesenchymal transition (EMT). Interestingly, tumors with high NBPF4 were sensitive to drugs that block one part of the MAPK pathway (JNK/p38) but resistant to another part (ERK), suggesting that the pathway had been rewired. Using this insight, computer-based drug screening and further testing identified MK-886 as a promising compound that could suppress NBPF4-promoted MAPK activation and tumor growth. Together, this work traces a complete path from a noninvasive imaging finding to a specific gene (NBPF4) and a potential treatment (MK-886). It establishes the NBPF4–MAPK–EMT axis as a key player in breast cancer metastasis and provides a general framework for turning imaging-based risk predictions into biological understanding and possible therapies.
D. Cai, Haoxuan Huang, Zijun Chen et al.· Research· 0 citations
Histopathologic evaluation remains central to cancer diagnosis and treatment planning, yet the molecular programs underlying distinct tissue morphologies aren't routinely accessible in clinical workflows. Spatial transcriptomic/proteomic platforms provide region-specific molecular measurements but are limited by cost, throughput, and scalability. Most computational pathology models rely on either bulk tissue-based gene expression or a focused gene/protein expression-panel prediction, thereby obscuring subregion-specific morpho-molecular relationships and limiting spatial interpretation of a wider gene/protein expression network. This limitation is particularly significant in triple-negative breast cancer (TNBC), which exhibits pronounced spatial heterogeneity across tumor, stroma, and immune compartments. We developed X-SPATIO, a spatially compatible computational pipeline designed to directly link hematoxylin and eosin (H&E) morphology with region-matched mRNA and protein expression. The model was trained on H&E-defined regions of interest paired with spatially-resolved omics data obtained from GeoMx Digital Spatial Profiling. Using a multiple-instance learning approach, X-SPATIO captures morpho-molecular associations, generating spatio-morphologic attention maps that indicate predictive tissue regions. X-SPATIO demonstrated strong performance across biologically relevant spatial biomarkers, achieving area under the curve values ranging from 0.79-0.97. Attention maps revealed spatial patterns consistent with known biology, indicating alignment between learned features and tissue organization. By integrating spatial molecular ground truth with routine histopathology, X-SPATIO enables cost-effective inference of spatial biomarker expression and establishes a foundation for biologically grounded discovery and precision oncology in TNBC.
Vibha R. Rao, Madhumala K. Sadanandappa, C. Black et al.· American Journal of Patholog...· 0 citations
Esophageal Cancer: Molecular Biology/Pathology
Spatial transcriptomic profiling was performed using the Xenium Prime 5K platform on treatment-naïve ESCC primary tumors (n=10) spanning mucosal to deep submucosal invasion, matched metastatic lymph nodes (n=11), and normal tissues (n=3). After QC filtering, 8,284,705 cells were classified into nine major cell types. Cancer cells were subclustered per sample using consensus non-negative matrix factorization (cNMF) to delineate spatially distinct subpopulations. To link transcriptomic phenotypes with clonal architecture, LCM-based whole genome sequencing was performed on 20 regions from 3 patients, enabling phylogenetic reconstruction across primary tumors and matched metastatic lymph nodes.
Cancer cell analysis revealed two independent axes of variation. The first axis stratified tumors into NRF2-high and NRF2-low subtypes regardless of invasion depth. NRF2-high tumors showed reduced immune infiltration, consistent with NRF2 pathway enrichment in non-responders from an independent neoadjuvant immunochemotherapy cohort. The second axis reflected invasion depth-dependent progression, where deeper tumors acquired EMT features alongside the emergence of COL10A1+COL11A1+ myCAFs that drove desmoplastic reactions at the tumor-stromal interface. Metastatic lymph nodes showed two distinct stromal patterns: those with desmoplastic reactions and those without. Phylogenetic reconstruction revealed that metastatic lymph nodes lacking desmoplastic reactions branched early from primary tumor clones, consistent with dissemination prior to stromal remodeling.
Integrating spatial transcriptomics with LCM-based WGS, we identified two axes underlying ESCC heterogeneity. NRF2-driven immune exclusion is relevant to immunotherapy resistance, while EMT-driven co-evolution with COL10A1+COL11A1+ myCAFs induces desmoplastic reactions. Phylogenetic analysis confirms that metastatic lymph nodes without desmoplastic reactions arose through early dissemination. These findings provide new insights into immune evasion mechanisms and metastatic timing in ESCC.
Integrating spatial transcriptomics with LCM-based WGS, we identified two axes underlying ESCC heterogeneity. NRF2-driven immune exclusion is relevant to immunotherapy resistance, while EMT-driven co-evolution with COL10A1+COL11A1+ myCAFs induces desmoplastic reactions. Phylogenetic analysis confirms that metastatic lymph nodes without desmoplastic reactions arose through early dissemination. These findings provide new insights into immune evasion mechanisms and metastatic timing in ESCC.
Jangwo N Oh, S. Park, Jong-Eun Park· Diseases of the esophagus· 0 citations
PURPOSE
Distant metastasis remains a major cause of treatment failure in head and neck (HN) cancer, highlighting the need for more accurate early risk stratification. This study developed and validated a deep radiomics framework to characterize tumor heterogeneity from pre-treatment CT images and improve prediction of distant metastasis-free survival (DMFS).
MATERIALS AND METHODS
This multicenter study included 3,421 HN cancer patients from four cohorts across twelve institutions (RADCURE, HN1, HN-PET-CT and TCGA-HNSC). Radiomics features were extracted from primary tumors and transformed into OmicsMaps, a structured representation that spatially organizes inter-feature relationships to facilitate learning of complex prognostic patterns. A convolutional neural network was trained to derive prognostic signatures, which were integrated with key clinical variables to construct an OmicsMap-clinical fusion model for patient risk stratification. Model performance was assessed using the concordance index (C-index) and time-dependent area under the receiver operating characteristic curve (AUC) in the RADCURE, HN1, and HN-PET-CT cohorts. Radiogenomic analyses using RNA-seq data were conducted in the TCGA-HNSC cohort to investigate biological characteristics associated with the imaging-defined risk groups.
RESULTS
The OmicsMap achieved C-index values of 0.742, 0.768, and 0.671 in the RADCURE, HN1, and HN-PET-CT cohorts, outperforming conventional radiomics approach by 5.84-6.37%. Incorporating clinical variables further improved generalizability, yielding C-index of 0.864 (HN1) and 0.730 (HN-PET-CT), with time-dependent AUCs of 0.727-0.895. The fusion model consistently stratified patients into distinct high- and low-risk groups for both DMFS and overall survival across cohorts (P < 0.01). Radiogenomic analyses revealed enrichment of immune-related pathways in the low-risk group, whereas the high-risk group exhibited a more aggressive phenotype enriched for proliferation, hypoxia, and epithelial-mesenchymal transition pathways, along with a fibrosis-prone tumor microenvironment characterized by extracellular matrix remodeling.
CONCLUSION
Modeling inter-radiomic feature relationships using the OmicsMap representation substantially improves CT-based prediction of DMFS and characterization of tumor heterogeneity in HN cancer, supporting precision risk stratification in clinical oncology.
R. Hou, Yujia Shen, Chuanbao Zhang et al.· International Journal of Rad...· 0 citations
Glioblastoma (GBM) is a highly aggressive brain tumor with an extremely poor five-year survival rate of 6.9%, largely attributable to the lack of reliable biomarkers. While competing endogenous RNA (ceRNA) and copy number variation (CNV) analyses each offer unique biomarker-identification potential, current approaches neglect the integration of multiple regulatory mechanisms for biomarker detection. To address this limitation, we applied relational graph convolutional networks (RGCNs) to ceRNA and CNV knowledge graphs through a novel late-fusion ensemble architecture. The proposed architecture outperformed baseline models and identified five novel biomarkers, including hsa-miR-196a and hsa-miR-224. Kaplan–Meier survival analysis and Cox regression indicated that the identified genes hold significant prognostic and diagnostic power, and the early stratification of the Kaplan–Meier curves indicates their potential for patient survival prediction. The results illustrate that a late-fusion RGCN ensemble effectively captures complex gene interactions, overcoming limitations of existing models and providing a framework for future biomarker discovery. The novel biomarkers serve as prospective targets for future GBM therapeutic development and candidates for non-invasive diagnostic assays.
Samarth Khandelwal, Nicholas Jarvis, J. Zhan· bioRxiv· 0 citations
BACKGROUND
Lung adenocarcinoma (LUAD) is a highly prevalent and lethal form of lung cancer. Brain metastasis (BrM) is a major cause of mortality in patients with LUAD. The tumor microenvironment (TME) of LUAD BrM remains incompletely characterized, and there is an urgent need to identify biologically relevant biomarkers and potential therapeutic vulnerabilities.
METHODS
Single-cell transcriptomic data from LUAD BrM samples were processed through dimensionality reduction, clustering, and cell annotation. BrM-associated core genes were identified from BrM scRNA-seq datasets using inferCNV and hdWGCNA. A prognostic model for LUAD was then constructed using 101 combinatorial approaches derived from 10 machine learning algorithms and validated in LUAD cohorts. We further performed supportive analyses including enrichment, pseudotime trajectory, cell-cell communication, immune infiltration, and single-cell-based drug sensitivity prediction to provide additional biological context. Key findings were validated by immunohistochemistry (IHC) in human tissues and by qRT-PCR and IHC in a mouse BrM model. Functional assays were performed in PC9-BrM cells to assess the roles of SEC61G and CFL1 in LUAD BrM.
RESULTS
Our study unveiled a comprehensive single-cell atlas of the BrM TME and established a robust 12-gene BrM-informed prognostic signature for LUAD, validated across multiple LUAD survival cohorts. Functional analysis revealed dysregulated cell cycle control in BrM progression. Immune profiling demonstrated significantly reduced CD4+ and CD8+ T-cell infiltration in high-risk subgroups. Cell-cell interaction analysis highlighted the MIF-CD74/CXCR4 and MIF-CD74/CD44 signaling axes as key regulators of TME remodeling. IHC in human tissues and validation using a mouse BrM model supported the biological relevance of SEC61G and CFL1 in LUAD BrM. qRT-PCR further showed elevated expression of SEC61G and CFL1 mRNA in brain-metastatic cells. Knockdown of SEC61G or CFL1 significantly inhibited PC9-BrM cell viability, proliferation, and migration.
CONCLUSIONS
This study identified genes associated with LUAD BrM and established a machine-learning-based prognostic signature for LUAD. Further experimental validation supports SEC61G and CFL1 as candidate biomarkers with potential therapeutic relevance.
Jing-Xiang Pang, Wentian Wu, Peiwen Zhu et al.· Cancer Cell International· 0 citations