Artificial intelligence in ovarian cancer drug resistance advanced 3PM approach: subtype classification and prognostic modeling

Abstract

Background

Ovarian cancer patients’ resistance to first-line treatment posed a significant challenge, with approximately 70% experiencing recurrence and developing strong resistance to first-line chemotherapies like paclitaxel.

Objectives

Within the framework of predictive, preventive, and personalized medicine (3PM), this study aimed to use artificial intelligence to find drug resistance characteristics at the single cell, and further construct the classification strategy and deep learning prognostic models based on these resistance traits, which can better facilitate and perform 3PM.

Methods

This study employed “Beyondcell,” an algorithm capable of predicting cellular drug responses, to calculate the similarity between the expression patterns of 21,937 cells from ovarian cancer samples and the signatures of 5201 drugs to identify drug-resistance cells. Drug resistance features were used to perform 10 multi-omics clustering on the TCGA training set to identify patient subgroups with differential drug responses. Concurrently, a deep learning prognostic model with KAN architecture which had a flexible activation function to better fit the model was constructed for this training set. The constructed patient subtype classifier and prognostic model were evaluated using three external validation sets from GEO: GSE17260, GSE26712, and GSE51088.

Results

This study identified that endothelial cells are resistant to paclitaxel, doxorubicin, and docetaxel, suggesting their potential as targets for cellular therapy in ovarian cancer patients. Based on drug resistance features, 10 multi-omics clustering identified four patient subtypes with differential responses to four chemotherapy drugs, in which subtype CS2 showed the highest drug sensitivity to all four drugs. The other subtypes also showed enrichment in different biological pathways and immune infiltration, allowing for targeted treatment based on their characteristics. Besides, this study applied the latest KAN architecture in artificial intelligence to replace the MLP structure in the DeepSurv prognostic model, finally demonstrating robust performance on patients’ prognosis prediction.

Conclusions

This study, by classifying patients and constructing prognostic models based on resistance characteristics to first-line drugs, has effectively applied multi-omics data into the realm of 3PM.

Supplementary Information

The online version contains supplementary material available at 10.1007/s13167-024-00374-4.

Introduction

Importance of drug resistance in ovarian cancer

Ovarian cancer is one of the three most common cancers of the female reproductive system. In 2020, there were 313,959 new cases of ovarian cancer globally, with 207,252 deaths that year [1]. Due to the deep development of ovarian cancer within the pelvic cavity and limited early clinical manifestations, most ovarian cancer patients are diagnosed only after significant disease progression. Moreover, because of its high recurrence rate after resection and widespread chemoresistance [2], the mortality rate of ovarian cancer is the highest among all malignant gynecological tumors [3]. Chemoresistance is a major challenge currently faced in the treatment of ovarian cancer [4]. The underlying mechanisms of chemoresistance in ovarian cancer are multifaceted, with cancer stem cells, non-coding RNAs, autophagy, DNA repair deficiencies, hypoxic conditions, and other alterations in the tumor microenvironment all contributing factor [5–10]. Although most patients initially respond to chemotherapy, about 70–80% of tumors recur and develop drug resistance [2, 11–14]. Therefore, from the perspective of traditional chemotherapy, there is an urgent need to explore new treatment strategies to combat drug resistance. Drug resistance resulted in the paradigm change from reactive medical services to a predictive approach, targeted prevention, and treatments tailored to the person (predictive, preventive, and personalized medicine, shortly PPPM/3PM) [15].

The role of artificial intelligence in cancer

Ongoing 3PM research toward cancer groups in the population promoted by the European Association for Predictive, Preventive and Personalised Medicine (EPMA) demonstrated artificial intelligence-based predictive models can be used for stratifying cancer patients, predicting diagnoses, assessing prognosis, and targeting medical services, achieving promising results in clinical diagnosis and treatment for patients as reported here by the EPMA expert group [15–18]. Over the past few decades, artificial intelligence has rapidly developed in the field of cancer research [19, 20]. AI-based biological analysis not only efficiently processed high-throughput, heterogeneous, and complex molecular data but also uncovered features or relationships within biological networks, enabling precision cancer treatment [21–23]. The BeyondCell algorithm was a computational framework based on single-cell RNA sequencing (scRNA-seq) data that predicted drug sensitivity by constructing a detailed interaction network of tumor cells. The core principle of this algorithm was to leverage the heterogeneity of gene expression at the single-cell level, identifying cell subpopulations related to drug response through machine learning models. Compared to traditional tissue-level analysis methods, BeyondCell can more finely reveal the heterogeneity and sensitivity within tumors from a cellular perspective. MOVICS integrated 10 clustering algorithms, including CIMLR [24], iClusterBayes [25], MoCluster [26], COCA [27], ConsensusClustering, IntNMF [28], LRAcluster [29], NEMO [30], PINSPlus [31], and SNF [32]. By integrating various omics data such as gene expression, protein interactions, and metabolite levels, MOVICS provided a more comprehensive view of the molecular classification of cancer and a more robust result. Unlike previous single-omics analysis methods, MOVICS can reveal the more complex molecular mechanisms of cancer from multiple angles, which helps discover new biomarkers and therapeutic targets. In addition to that, we intended to apply the latest KAN architecture from the field of deep learning to our subsequent prognostic model research. The KAN architecture, based on the Kolmogorov-Arnold representation theorem, constructed a learnable activation function that can fit more accurate functions and had stronger performance than traditional machine learning models.

Working hypothesis in the framework of 3PM

Within the 3PM framework, early detection and monitoring of drug resistance, genotype-guided prescription, and dosing also provided information to enhance the tolerance of cancer treatment. Adverse drug reactions and chemotherapy toxicity deaths were catastrophic events, and in a significant portion of patients, these events were preventable [33, 34]. We hypothesized that drug-resistant cells at the single-cell level of ovarian cancer are the main cause of its drug resistance response and could serve as new therapeutic targets. Using the Beyondell algorithm, we can accurately identify the reactions that first-line drugs for ovarian cancer patients, such as paclitaxel, have on cells. For paclitaxel, docetaxel, and doxorubicin, cells that exhibit similar drug resistance responses were identified as resistant cells. By conducting differential expression analysis with other cells, we can obtain the specific molecular characteristics expressed by drug resistant cells. The drug resistance characteristics of ovarian cancer can be patient stratification features and prognostic biomarkers within the 3PM framework [35–37]. By utilizing these molecular characteristics for multi-omics clustering, we can obtain subtypes with different drug response differences. Accordingly, a deep learning prognostic model based on drug resistance molecular characteristics can accurately predict patient outcomes, forecasting the overall survival status of patients post-diagnosis and treatment from the level of gene expression.

Study design

Based on the working hypothesis, we first identified cells related to drug resistance reactions at the single-cell level. Based on their specific gene expression, we conducted subtype classification and prognostic model construction at the BULK transcriptome level with clinical information. The final subtypes and prognostic models were then validated in three external validation sets to test our hypotheses. If our subtypes exhibit similar drug response differences in all three external validation sets, and the constructed KAN prognostic model shows the best performance, it would indicate that our hypotheses are valid. It would also demonstrate that our classification strategy and prognostic models have a certain level of translatability and can be applied to our clinical diagnosis and treatment. These models will have the potential for patient stratification, predictive diagnosis, prognostic assessment, and personalized medical services.

Methods

Data source

We have obtained a single-cell RNA sequencing (scRNA-seq) dataset, GSE184880, from the GEO database (https://www.ncbi.nlm.nih.gov/geo). After acquiring the corresponding characteristics at the single-cell level, this study then proceeded to associate patient clinical information to construct a relevant model. For this purpose, we retrieved multi-omics data (RNA-seq, methylation, and mutation data) of patients with ovarian cancer through the UCSC Xena platform (https://xenabrowser.net/), which served as our training dataset. This was aimed at identifying subtypes of ovarian cancer that exhibit differential responses to drug treatments and developing a prognostic model. Datasets GSE17260, GSE26712, and GSE51088, obtained from the GEO database, were utilized as validation datasets to affirm the strong universality of the subtypes and to validate the generalizability of the prognostic model.

Single-cell data quality control

At the single-cell level, we utilized the Seurat R package (version 4.3.0) in the R programming language to perform quality control on each sample. Referring to the standards set by Birthe Dorgau [38], this study removed cells with gene reads less than 1000, features more than 10,000 or fewer than 500, and mitochondrial gene percentages greater than 15%. Cells that exceed the standards for gene reads and features were defined as low-quality cells and cells with mitochondrial gene percentages exceeding 15% were considered more likely to be dead cells, which therefore need to be filtered out. Subsequently, the raw data were normalized using the default parameters of the function. The FindVariableFeatures function was employed to select 2000 genes with high variability. Then, the data were standardized using ScaleData based on the 2000 highly variable genes. Ultimately, we obtained a single-cell expression matrix of 22,841 * 21,937, displaying the expression of 22,841 genes in 21,937 cells. The RunPCA function was used to reduce the dimensionality of this high-dimensional data, a step that allowed us to extract the principal components that best represent the variability between samples. Considering batch effects from different samples, we applied the “harmony” dimensionality reduction method to eliminate batch effects [39]. The Clustree algorithm can display the distribution of cells at different resolutions, in which a resolution where cell clustering did not exhibit abnormal distribution was considered to be an appropriate clustering result. Referring to the Clustree algorithm, we ultimately determined a certain resolution for clustering. To visualize the relationships between cells in high-dimensional single-cell RNA sequencing data in low-dimensional space, we employed two dimensionality reduction techniques: t-SNE (t-distributed stochastic neighbor embedding) and UMAP (uniform manifold approximation and projection). Both methods aimed to preserve the local structure of the data, making similar cells closer in the reduced space. Subsequently, we annotated and validated cell clustering through two methods: one method used the R package “SingleR” based on “celldex” to automatically annotate cells based on machine learning, and the other was a manual annotation of some marker genes using specifically expressed genes from cellmarker (http://117.50.127.228/CellMarker/). Due to the specificity of cellular molecular characteristics being the focus of this research, the R function “FindAllMarkers” was used to identify differentially expressed genes between each cell cluster and other clusters.

Beyondcell for identifying drug-resistant cells

Beyondcell algorithm could estimate the drug response of each cell to the medication of interest in the study. It was utilized for analyzing scRNA-seq data to identify the sensitivity of cell subpopulations to various drugs. Specifically, the algorithm calculated the Beyondcell score (BCS), which estimated the degree of enrichment for each drug feature in a specified set within each cell in the preprocessed single-cell expression matrix. The BCS ranged from 0 to 1, effectively measuring the perturbation sensitivity of cells (using drug perturbation feature sets). The calculated BCS matrix allows for the determination of therapeutic clusters, which are defined as “groups of cells within a cell population that share a common response to a set of drugs.” To identify potential therapeutic clusters, clustering analysis was applied to the BCS matrix, grouping cells based on their differential responses to selected drugs. For each drug, Beyondcell also calculated a switch point, which reflected the homogeneity of drug responses within the scRNA-seq dataset. UMAP representations were then generated using the BCS matrix and switch points, visualizing therapeutic clusters that reflect the homogeneity of drug responses within the cell population and highlighting drug responses within the therapeutic clusters. Additionally, Beyondcell identified therapeutic differences between cell populations and generated a prioritized list of differentially sensitive drugs to guide drug selection.

Cellular communication and stemness index

Cellular communication patterns can influence the homeostasis of the entire tumor microenvironment, and the interactions between drug-resistance cells and others may significantly impact the patient’s drug-resistance response. After identifying the drug-resistance cells, this study utilized the R package “Cellchat” to elucidate how these cells communicated with other cells. By integrating intercellular interaction networks and gene regulatory networks, Cellchat employed machine learning algorithms and statistical models to predict the communication mechanisms between prognostic-relevant cells and other cell types, including secretion and reception of cytokines, as well as cytokine signaling pathways. R package “CytoTRACE” (https://github.com/seqyuan/CytoTRACE) was used to infer the cell stemness index. Using vina, molecular docking experiments based on the top five specifically expressed genes were conducted to test for the presence of binding sites with paclitaxel, docetaxel, and doxorubicin.

Cluster analysis in multi-omics features

Clustering results from multi-omics can evaluate the molecular characteristics of patients from various dimensions, thereby identifying the subtypes with the greatest differences in drug resistance. The multi-omics characteristics associated with drug-resistant cells were determined by intersecting the features produced by the Beyondcell algorithm with four included gene expression matrices. We performed unsupervised clustering analysis on the multi-omics data of ovarian cancer patients. Initially, the R function “getElites” was utilized to filter the features, where “Cox” was applied to RNA-seq, “PCA” to methylation data, and the frequency “freq” to binary data. Subsequently, the optimal number of clusters was obtained using the cluster prediction index (CPI) and the Gaps statistic. Based on the optimal number of clusters, this study conducted consistent clustering using 10 multi-omics clustering methods: CIMLR, iClusterBayes, MoCluster, COCA, ConsensusClustering, IntNMF, LRAcluster, NEMO, PINSPlus, and SNF, to identify subtypes of ovarian cancer through the constructed similarity matrix.

Drug sensitivity analysis

The Cancer Drug Sensitivity Genomics (GDSC) was a public database that contained information on the sensitivity of cancer cells to drugs, as well as molecular markers corresponding to the drugs applied. The core idea of drug sensitivity analysis was to use known gene expression matrices of cell lines and corresponding drug sensitivity data (such as IC50 values) as a training set, establish a predictive model using machine learning methods, and then apply this model to predict the response of new samples to specific drugs. We considered the differences among four small molecule compounds (paclitaxel, docetaxel, doxorubicin, and olaparib) of patient subtypes. Paclitaxel, as a first-line chemotherapy drug, inhibited mitosis in tumor cells by disrupting the polymerization of tubulin; docetaxel, also a microtubule inhibitor, showed good efficacy in certain subpopulations of ovarian cancer; doxorubicin, as a topoisomerase II inhibitor, had different responses to DNA damage in tumor cells of different patients; olaparib, as a PARP inhibitor targeting BRCA1/2 mutations, had performed well in ovarian cancer patients with specific genetic backgrounds and played an important role in the treatment of ovarian cancer. The above four drugs had all been confirmed to have potential efficacy for ovarian cancer patients, but there were significant individual differences in the response of patients to these drugs.

Biological pathway enrichment analysis

GSEA pathway enrichment analysis was used to evaluate whether the expression pattern of a predefined gene set under drug resistance differential subtypes was significantly different from a random distribution. This method sorted the genes in the gene expression dataset and calculated the enrichment score for each gene set to determine whether the gene set was enriched at the top or bottom of the sorting list. Evaluate the statistical significance of enrichment scores through permutation testing and apply multiple hypothesis testing corrections to control for false positive rates. GSEA can reveal the impact of synergistic changes in gene concentration on phenotype, providing insights into complex biological processes and disease mechanisms. The gene reference set used for GSEA pathway enrichment analysis in this study included two datasets, GO (Gene Ontology) pathway reference sets and Hallmark pathway reference sets. The GO pathway reference set was a widely used gene functional classification system that categorized genes according to their biological processes (BP), molecular function (MF), and cellular component (CC). The Hallmark pathway was a representative set of gene expression patterns defined by MsigDB (Molecular Signatures Database), which represented core features of certain biological processes or disease states. The Hallmark set contained a typical set of genes involved in key biological processes such as cell cycle, apoptosis, and immune response.

Immune infiltration analysis

The “CIBERSORT” algorithm (https://cibersort.stanford.edu/) built a model using linear SVR and compared the gene expression data of mixed samples with known cell-type gene signatures. Through deconvolution analysis, CIBERSORT attempted to find the optimal combination of weights that represented the relative abundance of different cell types in the mixed sample. The result showed the degree of infiltration of 22 immune cells among the four subtypes.

Construction and validation of KAN-deep learning prognostic model

Baseline model establishment

Based on the drug resistance feature, this study constructed prognostic prediction models for ovarian cancer patients by using a proportional hazards regression (COX) model and a random survival forest (RSF), survival support vector machine, and DeepSurv model, separately. This study divided the TCGA dataset into training and validation sets in an 8:2 ratio to train the TCGA validation model. To further validate the model’s generalization capability, this study used GSE 26712, GSE 17260, and GSE 51088 datasets for external validation.

Improving the effectiveness of DeepSurv

Regarding the performance of the baseline model, our primary focus was on the DeepSurv model, which allowed for model tuning by adjusting parameters. To enhance the accuracy of the model, this study first tried the R-Drop strategy, which was specifically designed for small sample sizes, to DeepSurv. The R-Drop strategy was a regularized dropout strategy that minimized the bidirectional Kullback Leibler divergence [40] between the output distributions of two sub-models sampled by dropout. This strategy reduced the inconsistency caused by dropout and can effectively improve the training performance of the model, especially for small sample datasets, which can significantly improve model performance. Secondly, this study drew inspiration from the DeepSurv model and tried the Kolmogorov Arnold network (KAN) to improve the model. The construction of KAN involved multiple KAN layers, referring to multi-layer perceptron [41]. KAN architecture constructed a learnable activation function that can better fit the model, thereby achieving higher predictive performance.

Evaluation of KAN-deep learning prognostic model

The trained model was employed to predict the external validations’ prognosis outcome and the C-index was used to assess the predictive performance of the model. In this study, the predictive capabilities of the baseline model, the R-Drop optimized DeepSurv model, and the KANSurv model were assessed by comparing their C-indexes. A higher C-index indicated superior model performance.

Statistical analysis

All the data processing and analyses were executed in R software (Version 4.2.2). t-test and Wilcoxon test were utilized to compare the differences between quantitative variables while chi-square test was employed in qualitative variables. Spearman correlation test was utilized to explore the relationships between two variables. Kruskal–Wallis rank sum test was employed on the difference of multiple subtypes. p < 0.05 was considered statistically significant in the whole process.

Results

Data processing

The main process of this study is specifically shown in Fig. 1. After excluding patients with missing information, a total of 360 ovarian cancer (OV) patients were included in the training set. Using data from 360 TCGA-OV patients for training classifiers and prognostic models, 152 ovarian cancer samples from GSE51088, 110 samples from GSE17260, and 153 samples from GSE26712 were used as three validation datasets.

Fig. 1.

The flow chart of this study

The original distribution and baseline indicators of each dataset are shown in Table 1 (due to different centers, some centers did not collect information such as age and tumor stage).

Table 1.

The summary characteristics of the included samples in this study

Datasets
TCGA-OV
Sources	Data types	Samples
TCGA	Multi-omics	360
	Range	Mean
Age	30.00–87.00	59.47
	No	%
Clinical stage
Stage I	0	0
Stage II	23	6.39
Stage III	284	78.89
Stage IV	53	14.72
Radiation therapy
Yes	2	0.56
No	358	99.44
GSE195832
Sources	Data types	Samples
GEO	gene expression RNAseq	152
	Range	Mean
Age	26.44–90.78	59.35
	No	%
Clinical stage
Stage I	22	14.47
Stage II	9	5.92
Stage III	101	66.45
Stage IV	15	9.87
Unknown	5	3.29
GSE17260
Sources	Data types	Samples
GEO	Gene expression RNAseq	110
	No	%
Clinical stage
Stage III	93	84.55
Stage IV	17	15.45
GSE26712
Sources	Data types	Samples
GEO	Gene expression RNAseq	154
Only record patient survival related data
GSE184880
Sources	Data types	Samples
GEO	single-cell RNAseq	21,937

Reducing batch effects by “harmony,” UMAP and t-SNE dimensionality reduction visualization shows 21,937 cell clusters in the 11 cell clusters. Based on the specific gene expression levels from Cellmarker and the annotation results from singleR, each cell cluster was ultimately annotated as a specific cell population (Fig. 2B). There were differences in the proportion of cells among different individuals, in which cancer patients have more immune cells compared to normal samples, and fewer epithelial cells and fibroblasts (Fig. 2A). To identify drug-resistant cells from cancer samples, we extracted them for secondary clustering and further analyzed each cell cluster (Figs. 2C and S1). We performed differential analysis, enrichment analysis, and visualization on various cells using the R package “scRNAtoolVis.” As shown in Fig. 2D, T/NK cell clusters (marked with XCL2, GZMA, and GZMB, Fig. S6) were mainly enriched in the basic pathways of T-cell activation and regulation in the immune system. Epithelial (marked with CD24, CDH1, CLDN4, DSP, EPCAM, KRT18, KRT19, and OLCN, Fig. S3) were mainly enriched in pathways that maintain homeostasis, respond to environmental signals, and ensure normal cell function. Macrophages (marked with AIF1, CD14, CD163, CD74, and VSIG4, Fig. S7) were mainly enriched in immune system pathways for recognizing and clearing pathogens. Fibroblasts (marked with COL1A1, FAP, DCN, and PDPN, Fig. S4) were mainly enriched in pathways related to immune response and complement activation. B cells (marked with CD19, CD79A, CD79B, IGKC, and Fig. S5) were mainly enriched in pathways related to immune response activation and B-cell activation. Endothelial (marked with MGP, CLDN5, SPARCL1, CD34, CDH5, and PECAM1, Fig. S2) were mainly enriched in pathways related to angiogenesis and vascular development. Beyondcell algorithm was used to calculate the BCS of each cell and was shown in the UMAP dimension (Fig. 2E). As shown in Fig. 2F–H, paclitaxel, doxorubicin, and docetaxel all had lower BCS scores than the switch point in endothelial cells, indicating that endothelial cells will develop drug resistance reactions when the three drugs were applied separately, and their molecular characteristics had potential clinical diagnostic and therapeutic value. In addition, we used the bc4Squares function to describe in detail the positions of 5201 drugs when comparing endothelial cells with other cells. As shown in Fig. 2I, paclitaxel and docetaxel were both differentially low-sensitive drugs for endothelial cells.

Fig. 2.

Beyondcell recognized drug-resistant cells at the single-cell level. A The proportion of cells among different individuals. B The distribution of cell clusters from the total samples in tSNE dimension. C The distribution of cell clusters from cancer samples in tSNE dimension. D The differential and enrichment analysis of cell clusters from cancer samples. E Beyondcell calculated the BCS of each cell in the UMAP dimension. F–H The BCS scores distribution of paclitaxel, doxorubicin, and docetaxel in the UMAP dimension. I The bc4Squares plot describes in detail the positions of 5201 drugs when comparing endothelial cells with other cells

Cellular communication and regulatory network in endothelial

We utilized CellChat to characterize the interactions between endothelial and other cell types. Figs. 3A and S8 display the quantity and strength of communication between endothelial and other cells. Considering the enrichment analysis of endothelial, three signaling pathways associated with drug-resistance cells, either as important senders or major receivers, are displayed in Fig. 3B–C. The drug-resistance cells primarily communicated with other cells through the TGFb, ncWNT, and NOTCH signaling pathway (Fig. 3D). The results of CytoTRACE showed genes highly correlated with Cytotrace and also displayed the highest stemness index for endothelial (Fig. 3E–G), indicating endothelial had the highest differentiation potential. Molecular docking experiments confirmed that there were no potential binding sites between the five specifically expressed molecules and the drugs paclitaxel, docetaxel, and doxorubicin, indicating that these molecules themselves were incapable of exerting a therapeutic effect.

Fig. 3.

Cellular communication and stem index. A The overall landscape of cellular communication involving endothelial. B The contribution of each L-R pair. C Three signaling pathways significantly associated with endothelial. D The specific communication patterns of the three signaling pathways between cells. E, F The predicted ordering of stemness by CytoTRACE. G The correlation with CytoTRACE

Cluster analysis

Based on the combined results of CPI and GAP statistics (sum of the two values), the optimal number of clusters inferred was 4 (Fig. 4A). The consistency heatmap described the robust pairwise similarity between four subtypes and visualized 10 clustering results using a similarity matrix (Fig. 4B). Multi-omics heatmaps were used to reveal clustering patterns between samples and provided insights into potential sample biases or other human factors (Fig. 4D). The average silhouette value of this clustering result was about 0.4, indicating good similarity among the samples within each subtype (Fig. 4C). In the agreement analysis, we found that the distribution of the four subtypes in OS status and clinical staging was relatively uniform and did not have a strong correlation (Fig. 4E, F). In addition, it also provided some phenotypic differences between the four subtypes, such as age and clinical stage. Based on subtype-specific biomarkers (Tables S1 and 2), we established an ovarian cancer classifier using nearest template prediction (NTP) to predict possible subtypes for each sample. The Kruskal–Wallis rank sum test showed significant differences in sensitivity to four small molecule compounds among the four subtypes, The CS2 subtype showed lower IC50 values, indicating greater sensitivity to these drugs. In the TCGA training set and three validation sets, the drug sensitivity of the two subtypes showed similar differences, which also verified the robustness of our classification strategy (Fig. 4G–J). In addition to that, we also paid attention to the large number of T/NK cells that were quality controlled in the Beyondcell algorithm. After clustering analysis of the markers for T/NK cells, the resulting subtypes did not show better clustering effects than the drug-resistant cell subtypes in the frontline drugs we were concerned with (Fig. S9). Therefore, we believed that the four subtypes derived from clustering with endothelial cell markers were a better typing strategy and diagnostic and treatment guideline.

Fig. 4.

Cluster analysis. A The number of multi-omics clusters. B A stable clustering result by applying hierarchical clustering. C The silhouette plot of four subtypes. D The heatmap of the overall clustering process. E, F The agreement analysis between subtypes and clinical stage, OS. G–J The comparison of two subtypes’ drug sensitivity: The estimated IC50 of paclitaxel, olaparib, doxorubicin, and docetaxel between two subtypes in TCGA cohort

Biological pathway enrichment analysis

Although there were slight differences in pathway-enrichment analysis results across different datasets, some biological pathways were also significantly expressed within the same subtype. Specifically, in Hallmarker pathway enrichment analysis, the pathways enriched in CS1 were all related to cellular energy metabolism and lipid metabolism, such as HALLMARK OXIDATIVE PHOSPHORYLATION, HALLMARK FATTY ACID METABOLISM, HALLMARK PEROXISOME; CS2 involved pathways related to changes in cellular behavior, response to pressure, angiogenesis and signal transduction in tumor development, and cell apoptosis. For example, HALLMARK EPITHELIAL MESENCHYMAL TRANSATION (EMT); HALLMARK UV RESPONSE DN-UV; HALLMARK ANGIOGENESIS; HALLMARK HYPOXIA; HALLMARK KRAS SIGNALING UP; CS3-enriched pathways were relatively few, only HALLMARK MYC TARGET V2, indicating the expression patterns of genes regulated by c-Myc (also known as MYC) protein appear in multiple datasets; CS4 was mainly enriched in HALLMARK MYC TARGETS V1-MYC, HALLMARK SPERMATOGENESIS, HALLMARK G2M CHECKPOINT, which involved the regulation of the cell cycle, DNA repair, and reproductive cell maturation (Fig. 5A). The enrichment characteristics of these pathways suggested that CS1 (presumably a cell state or subtype) may possessed strong metabolic capabilities and high energy production; CS2 might exhibited great invasive and metastatic potential, angiogenic properties, and signaling and oncogenic activation ability, all of which could contribute to a more aggressive and adaptable phenotype in the context of tumor development; CS3 may be characterized by its reliance on c-Myc for gene regulation and metabolic activity, promoting rapid growth and proliferation. CS4, on the other hand, might be distinguished by its strict cell cycle regulation, DNA repair capabilities, and a potential role in maturation processes, with implications for both genomic stability and cell differentiation. In GO pathway enrichment analysis, CS1 from all four datasets was enriched in pathways closely related to mitochondrial function, especially those related to energy metabolism, such as GO OXIDATIVE PHOSPHORYLATION, GO ELECTRON TRANSPORT CHAIN, and GO ATP SYNTHESIS COUPLED ELECTRON TRANSPORT. The pathways enriched in CS2 were all related to the formation and maintenance of the extracellular matrix (ECM), such as GO EXCELLELAR STRUCURE ORGANIZATION, GO COLLAGEN FIBRIL ORGANIZATION, and GO COLLAGEN METABOLIC PROCESS. CS3 was enriched in pathways related to the maturation and differentiation process of epithelial tissues like skin and hair, such as GO KERATINIZATION and GO CORNFICATION. CS4 were all related to gene expression in cells, DNA repair, and cell cycle regulation, such as GO RNA SPLICING RNA, GO RECOMBINATIONAL REPAIR, and GO REGULATION OF CELL CYCLE G2 M PHASE TRANSITION (Fig. S1). At the single-cell level, endothelial cells were mainly enriched in HALLMARK TGF BETA SIGNALING, HALLMARK WNT BETA CATENIN SIGNALING, HALLMARK NOTCH SIGNALING, HALLMARK MITOTIC SPINDLE, HALLMARK APICAL JUNCTION, and HALLMARK HEDGEOG SIGNALING (Fig. 5B), as evidenced by specific gene set distribution plot (Fig. 5C). This suggested that endothelial cells were very likely to communicate with other cells through these signaling pathways, thereby regulating the homeostasis of the tumor microenvironment. These signaling pathways had also become the main focus of our cellular communication analysis and had been correspondingly validated.

Fig. 5.

Biological pathway enrichment analysis. A Enrichment analysis of different subtypes in four datasets. B A stable clustering result by applying hierarchical clustering. C The silhouette plot of four subtypes

Immune infiltration analysis

We utilized 119 tumor microenvironment gene sets and 114 metabolic gene sets obtained from literature to assess the tumor microenvironment and metabolism of the four subtypes. The ssGSEA analysis concluded that, for the majority of gene sets, CS2 exhibited high scores in immune cells, followed by CS1, while CS3 and CS4 showed somewhat lower scores compared to the other subtypes (Fig. 7A). In terms of metabolic levels, CS1 often demonstrated high scores, succeeded by CS4 and CS3, with CS2 scoring the lowest (Fig. 7B). We also used the CIBERSORT algorithm to evaluate immune infiltration in subtypes, and the four subtypes showed differences mainly in T-cell and macrophage subpopulations. CS2 mainly showed infiltration in M2 macrophages and neutrophils, while C4 mainly showed infiltration in plasma cells, T-cell follicular helper cells, and activated NK cells (Fig. 6C). Correlation analysis found that there is a strong antagonistic effect between M2 macrophages and neutrophils, with a similar relationship between plasma cells, T-cell follicular helper cells, and activated NK cells, indicating that although these cells infiltrated each subtype separately, they also interacted with each other to maintain the homeostasis of the subtype immune microenvironment, without making some cells very prominent (Fig. 6D). There were two different immune escape mechanisms in tumors: on the one hand, some immunosuppressive factors can prevent T-cell infiltration; on the other hand, although some tumors had high levels of cytotoxic T-cell infiltration, T cells were in a state of functional inactivation. TIDE predicted the immune escape ability of tumors by comprehensively evaluating the activity of these two mechanisms. Higher TIDE scores were associated with poorer immune checkpoint suppression therapy. As shown in Fig. 6F, CS2 had a higher TIDE score, exclusion score, and dysfunction score, with a lower MSI score compared to other subtypes, indicating that the immune checkpoint inhibition therapy (ICI) of these patients may have poorer efficacy. Notably, CS1, CS2, and CS4 performed relatively well in this test and are potential subgroups for immunotherapy. In specific ICI analysis, different subtypes exhibited distinct treatment responses, with CS1 and CS3 showing significantly higher expression in CEACM1 than other subtypes, making them potential immunotherapy targets. Although CS2 had shown poor efficacy in immunotherapy in previous evaluations, its CD86 expression was significantly higher than other subtypes, making it a viable treatment option (Fig. 6E).

Fig. 7.

Marker gene set analysis. A Enrichment analysis of endothelial cell markers. B Mutation analysis of endothelial cell markers. C The correlation between gene sets and immune key genes. D The details of the correlation between gene set and ICAM1, ITGB2, SELP, CD28, CD80, ICOSLG, ENTPD1, TGFB1, and TLR4

Fig. 6.

Immune infiltration analysis. A Tumor microenvironment analysis of four types. B Metabolism analysis of four subtypes. C The CIBERSORT analysis showed the immune infiltration of four subtypes. D The correlation of different immune cells. E The specific ICI analysis of four types. F The TIDE algorithm evaluated four subtypes’ immunotherapy response

Marker gene set analysis

Enrichment analysis showed that endothelial gene set involved cell movement or positional changes, mainly enriched in ameboidal type cell migration, cell-substrate adhesion, tissue migration, endothelial cell migration, cell-substrate junction, actin binding, and cell junction assembly (Fig. 7A). These genes were mainly missense mutation in ovarian cancer patients, with DST being the gene with the most mutations in the sample containing mutation information (Fig. 7B). The correlation indicated that this set of genes has a strong synergistic effect with ICAM1, ITGB2, SELP, CD28, CD80, ICOSLG, ENTPD1, TGFB1, and TLR4 (Fig. 7D). Among these, ICAM1, ITGB2, and SELP acted as co-stimulatory factors in the immune response; CD28, CD80, and ICOSLG functioned as co-inhibitory factors; TLR4 was a receptor; TGFB1 was a ligand; and ENTPD1 fell into the category of “other” (Fig. 7C).

Construction of deep learning prognostic model

Based solely on the results from the TCGA validation set, the RSF model performed the best, with a C-index that had reached 0.5696, followed by the DeepSurv model. However, upon considering the results from the external validation set, the performance of the RSF model was not satisfactory, showing a C-index of 0.50 on the GSE 26712 dataset, a C-index of 0.49 on the GSE 17260 dataset, and a C-index of 0.55 on the GSE 51088 dataset. This further suggested that although the RSF model had excelled on the validation set, it had poor generalization capabilities. The DeepSurv model also showed poor performance (a C-index of 0.50 on the GSE 26712 dataset), indicating it suffered from underfitting issues as well. The Cox model and the SSVM model demonstrated stronger generalization performance than RSF and DeepSurve model, but have a relatively poorer consistency on the TCGA validation dataset.

Based on the performance of the models, it was inferred that there were two main issues: 1) the TCGA dataset used for training was too small to extract sufficient information, and the distributions among the four datasets were different, further complicating model predictions; 2) there was an overfitting problem with the DeepSurv model and the RSF model, while the SSVM model and the Cox model were underfitting. To address these two issues and improve the model’s performance, this study took the following two approaches: 1) Feature selection was conducted using the Cox model and the RSF model to extract more effective features; 2) the model was improved based on the DeepSurv model framework to enhance model performance.

The feature selection results are depicted in Fig. 8C, where the importance scores calculated by the Cox model had indicated that genes such as TSPAN13, GGT5, DYNLL1, and EGFL7 played a significant role in prognosis prediction. The importance scores computed by the RSF model had shown that genes like AP1S2, EMP1, QKI, and MAGED2 played a crucial role for prognosis prediction. Upon intersecting the top 40 features calculated by both methods, genes such as UNC5B, TSPAN13, FBLN5, and SPRY4 were found to have played a significant role in both computational approaches.

Fig. 8.

Construction of deep learning prognostic model. A The overall framework of DeepSurv-R-Drop. The left diagram shows an input passing through the DeepSurv model twice, while the right diagram illustrates two different sub-models generated by Dropout. B The overall framework of KANSurv. The data passes through the KAN model to obtain the estimated value of h(x). C The results of feature importance calculated using the Cox model and the RSF model. The genes in red indicate those with high importance in both methods. D The comparison results of the c-index using six models and different numbers of features

Based on all features and the features selected using the Cox model and the RSF model, two additional models were added to the baseline model for the construction of prognostic models, and the construction results are shown Fig. 8D. Integrating results from the TCGA validation set and the three independent validation datasets, feature selection with the Cox model and the RSF model both made the models more prone to overfitting. The R-Drop (Fig. 8A) strategy effectively enhanced the performance of the DeepSurv model, leading to a significant improvement in its training outcome (compared to the previous C-index of 0.54, the C-index after R-Drop optimization has increased significantly to 0.62). The KANSurv model, built based on KAN (Fig. 8B), performed the best in all datasets when using all features (Fig. 8D). It had the highest average C-index (C-index: 0.6) across four datasets and also exhibited the highest C-index in the TCGA validation dataset (C-index: 0.63), GSE26712 (C-index: 0.61), and GSE17260 (C-index: 0.6), indicating its superior predictive capability.

Discussion

The situation of drug response in ovarian cancer

Ovarian cancer, as the type of gynecological malignancy with the highest mortality rate, has seen its treatment strategy gradually shift from traditional chemotherapy to personalized targeted therapy [3]. Paclitaxel, as a first-line chemotherapeutic drug, shows significant therapeutic effects in ovarian cancer patients by promoting the polymerization of tubulin and disrupting cell division [3]. In addition, traditional chemotherapeutic drugs such as doxorubicin and docetaxel are also used in some combined chemotherapy regimens for the treatment of ovarian cancer, although their side effects need to be strictly managed [42]. In recent years, the development of targeted therapy has provided new treatment options for ovarian cancer patients. Poly-ADP-ribose polymerase inhibitors (PARPi), such as olaparib, have shown good efficacy and tolerability in the treatment of ovarian cancer patients with BRCA mutations [43]. PARPi induces the accumulation of DNA damage by inhibiting the activity of PARP enzymes, thereby triggering the apoptosis of cancer cells. Olaparib, as a first-line maintenance therapy in advanced ovarian cancer patients, has increasingly substantial evidence of efficacy, especially in patients with germline BRCA1/2 gene mutations [44]. However, the long-term efficacy and drug resistance of targeted therapy remain issues that need to be addressed in future research.

The role of endothelial in drug resistance

This study utilized the Beyondcell algorithm to identify at the single-cell level that endothelial cells exhibited drug resistance to paclitaxel, doxorubicin, and docetaxel, suggesting that these cells could be potential targets for targeted therapy in future drug response research. Previous studies have found that endothelial cells may play an important role in the treatment of ovarian cancer. Ovarian cancer endothelial cells modulated biological functions through their specific molecular expression, playing a significant role in drug-resistance responses [45, 46]. Jessica Hoarau-Véchot’s research has found that ovarian cancer cells activated Akt phosphorylation in endothelial cells, inducing resistance to bevacizumab and leading to an autocrine loop based on FGF2 secretion [47]. Abnormally activated endothelial cells were also found in subsequent studies to increase the proliferation of cancer cells in ovarian cancer organoids and their resistance to treatment [48]. The Notch signaling pathway also affected the proliferation, migration, and drug resistance of ovarian cancer cells [49], and had a close interaction with endothelial cells in the ovary [50, 51]. Hoarau-Véchot used a co-culture model to demonstrate that activated endothelial cells induce the proliferation of ovarian cancer cell lines and an increase in chemotherapy resistance by activating the Notch signaling pathway [52]. They found that in co-cultures and chemotherapy-resistant ovarian cancer, the expression and activation of Notch receptors increased, suggesting that this pathway was associated with contact-mediated chemotherapy resistance. Pretreatment with the γ-secretase inhibitor DAPT increased the sensitivity of PROC to platinum by downregulating the Notch pathway [53, 54]. The SIERRA open-label phase Ib trial (NCT01952249) observed the safety and efficacy of demcizumab (a potent inhibitor of the Notch pathway) in combination with paclitaxel for the treatment of drug-resistant ovarian cancer. The results showed that this combination has a manageable toxicity profile, with a clinical benefit rate of 42% in patients with ovarian cancer who have developed drug resistance [55]. Multiple secretory factors, including members of the VEGF and TGF-β families, also interacted with endothelial cells through Notch signaling transduction, thereby regulating their biological functions [56]. For instance, reduced expression of miR-30a can lead to upregulation of TGF-β and SMAD4, ultimately activating autophagy and mediating cisplatin resistance in ovarian cancer [57]. Furthermore, the TGF-β pathway played a crucial role in platinum resistance through its typical downstream EMT-related molecules [58]. The TGF-β pathway also suppressed immunity within the tumor microenvironment and contributed to chemotherapeutic drug resistance. Shi’s research indicated that the TGF-β signaling pathway and RTK pathways were enriched in endothelial cells and their interplay was crucial for the development of cancer [59]. They regulated each other and synergistically controlled cell survival and migration, epithelial-mesenchymal transition, and the tumor microenvironment to accelerate tumorigenesis and metastasis. Daniel Newsted and colleagues developed an inhibitory antibody (anti-TGFBR2) to block TGF-β signaling, demonstrating that this antibody enhanced the efficacy of chemotherapy and elicited a limited anti-tumor immune response [60]. Endothelial cells can also communicate with epithelial cells expressing WNT/β, which further facilitated tumor progression [61, 62]. Chen’s study pointed out that endothelial cells communicate with Sox2 + tumor cells through the WNT/β-catenin signaling pathway, activating the dormancy of Sox2 + tumor cells, thereby reducing their response to chemotherapy [63]. Research by Doo DW [64] and Cascio S [65] had shown that the Wnt/β-catenin pathway can also promote T-cell exclusion and resistance to checkpoint inhibitors.

Patient stratification for personalized medicine and targeted prevention within the frame of 3PM

Multi-omics clustering had already been applied in other types of cancer and had demonstrated stable effects. For example, Hu [66] utilized multi-omics data to identify two subtypes with different survival differences in gastric cancer, which were validated in four external validation cohorts; Zheng [67] proposed patient subtype strategies with different immune therapy responses for colorectal cancer patients based on the characteristics of the tumor microenvironment and confirmed their stability in three external validation sets. This study was the first to classify patient subtypes using multi-omics data in the field of ovarian cancer, aiming to establish a stable patient medication strategy. Specifically, this study categorized ovarian cancer patients into four subtypes, among which the CS2 subtype from four datasets showed the lowest IC50 values in drug sensitivity prediction for all four drugs, indicating that the CS2 subtype had the best drug response across all four clinical treatment strategies. Specifically, for paclitaxel treatment, CS2 had the best drug response, followed by CS1, while CS3 and CS4 might exhibit drug resistance. For olaparib treatment, CS2 and CS4 were more sensitive compared to the other two subtypes. For doxorubicin and docetaxel treatments, the four subtypes showed similar distributions, with CS2 having the best drug response, followed by CS1, and then CS3 and CS4. In subsequent bioinformatics pathway enrichment analyses, CS2 was found to be enriched in pathways related to the extracellular matrix and tumor angiogenesis. Paclitaxel, as a microtubule damaging agent, can limit tumor growth and metastasis by inhibiting angiogenesis. The statistical results of this study suggested that patients with the CS2 subtype showed sensitivity to drugs like paclitaxel because this subtype was highly prone to intrinsic characteristics of angiogenesis, which were the targets of drugs like paclitaxel, thus resulting in a sensitive drug response. The CS1 subtype showed enrichment in pathways related to mitochondrial function and metabolic function, suggesting that clinical researchers could focus on the metabolic reprogramming of ovarian cancer and intervene in mitochondrial function. The CS3 subtype showed enrichment in pathways related to tissue maturation and differentiation, protein regulation, and clinical personnel could continue to delve into the proteome and phosphoproteome of ovarian cancer patients to identify biomarkers and therapeutic targets associated with the disease. The CS4 subtype showed enrichment in pathways related to DNA repair and cell cycle regulation. Some drugs that affect cell cycle regulation, such as CDK inhibitors, could be used to prevent the proliferation of tumor cells.

Prognostic model for predictive diagnostics within the frame of 3PM

Published in February, 2024, R-Drop was a deep learning approach tailored for small-sample strategies, which simplified model optimization and helped avoid the issue of model overfitting [68]. In April, 2024, a new neural network architecture known as KAN appeared, boasting powerful model fitting and predictive capabilities. With just 200 parameters, it was able to replicate the research findings that DeepMind had achieved with a multi-layer perceptron using 300,000 parameters [69]. As these two methods had recently been introduced, there were few examples of their application in deep learning models within the medical field. This study first enhanced the performance of the DeepSurv prognostic model under the R-Drop strategy and replaced the original multi-layer perceptron architecture with the KAN structure. The KANSurv model (average C-index: 0.60), built based on all endothelial cell-specific features, demonstrated superior predictive efficacy compared to other models.

Conclusions and expert recommendation in the framework of 3PM

Drug resistance was a formidable challenge in the treatment of ovarian cancer and was a major cause of poor prognosis [45, 70]. This study identified endothelial cells’ resistance response to three clinical drugs, using the Beyondcell algorithm, at the single-cell level of ovarian cancer. Under 10 multi-omics clustering algorithms, ovarian patient subtypes were classified based on the endothelial cell-specific molecular characteristics in multi-omics samples. In the analysis of drug sensitivity among subtypes, the CS2 subtype, which was highly enriched for angiogenesis pathways, showed high sensitivity to all four drugs (with paclitaxel as a vascular inhibitor) in the training set and three external validation sets (p < 0.05). For other subtypes, clinical researchers could also use the corresponding biological characteristics derived from pathway enrichment to tailor treatments and develop different therapeutic strategies. In clinical practice, when clinicians obtain omics data or gene expression profile data from a patient with ovarian cancer, they can classify the patient into subtypes based on the stratification strategy provided by this study. Based on the biological characteristics and differences in drug sensitivity provided by the research, they can implement personalized clinical treatment strategies. Besides, this study tried to improve the capabilities of the DeepSurv model by applying R-Drop strategy and the KAN architecture. The constructed KANSurv model demonstrated robust performance among ovarian cancer patients, effectively predicting their prognosis. Clinical physicians can make a rough prediction of a patient’s prognosis during the initial diagnosis and treatment, which can significantly influence subsequent medical care and patient management.

Therefore, we strongly recommend focusing on the research and practice of drug-resistance responses in ovarian cancer. The in-depth study of artificial intelligence in ovarian cancer drug resistance’ advanced 3PM approach has achieved significant innovation in the following three aspects.

• i.Predictive diagnostics. The KANsurv model constructed in this study was able to predict patient prognosis with relatively high accuracy. The prognosis prediction for patients will influence the clinical doctors’ treatment plans for their patients, achieving the goal of early intervention. An in-depth study of the predictive factors, such as the genes with the highest contribution in the deep learning model, could provide non-invasive screening tools for cancer patients [16, 71]. It might offer new targets for clinicians to diagnose ovarian cancer patients and provided references for the treatment of advanced ovarian cancer patients.
• ii.Targeted prevention. This study discovered at the single-cell level that endothelial cells were drug resistant and found that their specifically expressed molecules did not have binding sites for three clinical drugs. The molecular characteristics and related biological pathways of endothelial cells were potential targets for preventing drug resistance in ovarian cancer patients [50, 51, 60]. Besides, this study employed 10 clustering methods to classify patients, and based on the characteristics expressed in the enrichment analysis, even though some subtypes demonstrated drug resistance to certain clinical medications, clinical researchers could prevent adverse reactions in treatment and avoid the further potential of tumor progression.
• iii.Personalization of medical services. Based on the characteristics of drug-resistant cells, four subtypes were constructed to administer personalized treatment to patients with ovarian cancer. These four subtypes exhibited differences in biological characteristics including targeted therapy, immune microenvironment, metabolic capacity, and immunotherapy. The distinct clinical responses expressed by these subtypes facilitated a 3PM approach to the treatment of ovarian cancer patients [72, 73]. Specifically, CS1 had greater potential for immunotherapy, CS2 was more suitable for treatment with anti-angiogenic inhibitors, CS3 could be treated with therapies affecting the phosphoproteome of ovarian cancer patients, and CS4 could be adjunctively treated with drugs that influence cell cycle regulation. By considering the gene expression profiles of their patients, clinicians could devise more precise treatment plans tailored to the unique characteristics of each subtype, thereby enhancing therapeutic efficacy [74].

In summary, drug resistance events were a frequent occurrence in the treatment process of ovarian cancer. Our artificial intelligence work in ovarian cancer patients with drug-resistance-identified drug-resistant cells established prognostic models based on drug-resistant cells, and stratified patients to predict and assess various biological functions such as drug response as well as some suggestions about personalize medical services and targeted therapies for ovarian cancer patients. This had made innovative contributions to the paradigm shift of ovarian cancer from reactive medicine to 3PM [75–77]. Of course, incorporating more comprehensive omics data and expression profile data with survival information would enhance the completeness and robustness of the findings of this study. This also necessitated greater participation from medical institutions and patients who were willing to engage in such research and provide their data for analysis.

Limitations and further research

This study also has certain limitations and room for improvement. Firstly, although we have confirmed at the pharmacological and single-cell levels that endothelial cells are drug-resistant cells for first-line clinical medication, our findings can still be validated in future clinical studies through additional experiments. Secondly, the deep learning model can still be enhanced. It is hoped that clinical researchers and medical institutions will provide more data with survival information of ovarian cancer patients in the future, so as to enable us to perform more in-depth and refined model tuning, thereby improving its generalization capabilities.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contribution

Conceptualization, C.Z.; methodology, C.Z. and J.Y.; software, C.Z. and S.C.; validation, C.Z., J.Y., and L.S.; formal analysis, C.Z. and J.Y.; resources, G.L. and K.L.; data curation, G.L.; writing—original draft preparation, C.Z.; writing—review and editing, X.Z., B.P., and B.X.; visualization, C.Z.; supervision, B.X.; project administration, X.Z.; funding acquisition, B.X. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by National Youth Science Foundation Project (grant number 82204159) and Science and Technology Research Program of Chongqing Municipal Education Commission (grant no. KJQN202300423).

National Natural Science Foundation of China,No.82273739,National Youth Science Foundation Project,82204159,Science and Technology Research Program of Chongqing Municipal Education Commission,No. KJQN202300423

Data availability

The datasets generated and analyzed during the current study are available in the Cancer Genome Atlas (TCGA) repository (https://portal.gdc.cancer.gov/) and GEO database (https://www.ncbi.nlm.nih.gov/geo).

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.