The original chip data for this experiment were provided by the Department of Neurosurgery, Southwest Hospital, PLA Military Medical University. SAH and sham-operated models were established; details are given in the Supplementary Material. Each experimental group included five mice, and the white matter area of the cerebral cortex was taken for gene chip testing. A total of 10 original chip samples were obtained from the SAH intervention experiments; these were divided equally into two groups as follows.

(1) SAH disease group: brain tissue in the white matter region of the cerebral cortex of SAH mice.

(2) Control group normal-1: brain tissue in the white matter region of the cerebral cortex of normal mice.

The chip was an Affymetrix GeneChip Mouse Gene 1.0 ST Array. Raw data included sample RNA extraction (white matter brain cells from the SAH model and from normal mice), sample RNA quality detection (total RNA>1 ug), cDNA synthesis, sense strand cDNA fragmentation, biotin labeling, chip hybridization, chip elution, and chip scanning. The raw data are available at http://www.ebi.ac.uk/arrayexpress/experiments/E-MTAB-8407.

We then carried out mass analysis and used the R Bioconductor package to perform quality control for each original chip (the SAH disease group and the control group normal-1). In the output gray scale image (Figure S1) for each chip sample, each chip name and the four corner patterns were very clear, and the contrast between light and dark was moderate.

The right panel of Figure 1A shows the Relative Log Expression (RLE) boxplot for these 10 chips. The center of each sample was close to the position RLE = 0. This indicates that the expression levels of most genes in the sample were consistent. In addition, Figure S2 describes a normalized unscaled standard errors (NUSE) detection (Marta and Marc, 2014). Since Figure S2 shows that the center of each sample is close to the position NUSE = 1, we consider that the samples are too stable to have obvious batch effect. Then, we used Robust Multi-chip Analysis (RMA) (Irizarry et al., 2003) for data preprocessing, including background and perfect match probes (PM) correction, normalization, and summarization, to obtain the probe expression data matrix (Table S1). Finally, clustering analysis (Liu et al., 2019; Xiao et al., 2019; Zhang et al., 2019c; Wu and Zhang, 2020) (Figure S3) shows that the major differences between the chip of each group comes from SAH.

Here, in order to interfere with the expression of LCN2, 2 μL of specific short interfering RNAs (siRNAs) was delivered into the lateral ventricle with a Hamilton syringe. The injection was performed 48 h before SAH and three groups were used, as described below. We detail the procedures in the Supplementary Material.

(1) SAH-siRNA-LCN2: the SAH model was established and treated with intrathecal injection of LCN2 siRNA, and two samples were taken on the first and third days after surgery.

(2) SAH-siRNA-NC: the SAH model was established and treated with intrathecal NC siRNA, and two samples were taken on the first and third days after surgery, which helped us to remove the interference factors associated with the siRNA vector.

(3) Control group normal-2: the brain tissue of the white matter region of the cerebral cortex without any treatment.

The total number of samples in all experiments was 25 (Table 1). RNA sequencing was performed on the samples and the raw data are available at https://www.ncbi.nlm.nih.gov/sra/PRJNA575372.

We used E-Bayes, one of the most commonly used methods for differential expression analysis (Edwards et al., 2005), to screen the differential genes by setting Fold change ≥ 1.5 and p-value < 0.05. Table S2 lists 2942 differentially expressed genes, accounting for 10.16% of the total number of genes (28,944). Among them, there were 1016 and 1926 genes with upregulated and downregulated expression (Figure S4), respectively.

We used Equation 1 and the data in Table S3 to explore related signaling pathways by carrying out Fisher's exact test (Xia et al., 2017) using Kobas 3.0 (Wu et al., 2006; Xie et al., 2011; Ai and Kong, 2018) for the differentially expressed genes from Table S2.

Here, N is the number of genes in the sample and n is the number of genes contained in the pathway. N_f is the number of differentially expressed genes and n_f is the number of differentially expressed genes included in the pathway.

The Fisher's exact test assumes H₀:p₁ = p₂; the alternative hypothesis is H₁:p₁ ≠ p₂. p₁is the probability that the differentially expressed gene will fall in the pathway, and p₂ is the probability that the non-differentiated gene does not fall in the pathway. The p-value (p_F) of Fisher's exact test was obtained by Equation 1.

Table S2 lists 70 signaling pathways for which the p-value was less than 0.001. LCN2 is a protein involved in MAPK signaling pathways that protects the CNS as part of the innate immune system (Warszawska et al., 2013). Previous studies have shown that LCN2 activates phosphorylation of p38 MAPK, which phosphorylates the Ser168 and Ser170 sites of NFATc4 and inhibits nuclear translocation of NFATc4 (Olabisi et al., 2008). NFATc4 is a key factor in remyelination and closely related to SAH, indicating that white matter damage after SAH is associated with remyelination (Kao et al., 2009; Guo et al., 2017).

Therefore, we hypothesize that LCN2 could promote the phosphorylation of transcription factor NFATc4 and inhibit its nuclear transcription by activating p38 MAPK, thereby preventing remyelination and causing white matter damage after SAH.

To prove our hypothesis, we designed a LCN2 intervention experiment (Figure 1B) to test whether LCN2 could affect SAH from the perspective of the differential expressed genes and the related signaling pathways.

First, we used the DESeq2 (Varet et al., 2016) method to select differentially expressed genes from SAH-siRNA-LCN2 and normal-2, SAH-siRNA-NC and normal-2, and SAH-siRNA-LCN2 and SAH-siRNA-NC groups on days 1 and 3, respectively (Table 1). The results are shown in Table 2, Table S4, and Figure S5.

Next, we used Kobas 3.0 (Wu et al., 2006; Xie et al., 2011; Ai and Kong, 2018) to carry out Fisher's exact test for the differential genes in Table 2, to identify related signaling pathways (Table S5). Next, we used the manually reviewed evidence (Osuka et al., 2006; Majdalawieh et al., 2007; Hanafy et al., 2010; Hao et al., 2014; Kwon et al., 2015; Yu et al., 2018) to cross-validate the SAH-related signaling pathways in Table S5. Table 3 lists the cross-validated SAH-related signaling pathways.

As shown in Table 3, all the experimental groups had SAH-related signaling pathways except the transcriptional misregulation in cancer signaling pathway (Lee and Young, 2013) in the SAH-siRNA-LCN2 (3 day) vs. SAH-siRNA-NC (3 day) experimental group. However, as one of the proteins from this pathway, Gzmb (Table S5), is closely associated with post-ischemic brain cell death (Chaitanya et al., 2010), we consider that it could be a new target for secondary brain injury inhibition (Armstrong et al., 2017). Therefore, we conclude that specific intervention for LCN2 is a promising SAH treatment strategy.

This study used three classification algorithms, LR (Hosmer et al., 2013), SVM (Suykens and Vandewalle, 1999), and Naive-Bayes (Wang et al., 2007) to develop the SAH prediction model, using the selected key genes as the respective classifiers. These three classic methods were then integrated into a novel ensemble learning model to improve the predictive accuracy.

Figure 3 shows the workflow of the SAH prediction model, based on our previous studies (Li et al., 2017; Xia et al., 2017; Zhang et al., 2019b). The key equations of the model are as follows.

Here, D_t(i) is the weight distribution, t is the iteration time, i is the index of the sample, and n is the number of the sample. ε_t and α_t are the error rate and weight of each weak classifier h_t, respectively. For a sample set S = { (x₁, y₁), (x₂, y₂), …, (x_n, y_n) }, x_n are the samples and y_n ∈ {0, 1} are the labels; y_i=0 indicates that x_i is not an SAH patient, and y_i=1 indicates that x_i is an SAH patient. H_m is the homomorphic integration for each weak classifier h_t; m is the index of the weak classifier, m = 1, 2, 3; T is the threshold of the iteration time; P_Hm is the predictive probability of disease; and E_Hmis the estimated probability of the model H_m. Y (x) is the result of the final classifier obtained by a voting method (Dietterich, 2000).

Figure 4A compares the classification performance for the LR, Naive-Bayes, SVM, and ensemble learning models, based on four commonly used classification measurements (Table S10) (Zhang et al., 2019b). The numerical values used in Figure 4A are listed in Table S11; these demonstrate that the ensemble learning method outperforms the other three methods with respect to accuracy, precision, sensitivity and specificity. The ROC chart plotted in Figure 4B compares the classification effects of LR, Naive-Bayes, SVM, and ensemble learning models. The classification effect of ensemble learning models is also superior to the other three.