Medicine
OPEN
Interaction between ENPP1 and homologous recombination deficiency defines distinct pan- cancer signatures
A retrospective observational study
Yong Min Kim, MDª, Sung Hak Lee, MD, PhDb, Woong Na, MD, PhDe, Il Ju Lee, MSd, Mihye Kwon, MD, PhDe, Oyeon Jo, BSª, Young Soo Song, MD, PhDa,f,*[D
Abstract
Ectonucleotide pyrophosphatase/phosphodiesterase 1 (ENPP1), 1st identified in breast cancer and subsequently in multiple other cancer types, is an innate immune checkpoint regulator that recently emerged as a promising biomarker and therapeutic target. Homologous recombination deficiency (HRD) has gained clinical relevance with therapeutic vulnerability, particularly in breast and ovarian cancers. Despite the increasing significance of ENPP1 and HRD in cancer biology and treatment, their potential relationships have not yet been comprehensively investigated. We analyzed the relationship between ENPP1 expression and HRD score across the Cancer Genome Atlas pan-cancer and individual tumor types using the Pearson and Spearman correlations. To account for heterogeneity, pan-cancer samples were clustered using linear regression into 3 groups based on Bayesian Information Criterion. Differential expression, functional enrichment, and survival analyses were performed for these clusters at both the pan-cancer and representative tumor type levels. Although the pan-cancer relationship between ENPP1 expression and HRD score was heterogeneous, significant correlations were observed in 11 tumor types. Linear regression-based clustering resolved this heterogeneity into 3 functionally and clinically distinct groups: Cluster 1 was characterized by proliferation programs; Cluster 2 by extracellular matrix remodeling, differentiation, and immune response; and Cluster 3, by metabolic reprogramming. Clinically, Cluster 3 was associated with better survival than Clusters 1 and 2 in a pan-cancer analysis (P < . 0001). At the individual tumor type level, these global cluster features were further modified in tissue-specific contexts, reflecting local microenvironment adaptation. Significant survival differences were observed in patients with adrenocortical carcinoma, chromophobe renal cell carcinoma, low grade glioma, and mesothelioma, further underscoring the tissue-specific modification of global cluster features. Our comprehensive pan-cancer analysis revealed the intrinsic heterogeneity of ENPP1 expression and HRD score, which may arise from complex and dynamic interactions with diverse cancer hallmarks, including proliferation, extracellular matrix remodeling, immune response, and metabolic reprogramming, and can be generalized into 3 clusters with distinct molecular and clinical characteristics. At the individual tumor type level, these global cluster features were further modified to adapt to a tissue-specific microenvironment, manifesting distinct tissue-specific patterns. Collectively, these findings provide a foundation for refining biomarker-driven precision medicine strategies for diverse tumor types.
Abbreviations: ACC = adrenocortical carcinoma, ADO = adenosing, DEG = differentially expressed gene, ECM = extracellular matrix, ENPP1 = ectonucleotide pyrophosphatase/phosphodiesterase 1, GO = gene ontology, HR = hazard ratio, HRD = homologous recombination deficiency, KEGG = Kyoto Encyclopedia of Genes and Genomes, KICH = kidney chromophobe renal cell carcinoma, LGG = low grade glioma, MESO = mesothelioma, TCGA = The Cancer Genome Atlas.
Keywords: ENPP1, functional enrichment analysis, HRD score, linear regression-based clustering, pan-cancer, survival analysis
YMK and SHL contributed to this article equally.
This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korean Government (MSIT) (2021R1A2C1094790).
The datasets generated during and/or analyzed during the current study are publicly available.
Supplemental Digital Content is available for this article.
a Department of Pathology, College of Medicine, Konyang University, Daejeon, South Korea, b Department of Hospital Pathology, College of Medicine, The Catholic University of Korea, Seoul, South Korea, ” Department of Pathology, H Plus Yangji Hospital, Seoul, South Korea, ª LabGenomics Clinical Research Institute, LabGenomics, SeongNam, South Korea, e Department of Internal Medicine, College of Medicine, Konyang University, Daejeon, South Korea, Myunggok Medical Research Center (Institute), College of Medicine, Konyang University, Daejeon, South Korea.
* Correspondence: Young Soo Song, Department of Pathology, College of Medicine, Konyang University, 158 Gwanjeodong-ro, Seo-gu, Daejeon 35365, South Korea (e-mail: lifen@konyang.ac.kr).
Copyright @ 2026 the Author(s). Published by Wolters Kluwer Health, Inc. This is an open access article distributed under the Creative Commons Attribution License 4.0 (CCBY), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
How to cite this article: Kim YM, Lee SH, Na W, Lee IJ, Kwon M, Jo O, Song YS. Interaction between ENPP1 and homologous recombination deficiency defines distinct pan-cancer signatures: A retrospective observational study. Medicine 2026;105:1(e47164).
Received: 26 September 2025 / Received in final form: 3 November 2025 / Accepted: 7 December 2025
http://dx.doi.org/10.1097/MD.0000000000047164
1. Introduction
Ectonucleotide pyrophosphatase/phosphodiesterase 1 (ENPP1), an innate immune checkpoint that regulates the cyclic GMP- AMP, which is a stimulator of interferon genes pathway, has recently emerged as a promising biomarker in immuno- oncology with therapeutic potential through selective enhance- ment of anti-cancer immunity.[1-11] Together with its therapeutic potential, the crucial features of ENPP1 were initially explored in breast cancer, and is now being actively investigated across multiple cancer types.[9,12] Therefore, the interaction between ENPP1 and other key biomarkers is emerging as a major theme in cancer research.[11]
Homologous recombination is a complex DNA repair sys- tem comprising numerous components and tumors harboring homologous recombination deficiency (HRD) has emerged as important therapeutic target in various cancers particularly breast and ovarian cancers.[13-23] The presence of HRD, mainly caused by germline or somatic mutations of HRD-related genes, is a crucial determinant in prescribing poly(ADP-ribose) poly- merase inhibitors, which can induce synthetic lethality and ulti- mately improve patient survival.[22,24,25] Although HRD can be assessed by various metrics, the HRD score is currently the most widely used in clinics.[15,20,23]
The potential interaction between ENPP1 expression and HRD score seems plausible, although, to our knowledge, a direct analysis of this relationship has not yet been reported. ENPP1 has been reported to be highly expressed in cancers characterized by genomic instability.[26] Moreover when ade- nosing (ADO) signaling was assessed using an ADO score to which ENPP1 contributed significantly, a significant association was observed between the ADO score and presence of HRD.[27] In contrast, HRD might attenuate ENPP1 activity by altering NAD+/ATP balance or by suppression of the cGAS-stimulator of interferon genes pathway.[5,28] Furthermore tumors with HRD often exhibit immune-hot microenvironment in which ENPP1 activity might be lowered.[29-31] In this context, to address the missing links between HRD and innate immune modulation, a comprehensive analysis of the relationships between ENPP1 expression and HRD score across multiple cancer types may be warranted.
In this study, using The Cancer Genome Atlas (TCGA) data- set,[32] we investigated the relationship between ENPP1 expres- sion and HRD scores across multiple cancer types and assessed their molecular and clinical significance. By applying linear regression-based clustering to pan-cancer data, the cases were naturally separated into 3 groups with distinct molecular and clinical features. These observations suggest that the interaction between innate immune regulation and DNA repair system is complex and widespread, and may hold promise as a novel bio- marker for cancer stratification.
2. Methods
2.1. Data collection and preprocessing
This retrospective observational study adhered to the STROBE guidelines. For a comprehensive analysis of the relationship between ENPP1 gene expression and HRD score across 33 cancer types, TCGA data (normalized pan-cancer ribonucleic acid-Seq, HRD score, and clinical data) were retrieved and downloaded using the UCSC Xena browser.[33] In accordance with TCGA polices, neither institutional ethics approval nor patient consent was necessary in this study. Because lists of TCGA sample IDs were not entirely consistent across data types, we initially created a pan-cancer dataset for ENPP1 expression and HRD score, and then survival data were assembled to this dataset. Consequently, 9828 samples were included in the anal- ysis of the relationship between ENPP1 expression and HRD score, and survival analyses were performed on 9625 patients
(Section 2.5). Molecular and survival analyses for individual cancer types were also performed using the same sample sets (see Table S1, Supplemental Digital Content, https://links.lww. com/MD/R134, which summarizes the number of samples per cancer type used in molecular and survival analyses).
2.2. Correlation between ENPP1 expression and HRD score
To evaluate the global pan-cancer relationship between ENPP1 expression and the HRD score, scatter plots were generated and the associations were quantified using both Pearson and Spearman correlation coefficients. The same analyses were independently performed for each tumor type. To enhance visualization of tumor type-specific territories within the pan- cancer scatter plot, we applied k-nearest neighbors approach (k = 50).[34] By employing both correlation metrics (Pearson and Spearman correlation coefficients) in a complementary manner, we aimed to capture more reliable and robust correlation trends across the heterogeneous pan-cancer landscape as well as within individual cancer types.
2.3. Linear regression-based clustering
Linear regression-based clustering is a statistical technique that partitions an entire data into distinct clusters and then fits a sep- arate linear regression model to each cluster.[35,36] This approach offers greater flexibility when a single global linear regression model cannot adequately capture complex or heterogeneous data structures. Because the distribution of ENPP1 expression and HRD score did not exhibit clearly separable clusters or a single global linear relationship, linear regression-based cluster- ing was considered an appropriate choice for this dataset. For implementation, we used the R package flexmix (version 2.3-20; Austria), which provides a general framework for fitting finite mixtures of regression models based on an expectation-max- imization algorithm.[37] The optimal number of clusters was determined using the Bayesian Information Criterion, as imple- mented in the same package.
To identify the unique characteristics of the clusters derived from linear regression-based clustering, 3 main analyses were performed. First, the information-theoretic framework was applied to quantitatively assess the dependence and indepen- dence relationships among the clusters, ENPP1 expression, and the HRD score. Second, the ENPP1 expression patterns for each cluster were analyzed in detail. Third, the results of this study’s clustering methods were compared with those obtained using k-means clustering, a commonly used unsupervised method.[38]
For the specific implementation of the information-theoretic framework, the infotheo R package was utilized to measure the entropy and mutual information of the clusters, ENPP1 expres- sion, the HRD score, and their multivariate relationships.[39] The presence of unique information in the clusters, independent of the individual effects of ENPP1 expression and the HRD score, was confirmed when the entropy of the clusters was greater than the mutual information between ENPP1 expression and the HRD score.[40]
2.4. Differentially expressed genes (DEGs) and gene set- based enrichment tests between clusters
To identify molecular features distinguishing the clusters derived from the linear regression-based clustering, DEGs were identified using the limma R package (version 3.60-6), apply- ing selection criteria of absolute log2-fold change >1 and adjusted P-value of < . 05.[41] The identified DEGs were subse- quently analyzed using gene ontology (GO) enrichment and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway
enrichment analyses implemented in the ClusterProfiler R pack- age (version 4.16-0).[42]
These DEG analyses and molecular profiling using GO and KEGG were performed for both the pan-cancer cohort and individual cancer types. For the latter, instead of analyzing all 33 cancer types, we focused on the representative ones that exhibited significant survival differences according to the clus- ter membership (Section 2.5). Subsequent DEG and enrichment analyses were conducted for these selected cancer types to eluci- date the biological mechanisms associated with their distinctive clinical outcomes.
2.5. Survival analysis
For the assessment of clinical relevance, survival analyses were performed using survival and survminer R packages.[43,44] A total of 9828 tumor samples from TCGA were initially included in the primary analyses of ENPP1 expression and HRD score (Sections 2.2-2.4). Clinical survival data (overall survival time and vital status) were retrieved from TCGA Pan-Cancer Atlas clinical datasets. Samples were included in the survival analy- sis if both molecular data (ENPP1 expression and HRD score) and corresponding clinical survival information were available. Samples lacking either ENPP1 expression, HRD score, or sur- vival data were excluded from further analysis.
After applying these inclusion and exclusion criteria, 9625 samples (patients) with complete molecular and survival infor- mation were retained, corresponding to 97.93% of the initial dataset. Because only 203 samples (2.07%) were excluded due to missing clinical data, the potential bias introduced by data incompleteness was considered minimal. This cohort of 9625 samples constituted the final dataset for all subsequent survival analyses.
A 2-tiered survival analysis was conducted across the pan- cancer cohort. First, in the pan-cancer cohort, Kaplan-Meier survival curves were compared among the established clusters derived from regression-based analysis. Statistical significance was assessed using the log-rank test. Second, Cox proportional hazards were applied to investigate the effect of cluster member- ship on survival. To control for potential confounding factors, tumor type or ENPP1 expression was included as covariates, allowing the evaluation of cluster-specific survival effects inde- pendent of tumor type and ENPP1 expression level.
To identify cancer types that were clinically aligned with the established pan-cancer-derived clusters, Kaplan-Meier survival analyses were performed for each of the 33 individual cancer types, following the same procedures applied to the pan-cancer dataset. Cancer type annotations were not available for 59 sam- ples; therefore, a total of 9566 samples (patients) were included in the cancer type-specific survival analyses. Only cancer types exhibiting statistically significant survival differences among clusters were selected for subsequent molecular characterization using GO and KEGG pathway enrichment analyses (Section 2.4). Because sample sizes varied substantially across cancer types, the purpose of this selection was not to exclude cancer types lacking statistical significance in the current dataset, but rather to highlight those with distinctive survival patterns war- ranting further molecular investigation.
3. Results
3.1. Landscape of ENPP1 expression and HRD score
We explored the landscape of ENPP1 expression and HRD score using a scatter plot (Fig. 1). The Pearson and Spearman correlation coefficients of the 9828 pan-cancer samples were -0.03 and -0.06, respectively (Fig. 1A). When the cancer type-specific distribution of ENPP1 expression and HRD score were delineated using the k-nearest neighbor approach, each
cancer type exhibited a unique territory (Fig. 1B). Nevertheless, mixed and heterogeneous regions were observed, particularly in low- to mid-level ENPP1 expression, and low HRD score range where prostate adenocarcinoma, colon adenocarcinoma, rectum adenocarcinoma, cervical squamous cell carcinoma, and kidney renal clear cell carcinoma were the predominant tumor types. Because the tumor type exerted a strong effect, correla- tions were also analyzed within each tumor type (see Table S2, Supplemental Digital Content, https://links.lww.com/MD/R134, which illustrates the results of correlation analysis between ENPP1 expression and HRD score for each tumor type). Eleven tumor types demonstrated statistically significant correlations in both the Pearson and Spearman correlation coefficients (Table 1 and Fig. 1C-F).
3.2. Linear regression-based clustering of HRD-ENPP1 data
As the relationships between ENPP1 expression and HRD score across pan-cancer samples appeared heterogeneous, we applied linear regression-based clustering to identify multiple plausi- ble linear patterns between ENPP1 expression and HRD score (Fig. 2A). The number of clusters was set to 3 according to the Bayesian Information Criterion (Fig. 2B). Cluster 1 exhibited a positive linear relationship between ENPP1 expression and HRD score, whereas Clusters 2 and 3 showed slightly negative relationships. Generally, the samples in Cluster 3 showed higher ENPP1 expression than those in Cluster 2. The composition of tumor types was also heterogeneous among the 3 clusters (Fig. 2C).
We investigated the uniqueness of the cluster patterns from the individual influences of ENPP1 expression and HRD score, and assessed whether these clusters could be reproduced by conven- tional clustering methods using 3 distinct analytic approaches.
First, in the information-theoretic analysis of the pan-cancer data, the entropy values of the clusters, ENPP1 expression, and HRD score were 0.80, 3.04, and 2.94, respectively, whereas the pairwise mutual information between the clusters and ENPP1 expression, and between the clusters and HRD score were 0.66 and 0.02, respectively. These findings suggest that the cluster assignments capture synergistic information arising from the combined relationship between ENPP1 expression and HRD score, which is not accounted by single-marker linear associations.
Second, despite the fact that the separation between Clusters 2 and 3 was largely driven by ENPP1 expression (with a rel- atively modest contribution from the HRD score), substantial overlap in expression was observed. This overlap highlights the internal heterogeneity with the clusters. Specifically, approx- imately 37.0% (1907/5149) of Cluster 2 samples exhibited ENPP1 expression levels higher than the minimum expression of ENPP1 observed in Cluster 3. Conversely, approximately 10.7% (470/4408) of Cluster 3 samples showed ENPP1 expres- sion lower than the maximum observed in Cluster 2. These quan- titative findings indicate that, despite the overall higher ENPP1 expression in Cluster 3, the distinction between Clusters 2 and 3 cannot be made by ENPP1 expression alone. Additionally Cluster 1 was found to be entirely independent from the associ- ations observed in Clusters 2 and 3.
Finally we compared our linear regression-based clustering results with those obtained using k-means clustering with the number of the clusters fixed at k = 3. The cluster assignments from k-means clustering showed substantial discordance with those from the regression-based clustering (see Table S3, Supplemental Digital Content, https://links.lww.com/MD/R134, which compares linear regression-based clustering results with those of k-means clustering).
Collectively, these findings confirm that the unique molec- ular and clinical characteristics captured by our linear
A
Pan-Cancer
B
Pan-Cancer, Smoothing
ENPP1 expression
15
R=
-0.031
ENPP1 expression
15
10-
10
5
5
0
0
Cancer Type
0
25
50
75
100
0
20
40
60
80
ACC
LUSC
HRD score
HRD score
BLCA
MESO
C
D
BRCA
OV
ACC
MESO
CESC
PAAD
CHOL
PCPG
ENPP1 expression
10
R =- 0.31
ENPP1 expression
10.0
R = 0.22
COAD
PRAD
8
DLBC
READ
7.5
ESCA
SARC
6
GBM
SKCM
4.
5.0
HNSC
STAD
:
KICH
TGCT
0
20
40
0
10
20
30
40
KIRC
THCA
HRD score
HRD score
KIRP
THYM
E
F
LAML
UCEC
UVM
LIHC
LGG
UCS
LIHC
UVM
ENPP1 expression
10.0
R =- 0.32
ENPP1 expression
LUAD
12
R =- 0.27
7.5
…
10
5.0
?
…
8
2.5
6
0
5
10
15
0
20
40
60
HRD score
HRD score
| Cancer type | Pearson correlation | Spearman correlation | ||
|---|---|---|---|---|
| Coefficient | P-value | Coefficient | P-value | |
| LUAD | 0.24 | <. 0001 | 0.23 | <. 0001 |
| TGCT | 0.22 | .0055 | 0.20 | .0107 |
| PAAD | -0.20 | .0120 | -0.22 | .0060 |
| LIHC | -0.27 | <. 0001 | -0.27 | <. 0001 |
| CESC | -0.12 | .0343 | -0.12 | .0360 |
| BRCA | -0.28 | <. 0001 | -0.22 | <. 0001 |
| MESO | 0.22 | .0443 | 0.23 | .0389 |
| STAD | 0.11 | .0305 | 0.14 | .0051 |
| HNSC | 0.11 | .0107 | 0.10 | 0234 |
| ACC | -0.31 | .0057 | -0.33 | .0033 |
| UVM | -0.32 | .0038 | -0.33 | .0027 |
ACC = adrenocortical carcinoma, BRCA = breast cancer, CESC = cervical squamous cell carcinoma, ENPP1 = ectonucleotide pyrophosphatase/phosphodiesterase 1, HNSC = head and neck squamous cell carcinoma, HRD = homologous recombination deficiency, LIHC = liver hepatocellular carcinoma, LUAD = lung adenocarcinoma, MESO = mesothelioma, PAAD = pancreatic adenocarcinoma, STAD = stomach adenocarcinoma, TGCT = testicular germ cell tumors, UVM = uveal melanoma.
regression-based clustering approach cannot be reproduced by single-marker analysis or conventional clustering methods. This highlights the distinctiveness and added value of our integrative strategy.
3.3. DEGs between the clusters
To identify significant molecular features among the clusters, we performed DEG analysis using the limma R package (Fig. 3). Between Clusters 1 and 2, 253 DEGs were identified, including SLC41A2, COL5A1, and FNDC1 (Fig. 3A). Between Clusters 3 and1,859DEGswereidentifiedincludingSLC41A2,MOXD1,and AEBP1 (Fig. 3B). Five DEGs, SLC41A2, HEPH, PDGFRB, and ENPP1, were identified in Clusters 2 and 3 (Fig. 3C). Notably ENPP1 itself was included in all pairwise comparisons, suggest- ing that cluster membership was highly dependent on ENPP1 expression.
3.4. Functional profiling using GO and KEGG enrichment test
We investigated the functional characteristics of the DEGs between the clusters using GO term and KEGG pathway analyses.
A
Linear Regression-Based Clustering
15-
Cluster
1
ENPP1 expression
2
10
3
Model
5.
Best
Others
0
Type
ACC
LUSC
0
25
50
75
100
BLCA
MESO
HRD score
BRCA
OV
B
C
CESC
PAAD
Selection of Number of Clusters
Distribution of Cancer Types
CHOL
PCPG
1.00
COAD
PRAD
DLBC
READ
ESCA
SARC
44700-
0.75
GBM
SKCM
HNSC
STAD
Value
KICH
TGCT
BIC
0.50
KIRC
THCA
KIRP
THYM
44650
LAML
UCEC
0.25
LGG
UCS
LIHC
UVM
LUAD
44600
0.00
1
2
3
4
5
6
total
1
2
3
Number of clusters (k)
Cluster
In the GO biologic process category, enriched terms between Clusters 1 and 2 included extracellular matrix (ECM) organi- zation, extracellular structure organization, and regulation of leukocyte migration (Fig. 4A; see Table S4, Supplemental Digital Content, https://links.lww.com/MD/R134, which illustrates the results of GO term enrichment analysis between Clusters 1 and 2). Between Clusters 1 and 3, the enriched terms included ECM organization, extracellular structure organization, and renal tissue development (Fig. 4B; see Table S5, Supplemental Digital Content, https://links.lww.com/MD/R134, which illustrates the results of GO term enrichment analysis between Clusters 1 and 3). Between Clusters 2 and 3, the enriched terms included symbiont adhesion to a host cell, negative regulation of gly- cogen biosynthetic process, and ECM organization (see Table S6, Supplemental Digital Content, https://links.lww.com/MD/ R134, which illustrate the results of GO term enrichment anal- ysis between Clusters 2 and 3). Collectively, the GO enrichment analysis result suggested that Clusters 2 and 3 were functionally more similar than to Cluster 1.
In the KEGG pathway enrichment analysis, enriched path- ways between Clusters 1 and 2 included ECM-receptor inter- action, proteoglycans in cancer, and viral protein interaction with cytokine and cytokine receptor (Fig. 4C; see Table S7,
Supplemental Digital Content, https://links.lww.com/MD/ R134 which illustrates the results of KEGG pathway enrich- ment analysis between Clusters 1 and 2). Between Clusters 1 and 3, enriched pathways included ECM-receptor interac- tion, proteoglycans in cancer, and calcium signaling pathway (Fig. 4D; see Table S8, Supplemental Digital Content, https:// links.lww.com/MD/R134 which illustrates the results of KEGG pathway enrichment analysis between Clusters 1 and 3). Between Clusters 2 and 3, enriched pathways included pro- tein digestion and absorption, starch and sucrose metabolism, and nicotinate and nicotinamide metabolism (Fig. 4E; see Table S9, Supplemental Digital Content, https://links.lww.com/MD/ R134 which illustrates the results of KEGG pathway enrich- ment analysis between Clusters 2 and 3). Consistent with the GO enrichment analysis, the KEGG pathway analysis sug- gested that Clusters 2 and 3 were functionally more similar to each other than to Cluster 1. However, they diverged in their dominant programs: Cluster 3 was enriched for multiple meta- bolic processes including glycogen and nucleotide metabolism, whereas Cluster 2 showed enrichment in the ECM and stro- mal organization, as well as host-symbiont interaction-related terms, which may reflect microenvironmental and defense- related biology.
A
Cluster 1 vs Cluster 2
B
Cluster 1 vs Cluster 3
C Cluster 2 vs Cluster 3
60
SLC41A2
SLC41A2
30
SLC41A2
ĻAMA4
100
COLP5A1
9
MOXD1
SFRP4
60
MYADM
HEPH8
HEPH
COL15A1
CHN1
MMD
SULF1
-Log 10 P
FBXL7
40
CFAP
ADAM12
20
MMP11
-Log 10 P
ECM2
-Log 10 P
PDGFRB.
C7prf58
GLT8D2
WISP1
TMEM176A
40
GTHRC1
TRPC4
·MSG
ITGAM
PRUNE2
FARPT
8000
SYNPO2
20
RHEBL 1
LTBP
GD3028
10
LRRG33
C10otf99
MFAP5
FADS3
AQP9
C10011990
KCNMA1
20
ETV5
CYST
ADAM6
POPDC2
CG119
LOA5
AMPID2
RNF150
L24
FDXR
GADD45B
0
0
0
-2
0
2
-2
0
2
4
-2
-1
0
1
2
Log2 fold change
Log2 fold change
Log2 fold change
NS
Log2 FC
p-value
p -value and log2 FC
3.5. Survival analysis
We assessed the clinical significance of the 3 clusters, which reflects the interaction between ENPP1 expression and HRD score, by performing survival analysis using the Kaplan-Meier curves and Cox regression models (Fig. 5A). In the pan-cancer analysis, patients in Cluster 3 demonstrated significantly bet- ter survival than those in Clusters 1 and 2 (P <. 0001). In the Cox regression models, cluster membership remained partially significant after adjusting for tumor type. Using Cluster 2 as a reference, Cluster 1 was associated with improved survival (HR = 0.63, P = . 004), whereas Cluster 3 was no longer statis- tically significant (HR = 1.09, P = . 07). Since ENPP1 expression has been reported to be a predictive factor in many tumor types, we included it as a covariate. Cluster membership remained sig- nificant after adjusting for ENPP1 expression. Specifically, com- pared with Cluster 2, Cluster 1 was associated with a reduced mortality risk (HR = 0.74, P = . 027), and Cluster 3 also showed significantly lower risk (HR = 0.87, P = . 007).
We further investigated the tumor types that showed signif- icant differences in survival according to cluster membership. Four tumor types: adrenocortical carcinoma (ACC; Fig. 5B), kidney chromophobe renal cell carcinoma (KICH; Fig. 5C), low grade glioma (LGG; see Figure S1, Supplemental Digital Content, https://links.lww.com/MD/R135, which illustrates survival analysis between clusters in LGG), and mesothelioma (MESO; Fig. 5D), exhibited significant differences in Kaplan- Meier curve analyses (P =. 013, P =. 0029, P <. 0001, and P = . 005, respectively).
3.6. Functional profiling of individual tumor types
To explore tumor type-specific molecular features associated with cluster membership, we performed GO term enrichment and KEGG pathway analyses for tumor types that showed sig- nificant survival differences (ACC, KICH, LGG, and MESO) using the same approach as in the pan-cancer analysis (Fig. 6).
Because of the extremely small number of Cluster 1 samples in these tumor types, comparisons were restricted to a com- parison between Clusters 2 and 3. Functional enrichment anal- ysis revealed distinct tumor type-specific patterns and shared processes. ACC was characterized by enrichment of pathways related to metabolic processes, immune responses, and hor- monal responses (Fig. 6A; see Table S10, Supplemental Digital Content, https://links.lww.com/MD/R134, which illustrates the results of GO term enrichment analysis between Clusters 2 and 3 in ACC, and Table S11, Supplemental Digital Content, https://links.lww.com/MD/R134, which illustrates the results of KEGG pathway enrichment analysis between Clusters 2 and 3 in ACC). In KICH, enriched pathways were predomi- nantly associated with carbohydrate and nucleotide metabo- lism, insulin signaling, and energy homeostasis (see Table S12, Supplemental Digital Content, https://links.lww.com/MD/ R134, which illustrates the results of GO term enrichment analysis between Clusters 2 and 3 in KICH, and Table S13, Supplemental Digital Content, https://links.lww.com/MD/ R134, which illustrates the results of KEGG pathway enrich- ment analysis between Clusters 2 and 3 in KICH). LGG was enriched for immune responses, ECM regulation, and neuron/ synapse programs (Fig. 6B and C; see Table S14, Supplemental Digital Content, https://links.lww.com/MD/R134, which illus- trates the results of GO term enrichment analysis between Clusters 2 and 3 in LGG, and Table S15, Supplemental Digital Content, https://links.lww.com/MD/R134, which illustrates the results of KEGG pathway enrichment analysis between Clusters 2 and 3 in LGG). MESO demonstrated enrichment in mesenchyme development, ECM organization and response, and the transforming growth factor-beta signaling pathway (Fig. 6D; see Table S16, Supplemental Digital Content, https:// links.lww.com/MD/R134, which illustrates the results of GO term enrichment analysis between Clusters 2 and 3 in MESO, and Table S17, Supplemental Digital Content, https://links.lww. com/MD/R134, which illustrates the results of KEGG pathway enrichment analysis between Clusters 2 and 3 in MESO).
A
GO, Cluster 2 vs Cluster 1
GO, Cluster 3 vs Cluster 1
KEGG, Cluster 2 vs Cluster 1
extracellular matrix organization
B
extracellular matrix organization
C
Cytoskeleton in muscle cells
ossification
muscle tissue development
Protein digestion and
absorption
cell-substrate adhesion
cell-substrate adhesion
ECM-receptor interaction
negative regulation of cellular response to growth
Malaria
factor stimulus
ossification
positive regulation of cell
adhesion
renal system development
Focal adhesion
collagen metabolic process
regulation of cellular response to growth factor
PI3K-Akt signaling pathway
stimulus
bone mineralization
muscle cell differentiation
Proteoglycans in cancer
regulation of vasculature
development
bone development
Hypertrophic cardiomyopathy
osteoblast differentiation
connective tissue development
Dilated cardiomyopathy
bone development
cartilage development
Human papillomavirus infection
0
10
20
30
40
50
0
25
50
75
0
10
20
30
Count
Count
Count
D
KEGG, Cluster 3 vs Cluster 1
E
KEGG, Cluster 2 vs Cluster 3
Cytoskeleton in muscle cells
Protein digestion and absorption
ECM-receptor interaction
Protein digestion and
Pantothenate and CoA biosynthesis
absorption
Hypertrophic cardiomyopathy
Nicotinate and nicotinamide metabolism
p.adjust
0.01 0.02 0.03 0.04
Focal adhesion
Starch and sucrose metabolism
Dilated cardiomyopathy
PI3K-Akt signaling pathway
Pyrimidine metabolism
Malaria
Nucleotide metabolism
Arrhythmogenic right
ventricular cardiomyopathy
AGE-RAGE signaling pathway
Purine metabolism
in diabetic complications
0
20
40
60
0.0
0.5
1.0
1.5
2.0
Count
Count
4. Discussion
In this study, we comprehensively investigated the relationship between ENPP1 expression and HRD scores, and their molec- ular and clinical significance in pan-cancer samples. To over- come the intrinsic heterogeneity of ENPP1 expression and HRD score, we applied a linear regression-based clustering approach that delineated 3 distinct and biologically plausible clusters. Molecular functional profiling revealed that these clusters, iden- tified from the pan-cancer samples, exhibited both cancer type nonspecific global characteristics and tumor type-specific fea- tures, thereby providing a framework for understanding how ENPP1 and HRD jointly shape tumor biology. This integra- tive analysis highlights the potential avenues for refined pre- cision medicine strategies beyond single-gene or single-feature assessments.
In the pan-cancer samples, Cluster 1 was characterized by tumor cell-intrinsic proliferation programs, reflected by enrich- ment of GO terms or the KEGG pathways closely associated with the cell cycle, DNA replication, ribonucleic acid process- ing, oxidative phosphorylation, and mitochondrial translation. Cluster 2 exhibited signatures of ECM remodeling, development and differentiation programs, and immune responses. Finally, Cluster 3 was distinguished by tumor metabolic reprogram- ming, specifically enriched in glycogen biosynthetic processes, glucose import regulation, phosphate ion homeostasis, nucleo- side triphosphate catabolic process, and triglyceride storage.
Although Cluster 1 could not be sufficiently characterized at the individual tumor type level owing to the extremely small number of samples, global molecular features of Clusters 2 and 3 were partly recapitulated in tumor type-specific analyses. In ACC, Clusters 2 and 3 diverged in pathways related to meta- bolic processes, immune responses, and hormonal regulation. While the enrichment of metabolic processes in Cluster 3 and immune response in Cluster 2 was largely consistent with the pan-cancer characteristics of these clusters, additional enrich- ment of hormonal regulation appeared to be tumor type- specific, suggesting that adrenal-specific metabolic and endocrine programs may uniquely manifest in ACC. Indeed, Wang et al recently reported that ACC exhibits pronounced metabolic heterogeneity, particularly in pathways such as the pentose phosphate and galactose metabolism, and that higher intratumoral metabolic variability correlated with poor survival outcomes.[45] Moreover, cortisol-secreting ACC has been shown to exert systemic immunosuppression, further linking endocrine activity to immune escape and progrnosis.[46-48] These findings further suggest that the observed cluster distinctions derived from ENPP1 expression and HRD score capture adrenal-specific tumor biology and may serve as clinically relevant biomarkers.
In KICH, Cluster 2 was preferentially enriched in immune- related pathways, whereas Cluster 3 was associated with meta- bolic processes, reflecting the global pan-cancer features of these clusters. These results suggest that tumor type-specific features
| 0 | 3000 | 6000 Time | 9000 | 12000 | |
|---|---|---|---|---|---|
| Number at risk | |||||
| Cluster 1 | 251 | 13 | 2 | 1 | 0 |
| Strata Cluster 2 | 5043 | 237 | 23 | 1 | 0 |
| Cluster 3 | 4331 | 290 | 31 | 5 | 0 |
| 0 | 3000 | 6000 | 9000 | 12000 | |
| Time | |||||
A
Pan-Cancer
B
ACC
1.00
1.00
Survival probability
0.75
Survival probability
0.75
0.50
0.50
0.25
0.25
p < 0.0001
p = 0.013
0.00
0.00
0
1000
2000
3000
4000
5000
Time
Number at risk
Strata
Cluster 2
56
30
13
4
1
0
Cluster 3
21
15
8
3
1
0
0
1000
2000
3000
4000
5000
Time
C
KICH
D
MESO
1.00
1.00
Survival probability
0.75
Survival probability
0.75
0.50
0.50
0.25
p = 0.0029
0.25
p = 0.005
0.00
0.00
0
1000
2000
3000
4000
5000
0
1000
2000
3000
Time
Time
Number at risk
Number at risk
Strata
Cluster 1
7
5
4
4
1
0
Cluster 1
1
0
0
0
Cluster 2
Strata
46
35
25
13
2
0
Cluster 2
40
11
4
0
Cluster 3
12
7
6
3
0
0
Cluster 3
39
2
0
0
0
1000
2000
3000
4000
5000
0
1000
2000
3000
Time
Time
between the clusters may further emerge once a sufficient num- ber of Cluster 1 samples become available in KICH, and more comprehensive pairwise comparisons can be made. Despite the lack of Cluster 1 samples, the distinct features of Clusters 2 and 3 in KICH appear robust. A recent integrative study iden- tified cell-cycle regulator CENPE and glycolytic enzyme LDHA as prognostic biomarkers in KICH, with LDHA expression strongly linked to metabolic rewiring and worse overall survival (HR = 1.10, P <. 0001).[49] These findings are consistent with our observation that metabolic dominance in Cluster 3 is asso- ciated with a poorer prognosis relative to the immune/ECM- enriched Cluster 2.
In the LGG, Cluster 2 was enriched in immune and inflamma- tory signaling, whereas Cluster 3 was enriched in neuronal and synaptic processes. These divergences suggest that the global pan-cancer immune/ECM-metabolic pattern is modified into a more brain tissue-specific representation of the cluster features,
reflecting the unique neural tumor-microenvironment context of LGG. Previous studies have demonstrated that neuronal sig- naling pathways can contribute to glioma progression through neuron-glioma synaptic interactions.[50] Thus, the balance between immune suppression and neural activity may underlie the distinct survival differences between clusters in LGG.
Finally, in MESO, Cluster 2 was enriched for immune acti- vation, whereas Cluster 3 displayed signatures linked to cell adhesion and ECM remodeling. These results provide another example of the modification of global to tumor type-specific cluster features, reflecting the stromal-rich biology of meso- thelioma, where ECM and transforming growth factor-beta pathway activation have been consistently implicated in tumor aggressiveness and immune exclusion.[51-53]
Although.Cluster 3 was associated with favorable sur- vival outcomes in the pan-cancer analysis, tumor type-specific analyses (ACC, KICH, LGG, and MESO) revealed divergent
A
GO in ACC
B
GO in LGG
cellular glucuronidation
synapse organization
flavonoid metabolic process
extracellular matrix
organization
retinoic acid metabolic
extracellular structure
process
organization
xenobiotic metabolic process
regulation of trans-synaptic signaling
cell activation involved in
immune response
positive regulation of cell
adhesion
positive regulation of cytokine production
chemotaxis
pigment metabolic process
leukocyte migration-
retinoid metabolic process
leukocyte cell-cell adhesion
positive regulation of cell
cell-cell adhesion via
adhesion
plasma-membrane adhesion
regulation of immune effector process
molecules
positive regulation of cell
activation
0
5
10
0
50
100
150
Count
Count
C
D
KEGG in LGG
GO in MESO
Neuroactive ligand-receptor
external encapsulating structure organization
interaction
JAK-STAT signaling pathway
endoderm formation
Influenza A
ossification
Toll-like receptor signaling
pathway
endoderm development
Autoimmune thyroid disease
cartilage development
Hormone signaling
connective tissue development
Epstein-Barr virus infection
axonogenesis
Cytokine-cytokine receptor
interaction
odontogenesis
MAPK signaling pathway
mesenchyme development
Chagas disease
regulation of neuron
projection development
0
20
40
60
0
10
20
30
40
Count
Count
p.adjust
0.01
0.02
0.03
0.04
patterns. For example, in ACC, patients in Cluster 3 showed superior survival compared with those in Cluster 2 (P = . 013), whereas the opposite trend were observed in KICH (P = . 003), LGG (P <. 0001), and MESO (P =. 005). This discrepancy may reflect tissue-specific modification of global clusters features, in which even subtle change may be sufficient to reverse survival outcomes. Furthermore, the heterogeneous distribution of clus- ter membership across tumor types may have contributed to these divergent survival patterns. These findings underscore that the biological and clinical implications of ENPP1-HRD-defined clusters are highly context-dependent and shaped by both global cancer hallmarks and tissue-specific programs.
This study has several limitations. First, linear regression- based clustering, which we used as the core strategy to resolve the heterogeneity of ENPP1 expression and HRD score, does not guarantee fully consistent cluster assignments because of
the stochastic nature of the algorithm. Consequently, the exact cluster membership of the individual samples may not be repro- ducible. Nevertheless, despite such stochastic variations, the overall clustering patterns were predominantly preserved, and only a small fraction of the samples showed inconsistent mem- bership across iterations. Further methodological advances are required to develop clustering algorithms that provide more stable and reproducible cluster construction. Second, owing to the inherent heterogeneity across tumor types and wide varia- tion in sample size among them, a small subset of tumor types with relatively large cohorts and strong influence may have disproportionately driven the clustering patterns. This imbal- ance could reduce the generalizability of our findings across all tumor types, limiting the degree to which the clusters repre- sent pan-cancer features rather than being dominated by high- impact tumor types. These issues may be partly mitigated if our
approach is applied to groups of tumors that are more homo- geneous in terms of sample size and underlying tumor biology. This refinement is a valuable direction for future studies.
Further investigation is required to translate our con- ceptual framework into clinically actional predictive or therapeutic biomarkers. These efforts include multi-omics characterization, evaluation of temporo-spatial tumor dynam- ics, and validation in clinically compatible materials such as formalin-fixed paraffin-embedded tissues. If ENPP1 expres- sion can be reliably quantified as RT-qPCR ACt values from formalin-fixed paraffin-embedded samples and integrated with the established HRD score (already implemented clini- cally in breast and ovarian cancers) our 3-cluster framework could be feasibly adapted for precision oncology applications. The relatively low levels of intratumoral spatial heterogeneity observed for HRD scores further supports their potential clin- ical applicability.[54]
The divergence of the 3 clusters (representing proliferation, ECM-immune, metabolic reprogramming phenotypes) also suggests differential cluster-specific responses for cytotoxic agents, immunotherapies,[12] and metabolism-targeted drugs.[55] Identification of reliable biomarkers reflecting these cluster- specific biological programs could enable rational design of clin- ical trials tailored to each subgroups. Importantly, predictive or therapeutic biomarkers are not expected to be limited to ENPP1 or HRD themselves, but may include additional molecules that represent the broader molecular network underlying the 3- cluster framework.
In conclusion, we comprehensively analyzed the rela- tionship between ENPP1 expression and HRD score across pan-cancer samples from 33 tumor types, which, to our knowledge, represents the first report on a pan-cancer scale. Furthermore, we addressed the intrinsic heterogeneity of this relationship by applying a linear regression-based clustering approach, which delineated 3 biologically and clinically dis- tinct clusters. Functional enrichment analyses revealed global pan-cancer characteristics and tumor type-specific modifi- cations in these clusters, highlighting the context-dependent interplay between DNA repair and innate immune regula- tion. The interaction between ENPP1 expression and HRD score appeared dynamic, context-dependent, and integrated into core cancer hallmarks rather than being fixed or isolated. Importantly, cluster membership was associated with distinct survival outcomes in both pan-cancer and individual tumor type analyses, underscoring their potential prognostic and bio- logical relevance. Taken together, our findings provide novel insights into the ENPP1-HRD axis and suggest that these clusters may serve as a foundation for refining biomarker- driven precision medicine strategies across diverse tumor types.
Author contributions
Conceptualization: Young Soo Song, Yong Min Kim, Sung Hak Lee, Woong Na, Oyeon Jo.
Data curation: Yong Min Kim, Sung Hak Lee, Oyeon Jo.
Formal analysis: Young Soo Song, Sung Hak Lee, Woong Na, Il Ju Lee.
Funding acquisition: Young Soo Song.
Investigation: Mihye Kwon, Oyeon Jo.
Methodology: Young Soo Song, Yong Min Kim, Sung Hak Lee, Woong Na, Il Ju Lee.
Project administration: Yong Min Kim, Il Ju Lee, Oyeon Jo.
Resources: Sung Hak Lee.
Supervision: Young Soo Song, Sung Hak Lee, Mihye Kwon.
Validation: Young Soo Song, Yong Min Kim, Sung Hak Lee, Oyeon Jo.
Visualization: Young Soo Song, Yong Min Kim, Il Ju Lee, Mihye Kwon.
Writing - original draft: Young Soo Song, Yong Min Kim, Sung Hak Lee, Woong Na, Il Ju Lee, Mihye Kwon, Oyeon Jo. Writing - review & editing: Young Soo Song, Yong Min Kim, Sung Hak Lee, Woong Na, Il Ju Lee, Mihye Kwon, Oyeon Jo.
References
[1] Cogan D, Bakhoum SF. Re-awakening innate immune signaling in can- cer: the development of highly potent ENPP1 inhibitors. Cell Chem Biol. 2020;27:1327-8.
[2] Gangar M, Goyal S, Raykar D, et al. Design, synthesis and biological evaluation studies of novel small molecule ENPP1 inhibitors for cancer immunotherapy. Bioorg Chem. 2022;119:105549.
[3] Gao S, Hou Y, Xu Y, et al. Discovery of orally bioavailable phosphonate prodrugs of potent ENPP1 inhibitors for cancer treatment. Eur J Med Chem. 2024;279:116853.
[4] He J, Ma X, Sun J, et al. Design, synthesis, and pharmacological evalu- ation of quinazoline and quinoline derivatives as potent ENPP1 inhibi- tors for cancer immunotherapy. J Med Chem. 2025;68:5856-73.
[5] Huang R, Ning Q, Zhao J, et al. Targeting ENPP1 for cancer immu- notherapy: killing two birds with one stone. Biochem Pharmacol. 2024;220:116006.
[6] Johnson RM, Wang S, Carozza JA, et al. ENPP1 inhibitor with ultralong drug-target residence time as an innate immune checkpoint blockade cancer therapy. bioRxiv. 2025. doi: 10.1101/2025.05.18.654655.
[7] Ruiz-Fernandez de Cordoba B, Moreno H, Valencia K, et al. Tumor ENPP1 (CD203a)/haptoglobin axis exploits myeloid-derived suppres- sor cells to promote post-radiotherapy local recurrence in breast cancer. Cancer Discov. 2022;12:1356-77.
[8] Takahashi RU, Miyazaki H, Takeshita F, et al. Loss of microRNA-27b contributes to breast cancer stem cell generation by activating ENPP1. Nat Commun. 2015;6:7318.
[9] Wang S, Bohnert V, Joseph AJ, et al. ENPP1 is an innate immune check- point of the anticancer cGAMP-STING pathway in breast cancer. Proc Natl Acad Sci U S A. 2023;120:e2313693120.
[10] You C, Fang Q, Xiao X, et al. ENPP1 promotes immune suppression, drug resistance, and adverse outcomes in bladder cancer: potential for targeted therapy. Cancer Genet. 2025;294-295:1-14.
[11] Zhao L, Zhang Y, Tian Y, et al. Role of ENPP1 in cancer pathogenesis: mechanisms and clinical implications (review). Oncol Lett. 2024;28:590.
[12] Pu C, Cui H, Yu H, et al. Oral ENPP1 inhibitor designed using gener- ative AI as next generation STING modulator for solid tumors. Nat Commun. 2025;16:4793.
[13] Chen KT, Madison R, Moore J, et al. A novel HRD signature is pre- dictive of FOLFIRINOX benefit in metastatic pancreatic cancer. Oncologist. 2023;28:691-8.
[14] Cruz GI, Kane KL, Matsuno RK, et al. Treatment, outcomes, and resource utilization among patients with metastatic breast and advanced epithelial ovarian cancer, by BRCA1/2 and HRD status. Oncologist. 2025;30:oyaf100.
[15] Daly GR, Naidoo S, Alabdulrahman M, et al. Screening and testing for homologous recombination repair deficiency (HRD) in breast can- cer: an overview of the current global landscape. Curr Oncol Rep. 2024;26:890-903.
[16] Fasching PA, Schmatloch S, Hauke J, et al. Neoadjuvant paclitaxel/ olaparib in comparison to paclitaxel/carboplatin in patients with HER2-negative breast cancer and HRD-long-term survival of the GeparOLA study. Clin Cancer Res. 2025;31:1596-604.
[17] Guffanti F, Mengoli I, Damia G. Current HRD assays in ovarian cancer: differences, pitfalls, limitations, and novel approaches. Front Oncol. 2024;14:1405361.
[18] Kekeeva T, Andreeva Y, Tanas A, et al. HRD testing of ovarian cancer in routine practice: what are we dealing with? Int J Mol Sci. 2023;24:10497.
[19] Kekeeva T, Dudina I, Andreeva Y, et al. Molecular subgroups of HRD positive ovarian cancer and their prognostic significance. Int J Mol Sci. 2024;25:13549.
[20] Leman R, Cherifi F, Leheurteur M, et al. Homologous recombination deficiency (HRD) tests for ovarian cancer: a multicenter French phase II study (HERO). BMC Cancer. 2025;25:1075.
[21] Rempel E, Kluck K, Beck S, et al. Pan-cancer analysis of genomic scar patterns caused by homologous repair deficiency (HRD). NPJ Precis Oncol. 2022;6:36.
[22] Richardson DL, Quintanilha JCF, Danziger N, et al. Correction: effec- tiveness of PARP inhibitor maintenance therapy in ovarian cancer by BRCA1/2 and a scar-based HRD signature in real-world practice. Clin Cancer Res. 2025;31:3599.
[23] Telli ML, Timms KM, Reid J, et al. Homologous recombination defi- ciency (HRD) score predicts response to platinum-containing neoadju- vant chemotherapy in patients with triple-negative breast cancer. Clin Cancer Res. 2016;22:3764-73.
[24] Kohn EC, Lee JM, Ivy SP. The HRD decision-which PARP inhibitor to use for whom and when. Clin Cancer Res. 2017;23:7155-7.
[25] Skelin M, Sarcevic D, Lesin Gacina D, Mucalo I, Dilber I, Javor E. The effect of PARP inhibitors in homologous recombination proficient ovarian cancer: meta-analysis. J Chemother. 2023;35:150-7.
[26] Li J, Duran MA, Dhanota N, et al. Metastasis and immune evasion from extracellular cGAMP hydrolysis. Cancer Discov. 2021;11:1212-27.
[27] Liang W, Zhou C, Wang J, et al. A prognostic signature based on adenosine metabolism related genes for ovarian cancer. Front Oncol. 2022;12:1003512.
[28] Onyedibe KI, Wang M, Sintim HO. ENPP1, an old enzyme with new functions, and small molecule inhibitors-A STING in the tale of ENPP1. Molecules. 2019;24:4192.
[29] Knijnenburg TA, Wang L, Zimmermann MT, et al. Genomic and molec- ular landscape of DNA damage repair deficiency across the cancer genome atlas. Cell Rep. 2018;23:239-54.e6.
[30] Xu Q, Wen Y, Huang T, et al. The distinct landscape of tumor immune microenvironment in homologous recombination deficient cancers. Biomark Res. 2025;13:108.
[31] Yang D, Huang FX, Wei W, et al. Loss of HRD functional phenotype impedes immunotherapy and can be reversed by HDAC inhibitor in ovarian cancer. Int J Biol Sci. 2023;19:1846-60.
[32] Cancer Genome Atlas Research Network. Comprehensive genomic characterization defines human glioblastoma genes and core pathways. Nature. 2008;455:1061-8.
[33] Goldman MJ, Craft B, Hastie M, et al. Visualizing and interpreting cancer genomics data via the Xena platform. Nat Biotechnol. 2020;38:675-8.
[34] Zhang Z. Introduction to machine learning: k-nearest neighbors. Ann Transl Med. 2016;4:218.
[35] Wang S, Huang GH, He L. Development of a clusterwise-linear- regression-based forecasting system for characterizing DNAPL dissolu- tion behaviors in porous media. Sci Total Environ. 2012;433:141-50.
[36] Zamaninasab Z, Najafipour H, Mirzaee M, Bahrampour A. A cluster-wise linear regression model to investigate the effect of demo- graphical and clinical variables on the average depression score. Med J Islam Repub Iran. 2022;36:116.
[37] Grün B, Leisch F. Fitting finite mixtures of generalized linear regres- sions in R. Comput Stat Data Anal. 2007;51:5247-52.
[38] Botia JA, Vandrovcova J, Forabosco P, et al. An additional k-means clustering step improves the biological features of WGCNA gene co- expression networks. BMC Syst Biol. 2017;11:47.
[39] Meyer PE. Infotheo: Information-Theoretic Measures [Computer Program]. Comprehensive R Archive Network (CRAN); 2023. https:// CRAN.R-project.org/package=infotheo.
[40] Na W, Yi K, Song YS, Park MH. Dissecting the relationships of IgG sub- classes and complements in membranous lupus nephritis and idiopathic membranous nephropathy. PLoS One. 2017;12:e0174501.
[41] Ritchie ME, Phipson B, Wu D, et al. Limma powers differential expres- sion analyses for RNA-sequencing and microarray studies. Nucleic Acids Res. 2015;43:e47.
[42] Xu S, Hu E, Cai Y, et al. Using clusterProfiler to characterize multiomics data. Nat Protoc. 2024;19:3292-320.
[43] Kassambara A, Kosinski M, Biecek P. Survminer: Drawing Survival Curves Using ggplot2 [Computer Program]. 2024. Comprehensive R Archive Network (CRAN); 2024. https://CRAN.R-project.org/ package=survminer.
[44] Therneau TM, Lumley T. Survival: Survival Analysis [Computer Program]. Comprehensive R Archive Network (CRAN); 2024. https:// CRAN.R-project.org/package=survival.
[45] Wang Q, Sun N, Meixner R, et al. Metabolic heterogeneity in adrenocortical carcinoma impacts patient outcomes. JCI Insight. 2023;8:e167007.
[46] Fassnacht M, Libe R, Kroiss M, Allolio B. Adrenocortical carcinoma: a clinician’s update. Nat Rev Endocrinol. 2011;7:323-35.
[47] Peppa M, Pikounis V, Papaxoinis G, et al. Adrenocortical carcinoma secreting cortisol, androgens and aldosterone: a case report. Cases J. 2009;2:8951.
[48] Shariq OA, Mckenzie TJ. Adrenocortical carcinoma: current state of the art, ongoing controversies, and future directions in diagnosis and treatment. Ther Adv Chronic Dis. 2021;12:20406223211033103.
[49] Wu HF, Liu H, Zhang ZW, Chen JM. CENPE and LDHA were poten- tial prognostic biomarkers of chromophobe renal cell carcinoma. Eur J Med Res. 2023;28:481.
[50] Venkatesh HS, Morishita W, Geraghty AC, et al. Electrical and synaptic integration of glioma into neural circuits. Nature. 2019;573:539-45.
[51] Fujii M, Toyoda T, Nakanishi H, et al. TGF-beta synergizes with defects in the Hippo pathway to stimulate human malignant mesothelioma growth. J Exp Med. 2012;209:479-94.
[52] Nel AE, Pavlisko EN, Roggli VL. The interplay between the immune system, tumor suppressor genes, and immune senescence in mesothe- lioma development and response to immunotherapy. J Thorac Oncol. 2024;19:551-64.
[53] Wen G, Hong M, Li B, et al. Transforming growth factor-beta-induced protein (TGFBI) suppresses mesothelioma progression through the Akt/mTOR pathway. Int J Oncol. 2011;39:1001-9.
[54] von Wahlde MK, Timms KM, Chagpar A, et al. Intratumor heterogene- ity of homologous recombination deficiency in primary breast cancer. Clin Cancer Res. 2017;23:1193-9.
[55] Furukawa T, Tabata S, Minami K, Yamamoto M, Kawahara K, Tanimoto A. Metabolic reprograming of cancer as a therapeutic target. Biochim Biophys Acta Gen Subj. 2023;1867:130301.