We query stroke in DisGeNET and obtain a total of 14 stroke-related IDs that are consistent with the IDs of international classification of diseases (ICD10). The IDs are as follows: C0948008, C1857287, C3178801, C0242129, C0262469, C1112433, C1299567, C1298680, C0265113, C0038454, C0740392, C0751956, C0553692, and C3536593. Among these IDs, C0553692 is related to hemorrhagic stroke. The genes obtained from these above 14 IDs were treated as pathogenetic genes of stroke (Supplementary Table S1).

To construct comprehensive weighted gene network of stroke, the protein–protein interactions (PPI) data were collected from public web servers CMGRN and PTHGRN (Guan et al., 2014a; Guan et al., 2014b). Pathogenetic genes with evidence number from DisGeNET (Pinero et al., 2017) were mapped to the PPI network and used in constructing the weighted gene regulatory network of stroke. Cytoscape (Version 3.7.2) was utilized to visualize the network.

All herbal components of XXMD were extracted from literatures and two published natural product databases: traditional Chinese medicine systems pharmacology database and analysis platform (TCMSP) (Ru et al., 2014) and the Encyclopedia of Traditional Chinese Medicine (ETCM) (Xu H.-Y. et al., 2019).

The potential active components of XXMD were selected from the TCMSP database and those detected in herbs. The properties of all components of XXMD were retrieved from TCMSP, including molecular weight (MW), oral bioavailability (O B) (Sugumaran et al.), Caco-2 permeability (Caco-2), and drug-likeness (Olson et al.) (Xu et al., 2012). Three absorption, distribution, metabolism, and excretion-related (ADME-related) models, including OB (Sugumaran et al.), Caco-2, and DL, were employed to screen the components. Components with the properties of OB ≥ 30%, Caco-2> −0.4, and DL ≥ 0.18 (Zhang et al., 2020c; Feng et al., 2021; Wang et al., 2021) were considered the potential active components. Owing to high concentration and high biological activities reported in literatures, some components in XXMD that did not meet ADME-screening criteria were also manually selected and used in conjunction with components from ADME screening in the following study (Table 1).

To obtain the targets of potential active components in XXMD, Open Babel toolkit (version 2.41) was used in the conversion of all chemical structures into canonical SMILES. After that, the commonly used tools, i.e., Similarity Ensemble Approach (SEA) (Tao et al., 2013), HitPick (Liu et al., 2013), and SwissTarget Prediction (Gfeller et al., 2014), were employed to predict targets of potential active components based on canonical SMILES.

Node importance is an important topological property and can be used to evaluate the influence of nodes among the network. The nodes whose node importance is larger than the average node importance of all nodes are treated as critical roles and hub nodes in the network (Liu et al., 2016). Here, we designed a novel node importance calculation method to figure out the importance and the influence of genes. According to this rule, these nodes with higher importance scores than average importance score are kept and integrated with their edges to form functional response space (FRS). The detail of the method is described as follows: Nim (s)   =   [ ∑ s   ≠ v   ≠ t   ∈ V σ v t   ( s ) σ v t   ]   × ∑ s ≠ x 1 d ( s ,   x ) 2

Nim represents the node importance; σ v t represents the number of the shortest path between node v and node t, σ v t   ( s ) is the number of the shortest path passing through node s from node v to t; x, v represent nodes (genes); d ( s , x ) represents the shortest distance (minimum number of edges) when s and x are connected.

After being quantized, Nim ( s )   was sorted from small to large, and was represented by a new variable Q. Q = [ Q 1 ,   Q 2 ,   Q 3 ,   … ,   Q |   V   | ] =   [ Nim ( s ) x ,   Nim ( s ) x + 1 ,   Nim ( s ) x + 2 ,   … ,   Nim ( s ) x + n ] ,    x ∈ [ 1 ,   |   V   | ]    and    x + n = |   V   |

The new variable G represented the nodes in the network. Each G responds to its unique Q. G   ∈   { G 1 ,   G 2 ,   G 3 ,   … ,   G |   V   | }   ⇌   [ Q 1 ,   Q 2 ,   Q 3 ,   … ,   Q |   V   | ]

FRS represent the set of genes that were selected with our proposed method from the unite of pathogenetic genes and XXMD targets. N represented natural number. g represented a node g. FRS   ∈ { { G g   ,   … , G ( |   V   | − 2 )   ,   G ( |   V   | − 1 )   ,   G |   V   | }   ⇌   [ Q ( |   V   | + 1 ) / 2   ,   … , Q ( |   V   | − 2 )   ,   Q ( |   V   | − 1 )   ,   Q |   V   | ] ,         |   V   | = 2 p + 1   a n d   p ∈ N     { G g   , … ,   G ( |   V   | − 2 )   ,   G ( |   V   | − 1 )   ,   G |   V   | }   ⇌   [ ( Q |   V   | / 2 +   Q ( |   V   | + 2 ) / 2 ) 2   , … ,   Q ( |   V   | − 2 )   ,   Q ( |   V   | − 1 )   ,   Q |   V   | ] ,         |   V   | = 2 p     a n d   p ∈ N

To optimize effective components and get the KFCG, which would be used to illustrate the potential molecular mechanism of XXMD in the therapy of stroke, we designed a components contribution ratio (CCR) model to select KFCG:

For x ∈ ( 1 : n ) do

if W k x = max   { W k 1   :   W k n } then ∪ k 1   :   k n =   ∪ k 1   :   k n −   k x ∪ k x =   ∪ k x +   k x   else if W_kx = W_ky = max { W_k1: W_kn } and ∪ k x ≥   ∪ k y   then ∪ k 1   :   k n =   ∪ k 1   :   k n −   ∪ k x + k y   ∪ k x =   ∪ k x +   k x   End if W ∪ k x   ≥ 90 % End if ∪ k 1   :   k n =   Ø Return KFCG =   ∪ k x   j represent genes in FRS; k represent the components, which is corresponding to genes in FRS; k_n represent the nth component; U represent unite; U_k represent unite of component responding to genes in FRS; U_kn represent the targets unite of k_n; W represent the coverage of targets in U_j.

To quantify the comprehensive function of topological affect and control affection of KFCGs in the C-T-P network, we designed a component potential effect score calculation model. Usually, higher degree of components represents greater influence. The higher degree of the neighbor nodes of a component represents the higher control affection of this component in the topology of the network. It means that a component has a stronger control affection in the network when this component has much more targets and signaling pathways. Based on the above properties, we designed a component potential effect score calculation model, which was considered with both the network topology importance of KFCGs and the functional control ability of KFCGs. The specific model is as follows: PES =   ( Σ ( k n ) )   ×   [ |   V   (   k ( p ) )   | |   V   |   ×   ∑ w ∈ k ( p ) (   Δ k ( p ) + 1 − d i s t ( p , w ) ) max   { d i s t ( p , w ) : w   ∈ k ( p ) } ] 2

We defined the 56 components of KFCG and its 585 targets as node n. Then, the neighborhood of a node n is the set of nodes sharing an edge with n. The connectivity of node n, k _n , is the size of its neighborhood. The degree of node n is the number of edges reaching n, which is equivalent to k _n . k _(p) represents the component k of KFCG which can reach the pathway p. The relationship between k and pathway p contains two cases that component k effect the pathway p through its directive targets and through the proteins which can interact with its targets indirectly. V represents the collection of nodes within the network. |V| represents the number of nodes. Δ k ( p ) is the maximum distance between component k and other genes passing through pathway p. dist _{(p, w)} represents the length of a shortest path between pathway p and component w. The dist _{(p, w)} is equal to infinite if C_(p) ≠ C_(w) , and it makes methods of this category cannot be applied to networks with disconnected genes. Finally, PESs were normalized into 0–1.

Computer-simulated modeling can contribute to the prediction in the likelihood of molecular interactions. We obtained the 3D conformer of XXMD KFCG from ZINC (https://zinc.docking.org /) and PubChem (https://pubchem.ncbi.nlm.nih.gov /) and obtained proteins coded by genes involved in the comprehensive pathways. The affinity method and pyMOL were conducted in docking simulations and graph creation, respectively.

For analyzing the main function of targets, the clusterProfiler package of R software was used to perform GO analysis (Yu et al., 2012) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis (Draghici et al., 2007). According to the reports in using clusterProfiler package, the cutoff of p-values < 0.01 has the higher acceptance, and was used in the following enrichments analysis (Yu et al., 2012). The ggplot2 package and the Pathview (Luo et al., 2017) of R software were used in the graph creation and gene annotation, respectively.

Fetal bovine serum (FBS) and RPMI-1640 were purchased from ThermoFisher Biochemical Products (Beijing) Co., Ltd. Hypoxic bags were purchased from Mitsubishi Gas Chemical Company, Inc. (Japanese). Ferulic acid (≥98% purity by HPLC), zingerone (≥98% purity by HPLC), and vanillic acid (≥97% purity by HPLC) were purchased from Jiangsu Yongjian Pharmaceutical Technology Co., Ltd. (Jiangsu, China). Caryophyllene oxide (≥99% purity by HPLC) and methylephedrine hydrochloride (≥98% purity by HPLC) were purchased from TargetMol (United States) and Shenzhen Polymeri Biochemical Technology Co., Ltd. (Shenzhen, China), respectively. Edaravone (≥99% purity by HPLC) was purchased from TargetMol (United States). Cell Counting Kit-8 (CCK-8) was purchased from Dojindo Laboratories (Japanese).

The PC12 cell line was obtained from CHI SCIENTIFIC (Shanghai, China) and cultured in RPMI-1640 with FBS, penicillin 100 U/mL, streptomycin 100 μg/ml, respectively, at 37°C in a fully humidified 5% CO₂ atmosphere.

Oxygen and glucose deprivation (OGD) is a well-established in vitro model in studying the pathology and pharmacology of ischemic damage (Gu et al., 2013; Guo et al., 2013; Morán et al., 2017; Tian et al., 2020). Considering the actual clinical situations, patients usually take medications of XXMD in the poststroke (Pan et al., 2017; Zhong et al., 2020); we study the effect of components in treating stroke-based PC12 cells in the case that cells were protected with components after OGD. During the OGD, the cells were incubated in the culture medium without FBS in a hypoxic bag at 0.1% O₂, 5% CO₂, and 37°C for 18 h (Gu et al., 2013; Morán et al., 2017).

To test the predictive power of our proposed model, the effective components of KFCGs (ferulic acid, zingerone, and vanillic acid) and two non-KFCG components (caryophyllene oxide and methylephedrine hydrochloride) were selected to validate our model in vitro experiments with PC12 cells. Edaravone that has been shown to reduce cell death was selected as a positive drug (Zang et al., 2018; Shou et al., 2019; Jiang et al., 2020). Effects of these components with different concentrations on cell viabilities were detected by CCK-8 assay.

CCK-8 assay was utilized to measure cell viability. Cells were seeded in 96-well plates (2 × 10⁴ per/well). After 24 h incubation, cells were treated without any components for 18 h in OGD. A control group without the treatments of OGD and components was taken at the same time. After the OGD period, cells were treated with 0.01, 0.1, 1, 10, 100, and 1,000 μM ferulic acid (Hassanzadeh et al., 2018; Moghadam et al., 2018; Nakayama et al., 2020), zingerone (Ho et al., 2013; Ruangsuriya et al., 2017), vanillic acid (Shen et al., 2018), caryophyllene oxide, methylephedrine hydrochloride, and 20 μM edaravone (Zang et al., 2018; Shou et al., 2019; Jiang et al., 2020), and were incubated with complete culture medium under normoxic condition for 18 h. Furthermore, a model group with OGD treatment and without components treatments was taken at the same time. Then, cells were changed to be cultured in a fresh complete culture medium with 10 μl of CCK-8 for a further 4 h. The absorbance was measured at 450 nm with a microplate reader (Infinite M200, TENAN, Switzerland). The experiments were repeated twice with three replicates each time.

All data were expressed as mean ± SEM. The differences between the model group and the control group were analyzed by Student t test. The differences between components treatments and model group were analyzed by one-way ANOVA for multiple comparisons. Results were considered statistically significant if the p-value was <0.05.

The process of stroke is related to a series of phenotypic changes accompanied by alterations of genes expression. These genes may be labeled as pathogenetic genes at both the diagnostic and intervention levels. The collection and analysis of pathogenetic genes are the basis and key steps in understanding the pathogenetic genes of stroke and providing intervention strategies. To obtain more comprehensive pathogenetic genes, we extracted pathogenetic genes from the proved literatures in the DisGeNET database, related to both ischemic and hemorrhagic stroke (Supplementary Table S1).

In total, 2,788 literatures provided conclusive evidence for further construction of intervention space, and in total 1,012 predicted target genes were retained as pathogenetic genes. All 1,012 genes were supported by at least one published report. Among these, 27 genes were supported by more than 15 published reports (Figure 2A).

To determine whether pathogenetic genes reported with higher number of evidence have more comprehensive function, we preformed KEGG and GO analysis, and got 138 signaling pathways and 2,802 GO terms (Figures 2E,F). The results showed that genes with higher number of evidence have more relation of signaling pathways and GO terms.

The genes supported with 20–39 evidences have the largest average number of signaling pathways and GO terms (Figures 2B,C). The top 10 genes supported with the largest number of evidence are MTHFR, PLAT, APOE, ACE, NOTCH3, PDE4D, F5, F2, CRP, and LPA (Figure 2D) and mainly associated with the lipid metabolism and cell activity of stroke. Methylene tetrahydrofolate reductase (MTHFR) and Apolipoprotein E (APOE) reduce the risk of atherosclerosis (AS) and stroke (Guo et al., 2019; Xie Q. et al., 2020; Ji et al., 2020; Mazdeh et al., 2020; Qin et al., 2020); angiotensin-converting enzyme (ACE) affects the renin-angiotensin system (RAS), regulates blood pressure, and affects the occurrence of ischemic stroke (IS) (Dušanović Pjević et al., 2019; Isordia-Salas et al., 2019); NOTCH3 affects the maturation and homeostasis of vascular smooth muscle cells, and causes cerebral ischemic events (Mukai et al., 2020; Young et al., 2020); phosphodiesterase 4D (PDE4D) promotes cyclic adenosine monophosphate (Liu et al.) degradation and apoptosis (Zhou et al., 2019) and increases the probability of stroke occurrence (Yue et al., 2019) and IS reperfusion injury (Lu et al., 2020).

Among the top 30 pathways of pathogenetic genes (Figure 2E), more than 10 pathways are related to stroke. For example, the pathway Cytokine-cytokine receptor interaction is related to cerebral ischemia-reperfusion injury (Wang et al., 2017) and is one of the important active pathways after stroke (Li L. et al., 2020); the HIF-1 signaling pathway is related to the neuromodulation (Xu H. et al., 2019; Yang W. et al., 2020), cellular activity (Sugumaran et al., 2020), ischemic angiogenesis (Liu Y. et al., 2019), and inflammation (Liu R. et al., 2019; Du et al., 2020). Furthermore, ischemic stroke has 14 similar pathogenetic genes, ACVRL1, APP, ABCC6, CST3, ENG, F7, JAK2, PLAT, PLAU, VHL, SH2B3, PDCD10, KIF1B, and CCM2, and has 48 similar signaling pathways to hemorrhagic stroke, whereas the stroke-related ID C0553692 is related to hemorrhagic stroke and to 29 pathogenetic genes and 48 signaling pathways. It indicated that ischemic stroke and hemorrhagic stroke are both important in the study of XXMD (Zhang et al., 2021). In addition, most of the top 30 GO terms of pathogenetic genes are closely associated with stroke, such as response to oxidative stress (GO:0006979) (Chen et al., 2020; Zhang et al., 2020d) and regulation of inflammatory response (GO:0050727) (Zhang et al., 2014; Xu S. et al., 2020).

The weighted gene regulatory network can contribute to understanding the pathogenetic genes and provide intervention strategies of stroke. To construct a weighted gene regulatory network, the comprehensive PPI network was combined from CMGRN and PTHGRN (Guan et al., 2014a; Guan et al., 2014b).

The pathogenetic genes were mapped to the PPI network to construct the weighted gene regulatory network of stroke. The network contains 949 nodes and 42,716 edges (Figure 3). To validate the reliability of the weighted gene regulatory network, we compared the consistency of the degree of nodes and weight of the nodes, and found that IL6, TNF, and VEGFA have the highest degrees with 374, 336, and 308, respectively, whereas the numbers of literatures supporting these genes are 34, 32, and 23, respectively. Numbers of literatures supporting NOS3, APOE, CRP, BDNF, and F2 are 31, 95, 36, 28, and 40, respectively. These genes also have large degrees in the network as 180, 167, 167, 166, and 152 edges, respectively, which are higher than the average degrees of all nodes in the network of 45.01. These results indicated that the weighted gene regulatory network is reliable for further analysis. Additionally, according to published reports, these genes are widely enriched in the pathways that are closely associated with stroke, such as Cytokine-cytokine receptor interaction (hsa04060) (IL6, TNF), PI3K-Akt signaling pathway (hsa04151) (IL6, NOS3, VEGFA), MAPK signaling pathway (hsa04010) (TNF, VEGFA), and Ras signaling pathway (hsa04014) (VEGFA). These results suggested that the weight gene regulatory network and weighted genes could reflect the pathogenetic genes of stroke, which also provided a reliable reference for the next step to construct the FRS.

By a systematic search for components from public databases, we obtained 1,490 components from 12 herbs in XXMD, including BS, CX, FF, FJ, FZ, GC, HQ, KXR, MH, RS, RG, and SJ (Supplementary Table S2). Meanwhile, we obtained additional 114 components from the 12 herbs according to literatures (Table 1).

We obtained 220 potential active components from these 1,490 components based on three ADME-related models including OB, Caco-2, and DL (Supplementary Table S3). Except ADME prediction, experimental chemical analysis also plays important roles in the study of substances basis and mechanism of herbs in the formulas. Thus, we regarded the additional 114 components collected from literatures with these 220 ADME-predicted components as potential active components of XXMD (Figure Figure4; table 2, Supplementary Table S4).

A total of nine components are contained in three or more than three herbs (Supplementary Table S5). These components are sitosterol (MOL000359), beta-sitosterol (MOL000358), stigmasterol (MOL000449), gallic acid (COM4), kaempferol (MOL000422), quercetin (MOL000098), mairin (MOL000211), (+)-catechin (MOL000492), and mandenol (MOL001494). Among them, beta-sitosterol is present in MH, BS, RS, FF, HQ, FJ, and SJ, and can exert anti-inflammatory (Navarro et al., 2001) and antioxidant efficacies (Cao et al., 2018; Yuan et al., 2019; Devaraj et al., 2020). It has been proved to be an important component in the prescriptions of Hua-Feng-Dan and Buyang Huanwu decoction in the treatment of stroke (Yang P. et al., 2020; Gao et al., 2021). Stigmasterol is present in KXR, MH, RS, HQ, and SJ, which can promote cholesterol secretion and reduce AS (Lifsey et al., 2020); (+)-catechind is present in KXR, MH, and BS and can regulate the expression of proteins p-Akt and p-GSK-3b (Wang J. et al., 2020) which can inhibit the production of cell toxicity when PC12 cells are under hypoxic conditions (Zhou and Li, 2020). Kaempferol is present in MH, GC, BS, RS and can improve neurological deficits in cerebral ischemia/reperfusion (Zhai, 2019).

Except the shared components, most of the herbs possess their unique components (Supplementary Table S6). For example, the numbers of unique components of FZ, GC, and RS are 26, 91, and 43, respectively. In these herbs, some unique ingredients have a special therapeutic effect on stroke. For example, luteolin (MOL000006) is an unique component in MH, has properties of anti-inflammatory, neuroprotective, anti-allergic, and vascular protection (Qiu et al., 2013; Pandurangan and Esa, 2014; Su et al., 2021), and has been proved to enhance mitochondrial function by increasing the transduction of Sirtuin 3 (SIRT3) through the SIRT3/AMPK/mTOR pathway and to reduce the infarcted area of middle cerebral artery occlusion (MCAO) rat model (Liu S. et al., 2020). It also reduces oxidative damage to cells (Zhang Z. et al., 2019) and enhances cell viability and downregulates apoptosis (Luo S. et al., 2019). Skullcapflavone II (MOL002927) is a unique component in HQ and has properties of anti-inflammatory (Jang et al., 2012). It inhibits the proliferation of tumor cells (Tayarani-Najarani et al., 2012) and cancer cells (Bonham et al., 2005) and maintains the integrity of extracellular matrix (Lee et al., 2019). Hesperetin (MOL002341) is a unique component in FJ and regulates lipid metabolism (Xiong et al., 2019), lowers blood pressure, and prevents endothelial dysfunction (Yamamoto et al., 2008; Morand et al., 2011). Therefore, these components can be considered curative elements in treating stroke.

To explore the potential mechanism of XXMD in the treatment of stroke, 334 potential active components and their 1,329 targets (Supplementary Table S7) are used in constructing the component-target (C-T) network. The top 50 components and top 50 targets valued by degree in the C-T network were shown in Figure 5. Several of these potential active components are related to multiple targets, resulting in 12,395 C-T associations between 334 active components and 1,329 targets. The average number of targets per component is 4.02, indicating that there are multi-targets characteristics of XXMD in the treatment of stroke. Among these components, vanillic acid (MOL000114) has the highest number of targets as 258, followed by components of L-tryptophan (COM15), ferulic acid (COM14), benzoic acid (COM1), and zingerone (MOL002516), etc. Most of these components are related to inflammation and neuroprotection in stroke. For example, dihydrocapsaicin is a kind of capsaicin compound which can mediate the generation and functional recovery of blood vessels after ischemic stroke (IS) (Wu et al., 2019; Janyou et al., 2020); quercetin is a flavonoid that protects the neurons of IS (Wang et al., 2018; Zhao et al., 2018). These results identified a range of plausible biological responses to XXMD, and demonstrated that the XXMD might work in a multi-target manner in the treatment of stroke.

Here, we constructed a disease-targets network based on the weighted gene regulatory network and C-T network. This network contains 1,836 nodes and 25,383 edges (Supplementary Table S8).

To evaluate the effectiveness of the FRS, we defined the intersection of pathogenetic genes and XXMD targets as the un-optimized effective targets (UETs) of XXMD and defined the genes that were included in the FRS as effective proteins. Then we evaluated the effective proteins with three evaluation indicators: 1). the proportion of effective proteins in the number of UETs; 2). the proportion of effective proteins in the number of UETs enriched pathways; 3). the proportion of effective proteins in the number of UETs enriched GO-terms.

We used four methods in the optimization of targets and pathogenetic genes. Among these four methods, the best optimal method was method 1, our proposed node importance calculation method (Figures 6D,F,H). Based on our proposed model, 918 nodes were filtrated as important nodes, and also defined as the effective proteins. The effective proteins and their interactions were used to construct the FRS. There were three subtypes of effective proteins in the FRS (Figure 6K). The first subtype represents essential common targets, which directly linked pathogenetic genes and herbs targets. The second subtype represents disease-specific targets. The third subtype represents component-specific targets. These three subtypes indicated that the effective proteins could represent the effect of pathogenetic genes and XXMD targets.

The numbers of effective protein-enriched pathways and GO-terms were 180 and 3,043, respectively (Figures 6L,M). Before the optimization, the numbers of XXMD targets and pathogenetic genes of stroke were 1,329 and 1,012, respectively (Figure 6C), the numbers of targets and pathogenetic genes enriched pathways were 179 and 138, respectively (Figure 6E), whereas the numbers of targets and pathogenetic genes enriched GO-terms were 2,149 and 2,802, respectively (Figure 6G). The enriched pathways of effective proteins accounted for 99.10% of UETs enriched pathways (Figure 6F). The GO terms accounted for 97.30% of UETs enriched GO Terms (Figure 6H). These results confirmed the reliability and accuracy of our approach in constructing FRS and further demonstrated that the effective proteins in the FRS played a leading role in the pathogenetic genes of stroke.

KEGG and GO enrichment analysis of the effective proteins in FRS showed that the effective proteins could play a role in treating stroke. Among the 180 signaling pathways of the effective proteins (Figure 6L), PI3K-Akt signaling pathway could increase the expression of anti-apoptotic protein (Bcl-2, Bcl-XL), inhibit the expression of apoptosis protein (caspase-3, Bax), and could reduce neuronal apoptosis during stroke or hypoxia and hypoglycemia reoxygenation (OGD/R) (Wu et al., 2019; Xie W. et al., 2020; Miao et al., 2020); Neuroactive ligand-receptor interaction could affect expressions of GRM5, GRIK1, GRIK3, GABRA3, ADRA2C, VIPR2 and other genes (Zhang S. et al., 2019), affecting the ability of nerve transmission in inhibiting stroke occurrence (Riccio et al., 2012; Takenouchi et al., 2014) and affecting the pro-inflammatory cytokine releasement (Olson et al., 2015). Among the 3,043 GO terms of the effective proteins (Figure 6M), some of these GO terms were closely associated with the occurrence and recovery of stroke, such as response to oxidative stress (GO:0006979) (Chen et al., 2020; Zhang et al., 2020d), regulation of lipid metabolic process (GO:0019216) (Mahrooz et al., 2019; Sharma et al., 2020), regulation of inflammatory response (GO:0050727) (Xu S. et al., 2020; Du et al., 2020), regulation of blood circulation (GO:1903522) (Du et al., 2020; Gutierrez-Vargas and Cardona-Gomez, 2020), and neuron death (GO:0070997) (Iadecola and Anrather, 2011).

The CCR model was established to optimize the effective components and get the KFCG that could be used to illustrate the potential molecular mechanism of XXMD in the treatment of stroke. According to the contribution accumulation score ranking, targets of the top seven components including vanillic acid, ferulic acid, L-tryptophan, Dihydrocapsaicin, zingerone, rhamnetin, and Norwogonin could cover 50.00% of effective proteins. For further analysis, targets of the top 56 components could cover 90.00% of effective proteins, and these 56 components were selected as KFCG (Figure 7; Table 3, Supplementary Figure S1). The high targets coverage of effective proteins of 90.00% proved that the KFCG of XXMD might play the leading role and generate combination effects in the therapy of stroke.

To analyze XXMD in the treatment of stroke at the functional level, we performed pathway analysis using KFCG targets. The number of KFCG targets enriched pathways is 166 which can cover 80.43% of pathogenetic genes enriched pathways (Figure 8).

These KFCG targets were involved in 166 pathways such as Retrograde endocannabinoid signaling (hsa04723), neuroactive ligand-receptor interaction (hsa04080), and cAMP signaling pathway (hsa04024). Most of these signaling pathways have been reported to be related to the occurrence of stroke or the recovery in poststroke. For example, neuroactive ligand–receptor interaction can affect the expression of genes such as GRM5, GRIK1, GRIK3, GABRA3, ADRA2C, and VIPR2 (Zhang S. et al., 2019), thereby affecting the ability of nerve transmission (Riccio et al., 2012; Takenouchi et al., 2014) and the releasing of pro-inflammatory cytokines (Olson et al., 2015); the cAMP signaling pathway can regulate the expression of Adora2a, Drd2, and Pde10a and the expression of DEG-targeted transcription factors (TFs), such as androgen receptor system (AR), contributing to the recovery of brain (Ito et al., 2018); PI3K-Akt signaling pathway (hsa04151) can increase the expression of anti-apoptotic proteins (Bcl-2, Bcl-XL), inhibit the expression of apoptotic protein (caspase-3, Bax), and reduce the neuronal apoptosis (Xie W. et al., 2020; Miao et al., 2020; Wu et al., 2020).

Stroke is classified into hemorrhagic stroke and IS, and has the characteristics of multisystem cross-effects, mainly involving the cerebrovascular system and nervous system. Studies have shown that the occurrence of stroke is related to cerebral microvessels, neural information transmission, neuronal activity (Wu et al., 2020), and inflammation (Olson et al., 2015). According to previous extensive studies, stroke has been confirmed to be associated with some pathways such as vascular endothelial growth factor (VEGF), signaling pathway (hsa04370), cAMP signaling pathway, MAPK signaling pathway (hsa04010), and PI3K-Akt signaling pathway. In our study, the numbers of KFCG genes enriched in the above four pathways were 29, 55, 62, and 60, respectively. The reports showed that the VEGF signaling pathway could affect the protein activity of VEGF, mediate in angiogenesis and blood supply (Wu et al., 2019), reduce the formation of brain edema in the poststroke (Peng et al., 2014; Hu et al., 2019), and could promote dendrite and synaptic plasticity and improve nerve recovery in stroke (Shim and Madsen, 2018; Xie et al., 2019). The cAMP signaling pathway could regulate the activities of Adora2a, Drd2, and Pde10a, contributing to the recovery in poststroke (Ito et al., 2018). It could also regulate the activity of DEG-targeted transcription factors (TFs), such as the androgen receptor system (AR), contributing to the recovery of brain (Zhang H. et al., 2019). The PI3K-Akt signaling pathway could reduce neuronal apoptosis during stoke or OGD/R (Xie W. et al., 2020; Miao et al., 2020; Wu et al., 2020). The MAPK signaling pathway could inhibit the apoptosis of nerve cells (Han et al., 2013; Han et al., 2020), increase the activities of ERK, JNK, and p38 (Hong et al., 2020), reduce the activities of PLA2, IL1, TNF, IL1β, IL6, and IL8 (Dong et al., 2019; Ding Y. et al., 2020), and regulate the synthesis of proteins such as ASPK/JNK, p38, and P-SAPK/JNK (Zeng et al., 2019).

To further explore the potential mechanism of XXMD in the treatment of stroke, we constructed a comprehensive pathway with these four pathways, VEGF signaling pathway (hsa04370), MAPK signaling pathway (hsa04010), PI3K-Akt signaling pathway (hsa04151), and cAMP signaling pathway (hsa04024) (Figure 9).

Combining KFCG targets and these four pathways, we speculated the potential mechanism of XXMD to be two important aspects: 1) KFCG can benefit for cell proligeration, migration, survival, and DNA repair. 2) KFCG can regulate the activities of proteins which can interact between multiple pathways to play a role in the treatment of stroke by XXMD.

1) KFCG can benefit for cell survival and DNA repairation.

In the cAMP signaling pathway, KFCG can activate the activities of PI3K (genes PIK3CA, PIK3CB, PIK3CD, PIK3R1, PIK3R2, PIK3R3) (Xin et al., 2020) and Akt (genes AKT1, AKT2, AKT3), thereby regulating the gene expression and proliferation through the PI3K-Akt signaling pathway. For example, the deletion of PI3Kγ can upregulate 3′,5′-cAMP signaling in microglia, resulting in an increased release of microglial MMP-9 and a decrease in microglial phagocytic capacity (Schmidt et al., 2016). KFCG can phosphorylate Raf1 (gene RAF1), MEK (genes MAP2K1, MAP2K2) (Namura et al., 2001), CREB (genes CREB3, CREB1, CREB3L4, CREB3L2, CREB3L3, CREB3L1, CREB5) (Liu J. et al., 2019; Gao et al., 2020b; Queme et al., 2020), CBP (genes CREBBP, EP300), and can activate the activities of fos (gene FOS) and jun (gene JUN), thereby inhibiting the hyperexcitability and regulating cell death/survival of hippocampal neuron. For example, mitogen-active protein/extracellular signal-regulated kinase (MEK) inhibition can protect the hippocampus against forebrain ischemia (Namura et al., 2001). High phosphorylation of cAMP response element binding protein (CREB) is the target of 16 components of KFCG, including MOL000114, MOL000006, MOL000098, etc. (Supplementary Table S10). It can represent the high activity of CREB and can increase the expression of brain-derived neurotrophic factor (BDNF) which can enhance the synaptic efficiency and structural plasticity effectively as a most important neuronal protective factor and a prime mediator of synaptic plasticity (Leal et al., 2017; Lin et al., 2018). In addition, KFCG can regulate the activities of phosphorylate CFTR (gene CFTR) and AMPAR (genes GRIA1, GRIA2, GRIA3, GRIA4), which were located on the cell membrane, thereby affecting the transport of Ca²⁺, K⁺, and Na⁺ in cell membrane and affecting the process of pancreatic secretion, bile secretion, and cardiac muscle contraction.

In the VEGF signaling pathway, KFCG can regulate the activities of VEGF (gene VEGFA), VEGFR2 (gene KDR), PLC (genes PLCG1, PLCG2), and PKC (genes PRKCA, PRKCB, PRKCG), thereby affecting the calcium signaling pathway. KFCG can regulate the activities of PKC (genes PRKCA, PRKCB, PRKCG), Ras (genes HRAS, KRAS, NRAS) (Ding M.-H. et al., 2020), Raf-1 (gene RAF1) (Ding M.-H. et al., 2020), phosphorylate SPK (genes SPHK1, SPHK2), and MEK (genes MAP2K1, MAP2K2) (Ding M.-H. et al., 2020), thereby affecting the arachidonic acid metabolism and cell proliferation through the MAPK signaling pathway. The reports have shown that Ras and MEK can play actions in Ras/Raf/MEK/ERK signaling pathways and can be regulated by some miRNAs such as miRNA-21 and miRNA-26a, which can induce angiogenesis to aid in blood vessel formation for vascular tissue engineering in ischemic diseases (Ding M.-H. et al., 2020).

In the MAPK signaling pathway, KFCG can regulate the activities of Ras (genes RRAS2, MRAS, HRAS, KRAS, NRAS, RRAS), MEK1 (gene MAP2K1), MEK2 (gene MAP2K2) (Akgun-Dogan et al., 2019) and c-fos (gene FOS), and can regulate the activities of MEK5 (gene MAP2K5) and Nur77 (gene NR4A1) to regulate cell proliferation and differentiation. At the same time, the reports also show that germline mutations of RAS superfamily (KRAS, NRAS, HRAS, RRAS) and MAPK cascade (MEK1, MEK2) can cause strokes (Garavelli et al., 2015; Akgun-Dogan et al., 2019). Besides, KFCG can also phosphorylate IKK (genes CHUK, IKBKB, IKBKG) and regulate the activities of NFκB (genes NFKB1, NFKB2, RELA, RELB), thereby regulating cell Proliferation, inflammation, and anti-apoptosis. In addition, KFCG can also regulate the activities of TNF (gene TNF) (Ma et al., 2020; Meng et al., 2020), TNFR (gene TNFRSF1A), AKT (genes AKT1, AKT2, AKT3) (Gao et al., 2008), PTP (genes DUSP3, PTPN7, PTPRR), PTPN5), AP1 (genes FOS, JUN), MAX (gene MAX), and other protein activities, thereby affecting the cell cycle comprehensively. At the same time, the reports also have shown that the lower activities of proinflammatory factor TNF can benefit against IS (Ma et al., 2020).

In the PI3K-AKT signaling pathway, KFCG can regulate the activities of TLR2/4 (genes TLR2, TLR4) (Guo and Zhu, 2019), Rac1 (gene RAC1), PI3K Class IA (genes PIK3CA, PIK3CB, PIK3CD, PIK3R1, PIK3R2, PIK3R3), PI3K Class IB (genes PIK3R5, PIK3R6, PIK3CG) (Kong et al., 2016; Li J. et al., 2019), AKT (genes AKT1, AKT2, AKT3) (Chen et al., 2003) and can phosphorylate GSK3 (gene GSK3B) (Krafft et al., 2012), thereby promoting the neurorestorative activity in poststroke. Besides, KFCG can also regulate the activities of Ras (genes HRAS, KRAS, NRAS), Raf-1 (gene RAF1) and can phosphorylate MEK (genes MAP2KA, MAP2K2) (Bai et al., 2015), thereby affecting cell proliferation angiogenesis DNA repair through VEGF and MAPK signaling pathways. The reports have also shown that the inhibition of MEK can inhibit the activity of VEGF and be helpful for the decrease in the occurrence of stroke (Bai et al., 2015). KFCG can phosphorylate IKK (genes CHUK, IKBKB, IKBKG) and regulate the activities of NFκB (genes NFKB1, RELA), thereby promoting cell survival by the NFκB signaling pathway.

2) KFCG can regulate the activities of proteins which can interact between multiple pathways to play a role in the treatment of stroke by XXMD.

For example, in the 585 KFCG target enrichments, four proteins existed in three pathways and whose activities could be affected by KFCG, and were as MEK (pathways of hsa04024, hsa04370, and hsa04151), NFκB (hsa04024, hsa04010, hsa04151), PI3K (hsa04024, hsa04370, hsa04151), and Ras (hsa04370, hsa04010, hsa04151) (Figure 9). Meanwhile, 18 proteins existed in two pathways and whose activities could be affected by KFCG, such as Akt, cPLA2, and CREB (Figure 9). Most of these proteins have been proved to be closely related to stroke. NFκB and AKT1 can both affect inflammation and anti-apoptosis in PI3K-Akt, cAMP, and MAPK signaling pathway. RAF1 can affect the activity of Raf1 in the cAMP signaling pathway and the MAPK signaling pathway, thereby affecting hyperexcitability, cell death/survival, proliferation, and differentiation. In addition, we also found that KFCG could affect many other signaling pathways indirectly such as insulin signaling pathway, ErbB signaling pathway, calcium signaling pathway, Rap1 signaling pathway, and NFκB signaling pathway through the four pathways that were shown in the compressed stroke pathways (Figure 9).

The results of the compressed stroke pathways indicated that the KFCG were important and effective in the stroke treatment, and these effects could be achieved through a multichannel biological process. This result suggested that we need to consider the relationship of components and targets and need to consider the interactions between different pathways in the treatment of stroke.

Based on the component potential effect score calculation model we constructed above, we calculated the PESs of 56 KFCG (Table 4). These PESs both considered about the network topology importance and the functional control ability of KFCGs. In the following analysis, we selected the top three components, vanillic acid, ferulic acid, and zingerone, for subsequent validations to test the predictive power of the model.

To validate the effect of KFCG on stroke and further validate whether the three components, ferulic acid, vanillic acid, and zingerone, could combine with targets effectively, we conducted molecular docking by 56 components with 3D conformer and 136 protein structures responding to 49 genes in comprehensive pathways and obtained 66,967 binding relationships in the docking results (Supplementary Table S11).

According to literatures, the lower value of affinity represents the more stable binding between protein and ligand and represent the better binding energy in protein–ligand interactions (Fu et al., 2018; Elhenawy et al., 2019). The information of binding relationships divided with binding affinity value has shown that all the 56 components of KFCG could bind the 49 genes of comprehensive pathways, whose affinities were equal to or lower than −6 kcal/mol, indicated that the 136 proteins searched from PDB could well represent the effectiveness of 49 genes in the comprehensive pathways, and confirmed that the components of KFCG could well target the proteins involved in comprehensive pathways (Figures 10A–C).

Among the 66,967 binding relationships, the three components, ferulic acid, vanillic acid, and zingerone, which were selected from the component potential effect score calculation model, showed good combinations with proteins.

Component COM14 can bind 37 genes and 68 proteins with the average binding affinity of −6.44 kcal/mol. It can bind best with protein structure 5ebz responding to gene CHUK with binding affinity of −7.6 kcal/mol, and be equal to COM14-1t46(KIT, −7.6 kcal/mol) and COM14-5m6v (PRKACA, −7.6 kcal/mol) (Figures 10D₁–F₂ ). Component MOL000114 can bind 25 genes and 37 proteins with the average binding affinity of −6.22 kcal/mol, and has a best binding relationship of MOL000114-5lcq (PRKACA, −7.2 kcal/mol) (Figures 10G₁, G₂ ). Component MOL002516 can bind 38 genes and 66 proteins with the average binding affinity of −6.40 kcal/mol, and has two best binding relationships of MOL002516-4xv9 (BRAF, −7.7 kcal/mol) and MOL002516-5m6v (PRKACA, −7.7 kcal/mol) (Figures 10H₁–I₂ ).

In addition, we found all the 56 KFCGs could have good binding ability with proteins. For example, MOL001755 can bind best with protein structure 5m6v responding to gene PRKACA with binding affinity of −11.40 kcal/mol (Supplementary Figures S2A₁,A₂ ), followed by bindings relationships of MOL010921-5m6v (PRKACA, −11.20 kcal/mol), COM52-3m2w (MAPKAPK2, −10.90 kcal/mol), and MOL000953-5m6v (PRKACA, −10.90 kcal/mol) (Supplementary Figures S2B₁–S2D₂ ). In depth, we analyzed the binding relationships whose binding affinity value is equal to or lower than −6 kcal/mol, obtained 32,928 bindings including 49 genes, 56 components, and 134 proteins, and focused on two questions including 1) which component can bind the most number of genes, 2) which genes can bind the most number of components. The results were as follows: 1) MOL000354, MOL000953, and MOL004951 can either bind 49 genes, whose average affinities are −7.30 kcal/mol, −7.33 kcal/mol, and −7.44 kcal/mol, respectively, whose optimal docking bindings are MOL000354-3zs5 (MAPK14, −10.40 kcal/mol), MOL000953-5m6v (PRKACA, −10.90 kcal/mol), MOL004951-5m6v (PRKACA, −10.70 kcal/mol), and MOL004951-5m6y (PRKACA, −10.70 kcal/mol), respectively (Supplementary Figures S2E₁–S2H₂ ). 2) Genes MAPK14 and NOS3 can both bind 56 components, whose average affinities are −7.39 kcal/mol and −7.59 kcal/mol, respectively, whose optimal docking bindings are MOL000006-3zs5 (MAPK14, −10.50 kcal/mol), MOL001689-3zs5 (MAPK14, −10.50 kcal/mol), MOL005842-3zs5 (MAPK14, −10.50 kcal/mol), and MOL010921-6pp1 (NOS3, −10.40 kcal/mol), respectively (Supplementary Figures S2I₁–S2L₂ ). The above results indicated that KFCG could effectively bind with the proteins involved in the comprehensive pathway, validating that KFCG play key roles of XXMD in the treatment of stroke.

To test the predictive power of our proposed model, we defined the 1,489 components that were eliminated by our models as the non-KFCG (Supplementary Table S9, Figure 11G). Three components of the KFCG (ferulic acid, zingerone, and vanillic acid), two components of the non-KFCG (caryophyllene oxide and methylephedrine hydrochloride), and a positive drug (edaravone) were performed in the experiments with PC12 cells (Figures 11F,G). The results showed that ferulic acid, zingerone, and vanillic acid could be protective for PC12 cells after OGD. In particular, compared with model groups, ferulic acid at 1–100 μM, vanillic acid at 0.001–10 μM, and zingerone at 1–10 μM could improve PC12 cell survival by 13.13, 17.17, 23.51, 15.02, 20.48, 26.46, 21.45, 18.60, and 21.14%, respectively, at a level comparable to edaravone treatment (Figures 11A–C). Relatively, the two components of the non-KFCG which were eliminated with our model, caryophyllene oxide and methylephedrine hydrochloride, could not protect PC12 cells (Figures 11D,E), while edaravone could significantly protect PC12 cells. At the same time, the results in the experiments with HT22 cells also showed that ferulic acid, zingerone, and vanillic acid could improve the survival of HT22 cells after OGD (Supplementary Figure S3, Supplementary Methods and Materials).

These results in the in vitro experiments proved that our proposed model could be used in selecting the KFCG of XXMD in treating stroke, effectively and accurately.