GSE64554 (14), GSE120774 (15), GSE192886 (16) and GSE24425 (17) were obtained from Gene Expression Omnibus (GEO, https://www.ncbi.nlm.nih.gov/geo). Array or sequencing data of paired EAT and SAT in GSE64554 (n=46), GSE120774 (n=36), and GSE24425 (n=12) were from patients undergoing cardiac valve or coronary artery bypass graft surgery. GSE192886 contained sequencing data of EAT from HF patients (n=5) and non-HF patients (n=5) undergoing coronary artery bypass graft surgery. Clinical characteristics for analyzed patients can refer to the original citations of these datasets and Tables S1 - S3 .

In the validation experiments, fresh EAT with paired SAT, heart, and peripheral blood samples were collected from HF patients undergoing heart transplantation in Wuhan Union Hospital. Peripheral blood samples were obtained before surgery. SAT samples were obtained from the suprasternal region, heart and EAT samples were obtained from the left ventricle. We obtained informed consent from all enrolled subjects. The experimental protocol and sample collection were in accordance with the Declaration of Helsinki and approved by the Medical Ethics Committee of Wuhan Union Hospital of Huazhong University of Science and Technology (METC number: 20200462). Information on the involved subjects was listed in Table S4 .

The data obtained from GSE64554 and GSE120774 were processed by log₂ transformation and quantile normalization via limma package (18) using R separately. The differential expression matrixes of the datasets were also identified by the limma package separately and P values were adjusted by the Benjamini-Hochberg method. We then applied the Robust Rank Aggregation (RRA) method (19) to filter the differential expression matrixes, so as to obtain the comprehensive differentially expressed genes (DEGs) across two different microarray platforms. DEGs with RRA score less than 0.05 were selected for further analyses.

Functional enrichment analyses were performed using the DAVID (20) by inputting the official gene symbols of obtained DEGs. Figures for functional enrichment analyses were plotted by R and Sangerbox (http://www.sangerbox.com/tool). Construction of the protein-protein interaction (PPI) network and identification of hub genes were performed as the previous description (21).

To explore the gene modules responsible for the phenotypic differences between EAT and SAT, we performed the Weighted gene co-expression network analysis (WGCNA)to identify co-expressed gene modules (22). First, we screened the top 25% of the genes in the variance variability between samples in a pooled matrix and used them as input data. Next, we obtained the soft threshold and set the minimum gene number in the module to 30 to get gene co-expression modules. By analyzing the correlation between each module with the EAT/SAT phenotypes, we screened out the gene modules that need further exploration. Finally, functional enrichment analyses were performed on the obtained modules, and the modules significantly related to the immune process were identified. By taking the intersection of immune-related key modules and DEGs identified by RRA, we obtained a set of key immune-related genes.

xCell (23) and CIBERSORT (24) are signature-based methods to infer the immune cell landscape according to expressional profiling. We performed immune cell infiltration analyses and obtained the immune cell landscapes for EAT and SAT based on the pooled matrix. “Lymphoid cells” and “myeloid cells and others” were categorized. Results were evaluated by t-test to determine the significance of differences. The correlation relationship between immune cell types, WGCNA modules, and target genes was evaluated by Pearson correlation coefficients.

Paired EAT, SAT, and heart samples were used for T cell receptor (TCR) repertoire sequencing. Tissue genomic DNA was extracted using Universal Genomic DNA Kit (CWBio, China). DNA quality was evaluated using Nanodrop2000 (Thermo, USA) with concentration >20ng/uL and OD260/280 between 1.7 and 2.0. Multiplex PCR reactions were run to specifically amplify the third complementarity-determining region (CDR3) of the TCRβ chain for libraries construction. The constructed libraries were deeply sequenced by Illumina NextSeq500. Primers and sequencing were provided by SEQHealth (China).

Raw sequences filtered by SOAPnuke (version 1.6.0) were used for TCR sequencing analyses, and the sequencing data were mapped to the ImMunoGeneTics (IMGT) database using MiXCR (version 3.0.3) to define the V, D, and J fragments and CDR3 sequence (25). The terms TCR clonotype and TCR clone describe the CDR3 sequence composed of a unique amino acid sequence and CDR3 sequence composed of unique V, D, and J fragments, respectively. Antigen matching analysis was performed via the IEDB database (http://www.iedb.org/).

Peripheral blood mononuclear cells (PBMCs) were isolated by density gradient centrifugation using lymphocyte separation medium (MPbio, USA). Fresh EAT samples were digested at 37°C in Hepes buffer containing collagenase D (1mg/mL, Sigma, USA) and dispase II (2mg/mL, Sigma, USA), and then filtered by 100μm and 40μm filters (Falcon, USA) sequentially to collect the stromal vascular fraction (SVF) for subsequent flow cytometric analyses. Memory phenotypes of T lymphocytes were categorized into naïve T cell (T_N, CD62L⁺CD45RA⁺), central memory T cell (T_CM, CD62L⁺CD45RA^-), effector memory T cell (T_EM, CD62L^-CD45RA^-) and CD45RA⁺ effector memory T cell (T_EMRA, CD62L^-CD45RA⁺). For the detection of interferon (IFN)-γ, cells were re-suspended in RPMI-1640 medium (Gibco, USA) with 10% heat-inactivated FBS (Gibco, USA) at a concentration of 10⁶ cells/ml and stimulated with Cell Stimulation Cocktail (eBioscience, USA). After 6 hours of stimulation, cells were harvested, permeabilized, and then stained with fluorescence-conjugated antibodies. Used antibodies were as follows: PE-Cy7-anti-human CD3(BD Biosciences, USA), PE-anti-human IFN-γ (BD Biosciences, USA), BV421-anti-human CD45RA (BD Biosciences, USA), APC-anti-human-CD62L (Biolegend, USA), Fixable Viability Stain 510 (BD Biosciences, USA). The stained cells were washed with Flow Cytometry Staining Buffer (eBioscience, USA) and fixed with IC Fixation Buffer (eBioscience, USA). Flow cytometry analyses were performed with a FACS Calibur flow cytometer (BD Biosciences, USA) and analyzed by FlowJo software.

For immunohistological or immunofluorescence staining, paired EAT and SAT samples were fixed in 4% paraformaldehyde at 25 °C for 24 hours and embedded in paraffin. Slides were sectioned in 5μm and blocked with 1% BSA PBS buffer and then stained with target antibodies and DAPI following routine procedures. The slides were scanned with a digital scanner (3D-HISTECH, Hungary). CaseViewer software was used for observation and statistics. For immunohistological statistics, 3 areas under the 20x field of view from each slide were randomly selected, and the average number of positive cells per mm² was calculated (3 slides included for each sample). Used antibodies were as follows: human CD3 antibody (Servicebio, China), human CCL5 antibody (R&D systems, USA), and human GZMK antibody (R&D systems, USA).

Data processing and analyses were performed using SPSS 22.0, GraphPad Prism, and R. Normality were evaluated by the Shapiro-Wilk test. Differences were evaluated using Student’s t-test and P < 0.05 was considered statistically significant unless indicated otherwise.

The DEGs between EAT and SAT in GSE64554 and GSE120774 were identified separately and shown in Figure 2A . Next, we applied the RRA algorithm to integrate DEGs of the two datasets and obtain a more comprehensive DEGs list. The RRA method identified 131 genes that were up-regulated in EAT compared to SAT, while 159 genes were down-regulated. DEGs identified by RRA presented significant differences (adjusted P value<0.05 and |log₂FC| ≥0.5) in at least one dataset, most of which (90%) showed consistent expressing trends across datasets. The top10 up- and down-regulated DEGs recognized by RRA were shown in Figure 2B . Next, we applied Gene Ontology (GO) enrichment analysis on the DEGs identified by RAA that were up- and down-regulated in EAT versus SAT to explore their potential functions, respectively. As shown in Figure 2C , the up-regulated DEGs in EAT were mainly enriched in complement activation and immune response, while the down-regulated DEGs were mainly related to embryonic skeletal system morphogenesis, suggesting immune activation in EAT compared to paired SAT. The PPI network of DEGs identified by STRING was further analyzed by cytoHubba to identify hub genes. As shown in Figure S2 and Table S5 , we obtained the top 10 hub genes including COL1A1, FGF2, BGN, C3, TIMP1, CD44, POSTN, COL3A1, CCL2 and APOB.

Further, we applied the WGCNA method to identify immune-related key gene modules associated with EAT. By filtering the expression profiles of the top 25% variance in all EAT and SAT samples, a total of 3869 highly variable genes were included in WGCNA. Then, filtered genes were clustered into 18 different modules based on WGCNA clustering ( Figure 2D ). The correlation analyses between all modules and EAT/SAT phenotype were carried out and 7 modules were found to be significantly associated with the EAT/SAT phenotype ( Figure 2E ). Functional enrichment analyses suggested that the blue and greenyellow modules were closely related to immune response ( Figure 2F ). The overlap of DEGs and the two modules were identified and 9 key genes were obtained for further analyses. Of the 9 key genes, all were up-regulated DEGs in EAT and listed in Table S6 , including SLCO2B1, F13A1, C1QA, C1QB, and C1QC from the blue module and IGLL1, GZMK, CCL5, and SLC38A1 from the greenyellow module.

We used xCell to explore the differences in the immune cell landscape between EAT and SAT. As shown in Figure 3A , EAT was infiltrated by more lymphocytes and dendritic cells (DC), while the abundance of macrophages and M1 macrophages showed no significant difference. In SAT, M2 macrophages, basophils, and mast cells showed higher degrees of infiltration. The correlation between different cell subtypes was calculated to infer their potential interaction. In Figure 3B , CD4⁺ T cells and CD8⁺ T cells presented a strong positive correlation (r=0.82), indicating that the two subtypes of T cells had a consistent tendency of infiltration.

Next, we analyzed the correlation between the infiltrated immune cells with hub genes, key genes and key modules identified in EAT from the previous PPI network and WGCNA analyses. As shown in Figure 3C , the expression of GZMK, CCL5, IGLL1, and SLC38A1 presented a strong positive correlation with CD4⁺ T cells, CD8⁺ T cells, and B cells, while the expression of SLCO2B1, BGN, C3, TIMP1, C1QA, C1QB and C1QC showed a strong positive correlation with DC. As shown in Figure 3D , the 7 key modules related to EAT and SAT obtained by WGCNA were all related to different subtypes of immune cells. In particular, the blue and greenyellow modules closely related to the immune process presented a strong positive correlation with lymphocyte abundance. The correlation coefficients between the greenyellow module and T or B cells were more than 0.8. Based on the above analyses, we concluded that EAT acts as a pro-inflammatory adipose tissue characterized by abundant lymphocyte infiltration compared with SAT.

To further explore the characteristics of EAT from HF patients, we analyzed a public dataset GSE192886 containing transcriptome profiles of EAT from 5 HF patients and 5 patients without HF as controls (CON). We obtained 196 up-regulated and 261 down-regulated DEGs in EAT from HF patients versus that from controls. Function enrichment analysis of up-regulated DEGs suggested immune activation, particularly lymphocyte activation in EAT from HF patients ( Figure 4A ). Up-regulated DEGs were mainly enriched in the lymphocyte activation pathway relative to the myeloid leukocyte activation pathway ( Figure 4B ). Next, we used CIBERSORT to compare the immune cell composition between EAT from HF patients and non-HF controls. As shown in Figure 4C , the frequencies of T cells and B cells were higher in EAT from HF patients compared to non-HF controls, indicating lymphocyte activation as the hallmark of EAT from HF patients.

We examined the expression levels of genes associated with the inflammatory characteristics of EAT (10 hub genes and 9 key genes identified above). As shown in Figure 4D and Figures S3 , S4 (Mann-Whitney test), of these genes (IGLL1 not included), the expression of CCL5, GZMK, and POSTN showed a further increase in EAT from HF patients indicating an enhanced degree of inflammation and fibrosis while CCL5 and GZMK presented the strongest positive correlation with infiltrated T lymphocytes in previous analyses ( Figure 3C ). Next, we examined the expressions of T cell-inflamed gene expression profiles (GEPs) between EAT from HF patients and non-HF controls. T cell-inflamed GEPs (composite genes listed in Table S7 ) has been reported to be associated with inflammatory T-lymphocyte infiltration and prediction of sensitivity to immunotherapy in tumors (26, 27). As shown in Figure 4E (Mann-Whitney test), the expressions of T cell-inflamed GEPs were higher in EAT of HF patients, providing further evidence of an enhanced T-lymphocyte response. In addition, the expression of CCL5 and GZMK were also strongly positively correlated with T cell-inflamed GEPs in EAT and SAT samples from validation dataset GSE24425 ( Figure 4F ).

Next, we identified the potential key TFs regulating the phenotypic transition of EAT from HF patients using the ChEA3 database (28) ( Figure 4G ). The PPI network of top 10 predicted key TFs suggested a crucial role of lymphocyte-specific TFs in EAT of HF patients, especially for those were differentially expressed including TBX21, PAX5, NFATC2, and STAT4 ( Figure 4H ).

The numbers of TCR clones and TCR clonotypes were higher in EAT than in paired SAT from HF patients, indicating enhanced T lymphocyte infiltration in EAT ( Figure 5A ). Next, we compared the distribution of the low (fraction>0.1%), middle (fraction>0.5%), and high (fraction>1%) frequency TCR clonotypes between EAT and paired SAT. The results suggested the enrichment of highly expanded TCR clonotypes in EAT compared to paired SAT ( Figure 5B ). Accordingly, the proportion of top 10 TCR clonotypes was higher in EAT than in SAT ( Figures 5C, D ). These results suggested that T lymphocytes from EAT of HF patients exhibited higher clonal expansion than those from SAT. TCR clones with high frequency in EAT were listed and evaluated by antigen matching analysis via the IEDB database ( Tables S8 , S9 ).

A relatively low proportion of shared TCR clonotypes was observed between EAT and paired SAT ( Figure 5E ). However, the degree of TCR clonotype sharing between cardiac tissue and EAT was higher than that between cardiac tissue and SAT ( Figure 5F ). Further, we found the Spearman’s correlation coefficients between frequencies of TCR clonotypes in cardiac tissue and paired EAT was higher compared to that of SAT ( Figure 5G ). Next, we examined the usages of TRBV-TRBJ fragments in EAT, SAT, and cardiac tissue ( Figure 5H ). For the frequency distribution of V-J fragments with an average frequency >1% in the heart (19 V-J fragments ranked in Figure S5 ), the correlation between cardiac tissue and EAT was greater than that between EAT and SAT while no obvious correlation was observed between cardiac tissue and SAT ( Figure 5I ). Thus, the above results suggested a similar antigenic microenvironment between the heart and adjacent EAT.

In order to verify the accumulation of T lymphocytes in EAT and the role of key genes, we collected EAT together with paired SAT and peripheral blood samples from HF patients undergoing heart transplantation. Immunohistomical staining showed abundant CD3-positive T lymphocytes in EAT compared to SAT ( Figures 6A, B ). Further flow cytometry showed that enriched T lymphocytes in EAT were mainly composed of T_EM expressing high levels of IFN-γ ( Figures 6C–F ). Further immunofluorescence staining results confirmed that CCL5 and GZMK were co-localized with CD3-positive T lymphocytes ( Figure 6G ). Taken together, we concluded that EAT from HF patients were populated by inflammatory T_EM cells expressing high levels of effector molecules including IFN-γ, CCL5, and GZMK and thus contributing to EAT pro-inflammatory conversion in HF patients.