Machine learning and deep learning to identifying subarachnoid haemorrhage macrophage-associated biomarkers by bulk and single-cell sequencing

Abstract

We investigated subarachnoid haemorrhage (SAH) macrophage subpopulations and identified relevant key genes for improving diagnostic and therapeutic strategies. SAH rat models were established, and brain tissue samples underwent single-cell transcriptome sequencing and bulk RNA-seq. Using single-cell data, distinct macrophage subpopulations, including a unique SAH subset, were identified. The hdWGCNA method revealed 160 key macrophage-related genes. Univariate analysis and lasso regression selected 10 genes for constructing a diagnostic model. Machine learning algorithms facilitated model development. Cellular infiltration was assessed using the MCPcounter algorithm, and a heatmap integrated cell abundance and gene expression. A 3 × 3 convolutional neural network created an additional diagnostic model, while molecular docking identified potential drugs. The diagnostic model based on the 10 selected genes achieved excellent performance, with an AUC of 1 in both training and validation datasets. The heatmap, combining cell abundance and gene expression, provided insights into SAH cellular composition. The convolutional neural network model exhibited a sensitivity and specificity of 1 in both datasets. Additionally, CD14, GPNMB, SPP1 and PRDX5 were specifically expressed in SAH-associated macrophages, highlighting its potential as a therapeutic target. Network pharmacology analysis identified some targeting drugs for SAH treatment. Our study characterised SAH macrophage subpopulations and identified key associated genes. We developed a robust diagnostic model and recognised CD14, GPNMB, SPP1 and PRDX5 as potential therapeutic targets. Further experiments and clinical investigations are needed to validate these findings and explore the clinical implications of targets in SAH treatment.

Keywords: deep learning; hdWGCNA; machine learning; single‐cell sequencing; subarachnoid haemorrhage rat model.

FIGURE 1

The flow chart of the study.

FIGURE 2

Bioinformatics analysis of single cell transcriptome. (A) The uniform manifold approximation and projection (UMAP) of cell types from SAH and sham samples. (B) A bar chart showing the proportion of various cell types from SAH and sham samples. (C) The uniform manifold approximation and projection (UMAP) of subclusters of Macrophages from SAH and sham samples.

FIGURE 3

Strength and pathways involved in cell–cell interactions from SAH samples. (A) Interaction net count plot of right temporal cortex cells from SAH samples. The interaction weight plot of each cells. The thicker the line represented, the more the number of interactions, and the stronger the interaction weights/strength between the two cell types. (B, C) Summary of selected ligand–receptor interactions between different cell clusters from SAH samples, respectively. p‐Values (permutation test) are represented by the size of each circle. The colour gradient indicates the level of interaction. (D) The analysis of intercellular communication networks using CellChat revealed the inferred interactions involved in CCL signalling pathways. The circle plot visualizes the intercellular communication network for these pathways, highlighting the ligand‐receptor pairs and their connections between different cell populations. (E) This analysis quantifies the relative importance of cell groups in CCL signalling networks using network centrality measures. Influencer cells regulate information flow, while gatekeeper cells control communication between cell groups. Importance is based on sender, receiver, mediator and influencer roles. Darker colours indicate greater involvement in these roles. (F) Heatmaps of different signals contributing mostly to outgoing or incoming signalling of each cell population. (G) The incoming and outgoing strength of each cell population under SAH samples.

FIGURE 4

Identification of 160 key genes highly correlated with SAH specific macrophage (SSM). (A) Determination of soft‐threshold power in the WGCNA. (B) Dot plot showing expression of genes in each module in different cluster of macrophage. The size of the dot indicates the percentage of cells within a cell type in which that marker was detected, and its colour indicates the average expression level. (C) Single‐cell sequencing analysis results show the expression in different module eigengenes in macrophages. (D, E) Cell trajectory maps of macrophages highlighting the contribution of cells coming from each state (D) and each cluster. (F) Heatmap showing expression of 160 key genes highly correlated with SSM across single cells. Colour key from blue to red indicates relative expression levels from low to high.

FIGURE 5

Protein–protein interaction network and Functional analysis. (A) PPI network of 160 key genes highly correlated with SAH specific macrophage (SSM) from hdWGCNA analysis. (B) Bubble graph for Gene Ontology (GO) enrichment. (C) Bubble graph for Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment (the bigger bubble means the more genes enriched, and the increasing depth of red means the differences were more obvious; q‐value: the adjusted p‐value).

FIGURE 6

Machine learning for diagnostic model. (A) LASSO coefficient profiles of the 10 SAH‐related genes. (B) A coefficient profile plot was produced against the log (lambda) sequence in the LASSO model. The optimal parameter (lambda) was selected as the first black dotted line indicated. (C) Mean Area Under The Curve (AUC) of the seven machine learning algorithms based on the training set. (D) Receiver operating characteristic (ROC) curve of the random forest algorithm based on the validation set. (E) qPCR showed the gene expression of CD14, SPP1, PRDX5 and GPNMB in the right temporal lobe cortex of the SAH group compared to the sham group. (p = 3, ****p < 0.001 vs. sham group, p > 0.05 vs. sham group; t‐test; mean ± SD).

FIGURE 7

Deep learning for diagnostic model. (A) Bar plots showing distributions of gene set score of 160 key genes related to SAH specific macrophage (SSM) from hdWGCNA analysis in normal and SAH samples. (B) Broad co‐expression network exists between the 10 key genes and the 10 cell types. Red shows a positive correlation and blue shows a negative correlation. *p < 0.05, **p < 0.01, ***p < 0.001. (C) Receiver operating characteristic (ROC) curve of the convolutional neural network algorithm based on the training set. (D) ROC curve of the convolutional neural network algorithm based on the validation set.

FIGURE 8

Molecular models of each target gene binding to its predicted drug targets. Acetaminophen target CD14 (A), Cyclosporine target CD14 (B), Hydrocortisone target CD14 (C), Vancomycin target CD14 (D), Azacitidine target GPNMB (E), Dexamethasone target GPNMB (F), Doxorubicin target GPNMB (G), Estradiol target GPNMB (H), Acetaminophen target SPP1 (I), Aspirin target SPP1 (J), Ethinyl Estradiol target SPP1 (K), Tamoxifen target SPP1 (L), Progesteron target PRDX5 (M).