Abstract

   Endothelial cells (ECs) are critical drivers of tumour progression, and
   their angiogenic process has been widely studied. However, the
   post-angiogenic transition of tip endothelial cells after sprouting
   remains insufficiently characterised. In this study, we utilised
   single-cell RNA sequencing analyses to identify a novel EC transition
   signature associated with endothelial permeability, migration,
   metabolism, and vascular maturation. Within the transition pathway, we
   discovered a critical EC subpopulation, termed tip-to-capillary ECs
   (TC-ECs), that was enriched in tumour tissues. Comparative analyses of
   TC-ECs with tip and capillary ECs revealed distinct differences in
   pathway activity, cellular communication, and transcription factor
   activity. The EC transition signature demonstrated substantial
   prognostic significance, validated across multiple cancer cohorts from
   TCGA data, particularly in bladder cancer. Subsequently, we constructed
   a robust prognostic model for bladder cancer by integrating the EC
   transition signature with multiple machine-learning techniques.
   Compared with 31 existing models across the TCGA-BLCA, [28]GSE32894,
   [29]GSE32548, and [30]GSE70691 cohorts, our model exhibited superior
   predictive performance. Stratification analysis identified significant
   differences between different risk groups regarding pathway activity,
   cellular infiltration, and therapeutic sensitivity. In conclusion, our
   comprehensive investigation identified a novel EC transition signature
   and developed a prognostic model for patient stratification, offering
   new insights into endothelial heterogeneity, angiogenesis regulation,
   and precision medicine.

1. Introduction

   Cancer remains a leading cause of mortality worldwide, with
   approximately 20 million new cases and nearly 9.7 million deaths
   reported in 2022 [[31]1]. In the United States alone, it is estimated
   that 2,041,910 new cases and 618,120 deaths will happen in 2025
   [[32]2].

   Despite advancements in innovative therapies, the inherent
   heterogeneity of tumours continues to pose significant challenges,
   limiting treatment efficacy [[33]3]. Traditional cancer studies
   focusing on single cancer types often fail to capture the complexity
   and diversity observed across different malignancies. Pan-cancer
   studies have thus emerged to overcome these limitations, providing a
   more comprehensive understanding by identifying common molecular
   drivers across multiple cancer types [[34]4]. Comparative analyses
   between primary and metastatic tumours across cancers further aid in
   revealing shared and unique features, facilitating better prediction of
   therapeutic outcomes and prognosis [[35]5].

   Investigating tumour heterogeneity at the cellular level necessitates
   high-resolution technologies, notably single-cell RNA sequencing
   (scRNA-seq), enabling detailed characterisation of the tumour
   microenvironment (TME) [[36]6]. Within the TME, endothelial cells (ECs)
   play crucial roles in tumour progression, primarily through their
   functions in angiogenesis and immune modulation [[37]7]. Based on their
   anatomical locations, ECs can be categorised into capillary, lymphatic,
   arterial, and venous ECs [[38]8]. Additionally, during angiogenesis, a
   specialised EC subpopulation called tip ECs guides the growth of new
   blood vessels and subsequently matures into stable ECs upon the
   formation of a new vascular network [[39]9]. Anti-angiogenic therapies
   targeting EC-driven angiogenesis have demonstrated survival benefits
   [[40]10]. However, their efficacy varies significantly across different
   cancer types, likely due to EC heterogeneity within tumours [[41]11].
   Although EC heterogeneity is recognised, the precise transition from
   tip ECs to mature capillary ECs remains inadequately explored,
   particularly from a pan-cancer perspective.

   Consequently, this study aims to bridge this critical gap by
   integrating published scRNA-seq datasets across multiple cancer types
   to systematically characterise the transition from tip ECs to capillary
   ECs and identify a robust endothelial transition signature.
   Subsequently, bladder cancer was specifically selected for further
   validation in light of the critical role of EC transition in it.
   Additionally, we developed a robust prognostic model based on the
   signature using multiple machine-learning algorithms, providing a novel
   tool for guiding patient stratification and personalised treatment
   strategies in bladder cancer.

2. Methods

2.1. Data Collection

   The single-cell data were downloaded from CellxGene, the Curated Cancer
   Cell Atlas, Tabula Sapiens, and [42]GSE210347, comprising 280 samples
   from bladder, breast, gastric, colorectal, lung, liver, ovarian,
   pancreatic, and prostate cancers and corresponding normal tissues
   ([43]Supplementary Table S1) [[44]12,[45]13]. Only 10x scRNA-seq
   datasets of human tissues were chosen to rule out possible
   technology-induced bias. Quality control initially excluded low-quality
   datasets with fewer than 10,000 detected genes or lacking expression
   data for mitochondrial genes or key endothelial markers (PECAM1, CDH5,
   VWF, and TIE1). Additionally, the bulk RNA sequencing samples of
   bladder cancer and associated survival information were obtained from
   four public datasets, including TCGA-BLCA (n = 412), [46]GSE32894 (n =
   308), [47]GSE32548 (n = 131), and [48]GSE70691 (n = 49).

2.2. Identification of EC Subpopulations

   The scRNA-seq data analysis was based on the Seurat R package (v4.1.3)
   [[49]14]. Samples with fewer than 1000 total cells or fewer than 100
   ECs were removed. Cells were further filtered based on the following
   criteria: total read counts less than 500 or more than 20,000, detected
   genes fewer than 500 or more than 5000, and mitochondrial gene
   percentage greater than 10%. Subsequently, all 148,864 ECs were
   extracted for further dimension reduction and clustering. The
   expression profiles were log-normalised, and the data were then scaled
   after identifying the 2000 most variable genes. Principal component
   analysis (PCA) was conducted, followed by batch effect correction using
   the Harmony R package (v1.2.0) [[50]15]. Dimensionality reduction was
   performed using UMAP with 40 dimensions, and clustering was performed
   at a resolution of 0.1. Some contaminating clusters, showing the
   signatures of epithelial cells (KRT19 and KRT8), fibroblasts (COL1A1
   and DCN), macrophages/monocytes (MS4A6A, S100A8, and CSTA), T cells
   (CD3D and CD3E), and smooth muscle cells (TAGLN and MYL9) were excluded
   from downstream analysis. Ultimately, five EC subpopulations were then
   identified, including arterial, vein, capillary, lymphatic, and tip
   ECs, using marker genes from published studies and the Cell Taxonomy
   database ([51]Table 1) [[52]16]. Tip ECs and capillary ECs were
   subsequently extracted and processed through the same workflow,
   resulting in ten clusters at a resolution of 0.2. Tissue preferences of