MRI-Based Deep Learning Guides Multi-Omics Discovery of NBPF4 as a Therapeutic Target for Breast Cancer Lymph Node Metastasis

MRI-Based Deep Learning Guides Multi-Omics Discovery of NBPF4 as a Therapeutic Target for Breast Cancer Lymph Node Metastasis

PDF

Dianqi Cai¹^,²^,³, Haoxuan Huang⁴, Zijun Chen¹, Chao Zhu⁵, Zehui Li¹, Ming Chen¹, Zhiyi Guo¹, Jie Liu⁶, Gengxi Cai⁷^,^*, Wenjun Mao⁸^,⁹^,^*, Jianguo Lai¹^,²^,^*

Research. Vol 9 Article ID 1402

Less

Research. Vol 9 Article ID 1402

• RESEARCH ARTICLE •

MRI-Based Deep Learning Guides Multi-Omics Discovery of NBPF4 as a Therapeutic Target for Breast Cancer Lymph Node Metastasis

Full

Dianqi Cai¹^,²^,³, Haoxuan Huang⁴, Zijun Chen¹, Chao Zhu⁵, Zehui Li¹, Ming Chen¹, Zhiyi Guo¹, Jie Liu⁶, Gengxi Cai⁷^,^*, Wenjun Mao⁸^,⁹^,^*, Jianguo Lai¹^,²^,^*

Affiliations

¹Guangdong Provincial People’s Hospital, Guangdong Academy of Medical Sciences, Southern Medical University, Guangzhou, China.

²Guangdong Provincial People’s Hospital Ganzhou Hospital, Ganzhou Municipal Hospital, Ganzhou Innovation Center, National Regional Medical Center, Gannan Medical University, Ganzhou, China.

³The Sixth Affiliated Hospital, Sun Yat-sen University, Guangzhou, Guangdong Province, China.

⁴Department of Urology, The Third Affiliated Hospital (the First Hosptial of Nanchang), Jiangxi Medical College, Nanchang University, Nanchang 330006, Jiangxi, China.

⁵Department of Clinical Laboratory, The First Affiliated Hospital of Nanchang University, Nanchang, China.

⁶Department of Breast Cancer, Foshan Women and Children Hospital Affiliated to Guangdong Medical University, Foshan, China.

⁷The First People’s Hospital of Foshan (Foshan Hospital Affiliated to Southern University of Science and Technology), School of Medicine, Southern University of Science and Technology, Guangdong, China.

⁸Department of Thoracic Surgery, The Affiliated Wuxi People’s Hospital of Nanjing Medical University, Wuxi People’s Hospital, Wuxi Medical Center, Nanjing Medical University, Wuxi, China.

⁹Wuxi College of Clinical Medicine, Nanjing Medical University, Wuxi, China.

About Author:

^† These authors contributed equally to this work.

Published: 2026-08-20 doi: 10.34133/research.1402

Outline

Abstract

Less

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.

Cite this Article

Dianqi Cai, Haoxuan Huang, Zijun Chen, Chao Zhu, Zehui Li, Ming Chen, Zhiyi Guo, Jie Liu, Gengxi Cai, Wenjun Mao, Jianguo Lai. MRI-Based Deep Learning Guides Multi-Omics Discovery of NBPF4 as a Therapeutic Target for Breast Cancer Lymph Node Metastasis[J]. Research, 2026 , 9 (8) : 1402 - . DOI: 10.34133/research.1402

Full Text

Less

Introduction

Less

Breast cancer (BRCA) axillary lymph node metastasis (LNM) is a well-established prognostic factor [1–4]. However, identifying high-risk patients preoperatively and developing effective therapies to counteract metastatic spread remain significant clinical challenges [4,5]. While deep learning models show increasing promise in predicting LNM from routine imaging [6–12], they often function as “black boxes”, failing to illuminate the underlying biological mechanisms, such as the cellular origins and molecular drivers, of the metastatic phenotypes they detect. In parallel, genomic and transcriptomic studies have identified many molecular alterations associated with metastasis [13–19], yet a major translational gap persists in refining these findings into biologically coherent, druggable targets and linking them back to clinically accessible imaging features.

To bridge this gap, integrative approaches are needed to systematically link noninvasive radiographic patterns with causal molecular mechanisms, thereby transforming imaging-based risk stratification into a discovery engine for novel biology and therapeutic targets. Here, we report a multi-stage radiogenomic pipeline designed to achieve this goal. This pipeline begins with a clinically motivated question: how to preoperatively identify patients at high risk of LNM. To address this, we first developed a deep learning Magnetic Resonance Imaging (MRI) model as a noninvasive screening tool to enrich for tumors with high metastatic potential. The imaging-defined high-risk phenotype then served as the entry point for multi-omics integration, which converged on NBPF4 as the central molecular driver. Subsequent functional validation and drug screening completed the discovery-to-therapy cycle. We first developed a deep learning model for noninvasive LNM prediction and then integrated this imaging phenotype with multi-omics data to prioritize candidate driver genes. This approach identified NBPF4 as a central hub gene supported by convergent evidence from transcriptomics, proteomics, and genetics. Mechanistically, NBPF4 activated the mitogen-activated protein kinase (MAPK) signaling pathway and induced epithelial–mesenchymal transition (EMT), and we ultimately identified MK-886 as a candidate therapeutic compound capable of suppressing this NBPF4–MAPK–EMT axis that promotes LNM.

This study provides an integrative framework for generating mechanistic hypotheses and therapeutic leads regarding the molecular basis of imaging-defined metastatic risk in BRCA. Our work establishes a methodology for converting radiographic phenotypes into discoverable biological mechanisms and viable therapeutic strategies.

Results

Less

Integrated radiogenomic pipeline predicts LNM and identifies a target gene

This study established and validated a deep learning model based on the MobileNet V3 Small (MV3S) architecture for predicting the risk of LNM from preoperative contrast-enhanced breast MRI data in a multicenter retrospective study (Fig. 1). Multi-omics analysis of the high-risk group identified by the model prioritized candidate genes associated with MAPK signaling and EMT, leading to a focus on NBPF4 for functional validation. Experiments demonstrated that overexpression of NBPF4 promotes tumor cell proliferation, migration, and lymphangiogenesis, mediated through activation of the MAPK pathway and induction of EMT. Further computational drug screening identified MK-886, which was shown to inhibit NBPF4-associated MAPK pathway activity and tumor growth, confirming its therapeutic potential. Causal inference using double least absolute shrinkage and selection operator (LASSO) and causal forest double machine learning (DML) confirmed a significant effect of NBPF4 expression on the imaging-defined high-risk phenotype, independent of genomic confounders, establishing NBPF4 as an independent risk factor for increased imaging risk. This research constructs a complete framework from imaging phenotype to molecular mechanisms and therapeutic strategies.

Deep learning predicts LNM and links imaging to pro-metastatic biology

To develop a noninvasive tool for BRCA risk assessment, we constructed a deep learning pipeline based on preoperative MRI images. Following standardized preprocessing (Fig. 2A), we implemented an MV3S convolutional neural network for feature extraction and prediction (Fig. 2B). The model achieved robust performance in stratifying patients into high- and low-risk groups across 3 independent cohorts [Duke, area under the curve (AUC) = 0.869; The Cancer Imaging Archive (TCIA), AUC = 0.838; Foshan, AUC = 0.756; Table S1 and Fig. 2C to E]. The predicted risk groups were strongly associated with LNM⁺ status, a key clinical prognostic factor (Fig. 2F to H). To ensure the robustness of this association, we performed propensity score matching (PSM) on the Duke cohort, which balanced potential confounders and yielded an even higher predictive accuracy (AUC = 0.895; Fig. 2I to L).

To identify the biological basis of the imaging-derived risk phenotype, we then sought to interrogate the transcriptomic correlates of radiomic. Integration of imaging and transcriptomic data (Fig. 2M) yielded a set of 51 highly expressed genes. To ensure the robustness of this signature, we performed a rigorous feature selection using the Boruta algorithm, which confirmed 49 of these genes as stable, important predictors (Fig. S1). Their differential expression distinguished the risk groups, as confirmed by principal components analysis (Fig. 2N and O). Functional enrichment revealed that these genes were significantly enriched for pro-tumorigenic pathways, notably MAPK signaling and EMT regulation (Fig. 2P). To independently validate the link to EMT, we performed gene set enrichment analysis (GSEA) using an EMT gene set on samples stratified by the expression of canonical EMT markers (CDH2 or VIM). The radiomics risk gene signature was significantly enriched in samples with high EMT marker expression (Fig. 2Q), corroborating the association. Finally, at single-cell resolution, we observed that EMT and MAPK pathway activity scores were significantly elevated in cancer clusters within LNM⁺ compared to LNM-BRCA (Fig. 2R), indicating that these pro-metastatic programs are activated in a malignant epithelial cell manner within the lymph node metastatic tumor ecosystem.

To quantitatively identify which imaging features drive the high-risk prediction, we performed SHapley Additive exPlanations (SHAP) attribution analysis on the radiomic feature set. The top-contributing features were measures of margin irregularity, intratumoral heterogeneity, and peritumoral enhancement, all of which are biologically plausible correlates of EMT and MAPK-driven aggressive behavior. Consistent with this, attention heatmaps revealed that our deep learning model primarily focuses on the tumor–stroma interface and heterogeneous tumor cores (Fig. 2S to U). These radiological “hotspots” correspond to the regions where the above high-contribution features are most salient.

This multi-modal analysis establishes a deep learning model for noninvasive LNM prediction and links the radiographic phenotype to MAPK/EMT signaling programs. In summary, the mapping from the MRI-derived risk phenotype to the transcriptomic gene signature and onward to causal inference proceeded as follows: (a) deep learning (DL)-based risk stratification of patients into high- and low-risk groups; (b) differential expression analysis between these groups using paired RNA-sequencing (RNA-seq) data; (c) Boruta-based feature selection to retain stable genes; (d) functional enrichment analysis linking the signature to MAPK/EMT pathways; and (e) causal inference using double LASSO to control for genomic confounders and Causal Forest DML to estimate the average treatment effect of NBPF4 on Radrisk [average treatment effect (ATE) = 0.037, 95% CI: 0.005 to 0.070], establishing a statistically significant causal link from NBPF4 expression to the imaging-defined high-risk phenotype.

Multi-omic convergence validates NBPF4 as an LNM-associated gene

To substantiate the biological relevance of the radiomics risk gene set, we validated it across orthogonal molecular layers. Immunohistochemistry identified a subset with medium-to-high protein expression in clinical samples (Fig. 3A), while a complementary set of genes showed low or absent protein levels (Fig. S2), reflecting the heterogeneity of protein abundance within the radiomics risk signature. Transcriptomic analysis independently confirmed a subset that was significantly up-regulated at the mRNA level in BRCA [logFC > 0.5, false discovery rate‌ (FDR) < 0.05] (Fig. 3B).

To establish a genetic link, we performed Bayesian colocalization analysis across 12 BRCA genome-wide association study (GWAS) datasets. This identified a set of radiomics risk genes, harboring genetic variants that jointly influence gene expression and disease risk (Fig. 3C). Intersection analysis formally nominated NBPF4 as the sole gene common to the radiomic, transcriptomic, proteomic, and genetic evidence layers (Fig. 3D). We further refined the genetic evidence, finding that the variant rs3870737 showed strong support (PP.H4 > 0.8) for being a shared causal variant across all 12 datasets (Fig. 3E), as detailed in the regional colocalization plots (Figs. S3 and S4). Detailed visualization at the NBPF4 locus in a representative cohort confirmed the spatial overlap of expression quantitative trait locus (eQTL) and GWAS association signals. The colocalization evidence threshold was defined as PP.H4.abf > 0.8, consistent with standard practice (Fig. 3F). Finally, in an independent clinical cohort, NBPF4 mRNA expression was significantly elevated in tumors from patients with pathologically confirmed LNM compared to those without metastasis (Fig. 3G).

Collectively, this sequential multi-omic integration, through a progression from radiographic phenotype to protein expression, transcriptomic dysregulation, population-level genetic association, and direct clinical validation, systematically nominates and validates NBPF4 as a robust, multi-layer biomarker candidate functionally linked to the risk of LNM in BRCA.

Core confounders were identified via strict double LASSO. From the initial 90 candidate genes, target-specific LASSO regularization paths were constructed for both Radrisk and NBPF4. Bootstrap stability analysis demonstrated that a strict 0.5 threshold effectively filtered unstable features (Fig. 3H). After applying the dual-filtering criteria, 17 core genes exhibited high robustness across the LASSO trajectories (Fig. 3I). These variables were retained as control variables for subsequent causal inference (Fig. 3J). Coefficient profiling revealed that genes such as ABAT and RAB30 strongly influenced the NBPF4 model, whereas RAB24 was the only gene that both greatly impacted the 2 models (Fig. 3K). We observed a significant positive causal effect of NBPF4 on Radrisk. After rigorously controlling for the 17 core confounding genes, the Causal Forest DML model demonstrated a significant causal impact of NBPF4 on Radrisk. The estimated ATE was 0.037 (95% CI: 0.005 to 0.070) (Fig. 3L). Because the 95% CI did not include zero, this causal effect was deemed statistically significant. These findings indicate that, independent of high-dimensional genomic confounders, elevated NBPF4 expression is an independent risk factor for increased Radisk.

NBPF4 promotes BRCA growth in vitro and in vivo

We next sought to characterize the functional role of NBPF4, a top candidate emerging from our multi-omics validation. Analysis of paired samples from the TCGA (The Cancer Genome Atlas)–BRCA cohort confirmed that NBPF4 mRNA is significantly up-regulated in breast tumors compared to matched normal adjacent tissue (Fig. 4A). This finding was validated in an independent set of 30 patient-matched clinical samples (Fig. 4B). NBPF4 expression was also elevated across a panel of BRCA cell lines compared to a nontumorigenic mammary epithelial cell line (Fig. 4C).

To assess its functional impact, we ectopically overexpressed NBPF4 in 2 BRCA cell lines, MDA-MB-231 and BT-549 (Fig. 4D). NBPF4 overexpression significantly enhanced cellular proliferation, as measured by both metabolic activity [Cell Counting Kit-8 (CCK-8) assay; Fig. 4E] and DNA synthesis [5-ethynyl-2′-deoxyuridine (EdU) incorporation; Fig. 4F and G]. The pro-tumorigenic function of NBPF4 was further substantiated in vivo. Xenograft tumors derived from NBPF4-overexpressing cells grew significantly faster and formed larger end-point masses compared to tumors from control cells (Fig. 4H and I).

To contextualize the role of NBPF4 beyond BRCA, we performed extensive pan-cancer analyses. NBPF4 was frequently up-regulated at the transcriptomic and proteomic levels across multiple malignancies (Fig. S5A and B) and showed genetic associations with cancer risk in several GWAS (Fig. S5C). Its expression displayed notable tissue specificity, with highest levels in testis among normal tissues, and was detectable in immune cells and various cancer cell lines (Fig. S6). Importantly, NBPF4 expression held prognostic value, capable of distinguishing tumor from normal tissue and associating with patient survival outcomes in a cancer-type-dependent manner, including in external BRCA cohorts (Fig. S7).

Collectively, these functional data move beyond association to establish a causal role for NBPF4. By demonstrating that its overexpression is sufficient to drive both in vitro proliferation and in vivo tumor growth, we define NBPF4 as a core effector molecule that mechanistically links the radiogenomically defined risk phenotype to accelerated tumor progression, nominating it as a direct therapeutic target within this aggressive disease axis.

NBPF4 orchestrates an LNM pro-metastatic program

We next investigated the functional mechanisms by which NBPF4 might promote LNM. Ectopic overexpression of NBPF4 significantly increased the migratory capacity of MDA-MB-231 and BT-549 cells in Transwell assays (Fig. 5A and B). Conditioned medium from NBPF4-overexpressing cells also potently enhanced the tube-forming capability of human lymphatic endothelial cells (HLECs) in vitro (Fig. 5C and D), suggesting a role in promoting lymphangiogenesis.

To understand the molecular basis of these phenotypes, we performed transcriptomic analyses. GSEA in 2 independent BRCA cohorts revealed that high NBPF4 expression was significantly associated with the enrichment of EMT scores (Fig. 5E and F). Broad pathway analysis further confirmed the association of high NBPF4 expression with multiple pro-tumorigenic processes, including EMT and inflammatory responses (Fig. 5G). This link to EMT was corroborated at the single-cell level, where a significant positive correlation was observed between the average expression of NBPF4 and the canonical EMT marker CDH2 (N-cadherin) across multiple independent single-cell RNA-seq datasets (Fig. 5H). At the protein level, NBPF4 overexpression led to a marked decrease in E-cadherin and concurrent increases in N-cadherin and Vimentin, confirming the induction of a full EMT program (Fig. 5I).

Collectively, this multi-faceted analysis demonstrates that NBPF4 orchestrates a pro-metastatic program in BRCA cells, encompassing enhanced migration, promotion of a lymphangiogenic niche, and activation of the EMT transcriptional network, thereby providing a mechanistic foundation for its role in facilitating LNM.

NBPF4 links to MAPK pathway activation and drug response

To explore the mechanistic basis of NBPF4’s pro-tumorigenic role, we investigated its relationship with key oncogenic signaling. In the TCGA-BRCA cohort, NBPF4 mRNA levels showed a significant positive correlation with a gene expression signature representing MAPK pathway activity (Fig. 6A). GSEA in 2 independent BRCA cohorts confirmed that tumors with high NBPF4 expression were significantly enriched for MAPK pathway gene sets (Fig. 6B and C), indicating pathway activation in this context.

We next investigated the therapeutic implications of the association between NBPF4 expression and MAPK pathway activity. Analysis of the GDSC1 drug sensitivity database using a ridge regression model (pRRophetic) revealed that BRCA patients with high NBPF4 expression showed distinct response patterns to MAPK pathway inhibitors. Specifically, the NBPF4-high group demonstrated significantly increased sensitivity to inhibitors targeting the c-Jun N-terminal kinase (JNK)/p38 branches (VX.702, BIRB.0796) but exhibited marked resistance to extracellular signal-regulated kinase (ERK) pathway inhibitors (CMK, XMD8.85) compared to the NBPF4-low group (Fig. 6D and E). This differential drug sensitivity pattern was particularly intriguing given our protein-level validation results. Western blot analysis confirmed that NBPF4 overexpression led to increased phosphorylation of key MAPK effectors, with particularly strong activation of ERK1/2, along with more moderate activation of p38 and JNK (Fig. 6F). The apparent paradox where strong ERK activation correlates with resistance to ERK-targeted therapy suggests that NBPF4-promoted tumors may undergo pathway rewiring, potentially developing alternative survival dependencies that reduce their reliance on the highly activated ERK branch while increasing vulnerability to JNK/p38 inhibition. These findings reveal that NBPF4 not only activates the MAPK pathway but also orchestrates a therapeutic vulnerability shift, suggesting that NBPF4 expression status could serve as a biomarker for guiding targeted therapy selection in BRCA, potentially favoring JNK/p38 inhibitors over ERK-targeted approaches in NBPF4-high tumors.

Collectively, these data establish a multi-level connection between NBPF4 and MAPK signaling, spanning from transcriptional correlation and pathway enrichment to functional protein activation, and reveal a paradoxical yet therapeutically relevant pattern: High NBPF4 expression predicts sensitivity to JNK/p38 inhibition but resistance to ERK inhibition. This branch-specific vulnerability positions NBPF4 as a potential biomarker for guiding personalized MAPK-targeted therapy selection.

NBPF4 correlates with an aggressive oncogenic state and LNM in vivo

To systematically evaluate the functional context of NBPF4, we analyzed its association with 14 predefined oncogenic states (e.g., metastasis, EMT, and angiogenesis) across 2 independent BRCA transcriptomic cohorts. In both the GSE9893 and GSE7390 datasets, NBPF4 expression showed significant positive correlations with multiple aggressive functional programs (Fig. 7A and B). The strongest and most consistent association was with the metastasis state (GSE9893: R = 0.58, P = 1.9e−15; GSE7390: R = 0.21, P = 0.0032), positioning NBPF4 within a transcriptional program conducive to tumor dissemination.

We next sought causal, functional evidence for this correlative link using an in vivo popliteal LNM model. Injection of NBPF4-overexpressing (OE) MDA-MB-468 cells into the footpads of nude mice resulted in significantly larger and heavier metastatic lesions in the draining popliteal lymph nodes compared to vector-control cells (Fig. 7C to F). The primary footpad tumors in the NBPF4-OE group also exhibited increased growth, as measured by both volume and weight (Fig. 7G to I), confirming the pro-tumorigenic role of NBPF4 observed in prior in vitro assays.

Collectively, this figure integrates computational correlation with direct in vivo validation, demonstrating that NBPF4 is transcriptionally linked to a metastatic functional state in human tumors and is sufficient to drive lymphatic metastasis and tumor growth in a preclinical model. This multi-modal evidence solidifies NBPF4 as a functional mediator of BRCA progression and metastasis.

Integrative discovery of MK-886 targeting the NBPF4-MAPK axis

To identify potential therapeutics targeting NBPF4-mediated oncogenicity, we performed a pan-cancer Connectivity Map (cMap) analysis using the optimal XSum algorithm. Screening 1,288 compounds against an NBPF4-high gene signature across 24 cancer types identified MK-886 as a candidate with broad activity (Fig. S8). Subsequent BRCA-specific analysis confirmed MK-886 as a top candidate predicted to reverse the NBPF4-associated transcriptional profile (Fig. 8A).

We first tested whether MK-886 could counteract the molecular effects of NBPF4. Treatment with MK-886 markedly reduced the elevated phosphorylation levels of ERK1/2, p38, and JNK induced by NBPF4 overexpression in both MDA-MB-231 and BT-549 cells (Fig. 8B and C), confirming its inhibitory effect on the NBPF4-activated MAPK pathway. Functionally, MK-886 treatment significantly decreased the enhanced DNA synthesis observed in NBPF4-overexpressing cells, as shown by EdU incorporation assays (Fig. 8D and E). Consistent with this, CCK-8 assays demonstrated a concentration-dependent suppression of cell proliferation by MK-886 (Fig. 8F and G). Finally, in a mouse xenograft model, administration of MK-886 potently attenuated the accelerated tumor growth driven by NBPF4 overexpression (Fig. 8H and I).

Together, this integrated computational and experimental pipeline demonstrates that MK-886 effectively suppresses NBPF4-driven tumorigenesis by inhibiting the associated MAPK pathway activation, providing a preclinical rationale for targeting this axis in NBPF4-high expression.

Discussion

Less

The transformation of imaging patterns into mechanistic understanding remains a fundamental challenge in oncology, particularly in the context of metastasis where cellular plasticity and molecular adaptation drive progression [7,20–24]. Although current deep learning models can predict the risk of metastasis or treatment response based on imaging features, their primary function is to establish the correlation between these features and clinical outcomes. A key limitation is their inherent inability to elucidate the underlying causal mechanisms, such as the dynamics of cellular plasticity or the specific regulation of molecular pathways [25–27]. While deep learning has successfully decoded prognostic information from medical images, this advancement has inadvertently reinforced a diagnostic paradigm centered on correlation rather than causation [19,27–30]. The resultant “black box” problem is not merely an interpretability issue [31,32]. It represents a deeper methodological gap that separates descriptive phenotyping from the functional biology required for therapeutic intervention [33,34]. This study was designed to bridge this gap by establishing a closed-loop pipeline that explicitly connects an image-derived prognostic label to a functionally implicated molecular node and a candidate therapeutic strategy.

Recent radiogenomic research has predominantly focused on establishing parallel associations, correlating imaging features with molecular data [19,35–37]. While valuable, this approach often yields lists of candidate genes or pathways without a clear hierarchy of causality or a defined functional anchor [38]. Our methodology diverges from this convention by utilizing the imaging-derived risk phenotype as a functional screening tool. In this framework, the deep learning model functions as a selective filter, enriching for a tumor subpopulation whose defining biology is subsequently decoded through integrated multi-omics analysis. This causal prioritization led us to NBPF4, a gene whose role in BRCA metastasis was previously obscure. The convergence of transcriptional, proteomic, and genetic evidence around NBPF4 validates this approach, demonstrating that imaging signatures, when properly leveraged, can pinpoint central regulatory nodes within a complex disease process. To further strengthen the causal link between the imaging phenotype and the molecular driver, we performed a double LASSO selection to control for genomic confounders and applied a Causal Forest DML model. The analysis revealed a significant positive association between NBPF4 and the imaging risk score, providing statistical evidence consistent with a causal hypothesis at the population level. This result links the MRI-defined high-risk phenotype to NBPF4–MAPK–EMT activity, although direct spatial co-registration or interventional validation would be needed to establish a definitive causal chain.

The MAPK pathway and the EMT program are widely recognized as 2 central pillars driving cancer metastasis [39,40]. Activation of the MAPK cascade, particularly through the ERK and p38 branches, promotes tumor cell migration, invasion, and lymphangiogenesis across various malignancies, including esophageal squamous cell carcinoma, clear cell renal cell carcinoma, BRCA, and melanoma [41–46]. Concurrently, the EMT program facilitates metastatic dissemination in cancers such as nasopharyngeal carcinoma, cervical squamous cell carcinoma, oral squamous cell carcinoma, and pancreatic cancer [47–50]. These 2 pathways are frequently interconnected, as MAPK signaling can directly phosphorylate key EMT transcription factors, and their cooperative regulation is evident in cancers like gallbladder carcinoma and clear cell renal cell carcinoma [44,51]. Notably, prior research on NBPF4 had overlooked its role in BRCA metastasis, a gap that highlights the novelty of our discovery.

In this context, our functional investigation reveals that NBPF4 operates as a potent orchestrator of metastatic competence. It enhances proliferation, migration, and lymphangiogenesis in vitro and, critically, promotes LNM in vivo. Mechanistically, it converges on the MAPK pathway and the EMT program. However, a more nuanced finding emerged from the drug sensitivity analysis: NBPF4-high tumors displayed a paradoxical profile of ERK pathway activation coupled with ERK inhibitor resistance, alongside heightened vulnerability to JNK/p38 inhibition. This suggests that NBPF4 does not merely amplify the MAPK cascade but actively rewires its signaling logic, potentially creating branch-specific oncogenic dependencies. This insight, which would be invisible to a purely correlative analysis, underscores the value of moving from phenotypic association to gene-centric functional dissection.

The translational potential of this integrated pipeline is exemplified by the identification of MK-886 as a candidate therapeutic agent. The ability of this compound to suppress NBPF4-induced phenotypes, including reduced MAPK phosphorylation and inhibition of in vivo tumor growth, provides a compelling preclinical proof-of-concept. More broadly, our work outlines a pragmatic translational strategy. This strategy involves initial population-wide risk stratification using a noninvasive imaging tool, followed by a focused molecular assay to identify the patient subset most likely to respond to a targeted agent such as MK-886. This tiered approach effectively balances clinical scalability with biological precision.

We acknowledge several limitations that chart a course for future work. A notable challenge is the high sequence homology between NBPF4 and its family member NBPF6, which hindered specific knockdown via small interfering RNA (siRNA)/short hairpin RNA (shRNA) and complicated definitive rescue experiments. Therefore, establishing NBPF4 as a strict genetic dependency will require more precise genetic tools, such as CRISPR-based approaches, in follow-up studies. Beyond this, the precise biochemical function of NBPF4 and its direct molecular interplay with the MAPK pathway require further elucidation. The efficacy and safety profile of MK-886 must be rigorously established in advanced preclinical models. Furthermore, investigating whether this NBPF4-MAPK axis is a common feature in other aggressive carcinomas could reveal a broader therapeutic paradigm. A further limitation is that our retrospective biobank retains routine formalin-fixed and paraffin-embedded (FFPE) blocks without spatial orientation relative to preoperative MRI; consequently, direct voxel-level spatial co-registration between MRI-attention regions and molecular/immunohistochemical markers (e.g., NBPF4, p-ERK/p-p38, EMT, and lymphangiogenesis markers) could not be performed. The proposed MRI–NBPF4–MAPK–EMT axis should therefore be regarded as a hypothesis-generating model rather than a validated causal chain. The current evidence relies on population-level statistical inference (double LASSO + Causal Forest DML) and biological plausibility of the model-attended regions (tumor–stroma interface, heterogeneous cores), but lacks direct voxel-level spatial co-registration or in vivo imaging validation. Prospective studies with orientation-tracked specimens and/or spatial transcriptomics will be required to close this spatial evidence gap. Additionally, we did not perform small-animal MRI to directly assess whether NBPF4 overexpression alters imaging phenotypes. Future studies incorporating in vivo imaging in xenograft models would provide more direct evidence for the causal link between NBPF4 and imaging features. Until such spatial and imaging-based functional evidence is obtained, the causal interpretation of the MRI–NBPF4–MAPK–EMT axis should be considered provisional.

Conclusion

Less

In conclusion, this study demonstrates a principled framework for converting imaging-based prognostication into biological discovery and therapeutic hypothesis. By tethering a deep learning phenotype to a prioritized candidate gene and a candidate drug, we provide more than a new biomarker. We offer a proof-of-concept framework for a new type of translational research that starts with the patient’s clinical image, proceeds through multi-omics convergence and causal inference to identify a molecular driver, and culminates in a mechanistically grounded therapeutic strategy.

Materials and Methods

Less

Study design

We implemented a multi-center retrospective study leveraging radiogenomics to assess the presence of LNM in BRCA in 3 steps. Firstly, we established and evaluated a DL model based on MRI to predict LNM in BRCA patients. Subsequently, multi-faceted genomic differential analysis and enrichment analysis were conducted on risk prediction groups based on this model.

Datasets and participants

Radiological and clinical data for the Duke cohort and TCIA cohort were obtained from TCGA [52]. The data of the Foshan cohort originated from the First People’s Hospital of Foshan. Radiological and clinical data for the Duke cohort and the imaging and transcriptomic data derive from the same patient cohort and are paired by TCGA ID, with TCGA providing the clinical prognostic and transcriptomic data and TCIA providing the corresponding imaging data [52]. The data of the Foshan cohort originated from the First People’s Hospital of Foshan. Inclusion criteria were as follows: (a) patients with pathological diagnosis of BRCA and (b) patients who have undergone preoperative enhanced MRI scans. Image processing was the same as in our previous study [19]. MRI scans were acquired on 1.5- or 3.0-T scanners (GE/Siemens/Philips) across sites; typical sequences were axial pre- and post-contrast T1-weighted [repetition time (TR) ~ 400 to 700 ms, echo time (TE) ~ minimal, slice thickness 1.0 to 2.0 mm with ≤0.2 mm gap]. To reduce intersite variability, all images were resampled to isotropic 1 × 1 × 1 mm voxels, intensity z-normalized (by masked tumor/breast region), and rigidly aligned to a common orientation before feature extraction or model input. The methodological framework of this study was illustrated in Fig. 1. The final import of images was illustrated in Fig. 2A. This study was approved by our hospital and conducted under the guidance of the Declaration of Helsinki. Given the retrospective study design, informed consent was waived in Duke and TCIA cohort. This Foshan cohort involving human participants was reviewed and approved by the Ethics Committee of our hospital (approval number: FSYYY-EC-SOP-008-02.0-A09). The work has been reported in line with the STROCSS criteria [53]. In the statistical analysis, the 95% confidence intervals for the AUC were calculated using the DeLong method.

Baseline characteristic

The baseline characteristic of the study was presented in Table S1. Our study analyzed datasets from 3 cohorts of BRCA patients: Duke cohort (n = 313), TCIA cohort (n = 85), and Foshan cohort (n = 101).

Construction of deep learning model

We engineered a DL model, dubbed MV3S. The structural design and data intake process of the model were depicted in Fig. 2B. The average of the predicted probabilities of the multi-layered and multi-processed images for each training sample is the model’s final risk prediction. Using the Youden’s index from the training data, we stratified risk across the datasets, enabling targeted intervention strategies. To confirm the independence of risk grouping as a predictive factor, we applied PSM to mitigate confounding factors between the risk groups, aiming to minimize intersample variability with the help of the “MatchIt” package (method = “nearest”, ratio = 1). Expanded to include: region-of-interest (ROI) annotation method (manual delineation on axial post-contrast T1 images), data split strategy (Duke for training/internal validation; TCIA and Foshan as external test sets), and threshold selection [Youden index on training receiver operating characteristic (ROC)]

Cell line and cell culture

Chinese Academy of Sciences Cell Bank in Shanghai provided several cells including MCF-10A, MDA-MB-231, MDA-MB-468, BT-549, and MCF-7. A Special medium for MCF-10A cells (Procell, China) was used to culture the MCF-10A cells. MDA-MB-231 and MDA-MB-468 cells were cultured in Dulbecco's modified Eagle's medium (DMEM; Gibco, China) supplemented with 10% fetal bovine serum (FBS; Cellmax, China), while BT-549 and MCF-7 cells were cultured in RPMI 1640 medium (Gibco, China) supplemented with 10% FBS (Cellmax, China). All cells were maintained at 37 °C in a 5% CO₂ atmosphere.

Plasmid and cell transfection

We purchased pcDNA3.1-NBPF4 from GenePharma (Suzhou, China) in order to achieve overexpression of the NBPF4. The cells were transfected with DNA at 60% to 80% confluency by using jetPRIME transfection reagent (Polyplus, USA). After 6 h of transfection, we will replace the culture medium and then cultivate in complete medium for 48 h.

Extraction of RNA and quantitative real-time polymerase chain reaction

In order to extract total RNA from cells, we used TRIZOL reagent (Absin, China) based on the manufacturer’s instructions. The reverse transcription was performed with PrimeScript RT Reagent Kit (Takara, China). Quantitative real-time PCR (qRT-PCR) was performed using Vazyme ChamQ Universal SYBR qPCR Master Mix. Relative expression levels were calculated using the 2^−ΔΔCt method, with β-actin serving as the internal control.

Western blot

We separated total protein extracts using 10% sodium dodecyl sulfate–polyacrylamide gel electrophoresis (SDS-PAGE). Electrophoretic gels were then converted into polyvinylidene difluoride (PVDF) membranes (0.45 μm, Millipore, USA), after that, add 5% nonfat dried milk and block for 2 h. Membranes were incubated with primary antibodies overnight at 4 °C, including mouse anti-glyceraldehyde-3-phosphate dehydrogenase (GAPDH) (1:5,000, Proteintech), rabbit anti-E-cadherin (1:2,500, Proteintech), rabbit anti-N-cadherin (1:1,000, ABconal), rabbit anti-Vimentin (1:1,000, ABconal), rabbit anti-ERK1/2 (1:2,000, Proteintech), rabbit anti-p38 (1:1,000, Proteintech), mouse anti-JNK (1:2,000, Proteintech), rabbit anti-p-ERK1/2 (1:1,000, Proteintech), rabbit anti-p-p38 (1:1,000, Proteintech), and rabbit anti-p-JNK1/2/3 (1:1,000, Proteintech). On the following day, we incubated the membranes at room temperature with a secondary antibody conjugated with horseradish peroxidase (HRP). In the next step, the membrane was washed 3 times for every 10 min. Images were acquired using the UVP/ChemStudio Imaging System (Analytik Jena AG, Germany).

Cell Counting Kit-8

The CCK-8 assay utilized 96-well plates containing 2,000 cells in 100 μl of medium. Adding 10 μl of CCK-8 (APExBIO) to the medium, it was incubated at 37 °C for 2 h, an then at 450 nm, the wells’ absorbances were measured.

5-Ethynyl-2′-deoxyuridine

Proliferation of cells can be measured with EdU Cell Proliferation Kit with Alexa Fluor 555 (CX003, CellorLab). We plated 50% of cells into 24-well plates on the first day and then conducted EdU experiments on the following day. EdU proliferating cells and DNA-stained cells were examined under a confocal microscope. ImageJ can be used to count the total number and EdU-positive cells, and then calculate the percentage of EdU-positive cells.

Transwell assay

Migration of cells was assessed in Transwell chambers (Corning, USA). On the top of the chamber, cells were plated in serum-free medium and on the bottom, where exist 600 μl medium with 20% serum. After 48 h, the sample was fixed for 1 h with 4% paraformaldehyde and then stained with 0.1% crystal violet for 30 min. The migrating cells are finally counted using ImageJ.

HLEC tube formation assay

HLEC cells are derived from a commonly used cell line routinely preserved in the laboratory, which are cultured adherently in a dedicated complete endothelial cell culture medium at 37 °C with 5% CO₂. The experimental procedures for HLEC are as follows: The Matrigel matrix is pre-thawed overnight at 4 °C, spread into a 96-well plate at 50 μl per well, and then placed in a 37 °C incubator to polymerize for 30 min for later use. Next, the treated cell supernatant is collected and centrifuged to obtain the conditioned medium, which is temporarily stored at 4 °C. Afterward, HLECs are resuspended in the aforementioned conditioned medium, seeded onto the pre-coated Matrigel wells at a density of 1 × 10⁴ cells per well, and cultured at 37 °C with 5% CO₂ for 6 to 24 h. Finally, images are captured in multiple randomly selected fields under an inverted microscope at 40× magnification, and the number of junctions of intact lumens is counted using ImageJ software as the core quantitative indicator of the experiment.

Statistical analysis

Continuous variables among groups were compared using the Welch’s t test, limma test, or Wilcoxon test. Categorical variables were analyzed with Fisher’s exact test for unpaired data and McNemar’s test for paired data. The AUC was utilized to evaluate the predictive performance for LNM. The DeLong test assessed differences between the 2 ROC curves. Bonferroni correction was applied for multiple comparisons to adjust the significance levels. Propensity scores were derived using logistic regression. All tests were 2-tailed, with statistical significance defined as a P value < 0.05. Statistical analyses were conducted using R (version 4.2.3) and Python (version 3.9.2).

Single-cell RNA sequencing and analysis

Single-cell RNA-seq libraries were prepared using microwell-based barcoding followed by reverse transcription and PCR amplification according to the GEXSCOPE Single Cell RNA Library Kit protocol [54]. Libraries were sequenced on an Illumina NovaSeq 6000 platform (150-bp paired-end reads). Raw reads were processed with CeleScope to remove low-quality sequences, and Cutadapt (v1.17) was used for adapter and poly-A trimming [55]. Following unique molecular identifier‌ (UMI) and cell-barcode extraction, reads were aligned to the GRCh38 reference genome using STAR (v2.6.1a) [56]. Gene-level UMI counts were generated with featureCounts (v2.0.1) to produce the expression matrix [57]. Cell-type annotation was performed by evaluating the expression of canonical marker genes derived from differentially expressed genes (DEGs) in the SynEcoSys database. Cluster identities and marker expression were visualized using Seurat (v3.1.2) functions (DoHeatmap, DotPlot, VlnPlot).

Pathway enrichment analysis

Functional enrichment of DEGs was performed using the “clusterProfiler” R package (v3.16.1) for Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) analyses [58]. Significantly enriched pathways were defined by an adjusted P value < 0.05.

Gene set scoring with AUCell

The UCell R package (v1.1.0) was employed to compute gene set activity scores based on the Mann–Whitney U statistic, which ranks genes by expression level per cell. This method supports robust scoring across diverse and large-scale single-cell datasets. Cell differentiation trajectories were reconstructed using Monocle2 [59], with highly variable genes selected for pseudotemporal ordering. DDRTree was applied for feature selection and dimensionality reduction. Genes exhibiting bimodal “on–off” expression patterns were assessed for switch points along the trajectory.

Mendelian colocalization analysis

Bayesian colocalization was used to evaluate whether genetic variants associated with 2 traits share a causal mechanism. Analysis was restricted to genomic regions harboring eQTLs for relevant genes, with evidence for colocalization defined as a posterior probability PP.H4.abf > 0.8. Results were visualized using the stack_assoc_plotfunction from the gassocplot2 package.

Pan-cancer gene expression landscape

Normalized expression data [normalized transcripts per million (nTPM)] for protein-coding genes across 50 tissues were integrated from the Human Protein Atlas (HPA v23.0) and GTEx projects. Consensus transcript expression values were summarized as the maximum nTPM per gene across datasets. Results were visualized using lollipop plots.

Drug sensitivity analysis

Drug response data were obtained from the Genomics of Drug Sensitivity in Cancer (GDSC) database, which includes 367 compounds screened across 987 cell lines. The oncoPredict R package was used to model relationships between gene expression and half-maximal inhibitory concentration (IC₅₀) values [60]. Negative correlations indicated that higher gene expression sensitized cells to a drug, whereas positive correlations indicated resistance.

Differential drug sensitivity by NBPF4 expression

The pRRophetic package implemented a ridge regression model to predict clinical chemotherapy response based on baseline gene expression and in vitro drug sensitivity data from the Cancer Genome Project (CGP) [61]. Model outputs included estimated IC₅₀ values for genes of interest (e.g., NBPF4) in TCGA samples, enabling stratification of high versus low expression groups for drug sensitivity comparisons.

Boruta feature selection

Boruta is an all-relevant feature selection wrapper algorithm that, by default, employs a Random Forest classifier. The method performs a top-down search for relevant features by comparing the importance of original attributes with the importance of “shadow” attributes—randomly shuffled copies of the original features—and iteratively eliminating irrelevant ones. In each iteration, the algorithm compares the importance of each real attribute with the maximum importance observed among the shuffled shadow attributes. Attributes with importance significantly lower than that of the shadow attributes are continuously removed. Conversely, attributes whose importance is significantly higher are labeled as “Confirmed”. The shadow attributes are recreated in every iteration. The algorithm terminates when only “Confirmed” attributes remain or when the maximum number of runs (maxRuns) is reached. In the latter case, some attributes may remain without a definitive decision and are labeled as “Tentative”. In summary, the Boruta algorithm enables the identification of the most diagnostically relevant features from a larger set of candidate genes.

Correlation of average gene expression

The average expression level of each of 2 specified genes was computed across multiple independent single-cell RNA-seq datasets. Pairwise association between the genes was then assessed using Spearman’s rank correlation coefficient [62,63].

Gene set variation analysis

Gene set variation analysis (GSVA) is a nonparametric, unsupervised method that transforms a gene-level expression matrix into a pathway-level or gene set-level activity matrix. It calculates an enrichment score for each predefined gene set within each sample, without the need for prior differential expression analysis between sample groups. This transformation from gene-centric to pathway-centric data enhances the biological interpretability of the results.

The core analysis was performed using the gsva() function from the GSVA R package. The algorithm provides several method options for calculating enrichment scores. In this study, the z-score method was employed. This method, as described by Lee et al., calculates a combined z score to reflect the activity of a given biological pathway or functional state.

The kcdfparameter, which defines the kernel type for the nonparametric estimation of the cumulative distribution function, was set to “Gaussian”. This setting is appropriate for log₂-transformed or otherwise continuous, normalized expression data.

Application to cancer functional states

Gene sets representing 14 distinct functional states of tumor cells (e.g., invasion, metastasis, and proliferation) were compiled from the CancerSEA database. The GSVA zscore method was applied to the gene expression matrix using these 14 gene sets to calculate a corresponding enrichment score (combined z score) for each functional state in each sample. Subsequently, these raw enrichment scores were standardized across samples using the R scale () function to generate a final “gene set score” for each functional state. Finally, Pearson correlation was computed between the expression level of a gene of interest (e.g., NBPF4) and each of the standardized gene set scores to assess their associations [64–66].

Ethical Approval

Less

This study involving human participants was reviewed and approved by the Ethics Committee of our hospital (approval number: FSYYY-EC-SOP-008-02.0-A09, RYE2024122302, and LW2025016HA). Written informed consent was obtained from the Foshan cohort, while the Duke and TCIA cohorts were exempted from informed consent requirements due to their retrospective and de-identified nature.

References

Less

Chang

, Leung

JWT

, Moy

, Ha

, Moon

.Axillary nodal evaluation in breast cancer: State of the art. Radiology. 2020;295(3):500–515.

Ullah

, Karthik

, Alkodsi

, Kjallquist

, Stalhammar

, Lovrot

, Martinez

, Lagergren

, Hautaniemi

, Hartman

, et al.Evolutionary history of metastatic breast cancer reveals minimal seeding from axillary lymph nodes. J Clin Invest. 2018;128(4):1355–1370.

Jin

, Gao

.Prediction of axillary lymph node metastasis in breast cancer using an ultrasonic feature- and clinical data-based model. Am J Cancer Res. 2024;14(12):5987–5998.

Zhao

, Xu

, Yang

, Ma

, Li

, Wu

, He

, Wang

, Liang

.Specific diagnosis of lymph node micrometastasis in breast cancer by targeting activatable near-infrared fluorescence imaging. Biomaterials. 2022;282:Article 121388.

, Tan

, Xie

, Hu

, Ouyang

, Chen

, Gu

, Li

, Lu

, He

, et al.Development and validation of a preoperative magnetic resonance imaging radiomics-based signature to predict axillary lymph node metastasis and disease-free survival in patients with early-stage breast cancer. JAMA Netw Open. 2020;3(13):Article e2028086.

Gong

, Wu

, Zhang

, Liu

, Zhu

, Cai

, Li

.Computer-assisted diagnosis for axillary lymph node metastasis of early breast cancer based on transformer with dual-modal adaptive mid-term fusion using ultrasound elastography. Comput Med Imaging Graph. 2025;119:Article 102472.

Zhang

, Dihge

, Bendahl

, Arvidsson

, Dustler

, Ellbrant

, Gulis

, Hjartstrom

, Ohlsson

, Rejmer

, et al.Deep learning on routine full-breast mammograms enhances lymph node metastasis prediction in early breast cancer. NPJ Digit Med. 2025;8:425.

Guo

, Liu

, Sun

, Zhang

, Wang

, Li

, Shi

, Wu

, Cui

, Zhang

, et al.Deep learning radiomics of ultrasonography: Identifying the risk of axillary non-sentinel lymph node involvement in primary breast cancer. EBioMedicine. 2020;60:Article 103018.

Chen

, Kong

, Lin

, Chen

, Guo

, Chen

, Cheng

, Chen

, Shi

, Xu

, et al.Development and validation of convolutional neural network-based model to predict the risk of sentinel or non-sentinel lymph node metastasis in patients with breast cancer: A machine learning study. EClinicalMedicine. 2023;63:Article 102176.

10.

van Hemert

AKE

, van Loevezijn

, Baas

MPD

, Stokkel

MPM

, Groen

, van der Noort

, Loo

, Sonke

, Russell

, van Duijnhoven

, et al.Iodine seed-marking protocol for response-guided axillary treatment after systemic therapy for node-positive breast cancer. JAMA Oncol. 2025;11(2):1204–1212.

11.

Zhu

, Huang

, Li

, Zhang

, Lin

, Cheng

, Wu

, Ye

, Lin

, Wang

.Multifactor artificial intelligence model assists axillary lymph node surgery in breast cancer after neoadjuvant chemotherapy: Multicenter retrospective cohort study. Int J Surg. 2023;109(11):3383–3394.

12.

Wei T-R, Yan Y. Multimodal medical imaging AI for breast cancer diagnosis: A comprehensive review. Intell Oncol. 2026;2(1):100037.

13.

Bao

, Qian

, Lyng

, Wang

, Yu

, Wang

, Zhang

, Yang

, Brunner

, Wang

, et al.Coexisting genomic aberrations associated with lymph node metastasis in breast cancer. J Clin Invest. 2018;128(6):2310–2324.

14.

Luo

, Cao

, Hu

, Liang

, Zhang

, Lai

.Lymphatic metastasis-associated circRNA–miRNA–mRNA network for exploring the pathogenesis and therapeutic target of triple negative breast cancer based on whole-transcriptome sequencing analysis: An experimental verification study. J Transl Med. 2022;20(1):508.

15.

Chatterjee

, Pai

, Hardiman

, Avery-Kiejda

, Scott

, Spencer

, Pinder

, Grigoriadis

.Molecular patterns of cancer colonisation in lymph nodes of breast cancer patients. Breast Cancer Res. 2018;20(1):143.

16.

Zhai

, Liu

, Zhao

, Zheng

, Chu

, Hong

, Cai

, Chen

, Liu

, Wang

, et al.Multi-omics analysis reveals comprehensive aberrant protein and phosphorylation characteristics in breast cancer and paired metastatic lymph nodes. Protein Cell. 2026;12:Article pwag002.

17.

Xie

, Liu

, Deng

, Wang

, Ou

, An

, Situ

, Yang

, Peng

, He

, et al.Paracrine orchestration of tumor microenvironment remodeling induced by GLO1 potentiates lymph node metastasis in breast cancer. Adv Sci (Weinh). 2025;12(32):Article e00722.

18.

Venet

, Fimereli

, Rothe

, Boeckx

, Maetens

, Majjaj

, Rouas

, Capra

, Bonizzi

, Contaldo

, et al.Phylogenetic reconstruction of breast cancer reveals two routes of metastatic dissemination associated with distinct clinical outcome. EBioMedicine. 2020;56:Article 102793.

19.

Lai

, Chen

, Liu

, Zhu

, Huang

, Yi

, Cai

, Liao

.A radiogenomic multimodal and whole-transcriptome sequencing for preoperative prediction of axillary lymph node metastasis and drug therapeutic response in breast cancer: A retrospective, machine learning and international multicohort study. Int J Surg. 2024;110(4):2162–2177.

20.

Perez-Gonzalez

, Bevant

, Blanpain

.Cancer cell plasticity during tumor progression, metastasis and response to therapy. Nat Cancer. 2023;4(8):1063–1082.

21.

da Silva-Diz

, Lorenzo-Sanz

, Bernat-Peguera

, Lopez-Cerda

, Munoz

.Cancer cell plasticity: Impact on tumor progression and therapy response. Semin Cancer Biol. 2018;53:48–58.

22.

Han

, Liu

, Zhu

, Miao

, Zhang

, Wang

, Lu

, Butch

, Cai

, Nie

, et al.Machine-learning-enabled Raman spectroscopy refines Indocyanine green fluorescence boundaries for precise glioblastoma margin delineation. Research. 2026;9:1216.

23.

Sun

, Sun

, Lan

, Luan

, Zhou

, Deng

, Yuan

, Wu

.Tumor cell plasticity in cancer: Signaling pathways and pharmaceutical interventions. MedComm. 2025;6(12):Article e70541.

24.

Peyvandi

, Bulliard

, Yilmaz

, Kauzlaric

, Marcone

, Haerri

, Coquoz

, Huang

, Duffey

, Gafner

, et al.Tumor-educated Gr1⁺CD11b⁺ cells drive breast cancer metastasis via OSM/IL-6/JAK-induced cancer cell plasticity. J Clin Invest. 2024;134(6):Article e166847.

25.

Castro

, Walker

, Glocker

.Causality matters in medical imaging. Nat Commun. 2020;11(1):3673.

26.

Luo

, Wang

, Lin

, Ma

, Tan

, Chan

, Vardhanabhuti

, Chu

, Cheng

, Chen

.Deep learning in breast cancer imaging: A decade of progress and future directions. IEEE Rev Biomed Eng. 2025;18:130–151.

27.

Tran

, Kondrashova

, Bradley

, Williams

, Pearson

, Waddell

.Deep learning in cancer diagnosis, prognosis and treatment selection. Genome Med. 2021;13(1):152.

28.

, Leng

, Tang

, Wang

, U

, Zhang

, Han

, Zhang

, Xiang

, Yang

, et al.Deep learning-enabled integration of histology and transcriptomics for tissue spatial profile analysis. Research. 2025;8:Article 0568.

29.

Zhang

, Isil

, Yang

, Li

, Elia

, Atlan

, Wallace

, Pillar

, Ozcan

.Deep learning-enabled virtual multiplexed immunostaining of label-free tissue for vascular invasion assessment. BME Front. 2026;7:0226.

30.

Sucre

, Calle Sanchez

, Perez-Herrera

, Vivanco

, Garcia-Gonzalez

, Lopez-Linares

, Calvo

, Garin-Muga

.Multimodal fusion strategies for survival prediction in breast cancer: A comparative deep learning study. Comput Struct Biotechnol J. 2025;27:4505–4516.

31.

Jiang

, Zhang

, Wang

, Huang

, Chen

, Xi

, Ahmad

, Ren

, Sang

, Xie

, et al.Biology-guided deep learning predicts prognosis and cancer immunotherapy response. Nat Commun. 2023;14(1):5135.

32.

Liu

, Wang

, Huang

, Pei

, Wang

, Cui

, Yan

, Yao

, Wang

, Zhu

, et al.Deep learning-based prediction of axillary pathological complete response in patients with breast cancer using longitudinal multiregional ultrasound. EBioMedicine. 2025;119:Article 105896.

33.

, Su

, You

, Zhang

, Xiao

, Jiang

, Shao

.Radiomics in breast cancer: Current advances and future directions. Cell Rep Med. 2024;5(9):Article 101719.

34.

Peng

, Wang

, Zhu

, Jiang

, Xia

, Cui

, Hong

, Zeng

, Li

, et al.Three-dimensional multimodal imaging for predicting early recurrence of hepatocellular carcinoma after surgical resection. J Adv Res. 2026;81:865–875.

35.

, Mayer

, Li

.Integrated imaging and molecular analysis to decipher tumor microenvironment in the era of immunotherapy. Semin Cancer Biol. 2022;84:310–328.

36.

Cai

, Zhu

, Huang

, Cao

, Cai

, Chen

, Feng

, Zhang

, Mao

, Lai

.Lymph node metastasis-associated spatiotemporal mapping of the TFF3-linked niche in breast cancer: Integrating radiogenomic signatures with immune-ecosystem remodeling. Research. 2026;9:1016.

37.

Fan

, Cui

, You

, Liu

, Gu

, Peng

, Bai

, Gao

, Li

.Radiogenomic signatures of Oncotype DX recurrence score enable prediction of survival in estrogen receptor-positive breast cancer: A multicohort study. Radiology. 2022;302(3):516–524.

38.

Shang

, Chen

, Gao

, Wan

, Yang

, Lei

, Chen

, Quan

, Bai

.Integrating clinical-pathological-MRI features to construct a prediction model for pathological complete remission of axillary lymph nodes after neoadjuvant therapy: A retrospective study. Int J Surg. 2025;111(11):7633–7646.

39.

Chou

, Kaller

, Rokavec

, Liu

, Hermeking

.AP4 induces JNK1 and a miR-22-3p/FOSL1 feed-forward loop to activate AP-1 and promote colorectal cancer metastasis. Cancer Commun. 2024;44(3):433–437.

40.

Ganesan

, Xu

, Wu

, Sui

, Du

, Zhang

, Gao

, Chen

, Tang

.Ononin inhibits tumor bone metastasis and osteoclastogenesis by targeting mitogen-activated protein kinase pathway in breast cancer. Research. 2024;7:0553.

41.

Huang

, Chen

, Zhao

, Wu

, Li

, Luo

, Su

, Zhang

, Liu

, Tang

.Dual regulation of sprouty 4 palmitoylation by ZDHHC7 and palmitoyl-protein thioesterase 1: A potential therapeutic strategy for cisplatin-resistant osteosarcoma. Research. 2025;8:0708.

42.

Chen

, Zhang

, Zhu

, Pang

, Wu

, Wang

, Zhan

.Focal adhesion kinase/Src family kinase axis-mediated tyrosine phosphorylation of metabolic enzymes facilitates tumor metastasis. Signal Transduct Target Ther. 2025;10(1):280.

43.

Howell

, Davies

, Clarke

, Appios

, Mesquita

, Jayal

, Ringham-Terry

, Boned Del Rio

, Fisher

, Bennett

.Localized immune surveillance of primary melanoma in the skin deciphered through executable modeling. Sci Adv. 2023;9(15):eadd1992.

44.

, Chen

, Liu

, Shi

, Liu

, Wu

, Fang

, Bao

, Jiang

, et al.Long noncoding RNA MRCCAT1 promotes metastasis of clear cell renal cell carcinoma via inhibiting NPR3 and activating p38-MAPK signaling. Mol Cancer. 2017;16(1):111.

45.

, Yang

, Chen

, Zhang

, Xia

, Wang

, Xu

.N6-methyladenosine modification of CENPF mRNA facilitates gastric cancer metastasis via regulating FAK nuclear export. Cancer Commun (Lond). 2023;43(6):685–705.

46.

Chumsri

, Larson

, Liu

, Tenner

, Adams

, Weidner

, Arnold

, Haley

, Advani

, Sideras

, et al.Pembrolizumab in combination with binimetinib in patients with unresectable locally advanced or metastatic triple-negative breast cancer. Clin Cancer Res. 2025;31(10):1885–1893.

47.

Qiang

, Li

, Sun

, Zheng

, Peng

, Yang

, Meng

, Lang

, Mei

, Xie

, et al.Along with its favorable prognostic role, CLCA2 inhibits growth and metastasis of nasopharyngeal carcinoma cells via inhibition of FAK/ERK signaling. J Exp Clin Cancer Res. 2018;37(1):34.

48.

Wang

, Luo

, Zhang

, Cao

, Huang

, Jiang

, Liu

, Su

, Qiu

.Hypoxic tumor-derived exosomal miR-301a mediates M2 macrophage polarization via PTEN/PI3Kgamma to promote pancreatic cancer metastasis. Cancer Res. 2018;78(16):4586–4598.

49.

Zhang

, Yuan

, Wen

, Shao

, Liao

, Jia

, Zhao

, Liao

, Xu

, Chen

, et al.LNMAC promotes cervical squamous cell carcinoma lymphatic metastasis via epigenetic regulation of FGF2-induced lymphangiogenesis. Adv Sci. 2024;11(38):Article e2404645.

50.

Dong

, Hu

, Li

, Pei

, Jin

, Zhang

, Chen

, Min

, Yin

.SPP1⁺ TAM regulates the metastatic colonization of CXCR4⁺ metastasis-associated tumor cells by remodeling the lymph node microenvironment. Adv Sci. 2024;11(44):Article e2400524.

51.

Shu

, Bao

, Jiang

, Wang

, Zhang

, Liang

, Li

, Ye

, Xiang

, et al.MicroRNA-29c-5p suppresses gallbladder carcinoma progression by directly targeting CPEB4 and inhibiting the MAPK pathway. Cell Death Differ. 2017;24(3):445–457.

52.

Clark

, Vendt

, Smith

, Freymann

, Kirby

, Koppel

, Moore

, Phillips

, Maffitt

, Pringle

, et al.The Cancer Imaging Archive (TCIA): Maintaining and operating a public information repository. J Digit Imaging. 2013;26(6):1045–1057.

53.

Mathew

, Agha

, Albrecht

, Goel

, Mukherjee

, Pai

, D’Cruz

, Nixon

, Roberto

, Enam

, et al.STROCSS 2021: Strengthening the reporting of cohort, cross-sectional and case-control studies in surgery. Int J Surg. 2021;96:Article 106165.

54.

Dura

, Choi

, Zhang

, Damsky

, Thakral

, Bosenberg

, Craft

, Fan

.scFTD-seq: Freeze-thaw lysis based, portable approach toward highly distributed single-cell 3’ mRNA profiling. Nucleic Acids Res. 2019;47(3):Article e16.

55.

Kechin

, Boyarskikh

, Kel

, Filipenko

.cutPrimers: A new tool for accurate cutting of primers from reads of targeted next generation sequencing. J Comput Biol. 2017;24(11):1138–1143.

56.

Dobin

, Davis

, Schlesinger

, Drenkow

, Zaleski

, Jha

, Batut

, Chaisson

, Gingeras

.STAR: Ultrafast universal RNA-seq aligner. Bioinformatics. 2013;29(1):15–21.

57.

Liao

, Smyth

, Shi

.featureCounts: An efficient general purpose program for assigning sequence reads to genomic features. Bioinformatics. 2014;30(7):923–930.

58.

, Wang

, Han

, He

.clusterProfiler: An R package for comparing biological themes among gene clusters. OMICS. 2012;16(5):284–287.

59.

Qiu

, Hill

, Packer

, Lin

, Ma

, Trapnell

.Single-cell mRNA quantification and differential analysis with Census. Nat Methods. 2017;14(3):309–315.

60.

Maeser

, Gruener

, Huang

.oncoPredict: An R package for predicting in vivo or cancer patient drug response and biomarkers from cell line screening data. Brief Bioinform. 2021;22(6):Article bbab260.

61.

Geeleher

, Cox

, Huang

.Clinical drug response can be predicted using baseline gene expression levels and in vitro drug sensitivity in cell lines. Genome Biol. 2014;15(3):R47.

62.

Sun

, Wang

, Han

, Dong

, Ge

, Zheng

, Shi

, Wang

, Li

, Ren

, et al.TISCH: A comprehensive web resource enabling interactive single-cell transcriptome visualization of tumor microenvironment. Nucleic Acids Res. 2021;49:D1420–D1430.

63.

Han

, Wang

, Dong

, Sun

, Liu

, Yue

, Wang

, Li

, Wang

.TISCH2: Expanded datasets and new tools for single-cell transcriptome analyses of the tumor microenvironment. Nucleic Acids Res. 2023;51:D1425–D1431.

64.

Yuan

, Yan

, Zhang

, Liu

, Deng

, Liao

, Xu

, Luo

, Yan

, Long

, et al.CancerSEA: A cancer single-cell state atlas. Nucleic Acids Res. 2019;47:D900–D908.

65.

Hanzelmann

, Castelo

, Guinney

.GSVA: Gene set variation analysis for microarray and RNA-seq data. BMC Bioinformatics. 2013;14:7.

66.

Lee

, Chuang

, Kim

, Ideker

, Lee

.Inferring pathway activity toward precise disease classification. PLOS Comput Biol. 2008;4(11):Article e1000217.

Appendix

Less

Year 2026 volume 9 Issue 8

PDF

153

Cite this Article

BibTeX

Article Info

doi: 10.34133/research.1402

Receive Date：2026-04-15
Online Date：2026-10-16
Published：2026-08-20

Article Data

Affiliations

History

Received：2026-04-15
Revised：2026-07-29
Accepted：2026-07-30

Affiliations