Cell communication pathway prognostic model identified detrimental neurodevelopmental pathways in neuroblastoma

Abstract

Neurodevelopmental cell communication plays a crucial role in neuroblastoma prognosis. However, determining the impact of these communication pathways on prognosis is challenging due to limited sample sizes and patchy clinical survival information of single cell RNA-seq data. To address this, we have developed the cell communication pathway prognostic model (CCPPM) in this study. CCPPM involves the identification of communication pathways through single-cell RNA-seq data, screening of prognosis-significant pathways using bulk RNA-seq data, conducting functional and attribute analysis of these pathways, and analyzing the post-effects of communication within these pathways. By employing the CCPPM, we have identified ten communication pathways significantly influencing neuroblastoma, all related to axongenesis and neural projection development, especially the BMP7-(BMPR1B-ACVR2B) communication pathway was found to promote tumor cell migration by activating the transcription factor SMAD1 and regulating UNK and MYCBP2. Notably, BMP7 expression was higher in neuroblastoma samples with distant metastases. In summary, CCPPM offers a novel approach to studying the influence of cell communication pathways on disease prognosis and identified detrimental communication pathways related to neurodevelopment.

Introduction

Neuroblastoma, a malignancy originating from neural crest cells, predominantly affects pediatric patients and demonstrates considerable heterogeneity in clinical outcomes, ranging from spontaneous regression in infants to aggressive disease progression in others [1]. Given the neural origin of neuroblastoma and the active neurodevelopmental processes involved in tumor formation, elucidating the intricate interplay among age, neuroblastoma, outcome, and neurodevelopment is paramount for optimizing treatment strategies and improving long-term outcomes for affected children [2,3]. However, the relationship between neurodevelopmental mechanisms and the outcome of neuroblastoma has not been fully explored [3,4].

During the process of neurodevelopment, coordinate cellular interactions are crucial for the proper formation and functioning of the nervous system [5,6]. These interactions involve communication between neural progenitor cells and adjacent support cells, as well as intercellular signaling among various neuronal, glial, and stromal cells [7]. These precise cell communication pathways are responsible for regulating the intricate process of neurodevelopment [8]. In neuroblastoma, these neurodevelopmental cell communication pathways may confer unique biological functions on tumor cells, such as neuronal differentiation and migration, thereby influencing the prognosis of neuroblastoma [9,10]. Consequently, it is essential to unravel the specific mechanisms underlying neurodevelopmental cell communication and its impact on the prognosis of neuroblastoma.

The commonly used approach in cell communication research is utilizing single-cell RNA sequencing (scRNA-seq) data to quantify the expression levels of communication ligands and receptors in each cell type to establishment of cell communication interactions numbers and strengths [11]. However, the high cost of scRNA-seq for large cohorts [12] poses challenges in studying the relationship between cell communication and prognosis. Although scRNA-seq data obtained from 16 human adrenal neuroblastoma samples can be used to study cell communication [13], it difficult to directly correlate the findings with long term prognosis. Moreover, commonly used methods like CellPhoneDB [14] and CellChat [15], based on scRNA-seq data analysis, cannot directly elucidate the cell communication post-effects on prognostic mechanisms. Hence, there is a need for a novel method that can establish the association between alterations in cell communication and prognosis, as well as clarify the impact of communication mechanisms on neuroblastoma prognosis.

In this study, we propose the cell communication pathway prognostic model (CCPPM), a comprehensive data analysis pipeline designed to assess cell communication mechanisms that are associated with disease prognosis. In CCPPM, all cell communication pathways are calculated through scRNA-seq data and subsequently input to the cohort bulk RNA-seq prognosis model to identify the prognosis related communication pathways, communication post-effects are calculated through the analysis of the receptor cascade transcription factors and target gene signaling transduction. Applying the CCPPM, we have revealed the neurodevelopmental communication pathway BMP7-(BMPR1B-ACVR2B) as significant contributors to neuroblastoma malignancy and overall survival (OS) of patients, communication post-effect analysis indicated that the BMP7-(BMPR1B-ACVR2B) pathway activated SMAD1 to maintain tumor cell migration in neuroblastoma. Overall, the CCPPM proposed in this study fills the gap in existing cell communication calculation methods that cannot correlate prognosis and analyze communication post-effects, highlights the impact of neurodevelopmental cell communication mechanisms on neuroblastoma prognosis and offers potential advancements in neuroblastoma risk stratification.

Methods

Data collection

The scRNA-seq dataset was downloaded from the NCBI Gene Expression Omnibus (GEO) database (accession number GSE137804). It comprises 16 treatment-naïve neuroblastoma samples, 4 fetal adrenal gland samples, and 2 embryo samples. The neuroblastoma and fetal adrenal gland sample were used to analyze the cell compositions and cell communication pathways [13]. For the analysis of cell communication pathways significantly associated with OS and the establishment of the prognostic model, bulk RNA-seq datasets GSE62564 [16] and TARGET-NBL were retrieved from the NCBI GEO database and The Cancer Genome Atlas Program (TCGA) database, respectively. The GSE62564 dataset contains INSS stage 1 to 4S patients and the dataset TARGET-NBL contains INSS stage 3 to stage 4S patients. The clinical information for these datasets was sourced from the original paper, as well as the TCGA and GEO databases.

Data processing and cell type identification

The scRNA-seq dataset was mainly processed by R package Seurat 4.0 [17]. Cells with gene counts exceeding 7,500 or falling below 500, as well as those with mitochondrial gene proportions exceeding 15 %, were deemed to be of low-quality and subsequently excluded. Gene expression in the qualified cells was normalized and the top 2000 highly variable genes were selected and scaled. Subsequently, the harmony batch effect correction was preformed, then the data were further processed with cell dimensionality reduction by t-distributed stochastic neighbor embedding (t-SNE) and clustering. Cell type was annotated according to the canonical markers obtained from the original text.

Cell communication pathways inference and comparison

The cell communication pathways of neuroblastoma and fetal adrenal gland samples were inferred by CellChat 1.4.0 [15]. The average strength of each cell communication pathway was calculated and shown in a chord diagram. Additionally, the roles of the cell communication source and target cells were depicted as proportions in the total cell communication strength.

Strengths of cell communication pathways calculation in the bulk RNA-seq

The bulk RNA-seq data were converted to Transcripts Per Million (TPM) values. Samples collected from relapsed tumor tissues, peripheral blood and patients whose OS equal to zero were removed; then the dataset GSE62564 contained 498 samples and TARGET-NBL contained 149 samples. Leveraging the cell communication pathways inferred from the scRNA-seq dataset of neuroblastoma, we calculated the strength of each communication pathway based on the law of mass action [15]. The relevant ligands and receptors were obtained from the CellChat database.

Prognostic model establishment and performance evaluation

The neurodevelopmental cell communication pathways were identified as the overlapping pathways observed in both neuroblastoma and fetal adrenal gland samples. The strengths of these cell communication pathways were utilized to establish a prognostic model. Initially, univariate Cox regression analysis was performed using the R package coxph to identify communication pathways significantly associated with the OS of neuroblastoma patients, using a significance threshold of P < 0.05. Subsequently, Lasso regression analysis using the R package glmnet was conducted to further refine the selected communication pathways and avoid the collinearity of independent variables. The remaining communication pathways were then subjected to stepwise multivariate Cox proportional hazards regression, and the Akaike information criterion (AIC) was employed to select the optimal model. The communication pathways, along with their corresponding coefficients in the model, were utilized to calculate a risk score using the following formula [1]. The training dataset was used to divide the patients into low-risk and high-risk groups based on the median risk score. The resulting risk scores, along with survival status, a heatmap of risk gene expression, Kaplan-Meier curves, and time-dependent receiver operating characteristic (ROC) analysis were subsequently validated using the TARGET-NBL dataset. Differences in survival rates between groups were assessed using the log-rank test, while the predictive accuracy was determined by calculating the area under the curve (AUC) using the R package survivalROC.

$$riskscore=\sum _{i=1}^{n}coefficien{t}_i\times communiactionstrengt{h}_i$$ (1)

Trajectory inference and branch-dependent genes analysis

The pseudo-time trajectory of the sympathoadrenal cells was inferred by R package monocle2 [18]. The genes used to order cells were identified by differentialGeneTest function, according to the gene varying with cell types. Then the method of discriminative dimensionality reduction with trees (DDRTree) was applied and the cells were visualized in two dimensions. The branch containing the most SCPs set as the root. To reveal the genes that influence cell type transition, the branched expression analysis modeling (BEAM) method was performed to obtain the branch-dependent genes. The gene enrichment analysis was performed by gene ontology (GO) to clarify the main biological functions of each interest gene set.

Identification of tumor cell subtypes in neuroblastoma samples by NMF

Tumor cells in neuroblastoma are known to exhibit diverse characteristics corresponding to different stages of peripheral nerve development. To classify these tumor cells into subtypes, we employed the non-negative matrix factorization (NMF) on the expression profiles of ordering genes obtained from the trajectory inference of sympathoadrenal cells. Specifically, the NMF algorithm was implemented using the "Brunet" method with a total of 100 iterations. To determine the appropriate number of clusters, we tested values ranging from 2 to 7. Upon finding that the optimal cluster number was 5, the resulting classification successfully grouped the different cell types of sympathoadrenal cells in the fetal adrenal glands. Consequently, the basis vectors obtained through NMF were utilized to identify the subtypes of tumor cells in each neuroblastoma sample. The similarity score for the neurodevelopmental subtypes of tumor cells was defined as the orthogonal projection of the ordering genes' expression profile onto these basis vectors.

Regulon activity calculation

The python package pySCENIC was implemented to get the activated transcription factors of tumor cells in neuroblastoma samples and sympathoadrenal cells in fetal adrenal gland samples [19]. The co-expression of target genes was analyzed and the regulon activity was calculated by the AUCell package.

Bayesian network establishment

To clarify the regulatory relationship, we extracted the activities of transcription factors that co-regulate the target genes of SMAD1, as well as the activities of transcription factors that regulate SMAD1 itself, and the expression profiles of target genes regulated by these transcription factors. Then we used hdWGCNA to get gene modules based on the target genes expression profiles [20]. The module eigengene levels, together with the activities of the transcription factors and the strength of the BMP7-(BMPR1B-ACVR2B) pathway, were used to infer the regulatory network structure. We incorporated a whitelist comprising known regulatory relationships between communication pathways and transcription factors. Finally, the Max-Min Hill-Climbing algorithm (MMHC) was utilized to obtain a robust network structure.

Immunohistochemistry of neuroblastoma in situ tissue

The 10 % formalin-fixed, paraffin-embedded neuroblastoma specimens were cut into 4 µm sections. After deparaffinization and dehydration, the sections were incubated in 3 % hydrogen peroxide at room temperature for 15 min to block endogenous peroxidase interference. For antigen retrieval, sections were steamed with citrate buffer (pH 6.0) for 40 min. After non-specific protein blocking, anti-BMP7 (1:200; Proteintech; Cat No: 12221-1-AP) was added for incubation at 4 °C overnight. Then, the sections were incubated with enzyme-labeled secondary antibody for 2 h at room temperature. Having been exposed to diaminobenzidine substrate (HRP; Wuhan Boster Biological Technology; Cat No: AR1027), the sections were counterstained with hematoxylin and sealed with resin to perform microscopic images.

For semi-quantitative analysis the BMP7 protein level, we used the Image-Pro Plus software to get its relative positive area in immunohistochemistry. After correction the system optical density, the measurement area was selected by the area of interest tool, and five measurement fields were selected for each slide. Then the accumulated optical density and positive area in the measurement region were calculated by software. Finally, the mean ratio of above two indicators in five measurement fields was calculated, then scaled by the mean value of group that neuroblastoma patients were without distant metastases.

Human in-situ neuroblastoma samples

The in-situ tissue specimens of neuroblastoma were collected from the Children's Hospital of Nanjing Medical University, which were the remaining samples of pathological examination. The study conformed to the standards set by the Declaration of Helsinki and was approved by the Ethics Committee of Children's Hospital of Nanjing Medical University (NO.202305006-1). The study was undertaken with the understanding and written consent of each participant.

Statistical analysis

All analyses were performed by R (4.0.5), and the packages used in each step were mentioned in the above methods. Data visualization was mainly performed by R packages ggplot2, pheatmap and ggvenn. P-values of multiple testing were adjusted by Bonferroni method. The significance level was set at p<0.05; p-values were shown in figures as *: p<0.05, **: p<0.01 and ***: p<0.005, ****: p<0.0001.

Results

Cell composition of human neuroblastoma samples and normal fetal adrenal samples

To elucidate the functions of neurodevelopmental communication pathways, the CCPPM data analysis pipeline was conducted as depicted in Fig. 1A. The pipeline involved several key steps, include cell communication analysis using scRNA-seq data, establish the association between changes in cell communication and prognosis using bulk RNA-seq, functional analysis of communication pathways, and post-effect analysis of cell communication. In the cell communication analysis section, scRNA-seq data from 16 treatment-naïve neuroblastoma samples obtained from the patient's adrenal gland and 4 fetal adrenal gland control samples were analyzed. After removing low-quality cells (cells with more than 7,500 genes or fewer than 500 genes, or more than 15 % of mitochondrial genes were be eliminated) and doublets, we obtained 164,931 cells of neuroblastoma samples and 50,324 cells of fetal adrenal gland samples and the batch effects between two samples were corrected.

Fig. 1.

Cell composition in neuroblastoma samples and fetal adrenal gland control samples. (A) The cell communication pathway prognostic model (CCPPM) workflow for the risk related communication pathway screening and post-effect inferring. (B) The major cell types of neuroblastoma samples were visualized in two-dimensional by t-distributed stochastic neighbor embedding (t-SNE). (C) Dot plot represents the average expression levels of canonical markers in each cell type. (D) Proportion of major cell types in each patient. (E) The major cell types of fetal adrenal gland control samples were visualized by t-SNE.

To annotate the cell types in neuroblastoma and control samples, linear and nonlinear dimensional reduction was performed. Seven different cell types were identified in neuroblastoma (Fig. 1B), including tumor cells (as identified by the presence of abnormal copy number variations compared to other control cells based on infercnv results), Schawann cells, myeloid cells, fibroblast, endothelium cells, T cells and B cells. The expression percentage and level of cell type annotation marker genes were plotted to indicate that the annotation of cell type was accurate (Fig. 1C). The proportion of seven types of cells in neuroblastoma samples is basically similar, with tumor cells occupying the majority (Fig. 1D). As for the health control sample, fetal adrenal gland samples contain nine types of cells (Fig. 1E and Supplemental Fig. S1), including Schwann cell precursors (SCPs), sympathoblasts, chromaffin cells, endothelial cells, fibroblasts, capsular cells, myeloid, steroidogenic cells and T cells, which correspond well to the cell types in the neuroblastoma sample.

Identification of neurodevelopmental cell communication pathways in neuroblastoma

In the CCPPM framework, we utilized scRNA-seq data to explore the cell communication pathways in both neuroblastoma and fetal adrenal gland samples as depicted in Fig. 2A. Cell communication pathways were visualized using a weighted directed cyclic network, where each edge represents a communication relationship and points to the receiving cell (directed), and the edge width represents the strength of the communication (weighted). In the fetal adrenal gland samples, the cell communication primarily occurred between SCPs, capsular cells, sympathoblasts, and fibroblasts. Furthermore, the distribution of communication strength between these cell types appeared to be well-proportioned (Fig. 2B, top). Conversely, in neuroblastoma, the dominant cell communications were observed between fibroblasts, Schwann cells, and tumor cells (Fig. 2B, bottom). This finding suggests a notable shift in the pattern of cell communication in neuroblastoma compared to the fetal adrenal gland samples. This analysis provides valuable insights into the alterations in cell communication patterns associated with neuroblastoma, highlighting the potential role of fibroblasts, Schwann cells, and tumor cells in driving the disease.

Fig. 2.

Identification of neurodevelopmental cell communication in neuroblastoma samples and fetal adrenal gland control samples. (A) CCPPM pipeline step 1 to 3: using scRNA-seq data to identify cell communication pathways of interest. (B) The communication relationships between the cell types in two conditions represented by weighted directed cyclic graph. (C) Bar plots illustrated the number and weight variation of inferred interactions in the two groups. (D) Distribution of separated and shared communication pathways in the two groups. (E) Word cloud visualized the main components of the shared pathways, with font size indicating the average weight of the pathways. (F) The biological process of shared pathways was enriched by gene ontology (GO), the items of p-value less than 0.05 was shown in bar plot. (G and H) Heatmap indicated scaled interaction strength of source and target cells in each neuroblastoma sample and fetal adrenal gland control sample.

To provide a more accurate quantitative evaluation of changes in cell communication in neuroblastoma, we counted the mean number and strength of communications in neuroblastoma and fetal adrenal gland samples (Fig. 2C). We found that both the mean number and strength of communications in neuroblastoma were higher than those in fetal adrenal gland samples. Considering the strong pathological correlation between neuroblastoma and neurodevelopment [21,22], as well as the large number of developing peripheral nervous systems in the fetal adrenal gland [23,24], identify shared cell communication pathways between neuroblastoma and fetal adrenal gland will help to analysis neurodevelopmental mechanisms in neuroblastoma, rather than solely focusing on differential pathways. We identified a total of 496 communication mechanisms in neuroblastoma and 295 in the fetal adrenal gland. After calculating the intersection of the two datasets, we discovered 240 shared communication pathways (Fig. 2D) (Supplemental Table S1), which were mainly involving immune and neurodevelopmental pathways, such as MIF and MPZ (Fig. 2E). The functional enrichment analysis of these 240 intersecting communication pathways revealed their involvement in the regulation of neural development (Fig. 2F), particularly in processes related to the development and migration of neural crest cells, which directly contribute to the pathogenesis of neuroblastoma.

We conducted network topology analysis of the 240 communication pathways aimed to examine the source and target cells involved in the cell communication pathways related to neurodevelopment. In neuroblastoma, the source cells were identified as fibroblasts, Schwann cells, and tumor cells, with tumor cells serving as the target cells (Fig. 2G). In fetal adrenal glands, capsular cells, chromaffin cells, fibroblasts, SCPs, and sympathoblasts acted as either the main source cells or target cells (Fig. 2H). The normal development of fetal adrenal sympathetic neurons depends on the cell communication pathways depicted in Fig. 2H. However, in the context of neuroblastoma, the target cells that play a role in these pathways transition to tumor cells. Consequently, this research sheds light on the potential role of cellular communication pathways related to neurodevelopment in the pathogenesis of neuroblastoma.

Prognostic model in CCPPM utilize cohort bulk RNA-seq data elucidates cell communication pathways influence the survival of neuroblastoma patients

Despite many research inferring the impact of cell communication pathways on prognosis by comparing the existence and strength changes of communication pathways in disease stages indirectly using scRNA-seq. However, the small sample size (mostly n < 10) and patchy clinical survival information of scRNA-seq datasets limited the screening of the pathways directly related to prognosis from numerous communication pathways. To solve this problem, prognostic model in CCPPM provide a method to identify whether communication pathways influence the survival of patients as Fig. 3A shown.

Fig. 3.

Prognostic model based on shared cell communication pathways. (A) CCPPM pipeline step 4 and 5: cell communication prognostic model using cohort bulk-RNA-seq data. (B) Hazard ratio of candidate communication pathways in the prognostic model. (C) Heatmap illustrates the scaled strength of candidate cell communication pathways in the training dataset. (D) Distribution of risk score in high and low risk groups of the training dataset. (E) Distribution of decreased patients with risk score of the training dataset. (F) Kaplan-Meier curves representing the overall survival in high and low risk groups of training dataset. (G) The 1-, 3-, and 5-year ROC curves indicates the sensitivity and specificity of the model in the training dataset. (H-K) The performance of prognostic model applied in validation dataset, similar with the training dataset in Fig 3D-G.

Before construct the prognostic model, CCPPM analyzed the cell communication pathways which impact the OS of patients in neuroblastoma cohort with bulk RNA-seq dataset, then the risk related communication pathways used to establish the prognostic model. Here we calculated the strengths of the 240 neurodevelopmental cell communication pathways in bulk RNA-seq datasets and analyzed the impact of the neurodevelopmental cell communication pathways on the OS of neuroblastoma patients. Firstly, the communication strength of all 240 neurodevelopmental cell communication pathways was calculated by the law of mass action use bulk RNA-seq datasets. Subsequently, univariable Cox regression and lasso regression were carried out to narrow the filtered cell communication pathways. To select the optimal model, a stepwise multivariate Cox proportional hazards regression analysis was performed, guided by the AIC. As a result, ten cell communication pathways were included in the final model. The hazard ratio of each pathway is presented in Fig. 3B. The C-index of this model was 0.79.

The prognostic model of CCPPM include (step1 obtain the risk related cell communication pathways by univariable Cox regression, step2 reduce the collinearity of the cell communication pathways by lasso regression, step3 establish prognostic model by stepwise multivariate Cox proportional hazards regression). Here, a cohort neuroblastoma bulk RNA-seq dataset, GSE62564 (n = 498), was used as the training dataset, and another cohort neuroblastoma bulk RNA-seq dataset, TARGET-NBL (n = 149), was used as the validation dataset. The risk score of each patient was calculated according to the formula [1] (see Methods). The strength of protective cell communication pathways was higher in the low-risk group, and those of the detrimental cell communication pathways were higher in the high-risk group (Fig. 3C). Based on the prognostic model of CCPPM, the GSE62564 cohort was divided into a low-risk group and a high-risk group using the median risk score as the threshold (Fig. 3D). It was observed that deceased patients were predominantly distributed in the high-risk group (Fig. 3E). Furthermore, the high-risk group exhibited significantly lower survival probabilities compared to the low-risk group as depicted in the long-term outcome analysis (Fig. 3F). In addition, the time-dependent ROC curve was generated to assess the sensitivity and specificity of the model in predicting 1-year, 3-year, and 5-year OS, the areas under the curve (AUCs) exceeding 0.8 indicates that the model has good performance (Fig. 3G). The validation dataset showed similar results to the training set (Fig. 3H to K and Supplemental Fig. S2).

Communication pathways’ function analysis reveals how changes in communication signals affect the prognosis of neuroblastoma patients

On the basis of clarifying the impact of ten pathways of neurodevelopmental cell communication on the prognosis of neuroblastoma, we have further expanded the capabilities of CCPPM to include the analysis of communication pathways' functions as Fig. 4A shown. This computing module include two aspects: (i) communication pathways’ function analysis, (ii) analysis of pathways specific to tumor subtypes.

Fig. 4.

The main functions of candidate risk related communication pathways. (A) CCPPM pipeline step 6: function analysis of risk related cell communication pathways. (B) The left panel indicates the source and target cells of ten candidate risk related communication pathways in neuroblastoma samples, the linewidth represents average weight of communication pathways, and the color of line represents the source cell. The right panel is the bar plot of enrichment biological functions obtained by gene ontology (GO), and the grey line indicates the correlation between communication pathways and biological functions. (C) Source and target cells of ten candidate risk related communication pathways in fetal adrenal gland control samples. (D) Evolutionary trajectory of SCPs, sympathoblasts and chromaffin cells in fetal adrenal gland control samples. (E) Variation of branch-dependent genes was visualized by heatmap, and the functions of each gene clusters were shown in bar plot. (F) Biological attributes of five basis vectors from NMF of sympathoadrenal cells. (G) Heatmap indicates the similar scores, defined as the average inner products of tumor cells in each neuroblastoma samples with these five basis vectors. (H) The pearson correlation between the neurodevelopmental communication pathways and the differentiation stage of sympathoadrenal cells.

In communication pathways’ function analysis of neuroblastoma, the mean strengths of the top ten communication pathways were calculated through scRNA-seq data and the functional enrichment was preliminary performed to the ligands and receptors of these ten communication pathways (Fig.4B), and the main functions include axon genesis, regulation of nervous system development, and regulation of neuron projection development; some communication pathways were related to immune regulation and cell adhesion. To confirm whether the above functions have changed in neuroblastoma, a comparison is made between communication pathways in neuroblastoma samples and those in fetal adrenal gland samples (Fig. 4C). In fetal adrenal glands, pathways like MDK-PTPRZ1, EFNB2-EPHA4, EFNB3-EPHB6, SEMA6D-PLXNA1, COL6A2-(ITGA1-ITGB1) were the high-strength communication pathways, and these interactions mainly occur in SCPs, sympathoblasts, chromaffin cells, fibroblasts, and endothelial cells. For these pathways, some could promote nervous system development, and others such as SEMA6D-PLXNA1 indicated that the neuron has mature biological functions. However, in neuroblastoma, the high-strength interactions were changed into the CD46-JAG1, ADGRE5-CD55, MDK-PTPRZ1, and BMP7-(BMPR1B-ACVR2B), the target cells were changed into tumor cells, T cells, fibroblasts, and endothelial cells. The strengths of the protective communication pathways diminish, potentially revealing a mechanism that underlies the less favorable prognosis observed in neuroblastoma.

Considering the tumor heterogeneity and multiple subtypes of neuroblastoma [13], we wondered whether the function of these cell communication pathways was active in the specific subtype of neuroblastoma using the tumor subtype specific cell communication pathways’ function analysis. The implementation of the analysis is to construct the evolution trajectory between various cells in the normal control sample, and then evaluate which evolution stage the tumor cells belong to base on similarity score. The cell development trajectory constructed using SCPs, sympathoblasts, and chromaffin cells from fetal adrenal tissue (Fig. 4D), the trajectory branch containing the most SCPs was set as the development root according to known consensus [24]. Along the trajectory, the branch point 1 was most important for SCPs developed to sympathoblasts and chromaffin cells, the genes significantly branch-dependent was calculated using branched expression analysis modeling (BEAM), a heatmap with three similar lineage-dependent expression clusters demonstrates the bifurcation of gene expression along SCPs evolve to sympathoblasts and chromaffin cells (Fig. 4E). Gene enrichment analysis was performed for the branch-dependent genes in each cluster, showed that the evolution of SCPs was regulated by the nervous system development and neuronal migration related process.

To assess the developmental similarity between tumor cells and sympathoadrenal cells at different stages in fetal adrenal glands, we employed a similarity score that quantified the resemblance. The basis for this score was a set of basis vectors designed to capture the key features of sympathoadrenal cells. To obtain these basis vectors, we utilized non-negative matrix factorization (NMF) to decompose the branch-dependent gene expression profiles of sympathoadrenal cells and obtained five basis vectors (SCPs-like, sympathoblasts-like, and chromaffin cells 1 to 3-like) (Fig. 4F). As we progressed from the SCP-like vector to the chromaffin cells 3-like vector, we observed a gradual decrease in the potential for cell differentiation and development, coupled with an increase in cell proliferation and acquisition of specific biological functions. Further, we extracted the branch-dependent gene expression profiles of tumor cells from neuroblastoma samples and calculated the orthogonal projection on the five basis vectors as similarity score (Fig. 4G), the tumor subtypes of all neuroblastoma samples were close to the chromaffin cells1-like and chromaffin cells2-like. Then we calculated the Pearson correlation between the tumor cell attributes in each neuroblastoma sample and the main cell communication pathways related to neuron development (Fig. 4H). Altogether, communication pathways’ function analysis revealed that the activity of protective and detrimental cell communication pathways was different along the tumor cell development status. Specifically, the protective pathways exhibited greater strength at more mature stages, suggesting a potential influence on the prognosis of neuroblastoma. Consequently, the intensity of communication pathways with positive or negative effects dynamically varies at each stage of neuroblastoma, leading to divergent prognoses.

Cell communication post-effect analysis reveal BMP7 signaling is enhanced in neuroblastoma stage 4 and promotes the tumor migration

According to the filtrate through CCPPM prognostic model and communication pathways’ function analysis, among all ten communication pathways in Fig.3B, we further analyzed the function of BMP7-(BMPPR1B-ACVR2B) to explore whether it can accelerate the malignancy of neuroblastoma. To address the limitations of commonly used cellular communication analysis tools in analyzing communication mechanisms and their biological effects, we propose a computational pipeline specifically designed for analyzing the post-effects of specific cell communication pathways (Fig. 5A). Post-effect analysis for one specific communication pathway of CCPPM consists of three steps: (i) examination the extent to which communication pathways affect the disease prognosis; (ii) construction the signal transduction route of communication receptors based on its transcription factors and target genes of transcription factors; (iii) analysis the target effector genes in the cascade of target transcription factor.

Fig. 5.

The post-effect of BMP7-(BMPR1B-ACVR2B) communication pathway. (A) CCPPM pipeline step 7: Cell communication post-effect analysis. (B) Kaplan-Meier curves representing the overall survival according to the strength of BMP7-(BMPR1B-ACVR2B) communication pathway in the training (left) and validation dataset (right). (C) Violin plot displayed the activity of SMAD1 regulon in neuroblastoma samples and fetal adrenal gland control samples. (D and E) Hierarchical clustering tree in tumor cells of neuroblastoma samples and fetal adrenal gland control samples by hdWGCNA. (F) Bayesian network of transcription factors and target genes in neuroblastoma samples (left) and fetal adrenal gland control samples (right) inferred by MMHC. The edge of BMP7 communication pathway and SMAD1 set as the whitelist. (G) Intersection of turquoise module from neuroblastoma samples and blue module from fetal adrenal glands samples. (H) The enriched biological process of intersected genes was shown in the bar plot. (I) The scatter plots indicate the correlation between SMAD1 activity and UNK (left), MYCBP2 (right) expression.

Firstly, we select two neuroblastoma cohorts GSE62464 and TARGET-NBL to examining whether the communication between BMP7 and BMPPR1B-ACVR2B are related to neuroblastoma prognosis. Survival analysis revealed patients who with higher communication strength of BMP7- (BMPPR1B-ACVR2B) has significantly lower survival probability (Fig. 5, Fig. 5). Secondly, we try to figure out the downstream biology function of tumor cells regulated by this communication pathway. According to Cellcall database, we obtained the potential influenced transcription factors and a cascade of target genes of BMP7- (BMPPR1B-ACVR2B). The potential influenced transcription factors were the SMAD transcription factor family, thus we applied the pySCENIC analysis to calculated the activity of each regulon in SMAD transcription factor family. It was SMAD1 regulon was active in tumor cells of neuroblastoma and sympathoadrenal cells of fetal adrenal glands (Fig. 5D), but SMAD5 regulon was only active in fetal adrenal glands (Supplemental Fig. S3), suggest that SMAD1 is more likely involved in the disease mechanism of neuroblastoma. Thirdly, target effector genes in the cascade of SMAD1 was calculated. All activated transcription factors and the cascade target genes of SMAD1 was output to high dimensional weighted correlation network analysis (hdWGCNA) to get gene modules with co-expression relationship of genes (Fig. 5, Fig. 5). Subsequently, we extracted the eigengene levels of the gene modules and the activity of transcription factors in individual cells. To inferring the gene modules which regulated SMAD1, the module-phenotype network (MPN) [25], a omics data analysis method based on Bayesian network structure learning was employed (Fig. 5G). In neuroblastoma, the SMAD1 was regulated by the BMP7 communication pathway and TGFβ, then it regulated the turquoise gene module. In fetal adrenal glands, SMAD1 was regulated by the BMP7 communication pathway, then it regulated the blue gene module.

Through communication post-effect analysis, we identified a gene module in tumor cells regulated by the transcription factor SMAD1, which is mediated by the cascade effect from BMP7 binding to its receptor BMPPR1B-ACVR2B. By conducting enrichment analysis of the genes within this gene module, we can gain insight into the specific biological functions of post-effect. To further capture the target effect genes, we intersected the genes of the turquoise module from neuroblastoma and the blue module from fetal adrenal glands (Fig. 5H). Since we assume that the risk-related communication pathways involved neurodevelopment, the downstream genes should be shared between neuroblastoma and fetal adrenal glands.

We identified 112 shared genes, which were found to be enriched in biological functions such as neuronal migration, axon development, and neuroblast proliferation (Fig. 5I). Next, we intersected the target genes regulated by the SMAD1 regulon in neuroblastoma and fetal adrenal glands, and found two genes, namely UNK and MYCBP2, among the 112 shared genes in the intersection. The correlation between the activity of SMAD1 and the expression of UNK, MYCBP2 is shown in Fig. 5J, supporting the potential regulatory relationship. Considering that the expression level of BMP7 can affect the survival of patients with neuroblastoma, we further analyzed how BMP7-(BMPPR1B-ACVR2B) mediated cell communication and its post-effects impact patient survival. Analysis of two neuroblastoma cohort datasets revealed that the expression level of BMP7 gradually increased during tumor metastasis from stage 1 to stage 4 (Fig. 6A and 6B). BMP7 protein levels in two groups of in situ tumor samples (with and without distant metastasis) were evaluated by immunohistochemistry. As shown in Fig. 6C and D, there were large BMP7-positive area of neuroblastoma patients with distant metastases and the semi-quantitative result of relative BMP7-postive area was significantly higher in patients with distant metastases. Neuron migration marker genes UNC5C, UNC5A, and CORO2B also exhibit higher expression in patients with higher BMP7-(BMPPR1B-ACVR2B) communication strength (Fig. 6E). The detailed clinical information is provided in the Supplemental Table S2.

Fig. 6.

Validating the function of BMP7-(BMPR1B-ACVR2B) communication pathway. (A and B) Comparison of the BMP7 expression levels in different INSS stages in the training dataset (A) and validate dataset (B). (C) The BMP7 immunohistochemistry results of neuroblastoma samples without distant metastases group (top) and with distant metastases group (bottom). (D) Boxplot compared the relative BMP7-positive area in immunohistochemistry of these two groups. (E) Comparison the expression levels of the neuron migration marker genes UNC5C, UNC5A, and CORO2B in the training and validation datasets according to the strength of BMP7-(BMPR1B-ACVR2B) communication pathway.

Discussion

In this study, we constructed a data analysis process CCPPM, using scRNA-seq and bulk RNA-seq data to calculate key cell communication pathways and their post-effects on prognostic indicators in diseases. We addressed the limitations of commonly used cell communication algorithms that cannot associate with prognostic indicators. Though the data of bulk RNA-seq were just about the mean gene expression profiles of several cell types in the tissue, and the cell communication only occurred when the ligands and receptors bind, thus we inferred the communication pathways from the scRNA-seq data. In addition, in this study, the training dataset and validation dataset had a large sample size (over 100), and these risk communication pathways were stable in the model, ensuring the plausibility of cell communication pathways to be constituted into the prognostic models. In future studies, we will integrate the strengths of communication pathways with the activities of transcription factors and the expression profiles of target genes to improve the inference robustness of communication pathways based on large-sample bulk RNA-seq datasets.

Neuroblastoma is characterized by the origin of tumor cells from neural crest cells, which can differentiate into sympathoblasts and chromaffin cells of the peripheral nervous system. Cell communication pathways play a crucial role in regulating neuronal differentiation [26]. Through CCPPM, we mainly explored the role of neurodevelopmental communication pathways in the neuroblastoma progression. By combining scRNA-seq and bulk RNA-seq datasets, we established a prognostic model based on the neurodevelopmental cell communications, and previous studies have proposed these communication pathways can promote the peripheral and central nervous system development [27], [28], [29]. Our results indicated that the strengths of these communication pathways decrease in tumors, compared to those in healthy children at the same developmental stage. We can further explore whether the regulation of cell communication related ligands can promote tumor cell differentiation in neuroblastoma.

Neuronal migration is a critical process in nervous system development, driven by complex mechanisms. Disruptions in these mechanisms by tumor cells may exacerbate tumor progression [30]. However, how these mechanisms work has not been fully elucidated. Several studies have revealed the relationship between neuronal migration and the invasive and metastatic ability of tumor cells. Therefore, the communication between tumor cells and other cells in the nervous system may partially elucidate the pathogenesis of nervous system tumors. In the present study, we found that the BMP7-(BMPR1B-ACVR2B) pathway could accelerate the progression of neuroblastoma, and the BMP7 was expressed highly in the stage 4 neuroblastoma samples. We further clarified that this communication pathway could activate SMAD1, a downstream transcription factor of BMP7-(BMPR1B-ACVR2B) pathway, to promote the migration of tumor cells. The expression level of ligand BMP7 was also validated in situ samples of neuroblastoma patients. The results supported our speculation that this communication pathway can promote neuronal migration.

Conclusion

In summary, CCPPM identified potential communication pathways based on scRNA-seq dataset, integrated with large cohorts of bulk RNA-seq datasets to assess the impact of communication pathways on disease prognosis. and further combined with the dynamic changes of genes in target cells in the scRNA-seq dataset to predict cascades of regulatory mechanisms. The establishment of this general model improves the robustness and scalability of the communication pathways inferred from scRNA-seq data.

Ethics approval and consent to participate

The study conformed to the standards set by the Declaration of Helsinki and was approved by the Ethics Committee of Children's Hospital of Nanjing Medical University (NO.202305006-1). The study was undertaken with the understanding and written consent of each participant.

Consent for publication

Not applicable.

Availability of data and materials

The data that supports the findings of this study are available from the corresponding authors upon reasonable request.

Funding

This study was jointly supported by the National Natural Science Foundations of China (82204352), the Guiding Project of Jiangsu Provincial Commission of Health and Family Planning (Z201611), the Jiangsu Funding Program for Excellent Postdoctoral Talent (2023ZB043 and 2022ZB431), Nanjing Postdoctoral Research Funding Program (323721), Nanjing Medical Science and Technology Development Project (YKK18138), Medical Project for Key Youth Talents of Jiangsu Province (QNRC2016085), Technology Development Foundation of Nanjing Medical University (NMUB20220015 and NMUB20210361), and Women and Children Key Discipline on Children's Hematology Oncology of Jiangsu Province (FXK201742).

CRediT authorship contribution statement

Jiali Wang: Conceptualization, Data curation, Formal analysis, Funding acquisition, Methodology, Software, Writing – original draft. Huimin Li: Formal analysis, Investigation, Validation. Yao Xue: Data curation, Investigation, Validation. Yidan Zhang: Investigation, Validation. Xiaopeng Ma: Investigation, Validation. Chunlei Zhou: Investigation, Validation. Liucheng Rong: Conceptualization. Yixuan Zhang: Supervision, Visualization, Writing – original draft, Writing – review & editing, Project administration, Conceptualization, Funding acquisition. Yaping Wang: Conceptualization, Project administration, Supervision, Writing – original draft, Writing – review & editing. Yongjun Fang: Project administration, Writing – review & editing.

Declaration of competing interest

The authors declare no competing interests.