Ultrasound Genomics Reveals a Signature for Predicting Breast Cancer Prognosis and Therapy Response

Abstract

Background:

Breast cancer (BC) in women is the most common malignancy worldwide, but there is still a lack of validated tools to accurately assess patient prognosis and response to available chemotherapy treatment regimens.

Method:

We collected ultrasound images and transcriptome data of BC from our breast center and public database. Key ultrasound features were then identified by using the support vector machine (SVM) algorithm and correlated with prognostic genes. Long-term survival-related genes were identified through differential expression analysis, and a prognostic evaluation model was established by using Cox regression. In addition, VPS28 from the model was identified as a promising biomarker for BC.

Results:

Using univariate logistic regression and SVM algorithms, we identified 12 ultrasound features significantly associated with chemotherapy response. Subsequent correlation and differential expression analyses linked 401 genes to these features, from which five key signature genes were derived using Lasso and multivariate Cox regression models. This signature not only facilitates the stratification of patients into risk-specific treatment pathways but also predicts their chemotherapy response, thus supporting personalized medicine in clinical settings. Notably, VPS28, in the signature, emerged as a significant biomarker, strongly associated with poor prognosis, greater tumor invasiveness, and differing expression across demographic groups.

Conclusion:

In this study, we use ultrasound genomics to reveal a signature that can provide an effective tool for prognostic assessment and predicting chemotherapy response in patients with BC.

Introduction

As the most common malignant tumor in women around the world, research on breast cancer (BC) treatment and diagnostic methods has always been an important topic in the medical field.¹ Traditional BC treatments mainly include surgery, radiotherapy, chemotherapy, and endocrine therapy.^2
–4 Although these clinical decisions can significantly improve the survival benefit of patients, there is a lack of tools to effectively assess the prognosis of patients and their response to existing chemotherapy regimens.

Ultrasonomics is an emerging technology in recent years, which analyses quantitative data in ultrasound images and extracts a large number of ultrasound features through deep learning techniques. These ultrasound features include morphological features, texture features and waveform features, etc., which can be used to accurately describe the biological characteristics and behaviors of tumors. In the evaluation of neoadjuvant chemotherapy, ultrasonomics is able to provide timely and accurate feedback on the treatment effect through the comparative analysis of pre and postchemotherapy images and detailed documentation of changes in the tumor and its microenvironment.⁵ Some researchers have demonstrated high accuracy when predicting HER2+ and HER2− BC subtypes based on ultrasonomics.⁶ This result could identify more aggressive BC subtypes in the early stages of the diagnostic process and provide a justification for early clinical intervention. Ultrasound genomics, which is derived from transcriptome sequencing data, can take advantage of the image and genetic data characteristics to more accurately reflect oncological characteristics. However, there are fewer studies in this area, especially based on ultrasound genomics and prediction of prognosis and chemotherapy response in BC patients.

In our study, we performed feature extraction and analysis of ultrasound images combined with correlation analysis of transcriptome data to reveal the set of genes associated with BC ultrasound signature. Subsequently, a signature model was constructed based on this gene set, which provided significant value in the prognostic assessment and efficacy prediction of BC, and provided more accurate and effective clinical decisions for BC patients.

Materials and Method

Data collection and preparation

We collected ultrasound image data of chemotherapy-sensitive and nonsensitive patients from the university clinic breast center and obtained the RNA sequence data from the public database. All image data were normalized before further processing. For the RNA sequence data, we only keep the mRNA data, and the mRNA expression of less than 50% of samples was also removed. The clinical data between the breast center and public sources was matched using the propensity score matching method. The workflow is shown in Figure 1.

FIG. 1.

Workflow of this study.

Ultrasound feature selection and building of the model

We utilized 3D slice to identify and distinguish between sensitive and nonsensitive tumor areas. Subsequently, we extracted ultrasound features from these images. We applied uni-logistic and support vector machine (SVM) algorithms to identify important features and used multilogistic regression to construct a signature related to ultrasound features.

Features-related genes and correlation with ultrasound features related signature

We used the K–M method to filter the survival-related genes, and we identified significant results with a p-value < 0.05. Based on the median expression value of ultrasound features-related signature, patients were divided into low and high groups. We conducted a correlation analysis between prognostic genes and the high group, enrolling those with a p-value < 0.05 and an r-value > 0.1 into the next procedure.

Differentially expressed gene analysis between short- and long-term overall survival

After the analysis above, we identified the candidate ultrasound-related genes. Using these genes and tumor barcodes, we created a new expression matrix. Next, we planned to divide the patients into two groups based on whether their survival time is greater than 5 years or not, and then conducted differentially expressed gene (DEG) analysis between the two groups. Genes with absolute value of log-fold change (logFC) > 1 and p < 0.05 were defined as DEGs.

Prognosis gene model build

We used the Lasso regression model to select important genes, and then input these genes into the multivariate Cox regression model to build the prognostic gene-related signature. The gene expression value and corresponding regression coefficients are two important parameters for model building. The sum of the products of these two parameters for each gene is the score of the model after evaluating the patient. The log-rank test was used to evaluate the survival difference between different risk groups.

Predict the chemotherapy response

We downloaded all of the Food and Drug Association (FDA)-approved chemotherapy drugs from a public database and then used the R package to conduct the drug sensitivity analysis. The-value < 0.05 was identified as sensitive. The top sensitive drug was shown using a Violin plot.

Identify model novel gene and validation clinic value

We reviewed the lecture and discovered a new biomarker for BC. We validated its expression in both normal and tumor tissues, using both unpaired and paired groups. In addition, we examined the survival differences between low- and high-expression groups of the gene. Furthermore, we analyzed the relationship between age, race, menopause status, and American Joint Committee on Cancer (AJCC) stage of this gene.

Results

Ultrasound features select and model build

After marking the target area, we input the mask file and the original ultrasound image into the software to extract the image features. Then, we used univariate logistic regression and SVM algorithms to identify the ultrasound features related to chemotherapy. The results showed that 12 features are associated with chemotherapy response (Fig. 2A). Next, we conducted a multivariate logistic regression model to build the ultrasound features-related signature.

FIG. 2.

SVM discovers the 12 important features of ultrasound images (A). Ultrasound feature-related model associated with prognostic genes (B). Volcano plot for short and long-term survival (C). Lasso regression model identified hub survival-related genes (D and E). The five-gene signature mode was built (F). SVM, support vector machine.

Identified ultrasound genomics-related gene

We divided the model into low- and high-risk groups and conducted correlation analysis with 1965 prognostic genes obtained from batch survival analysis. The results indicate that a total of 401 genes are associated with ultrasound features, with the top 20 genes listed in Figure 2B. In addition, we analyzed the DEGs between short- and long-term groups using the above gene sets and identified 37 DEGs associated with survival duration (Fig. 2C).

Ultrasound genomics-related signatures could guide patient stratification

We used the R package “cart” to randomly split patients into training and test datasets. Then, we performed Lasso regression and multivariate Cox regression analysis to identify five candidate signature genes and their related coefficients (Fig. 2D–F). The dot plot and heat map of the model in the training dataset are depicted in Figure 3A. Survival analysis indicates that high-risk patients have a poor outcome, whereas low-risk patients experience the opposite prognostic events (Fig. 3B). The validation results in the test dataset are consistent with those in the training set (Fig. 3C and D).

FIG. 3.

The dot plot and heatmap of five signature genes in the training datasets (A). The survival curve of low and high-risk signature scores in the training datasets (B). The dot plot and heatmap of five signature genes in the test datasets (C); and the survival curve of low and high-risk signature scores in the test datasets (D).

Ultrasound genomics signature could predict the chemotherapy response

Chemotherapy is a crucial component of clinical decision-making, and predicting the response to therapy is beneficial for personalized treatment. Our analysis reveals that a total of 38 FDA-approved drugs are sensitive to both low and high-risk patient groups, with the top five drugs shown using violin plots (Fig. 4A–E). We found that low-risk patients are more sensitive to traditional chemotherapy drugs such as cisplatin and gefitinib and the results are also visualized as a matrix diagram (Fig. 4F).

FIG. 4.

The top five chemotherapy drugs for low risk of signature are temsirolimus (A), cisplatin (B), lenalidomide (C), vorinostat (D), and gefitinib (E) and results are also visualized as a matrix diagram (F).

VPS28 shows high expression in BC and is associated with prognosis

In our study, VPS28 is identified as a novel biomarker for BC. The body heatmap indicates a high expression of VPS28 in BC (Fig. 5A), and this conclusion is supported by both unpaired and paired group analyses (Fig. 5B and C). Survival analysis demonstrates that the high expression of this gene leads to poor overall survival (Fig. 5D). However, no significant differences were observed in terms of disease-specific survival (DSS) and progression-free interval (PFI) between high and low-expression patients (Fig. 5E and F).

FIG. 5.

VPS28 is high expression in breast cancer from body heatmap, unpaired samples and paired samples (A–C). High expression of VPS28 is associated with poor overall survival (D). No significant between high- and low-expression groups in disease-specific survival (E) and progress-free interval (F).

VPS28 is associated with higher tumor invasiveness

To evaluate the clinical value of this gene, we analyzed the correlation between this gene and common clinical factors. We found that patients of old age exhibit high expression of this gene (Fig. 6A). Compared with White and other races, the Asian race shows a low expression of VPS28 (Fig. 6B). In addition, postmenopausal patients also exhibit a high expression of VPS28 (Fig. 6C). Moreover, we found that VPS28 is associated with higher tumor invasiveness, meaning that these patients have bigger tumor sizes (Fig. 6E), more lymph node infiltration (Fig. 6F), and advanced tumor stage (Fig. 6G).

FIG. 6.

Patients of old age have high expression of VPS28 (A). Compared with the White and other races, the Asian race shows a low expression of VPS28 (B). Postmenopausal patients also exhibit a high expression of VPS28 (C). VPS28 is associated with higher tumor invasiveness, meaning that these patients have bigger tumor sizes (D), more lymph node infiltration (E), and advanced tumor stage (F).

Discussion

BC continues to be a leading cause of cancer-related death among women, largely due to the disease’s significant heterogeneity, propensity for metastasis, and resistance to therapy.⁷ The treatment of BC involves a variety of approaches, such as surgery, chemotherapy, radiotherapy, endocrine therapy, targeted therapy, and immunotherapy, all of which require the coordinated efforts of multiple subspecialties.⁸ Despite the significant improvements in overall patient outcomes due to existing treatments, further efforts are still required to tailor appropriate therapeutic strategies for this heterogeneous disease.⁹ Over the past decade, the large-scale integration of genomic and transcriptomic data has revealed distinct cancer subtypes, providing a foundation for personalized precision therapy.^10,11 In our study, we introduce, for the first time, an ultrasound genomics-based model that identifies a signature with the ability to accurately predict BC patient outcomes and their responsiveness to these different therapeutic strategies, including commonly used chemotherapy drugs.

Ultrasound genomics, which is favored with its noninvasive feature, combines the analysis of ultrasound imaging features with genomic data, integrating with radiomics and advanced machine learning algorithms, to enhance the detection, characterization, and prognosis of cancers,¹² has been applied in cancer studies.^13,14 However, research on ultrasound genomics in BC is still in its early stages. Some scholars have utilized ultrasound to guide BC management, and they found that a cut-off of ≤5 ultrasound-detected abnormal axillary nodes can accurately identify BC patients with a limited nodal burden, allowing for the potential omission of chemotherapy in favor of upfront surgery followed by gene assay testing. This approach may help guide treatment decisions, particularly between immediate surgery and the need for neoadjuvant chemotherapy.¹⁵ Another study demonstrated an association between vascular ultrasound features and DNA sequencing in BC, revealing that specific ultrasound characteristics are significantly correlated with certain SNPs related to angiogenesis and prognosis. These findings suggest that vascular ultrasound could potentially reflect underlying genomic changes, aiding in the prediction of disease outcomes.¹⁶ While previous studies have provided significant insights into the prognosis and treatment guidance of BC, no research has yet explored the use of combined ultrasound genomics to assess treatment response in BC. As shown in the results, our ultrasound genomics model effectively predicted the prognosis and treatment response of patients with BC in TCGA cohort, highlighting that ultrasound genomics, and by extension radiogenomics, holds critical value not only in diagnosis but also in predicting treatment response, underscoring its broader applicability in personalized cancer care.

In our ultrasound-genomics model, a six-gene signature comprising KLF11, FREM1, VPS28, PHKB, and CCDC9B was constructed, with VPS28 identified as the hub gene. Kruppel-like factor 11 (KLF11), a mediator of TGFβ signaling, plays a role in regulating growth inhibition in untransformed epithelial cells. In pancreatic cancer, KRAS-driven ERK pathway activation phosphorylates KLF11, disrupting its interaction with SIN3a and leading to SMAD7 transcriptional activation, which promotes tumor progression.¹⁷ However, the specific role of KLF11 in BC remains unexplored. FRAS1-related extracellular matrix 1 (FREM1) is a gene that encodes a protein involved in the formation of the extracellular matrix and is part of the FRAS1-related extracellular matrix protein family.^18,19 Although its role in tumors remains unclear, it may contribute to tumor progression via the extracellular matrix. Phosphorylase kinase regulatory subunit beta (PHKB) has been reported to play a role in chronic myelogenous leukemia, where it is positively regulated by the enhancer RNA Hmrhl, contributing to leukemia progression.²⁰ However, its involvement in BC remains unexplored. In a recent study, coiled-coil domain containing 9B (CCDC9B) was identified as part of a six-gene risk signature that effectively predicts BC prognosis, especially in triple-negative BC (TNBC) patients, and could be valuable in guiding personalized treatment strategies.²¹ This finding is consistent with our results. Although vacuolar protein sorting 28 (VPS28), as the core gene of this signature, is well-known for its role in Candida albicans virulence and HIV-1 budding,²² recent research has begun to focus on its function in BC. For instance, Shi et al.²³ found that the expression of VPS28 is associated with poor prognosis and shorter survival of BC patients, and miR-491-5p can suppress BC progression by reducing VPS28 expression, highlighting its potential as a therapeutic target. Another study revealed the coexpression of EXCSC4 with VPS28 and Myc, showing elevated serum levels that are risk factor of OS among BC patients.²⁴ The differential expression of VPS28 across various subgroups in our study also suggest VPS28 as a risk factor in BC patients. Interestingly, our study revealed that high VPS28 expression is associated with poorer OS but does not impact DSS or PFI. Higher expression of VPS28 was found primarily impact late-stage disease might explain this observation.

Despite advancements in BC therapy, how to identify potential patients who could benefit from these new drugs remains critically challenging. In addition to its diagnostic and prognostic roles, our ultrasound genomics signature can also predict sensitivity to common therapy drugs, such as cisplatin and gefitinib, which are commonly applied in BC therapies. Notably, lenalidomide is an immunomodulatory drug that modulates the immune system and exhibits antiangiogenic properties, primarily used in hematological malignancies. The application of lenalidomide in BC remains in the early exploratory stage.²⁵ A previous study has shown that lenalidomide enhances the effectiveness of cisplatin in the TNBC cell line by significantly reducing cell viability, inducing apoptosis, and modulating key molecular pathways, suggesting potential therapeutic synergy in TNBC.²⁶ Vorinostat, a histone deacetylase inhibitor, has shown promise in preclinical and clinical trials for BC treatment, either alone or in combination with other anticancer agents.^27,28 All these results suggest our ultrasound genomics signature offers a promising tool for predicting drug sensitivity and could potentially enhance the BC therapy response.

Conclusion

In our study, our ultrasound genomics-revealed signature can provide an effective tool for prognostic assessment and chemotherapy response prediction in patients with BC. Application of this model not only improves the personalization of treatment but also provides more accurate and effective clinical decision-making for patients with BC.

Footnotes

Authors’ Contributions

Conceptualization: Q.L., W.W., and R.L.D.W. Methodology: Q.L., X.C., and G.W. Formal analysis: B.C., Z.G., and L.A.T.R. Data preparation: G.W. and R.Z. Writing—original draft: Q.L. and B.C. Supervision: R.L.D.W., X.C., and W.W. Writing—revision: R.L.D.W., L.A.T.R., B.C., and Z.G. Project administration and funding acquisition: W.W. and X.C. All authors have reviewed and approved the final version of the article.

Data Availability

All data can be obtained from the corresponding author.

Ethics Approval Statement

Ethics approval was obtained from the Second Affiliated Hospital of Guilin Medical University (HXXM-2022001).

Patient Consent Statement

Patient consent permission form.

Permission to Reproduce Material from Other Sources

All code with analysis can be obtained from the corresponding author.

Disclosure Statement

All authors declare no conflicts of interest.

Funding Information

This study supported by Wu Jieping Medical Foundation (320.6750.2022-19-40).

References

Chen

, Xu

, Shi

, et al. Trends of female and male breast cancer incidence at the global, regional, and national levels, 1990–2017. Breast Cancer Res Treat, 2020; 180(2):481–490; doi: 10.1007/s10549-020-05561-1

Shi

, Li

, Wang

, et al. Innovative modified T-shape oncoplastic technique for early-stage breast cancer: Multicenter retrospective study. Front Oncol, 2024; 14:1367477; doi: 10.3389/fonc.2024.1367477

Shi

, Luo

, Zhao

, et al. Evaluation of the benefit of post-mastectomy radiotherapy in patients with early-stage breast cancer: A propensity score matching study. Oncol Lett, 2019; 17(6):4851–4858; doi: 10.3892/ol.2019.10197

Krauss

, Stickeler

. Endocrine therapy in early breast cancer. Breast Care (Basel), 2020; 15(4):337–346; doi: 10.1159/000509362

Baumgartner

, Tausch

, Hosch

, et al. Ultrasound-based prediction of pathologic response to neoadjuvant chemotherapy in breast cancer patients. Breast, 2018; 39:19–23; doi: 10.1016/j.breast.2018.02.028

Ferre

, Elst

, Senthilnathan

, et al. Machine learning analysis of breast ultrasound to classify triple negative and HER2+ breast cancer subtypes. Breast Dis, 2023; 42(1):59–66; doi: 10.3233/BD-220018

Siegel

, Miller

, Wagle

, et al. Cancer statistics, 2023. CA Cancer J Clin, 2023; 73(1):17–48; doi: 10.3322/caac.21763

Wang

, Wu

S-G

. Breast cancer: An overview of current therapeutic strategies, challenge, and perspectives. Breast Cancer (Dove Med Press), 2023; 15:721–730; doi: 10.2147/BCTT.S432526

Hashim

, Boffetta

, La Vecchia

, et al. The global decrease in cancer mortality: Trends and disparities. Ann Oncol, 2016; 27(5):926–933; doi: 10.1093/annonc/mdw027

10.

Yin

, Liu

, Yu

, et al. Genomic and transcriptomic analysis of breast cancer identifies novel signatures associated with response to neoadjuvant chemotherapy. Genome Med, 2024; 16(1):11; doi: 10.1186/s13073-024-01286-8

11.

Pradat

, Viot

, Yurchenko

, et al. Integrative pan-cancer genomic and transcriptomic analyses of refractory metastatic cancer. Cancer Discov, 2023; 13(5):1116–1143; doi: 10.1158/2159-8290

12.

Sushanki

, Bhandari

, Singh

. A review on computational methods for breast cancer detection in ultrasound images using multi-image modalities. Arch Computat Methods Eng, 2024; 31(3):1277–1296; doi: 10.1007/s11831-023-10015-0

13.

Mendes Serrão

, Klug

, Moloney

, et al. Current status of cancer genomics and imaging phenotypes: What radiologists need to know. Radiol Imaging Cancer, 2023; 5(6):e220153; doi: 10.1148/rycan.220153

14.

Demetriou

, Lockhat

, Brzozowski

, et al. The convergence of radiology and genomics: Advancing breast cancer diagnosis with radiogenomics. Cancers (Basel), 2024; 16(5):1076; doi: 10.3390/cancers16051076

15.

Lim

, Allen

, Lee

, et al. Use of axillary ultrasound to guide breast cancer management in the genomic assay era. Breast Dis, 2024; 43(1):19–23; doi: 10.3233/BD-230032

16.

Han

M-R

, Park

, Seo

, et al. Association between vascular ultrasound features and DNA sequencing in breast cancer: A preliminary study. Discov Oncol, 2023; 14(1):52; doi: 10.1007/s12672-023-00657-8

17.

Buck

, Buchholz

, Wagner

, et al. The tumor suppressor KLF11 mediates a novel mechanism in transforming growth factor β–Induced growth inhibition that is inactivated in pancreatic cancer. Mol Cancer Res, 2006; 4(11):861–872; doi: 10.1158/1541-7786.MCR-06-0081

18.

Hahn

, Sundar

. Current perspective on the role of the circadian clock and extracellular matrix in chronic lung diseases. Int J Environ Res Public Health, 2023; 20(3):2455; doi: 10.3390/ijerph20032455

19.

Rike

, Stern

. Proteins and transcriptional dysregulation of the brain extracellular matrix in Parkinson’s disease: A systematic review. Int J Mol Sci, 2023; 24(8):7435; doi: 10.3390/ijms24087435

20.

Fatima

, Choudhury

, Divya

, et al. A novel enhancer RNA, Hmrhl, positively regulates its host gene, phkb, in chronic myelogenous leukemia. Noncoding RNA Res, 2019; 4(3):96–108; doi: 10.1016/j.ncrna.2019.08.001

21.

Xing

, Lin

, Hong

, et al. Construction of a prognostic 6-gene signature for breast cancer based on multi-omics and single-cell data. Front Oncol, 2023; 13:1186858; doi: 10.3389/fonc.2023.1186858

22.

Dai

, Feng

, Liao

, et al. ESCRT machinery and virus infection. Antiviral Res, 2023; 221:105786; doi: 10.1016/j.antiviral.2023.105786

23.

Shi

, Hu

, Xing

, et al. Deciphering the oncogenic role of VPS28 modulated by miR-491-5p in breast cancer cells using in silico and functional analysis. Front Mol Biosci, 2021; 8:634183; doi: 10.3389/fmolb.2021.634183

24.

Bera

, Karaian

, Subramanian

, et al. EXOSC4, a novel gene at chromosome 8q24 loci is linked with breast cancer progression and is a prognostic marker for breast cancer survival. Cancer Research, 2020; 80(16_Supplement):4322–4322; doi: 10.1158/1538-7445

25.

Kotla

, Goel

, Nischal

, et al. Mechanism of action of lenalidomide in hematological malignancies. J Hematol Oncol, 2009; 2:36–10; doi: 10.1186/1756-8722-2-36

26.

Yin

L-L

, Wen

X-M

, Lai

Q-H

, et al. Lenalidomide improvement of cisplatin antitumor efficacy on triple-negative breast cancer cells in vitro. Oncol Lett, 2018; 15(5):6469–6474; doi: 10.3892/ol.2018.8120

27.

Kearney

, McGuinness

, Kalinsky

. Clinical trial data and emerging immunotherapeutic strategies: Hormone receptor-positive, HER2− negative breast cancer. Breast Cancer Res Treat, 2021; 189(1):1–13; doi: 10.1007/s10549-021-06291-8

28.

Nouri Emamzadeh

, Word

, Cotton

, et al. Modulation of estrogen α and progesterone receptors in triple negative breast cancer cell lines: The effects of vorinostat and indole-3-carbinol in vitro. Anticancer Res, 2020; 40(7):3669–3683; doi: 10.2187/3/anticanres.14356