Radiogenomics Analysis Reveals Tumor Heterogeneity of Triple-Negative Breast Cancer

Triple-negative breast cancer (TNBC) is a subset of breast cancer with adverse prognosis and signicant tumor heterogeneity. Here, we used MRI images from a breast cancer cohort consisting of 860 patients to construct a radiomic signature that could identify TNBC with an AUC of 0.92 (95% CI: 0.887 to 0.953) and validated in another cohort. Moreover, we developed radiomic signatures to distinguish TNBC subtypes with moderate ecacy. Furthermore, we identied peritumoral dependence nonuniformity of the gray level dependence matrix, which captures the intratumor heterogeneity in the tumor boundary, as the most signicant prognostic factor (P = 0.04 for recurrence-free survival and P = 0.02 for overall survival). The integration of transcriptomic and metabolomic data indicated that high peritumor heterogeneity was related to immune suppression and enhanced metabolism. Our ndings suggest that radiomics data can serve as a noninvasive predictor for molecular subtyping and clinical outcome in patients with TNBC. Abbreviations: HR, hazard ratio; OS, overall survival; Peri_V_DN, peritumoral variance in dependence nonuniformity of peritumoral regions; RFS, recurrence-free survival. Identication of differentially expressed pathways and metabolic ux analysis. GSEA revealed that metabolic pathways were upregulated and immune-related pathways were downregulated in the high group compared low differential (A) Differences in the abundance of immune cell types revealed an immunosuppressive tumor microenvironment in the high Peri_V_DN group compared with the low Peri_V_DN immune signature stromal signature inferred by ESTIMATE supported an immunosuppressive tumor microenvironment in the high Peri_V_DN showed higher inammasome,

In the past decade, radiomics has been an emerging eld that transforms medical images into mineable data by acquiring multiple quantitative image features (12,13). Compared to conventional invasive biopsies, the radiomic approach has several advantages. First, radiomics is a non-invasive method to infer tumor characteristics and can be carried out several times during treatment (11,14,15). Moreover, compared to genomic sequencing, which selects only a small part of the tumors, radiomics elucidates the landscape of a tumor and is not subjected to selection bias, and thus could comprehensively explore tumor heterogeneity (16)(17)(18). Previous studies focused on radiomic texture analysis have quanti ed tumor heterogeneity and suggested its associations with unfavorable prognosis in breast cancer (19,20).
These results warrant further research into evaluating tumor heterogeneity via a radiomic approach.
However, a multiomics TNBC dataset containing radiomics data with a large sample size has yet to be reported, and the correlation between radiomic features and genomic alterations remains largely unknown. In this study, we performed radiomic pro ling based on contrast-enhanced magnetic resonance imaging (CE-MRI) images from 860 Chinese breast cancer patients, which composed the largest breast cancer MRI radiomics dataset to date. This cohort comprised 246 TNBC patients, including 202 samples with genomic, transcriptomic or metabolomic data, and aimed to build a radiomics model for noninvasive TNBC classi cation and patient outcome strati cation. We also integrated radiomic data with our transcriptomic, metabolomic and clinical data to illustrate the biological basis of prognostic radiomic features in TNBC.

Results
Identi cation of TNBC from the Whole Breast Cancer Cohort via a Radiomics Approach To explore whether a non-invasive radiomics approach could be used to identify TNBCs from non TNBCs, we rst established a FUSCC breast cancer radiomics cohort that included all subtypes of breast cancer. A total of 860 patients from FUSCC with baseline breast CE-MRI images for radiomics analysis between August 2009 and June 2019 were retrospectively enrolled ( Figure 1A). Hormone receptor positivity was observed in 468 patients, HER-2 overexpression was observed in 268 patients, and 246 patients had TNBC ( Figure 1B). This radiomic cohort was further randomly divided into 50% training and 50% validation cohorts ( Figure 1A). Another 54 patients from IPMCH were recruited as an independent external validation cohort. Four types of ROIs were employed: tumoral, peritumoral, intratumoral and tumor-peritumoral. To the best of our knowledge, our cohort provides the largest breast cancer radiomic dataset to date ( Figure 1C). The analysis plan of this radiogenomics study is shown in Figure 1D.
Using the training set in the FUSCC breast cancer radiomics cohort, we developed radiomic prediction models through feature selection and model building. Hormone and HER-2 receptor status determined by IHC and uorescence in situ hybridization were regarded as the ground truth to evaluate the prediction e cacy. Using the 10-fold cross validation LASSO model (α = 0), 11 variables were retained for the development of TNBC prediction in the training cohort (Supplementary Figure 1). These radiomic features are presented in Supplementary Table 1. Logistic regression (LR) and support vector machine (SVM) were applied to establish prediction models based on retained features, and radiomics scores were calculated for each patient in both the training and FUSCC validation cohorts using these two methods.
Using the LR-based model, the ability of the radiomic signature to classify TNBC versus non-TNBC was  (Figure 2A). We further calculated that the Jaccard similarity index was 0.881 and 0.894 in the validation cohort for the LR and SVM models, respectively. To validate the effectiveness of this radiomic signature in images from different MRI machines, we calculated the accuracy of these prediction models separately according to MRI machines. To expand the usage of this radiomics signature to a larger extent, we further validated the e cacy of the prediction models in the IPMCH validation cohort. The LR and SVM models yielded AUCs of 0.723 (95% CI: 0.552-0.894) and 0.704 (95% CI: 0.534-0.874), respectively ( Figure 2B). Among these retained features, the mean of tumor zone entropy, which characterizes the mean heterogeneity during different MRI sequences, was signi cantly higher in TNBC ( Figure 2C, P < 0.001), and post hoc analysis indicated that tumor heterogeneity was highly distinctive in identifying TNBC patients, while tumor heterogeneity was not signi cantly different between luminal and HER2-enriched tumors. In summary, these results implied a satisfactory separation of TNBC versus non-TNBC via a radiomics approach and suggested that high intratumoral heterogeneity was a crucial feature of TNBC.

Predictive Value of Radiomics in Distinguishing TNBC Molecular Subtypes
Since the radiomic signature could identify TNBC with high accuracy, we further explored whether radiomic signatures could distinguish different TNBC molecular subtypes. We employed CE-MRI images from 202 patients who were enrolled in the FUSCC breast cancer radiomics cohort as well as the FUSCC TNBC cohort, which is the largest TNBC multiomics cohort to date (n = 465), to validate this hypothesis. The baseline characteristics of this subcohort are shown in Supplementary Table 2. A total of 167 samples with radiomic data also had transcriptomic data, and 138 samples also had metabolomic data (Supplementary Figure 2). Transcriptomic TNBC subtypes were regarded as the ground truth to evaluate the prediction e cacy (7).  Table 3. LR and SVM were used to construct prediction models in the training cohort and validation cohort based on the selected features. Radiomics scores were calculated for each patient using these three methods.
The AUCs and con dence intervals of the prediction models for each TNBC subtype using different methods are shown in Figure  A previous study developed an IHC-based method that speci cally used the staining results of AR, CD8A, FOXC1 and DCLK1 to classify TNBC molecular subtypes (21). We further explored the discriminatory power of prediction models combining radiomic features and IHC data. Since IHC alone could identify the LAR subtype with outstanding e cacy (AUC = 0.932) (21) and a satisfactory radiomic model was established for MES subtype identi cation, we built combined signatures for identifying BLIS and IM subtypes. For predicting BLIS subtype, the SVM-based combined prediction model reached an AUC of 0.975 (95% CI, 0.906-1) in the validation cohort ( Figure 3E-3F). For predicting IM subtype, the LR-based combined prediction model displayed an AUC of 0.731 (95% CI, 0.373-1) in the validation cohort ( Figure 3G-3H). Combined signatures showed better performance in the BLIS subtype than individual IHC and radiomics-based models, but no statistical signi cance was found in the IM subtype. The limited sample size (n = 42 in the training cohort and n = 14 in the validation cohort with radiomics, transcriptomics and IHC staining data) restricted the performance of this combined signature and warrants further research with a larger sample size.

Prognostic Value of Peritumoral Heterogeneity Derived from Radiomics
With the detailed clinical follow-up data of our TNBC cohort, we evaluated robust prognostic radiomic features. According to stringent ltering criteria, variance among the MRI sequences of dependence nonuniformity extracted from peritumoral ROIs (Peri_V_DN), a feature from the gray level dependence matrix group, was identi ed ( Figure 4A). Typical breast CE-MRI images with high and low Peri_V_DN values are shown in Figure 4B. The strati cation of patients with survival differences using the median value as the cutoff was veri ed in both the training cohort and validation cohort ( Figure  4C). The multivariate Cox proportional hazards model also revealed that low Peri_V_DN independently predicted better recurrence-free survival (RFS) and overall survival (OS) in TNBC patients ( Table 1). As a risk-increasing feature, Peri_V_DN represents the variation pattern of peritumoral heterogeneity through different sequences, with a lower value indicating less change in peritumoral heterogeneity among the sequences of the image.
Moreover, we systematically analyzed the correlation between the Peri_V_DN value and tumor characteristics. We observed larger tumor sizes and more pathologically con rmed metastatic lymph nodes (P = 1.8×10 -8 and P = 0.00661, respectively) in the high Peri_V_DN group than in the low Peri_V_DN group ( Figure 4D). The high Peri_V_DN group included more patients with the BLIS subtype, and the low Peri_V_DN group comprised more patients with the IM subtype (P = 0.02), while the distribution of the PAM50 subtypes was balanced ( Figure 4E). We analyzed the correlation between the Peri_V_DN value and TNBC microenvironment clusters according to TNBC microenvironment subtypes (22). The results revealed a tendency for the high Peri_V_DN group to consist of more "immune-desert" cluster 1 tumors, while the low Peri_V_DN group included more "immune-in amed" cluster 3 tumors (P = 0.09, Figure 4E). Fibrosis and necrosis grade evaluated by H.E. sections showed no difference between the two Peri_V_DN groups (Supplementary Figure 5A). Other molecular biomarkers that putatively predict TNBC precision treatment e cacy, including stromal tumor-in ltrating lymphocytes (TILs), IHC CD8 readings, tumor mutation burden (TMB) and homologous recombination de ciency (HRD) score, displayed balanced distributions between the Peri_V_DN groups as well (Supplementary Figure 5B-5E). Overall, we demonstrated that a high value of the radiomic feature Peri_V_DN predicted poor prognosis for TNBC and more aggressive tumor characteristics.

Integrative Analysis Elucidated Metabolic Reprogramming in High Peri_V_DN Patients
We further investigated the molecular characteristics associated with Peri_V_DN. Using paired transcriptomics (n = 167) (7) and metabolomics (n = 138) data, metabolite abundance and gene Previous differentially abundant metabolites and pathway enrichment analyses revealed that metabolic reprogramming was related to high Peri_V_DN. On this basis, we performed DA score analysis between the Peri_V_DN groups. Among 53 pathways in which more than three metabolites were annotated, 21 were upregulated and four were downregulated in high Peri_V_DN patients (Supplementary Table 9). Three pathways were upregulated with DA scores no less than 0.25, namely, fatty acid biosynthesis, primary bile acid biosynthesis and glutathione metabolism ( Figure 5B). This result was consistent with that of a previous analysis and further highlighted fatty acid metabolism reprogramming in high Peri_V_DN. We conducted a transcriptomic-metabolomic integrative analysis to depict a more meticulous fatty acid metabolism alteration in this population and to reveal pivotal metabolic targets. Metabolic ux analysis of fatty acid metabolism demonstrated that the initial step of fatty acid synthesis was signi cantly upregulated, suggesting that FASN played a crucial role in fatty acid synthesis rewiring ( Figure 5C). Acyl carnitines were also highly enriched in high Peri_V_DN and thus sustained an active βoxidation process. Taken together, these results demonstrated that vigorous de novo fatty acid synthesis was closely related to a high Peri_V_DN phenotype.

Distinct Tumor Microenvironments in the Different Peri_V_DN Groups
The landscape of the TNBC microenvironment has been delineated, and distinct "hot tumors" and "cold tumors" have been described (22)(23)(24). In this study, we estimated the cell subset composition of the tumor microenvironment by a published gene signature using transcriptomic data (22). We identi ed 24 different cell types, including immune and stromal cells. RNA-based immune cell deconvolution signatures revealed a major difference in the microenvironment between tumors with high and low peritumor heterogeneity ( Figure 6A-6B). Low Peri_V_DN was characterized by a stronger signature for CD8+ T cells, naïve CD4+ T cells, γδ T cells, activated NK cells, M1 macrophages and regulatory T cells. A comparison of the estimated cell number between the tumor sites and paired adjacent normal sites also suggested that tumors with low peritumor heterogeneity were composed of more immune cells, both immune-promoting and immune-inhibitory cells (Supplementary Figure 7A-7B). Cytolytic activity, which inferred the activity of effector immune cells (25), was lower in high Peri_V_DN cases (P = 0.01, Figure  6C). These results con rmed that high Peri_V_DN was associated with a suppressed immune response. Moreover, we investigated the possible immune escape mechanisms of both types of tumors. In addition to the previously described enrichment of regulatory T cells in low Peri_V_DN cases, another inhibitory immune cell type, myeloid-derived suppressor cells (MDSCs), also had a relatively higher abundance in low Peri_V_DN cases (P = 0.05, Figure 6D) (26). The expression levels of a wide range of immune coinhibitors and costimulators, including multiple immune checkpoints, were investigated, and a more inhibitory immune context was found in low Peri_V_DN cases ( Figure 6E). Overall, the delineation of the tumor microenvironment implied that the low Peri_V_DN group tended to have hot tumors and might escape immune surveillance by higher inhibitory immune cell in ltration and stronger immune checkpoint molecule expression.
Furthermore, the comparison of the two common innate immunity sensing pathways, cGAS-STING (27) and NLRP3 in ammasome (28), demonstrated weaker immunity activation in high Peri_V_DN cases (P = 0.03 and P = 0.02, respectively, Figure 6F). We also analyzed tumor immunogenicity through MHC molecule expression comparison. Reduced expression of MHC molecules in the high Peri_V_DN group is demonstrated in Figure 6G. In summary, the high Peri_V_DN group exhibited a cold tumor phenotype, and its potential escape mechanisms included a reduction in innate immune sensing and rejection of immune in ltration.

Discussion
Recent studies have revealed evident tumor heterogeneity of TNBC, and precision treatment based on molecular pro ling has achieved preliminary progress (5,7,29,30). These promising results encouraged molecular subtyping and genomic sequencing for TNBC in clinical practice, which is traditionally conducted by invasive biopsies. Herein, we developed a non-invasive radiomic approach for the identi cation and molecular classi cation of TNBC. In addition, we identi ed prognostic radiomic features with underlying biological properties. These results demonstrated the potential role of a surrogate radiomic approach in distinguishing TNBC patients and further differentiating TNBC into different subtypes and clinical outcomes.
In this study, we rst investigated the value of radiomics to distinguish TNBC from other subtypes of breast cancer. We concluded that the non-invasive radiomic approach could identify TNBC with an AUC of 0.922 in the FUSCC cohort and 0.723 in the IPMCH cohort. This conclusion not only informed us of the potential of an MRI-based radiomics approach to identify TNBC with high e cacy but also warned us that the generalization of radiomic models into medical images acquired from multiple medical centers and MRI machines remained an important issue to handle. Other studies also explored the value of radiomics to classify TNBC. Calstaldo and colleagues (31) found that radiomic signatures could identify TNBC with an AUC of 0.91 in MRI images derived from 91 patients, while Leithner and colleagues (32) identi ed TNBC with an accuracy of 0.736 using a radiomic approach in a cohort consisting of another 91 patients. These results together demonstrated that MRI-based radiomic signatures could classify TNBC with high accuracy and were consistent with the results of our study.
Previous studies have shown that peritumoral heterogeneity can be used to predict the clinical outcomes of several types of cancer (33)(34)(35). Herein, we found that Peri_V_DN, which represents the variance in peritumoral heterogeneity on MRI sequences, was associated with adverse clinical outcomes. This conclusion pointed out the unique signi cance of the peripheral tumor phenotype. Furthermore, to decipher the molecular mechanisms underlying the distinct radiomic patterns, the connection between a radiomic phenotype and multiomics data needs to be determined, rather than a single driver mutation (11,36,37). Several studies have made efforts to explore the association of radiomic features and transcriptomic data (18,38). Lee et al (16) found that a four-feature radiomic signature could predict the clinical outcomes of pathological T1 renal cell carcinoma and was associated with the abundance of certain immune cell types. This was consistent with our ndings, but the exact mechanisms of forming an immunosuppressive microenvironment in tumors with low peritumoral heterogeneity were not explored in this article.
Our study has several limitations. First, more re ned models are needed to further improve the prediction e cacy of the radiomic signatures, especially for predicting TNBC molecular subtypes. Second, most patients were recruited from a single institution, and the sample size of the independent external validation cohort was limited (n = 54). Third, the biological characteristics associated with the Peri_V_DN feature were subjected to the nature of exploratory analysis.
In conclusion, we presented a radiomic dataset originating from the largest breast cancer radiomics cohort to date. The radiomics approach showed promising e cacy in identifying TNBC and predicting TNBC molecular subtypes in a non-invasive way. Additionally, peritumoral heterogeneity quanti ed by radiomics strati ed patient outcomes and represented distinct tumor metabolism and immune response patterns. These results demonstrated the potential application of radiomics analysis in the analysis of tumor heterogeneity and the clinical management of TNBC.

Patient Cohorts and Data Sets
We retrospectively recruited patients diagnosed with malignant breast cancer whose baseline breast CE-MRI images were suitable for radiomics analysis. A total of 860 Chinese patients who were treated at Fudan University Shanghai Cancer Center (FUSCC) from 1 August 2009 to 31 May 2015 were enrolled according to the following de ned criteria: 1) female patients diagnosed with unilateral invasive ductal carcinoma with known ER, PR and HER2 phenotypes (39); 2) no evidence of distant metastasis at diagnosis.
To further explore whether radiomics could distinguish different TNBC molecular subtypes, 202 TNBC patients was selected from the FUSCC breast cancer radiomics cohort. These patients were also included in our TNBC multi-omics dataset, which has been published recently (7). Among the 202 patients eligible for the radiomics analysis, transcriptomic sequencing (n = 167), metabolomics (n = 138), hematoxylin and eosin (H.E.)-stained sections with IHC staining (n = 56) for the expression of AR, CD8A, FOXC1 and DCLK1, and clinical follow-up data were also available. The studies were conducted in accordance with the Declaration of Helsinki. All analyses were approved by the independent ethics committee/institutional review board of Fudan University Shanghai Cancer Center, and written informed consent was obtained from each patient.
To evaluate the ability of radiomic signatures to generalize to additional populations, we collected CE-MRI images and ER, PR and HER2 phenotype information from 54 patients diagnosed with malignant breast cancer at the International Peace Maternal and Children Hospital (IPMCH). These patients were recruited from 1 January 2013 to 31 June 2019. Data collection was under approval of the IPMCH institutional review boards.

CE-MRI and Regions of Interest (ROI) Delineation
All patients underwent breast MR examination before biopsy, and CE-MRI images were used for radiomic analysis in this study. The imaging parameters are listed in Supplementary Table 10.
ROIs were delineated semiautomatically on the peak enhanced phase of CE-MRI by 3D slicer software (https://www.slicer.org/). ROIs were placed on all slices that contained the whole tumor or the largest lesion (in the case of multicentric or multifocal tumors). Two radiologists (C.Y. and D.D.Z. with 9 and 4 years of experience in breast MRI, respectively) were blinded to the pathological and biochemical ndings of each patient and were chie y responsible for the evaluation of the ROIs. The interand intraobserver reproducibility of the ROIs and radiomic feature extraction were initially analyzed with the CE-MRI data of 60 randomly selected patients in a blinded fashion by two radiologists. To ensure reproducibility, one radiologist repeated the ROI drawing twice with an interval of at least 1 month and generated radiomic features following the same procedure. Intraclass correlation coe cients (ICCs) were utilized to evaluate the intra-and interobserver agreement in terms of feature extraction. Interand intraobserver reproducibility and radiomic feature extraction achieved substantial agreement with ICC > 0.6 both among the ROIs from the two radiologists and between the ROIs from the same radiologist (40). Furthermore, the peritumoral area was obtained by expanding the tumor outward with a 2-pixel width and subtracting the tumor area, while the intratumoral area was obtained by shrinking with a 2-pixel width. In addition, tumor and peritumoral regions were integrated as another region. In total, four sets of ROIs, including tumor, peritumor, intratumor and tumor-peritumor regions, were used in the radiomics feature extraction.

Radiomics Feature Extraction
This study extracted the radiomic features of the contrast-enhanced phase, including shape features, rst-order features, text features and sequential features, using the PyRadiomics package (41). Fourteen shape-based characteristics, which describe the difference in shape between different types of tumors, were calculated using the rst postcontrast phase of CE-MRI. Ninety-one rst-order features were calculated from the four phases individually, and 364 features were acquired in total; these features describe the distribution of voxel intensities. Textural features were calculated based on ve textural matrices to describe the radiological pattern of the ROI. Moreover, sequential features were calculated to consider time dimension information. The extracted sequential features were mainly composed of two parts: (a) the feature discrepancy between four different phases on the timeline including 364 features and (b) the mean, variance, kurtosis and skewness of the time curve extracted from each individual patient, including ninety-one in each group. As described in the ROI delineation, feature extraction was performed in the tumoral, peritumoral, intratumoral and tumor-peritumoral regions.

Feature Selection and Radiomics Model Building
The least absolute shrinkage and selection operator (LASSO) method was used to select the most useful predictive features from the training cohort (glmnet R package) (42). Tuning parameter (λ) selection in the LASSO model used 9-fold cross-validation. Radiomics scores were calculated for each patient via three different methods: 1) multivariate linear regression (glm R package), 2) support vector machine (SVM; e1071 R package), and 3) deep neural network (DNN; sklearn Python package). The ability to predict TNBC molecular subtypes was assessed by the area under the receiver operating characteristic curve (AUC) of the receiver operator characteristic (ROC) curve via the pROC R package (43). Con dence intervals of AUCs were calculated using the Delong method.

Radiomics Model Validation
The radiomics prediction models were validated internally and externally. First, the trained classi ers were assessed by cross-validation via the glmnet R package (42). Then, the trained classi ers were further tested in the validation cohort in terms of the AUC and its con dence intervals of the ROC curve.

Collection and Analysis of Metabolomics and Lipidomics Data
Samples in our multiomics TNBC cohort with adequate tissues for polar metabolites and lipids were collected. In total, 138 TNBC samples were selected for further metabolomics and lipidomics analysis. Acetonitrile: methanol: water = 2: 2: 1 solution and MTBE: MeOH= 5: 1 solution were applied to extract polar metabolites and lipids, respectively. An equal volume (10 μL) of each sample was mixed for quality control sample preparation. A BEH amide column (2.1 * 100 mm, 1.7 μm, Waters) or Kinetex C18 column (2.1 * 100 mm, 1.7 μm, Phenomen) coupled with a Triple TOF 6600 mass spectrometer or AB triple TOF 5600 mass spectrometer was deployed to conduct LC-MS/MS experiments for polar metabolite and lipid detection. Detailed information on sample processing, metabolomics and lipidomics data generation is contained in the supplementary materials and methods.

Analysis of Differentially Abundant Metabolites and Differentially Expressed Genes
The differential abundance of metabolites was calculated by performing Mann-Whitney U tests for all detected metabolites. The differential expression of genes was determined via the limma R package (44). Subsequent gene set enrichment analysis (GSEA) was performed using the clusterPro ler R package (45). The differential expression analysis outputs of limma were used to generate the ranked list le. One thousand total permutations were used. Differential Abundance (DA) Score The DA score was calculated rst by determining which metabolites were signi cantly increased/decreased in abundance, as described above. Then, the DA score was de ned as follows: DA = (Number of metabolites increased -Number of metabolites decreased)/Number of measured metabolites in that pathway (46) Thus, the DA score ranges from -1 to 1. A score of -1 indicates that all metabolites in a pathway decreased, while a score of 1 indicates that all metabolites increased in abundance.

Calculation of Microenvironment Cell Abundance
A signature containing 364 genes representing 24 microenvironment cell types was obtained from one published high-quality paper (22). This signature modi ed the CIBERSORT and MCP-Counter signatures and represented a more comprehensive landscape of the TNBC microenvironment.
Subsequently, we used single-sample gene set enrichment analysis (ssGSEA, "GSVA" function in GSVA R package) (47) to calculate the abundance of each cell subset in each sample with expression data. Additionally, we referred to the xCell signature to further validate the composition of the tumor microenvironment (48).

Statistical Analysis
Student's t test, Wilcoxon's test and Kruskal-Wallis test were used to compare continuous variables. Prior to the comparisons, the normality of the distributions was tested with the Shapiro-Wilk test. Pearson's chisquare test and Fisher's exact test were employed for the comparison of unordered categorical variables. To explore the association between radiomics features and survival, Kaplan-Meier analysis and a Cox proportional hazards model were employed in the training and validation cohorts. The comparison of survival between groups was conducted via the log rank test. All the tests were two sided, and P < 0.05 was regarded as indicating signi cance unless otherwise stated. In multiple hypothesis testing, false discovery rate (FDR) correction was used to decrease false positive rates. All statistical analyses were performed with R software (version 3.6.1, http://www.R-project.org).     Distribution of the TNBC mRNA subtypes, PAM50 subtypes and TNBC microenvironment clusters between the high and low Peri_V_DN groups. Abbreviations: HR, hazard ratio; OS, overall survival;

Declarations
Peri_V_DN, peritumoral variance in dependence nonuniformity of peritumoral regions; RFS, recurrence-free survival. Figure 5 metabolic pathways were upregulated and immune-related pathways were downregulated in the high Peri_V_DN group compared to the low Peri_V_DN group. (B) The differential abundance score illustrated major differential metabolic pathways utilizing metabolomic data. (C) Metabolic ux analysis combining transcriptomic and metabolomic data showed reprogramming of the fatty acid biosynthesis pathway between the Peri_V_DN groups. (D) GSVA revealed differences in multiple lipid metabolism pathways. In total, 167 samples with transcriptomics data and 138 samples with metabolomics data were included for analysis. Abbreviations: FA, fatty acid; GSEA, gene set enrichment analysis; GSVA, gene set variation analysis; NES, normalized enrichment score; Peri_V_DN, peritumoral variance in dependence nonuniformity of peritumoral regions; ssGSEA, single-sample gene set enrichment analysis. Figure 6