The GEO database was searched for intracerebral hemorrhage-related RNA transcriptomic datasets based on the following keywords: “intracerebral haemorrhage” and “brain haemorrhage”. Exclusion criteria were set as follows: (1) transcriptomic data from animal models or knockout animals; (2) brain hemorrhage caused by vascular malformations, aneurysms, etc.; and (3) drug experiments designed. A dataset containing mixed plasma samples from 15 patients with ICH and eight healthy controls (GSE43618 dataset) was filtered based on the filtering criteria (33). Peripheral blood mRNA transcriptomic data were obtained from GSE125512 for 11 patients with ICH at the onset and 3 days after onset (34). These two datasets were used to identify differentially expressed miRNAs (DEmiRNAs) and differentially expressed mRNAs (DEmRNAs). These DEmRNAs were derived from an analysis of differences between peripheral blood transcriptome expression profiles 3 days after and at ICH onset. The transcriptome data were log-2 transformed with different unit formats, and de-batching between samples was performed. Finally, we collected peripheral blood samples from 20 patients with an ICH and 17 healthy control volunteers in order to validate the genes related to inflammation and hypoxia identified by the bioinformatics analysis. Included patients must meet the following criteria: (1) experience acute cerebral hemorrhage within 3 days; (2) have clear diagnostic imaging and laboratory results; (3) have no vascular malformations, coagulation disorders, or other causes of bleeding; and (4) be free of malignant tumors and other serious diseases. Ethics approval for this study was obtained from the General Hospital of Ningxia Medical University. All the participants provided written informed consent.

In order to select a broad list of candidate gene sets for inflammation- and hypoxia-related genes, we searched the Kyoto Encyclopedia of Genes and Genomes (KEGGs) database (www.kegg.jp). Furthermore, the PubMed and Web of Science databases were searched for inflammation- or hypoxia-related gene complements. Ultimately, a total of 50 hypoxia-related genes and 200 inflammation-related genes were identified. These genes are mainly involved in inflammatory or hypoxic response processes during disease development.

First, the limma R package was used to identify DEGs, including DEmiRNAs and DEmRNAs, in the occurrence of ICH (35). We identified mRNAs that are up- and down-regulated in patients with ICH 3 days after cerebral hemorrhage compared with those at the onset of cerebral hemorrhage using differential expression analysis of the peripheral blood transcriptome. DEGs were filtered using p < 0.05 as the threshold. Data were then filtered by genes using Perl (https://www.perl.org/) to obtain dysregulated genes associated with inflammation or with hypoxia and the corresponding gene expression matrix.

Functional pathway analysis of up- or down-regulated DEmRNAs using the Kyoto Encyclopedia of Genes and Genomes (KEGGs) and Gene Ontology (GO). GO terminology is described in three parts: biological processes (BPs), cellular components (CCs), and molecular functions (MFs). The clusterProfiler R package was used to complete GO and KEGG analysis as previous research (36, 37). Further, GSEA is used to analyze functional pathways of dysregulated biology involving key genes (38). Selected functional pathways for differential analysis are referenced from “c2.cp.v7.2.symbols.gmt [Curated]” in MSigDB collections (https://www.gsea-msigdb.org/gsea/msigdb/) gene set. The threshold used to identify dysfunctional pathways was set at a false discovery rate of <0.25 and adjusted p of <0.05.

According to the previous study, the miRNA–TF–mRNA network was constructed (39–42). First, TF–mRNA relationship pairs were predicted in DEmRNAs using the TRRUST (v2) database (43), where mRNAs were associated with hypoxia or inflammation. To explore regulatory relationships between DEmiRNAs and DEmRNAs (including TFs), miRWalk was used (http://mirwalk.umm.uni-heidelberg.de/>). R software was used to match the interrelationships between DEmiRNAs and DEmRNAs. Among the predicted miRNA–mRNA molecular pairs, only those with opposite regulatory directions were subjected to further analysis. Furthermore, Cytoscape was used to visualize the entire miRNA–TF–mRNA interaction network and to identify the hub genes in the network based on the number of connections in each node (44).

Intelligent machines are converging with advancing biotechnologies to shape the future of medicine (45). Machine learning is used to screen genes associated with the progression of ICH. Support Vector Machine–Recursive Feature Elimination (SVM–RFE) was used to investigate genes associated with hypoxia and inflammation (46). SVM is excellent at handling small datasets and shows good classification performance. Redundant genes are filtered using the iterative algorithm of SVM–RFE, resulting in genes highly correlated with the outcome. Furthermore, the least absolute shrinkage and selection operator (LASSO) is used for gene screening. As previous research, LASSO analysis was performed using the “glmnet” R package (42, 47). ROC curves were used to assess the predictive ability of core genes to distinguish between patients with ICH at different progression stages, thereby testing their reliability for the outcome prediction. Principle component analysis and t-distributed stochastic neighbor embedding (t-SNE) (48) were used to demonstrate the ability of screened core genes to classify patients with ICH at different developmental stages. The R package “Rtsne” was used to implement the t-SNE algorithm based on non-linear dimensionality reduction.

Total RNA was extracted using TRIzol (TaKaRa Bio, Shiga, Japan) and reverse-transcribed into cDNA using PrimeScript RT Master Mix (TaKaRa Bio, Shiga, Japan). According to the manufacturer's instructions, Real-time PCR was performed using SYBR Green PCR Master Mix (Takara). Using GAPDH as a reference, the 2^−ΔΔCt method was applied for the relative quantification of core gene expression levels and normalized.

All the drawings are performed using the R software (version 4.0.5). The “VennDiagram” R package is used to create a Venn diagram to present the results of the gene intersection analysis. Spearman's correlation test was used to assess the correlation between key genes, and correlation coefficients of >0.3 were considered to be co-expression relationships. Differential analysis of gene expression between the two groups was performed using the Wilcoxon rank-sum test. Unless otherwise indicated, p < 0.05 was considered a statistically significant difference.

To obtain biomarkers and pathways associated with the occurrence or progression of ICH, we performed a differential expression analysis. First, 46 DEmiRNAs were identified by differential analysis between ICH and healthy control (HC) (Figure 1A). Based on the transcriptomic data obtained from the peripheral blood of ICH at different development stages, 914 DEmRNAs were identified. Compared to the onset of ICH, 444 genes were upregulated and 470 genes were downregulated 3 days after the onset of ICH (Figure 1B). These up- and downregulated genes were then subjected to separate functional pathway analyses. The main pathways shown to be upregulated in ICH by GO and KEGG included neutrophil-mediated immunity, secretory granule lumen, cell adhesion molecule binding, and regulation of actin cytoskeleton (Figure 1C). In addition, the pathways enriched by the downregulated genes contained RNA splicing, nuclear speck, Herpes simplex virus 1 infection, and condensed chromosome (Figure 1D). These results suggest that neutrophils play an important role in the development of ICH.

To explore the possible role of inflammation- or hypoxia-related genes in ICH development, we further screened the inflammation- and hypoxia-related genes separately in the DEmRNAs. We identified 15 genes associated with hypoxia; their relative expression between ICHs is shown in Figure 2A. The number of genes related to inflammation was 16 (Figure 2B). The corresponding heat map showed a clear boundary between hypoxia or inflammation-related DEGs in patients with different stages of ICH. This suggests that the genes associated with hypoxia or inflammation play an important role in ICH development.

From all the DEmRNAs, eight differentially expressed transcription factors (DETFs) were identified, which can regulate some of these genes in DEGs associated with hypoxia or inflammation. Furthermore, 10 DEmiRNAs were predicted to potentially act on these DEmRNAs (or DETFs) associated with hypoxia or inflammation. And the relationship pairs between these DEmiRNAs and DEmRNAs (including DETFs) were constructed as an miRNA–TF–mRNA interaction network (Figure 2C). From this molecular interaction network, SP3 was identified as the hub gene in the network, as it has the highest number of linkages (Figure 2D). From the constructed molecular interaction network, we further identified a possible correlation between SP3 and TIMP1 and PLAUR. TIMP1 is an inflammation-related gene, while PLAUR is a gene associated with hypoxia and inflammation. Therefore, the hub gene SP3 might be a transcription factor associated with hypoxia and inflammation.

To evaluate the genes associated with hypoxia or inflammation that are closely related to ICH development, SVM–RFE and LASSO analyses were applied to screen DEGs. SVM–RFE showed minimal error in outcome prediction when all 31 hypoxia- or inflammation-related DEmRNAs were selected (Figure 3A). In a subsequent analysis, the Lasso analysis identified 13 hypoxia- or inflammation-related genes that were strongly associated with ICH development (Figure 3B). Figure 3C shows the co-expression network of these 13 genes, and TIMP1 and PLAUR were found to be co-expressed with several genes. Moreover, PCA and tSNE visualizations demonstrate that these 13 genes can be used to distinguish patients with ICH at different stages (Figures 3D,E).

An intersection analysis of hypoxia- or inflammation-related DEGs obtained from machine learning and PPI networks was performed, and four core inflammation- or hypoxia-related DEGs (TIMP1, PLAUR, DDIT3, and CD40) were identified (Figure 4A). The relative expression profiles of these four core genes in patients with ICH are shown in Figure 4B. The ROC curves for PLAUR (AUC = 0.777), DDIT3 (AUC = 0.669), CD40 (AUC = 0.727), and TIMP1 (AUC = 0.719) for the different stages of ICH development are shown in Figures 4C–F, respectively, indicating their moderate predictive power. PCA and tSNE visualization showed that patients with different stages of ICH can be distinguished based on these four genes (TIMP1, PLAUR, DDIT3, and CD40) (Figures 4G,H). Therefore, four biomarkers (TIMP1, PLAUR, DDIT3, and CD40) were screened for their association with inflammation or hypoxia by machine learning.

In the regulatory network, we discover that RUNX1 and SP3 may act as transcriptional regulators of TIMP1 (Figure 2C). RUNX1 is regulated by hsa-miR-940, and SP3 is regulated by hsa-miR-571 (which has the most junctions). The magnitudes of the fold change values for hsa-miR-940 and hsa-miR-571 are shown in Figure 5A. The relative ranking of the fold change values for TIMP1, PLAUR, DDIT3, and CD40 are shown in Figure 5B. Further correlation analysis showed a positive correlation between RUNX1 and TIMP1 (r = 0.24, Figure 5C) and a negative correlation between SP3 and TIMP1 (r = −0.381, Figure 5D). Although there was no statistically significant difference, more cellular research are needed. Therefore, we further predicted two miRNA–TF–mRNA axes around the inflammation-related gene TIMP1, namely, hsa-miR-940/RUNX1/TIMP1 and hsa-miR-571/SP3/TIMP1 (Figure 2C). Validation in an independent clinical cohort showed statistically significant differences in TIMP1 expression between ICH and HC (p < 0.01, Figure 5E).

In the molecular interaction network (Figure 2C), the transcription factors of the hypoxia-related gene PLAUR were SP3 and FOS. SP3 also exhibited a regulatory effect on the inflammation-related gene TIMP3. Based on a condition of >0.3 correlation coefficient, we found no co-expression between FOS and PLAUR (r = 0.08, p = 0.282, Figure 5F), while there was some negative correlation between SP3 and PLAUR (r = −0.348, Figure 5G). Therefore, we identified SP3 as a key hypoxia- and inflammation-related transcription factor. In addition, hsa-miR-571/SP3/PLAUR was constructed around SP3 and PLAUR as an axis miRNA-TF-mRNA (Figure 2C). There was a significant difference in PLAUR expression between ICH and HC based on the data from an independent clinical cohort (p < 0.01, Figure 5E).

We performed GSEA analysis of the transcription factors SP3 and RUNX1 to explore the pathways in which they might be involved. GSEA showed that the signaling by interleukins, hemostasis, platelet activation signaling and aggregation, and mitotic prometaphase pathways were upregulated in patients with ICH with SP3 downregulation (Figure 6A). In patients with ICH with elevated RUNX1 expression, elastic fiber formation was downregulated, while the transcriptional regulation of the granulopoiesis pathway was upregulated (Figure 6B). The aforementioned results indicate that SP3 may be involved in interleukin signaling and platelet activation for hemostasis, while RUNX1 may be involved in granulopoiesis after ICH onset.

In patients with ICH with PLAUR upregulation, mitotic prometaphase and cell cycle checkpoints were downregulated, while interferon signaling and hippoyap signaling pathway were upregulated (Figure 6C). In patients with ICH with elevated expression of TIMP1, the platelet activation signaling and aggregation, insulin signaling pathway, and angiogenesis pathways were upregulated (Figure 6D). PLAUR and TIMP1 were both upregulated 3 days after ICH onset. Therefore, it is hypothesized that PLAUR, a hypoxia and inflammation-related gene, may be involved in upregulating the interferon signaling pathway and hippo pathway, while TIMP1 may be involved in platelet activation and aggregation, activating insulin signaling pathways, and promoting angiogenesis.