The 4 text mining methods extracted 4 different sets of drug-related genes. Through these gene-drug relations, a bridge between the genes and the drug was established. The aim of this study was to investigate how a drug is associated with its indication through the gene. The next part was to establish the bridge between these genes and the indication. Mature gene-disease relations were easily accessed through KEGG in the form of the KEGG pathway. The KEGG pathway is a collection of manually drawn pathway maps representing the knowledge on the molecular interaction, reactions, and relation network. Thus, a bridge between the genes and the drug was established via KEGG. The whole path in that mechanism was addressed by finding a gene bridge between the drug and its indication. The next step was to evaluate this strategy. We paid attention to which text mining method is more suitable in this strategy. We focused on the drug-related gene set extracted by the text mining method in terms of the quantity and importance. Thus, we needed to define the importance standard of the gene to the indication. The standard of the gene to the indication in this case is based on the KEGG pathway information. One gene specifically shows up in a specific pathway, which means that this pathway can be identified with this gene. In other words, the less pathways a gene appears in, the more important it is to its related pathway. To calculate this situation, we give a value IPF.

Where P={p₁,p₂…..p_M} refers to all KEGG pathways, where M=#{P} is the number of pathways in the KEGG database.

{p_m|gene_i ∈ p_m} refers to a pathway that contains the i-th gene, denoted as gene_i. Thus, every gene in the KEGG database receives a basic score. Simply adding all the gene scores together is unfair. Because all pathways show up in KEGG in the form of a map, each map consists of a set of node boxes and severe edges instead of genes and edges. Therefore, we need to figure out how to calculate that score that one text mining method receives from all node boxes in a specific pathway.

Assume T_t is a gene set that contains all of the genes mined by the t-th text mining method. In order to evaluate how T_t genes are enriched in a specific pathway, P_m, we define

Where IPF_gene_Tt,Pm considers the number of genes that exist in a pathway as well as the weight of each gene. The sum of the IPFs can be used to evaluate the association of the group of genes to a pathway. By doing this, cumulative associations along with gene weights are represented.

In KEGG, a node box in some cases represents 1 set of homologous genes, instead of 1 separate gene. Generally, although there exists more than 1 gene, these genes play the same role. Therefore, even the text mining method digs more than one gene belonging to this pathway but they play the same role in the same node box. We only applied the max gene score to represent the score that this text mining method receives in this node box in this pathway. If node_j is a single node,

where

If node_j has E subnodes,

Where g_i ∈ {N_nodej ∩ T_t}, g_i = g_max

For each gene_i, which belongs to gene set node_j as well as T_t, the maximum IPFgene_i is assigned, which means gene_i belongs to gene set N_nodej.

It is noted that a node box sometimes represents 1 set of protein complex genes that need to work together to play a role in the pathway. Therefore, we applied the sum of all the gene scores that the text mining method received in this node and multiplied it with a coefficient to represent the score that this text mining method receives in this node box in this pathway.

where ∣N_nodej∣ means the gene number of gene set N_nodej, while ∣g_i ∈ {node_j ∩ T_t}∣ means the gene number of the union of gene set N_nodej and gene set T_t.

Besides the inclusion of genes in 1 node, the graph theory of the node in the pathway should be taken into consideration. In graph theory, the degree of a vertex is the number of edges associated with the vertex. In a pathway graph, one node holding a high degree indicates that this node connects with more vertices. In term of gene, this gene is associated with many genes. Mutations and regulation of the gene affect more genes. In 1 pathway, the more a node shows up in the shortest path between the 2 genes, the more important this gene is in this pathway.

First, assume SP_noder,nodes refers to the shortest path between 2 arbitrary nodes, that is, node_r and node_s in pathway P_m, then, we count the occurrence of node_j in SP_noder,nodes with respect to P_m.

Count_nodej,Pm = #{SP_noder,nodes|node_j ∈ SP_noder,nodes} (9)

In addition, NShortPath_{noder,nodej,nodek} is a binary value, which denotes whether or not node_j appears in the shortest path between node_r and node_k.

Thus, each node in the pathway holds a “count” score. To compare the importance of a node among all the nodes in one pathway, softmax function is applied to NShortPath_{noder,nodej,nodek}. Here, the softmax function is the gradient logarithmic normalization of the discrete probability distribution of finite terms. The result of softmax is suitable for describing the importance of 1 node in 1 pathway.

Then, we added all IPF_nodej to represent the total score that the text mining method receives in this pathway,

where

Based on the above discussion on IPF_gene (equation 4), IPF_node (equation 8), and IPF_shortpath (equation 11), we formulate a generalized formula for IPF_node_Tt,Pm.

Here, equation (13) summarizes all the above metric considerations and proposes a generalized form of IPF metrics. For instance, IPF_gene in equation (4) holds if 1 is assigned to Weight_nodej,Pm. Equation (12) is assigned to Score_Tt,nodej Score(T_t, node_j) and equation (3) to Weight_genei. The full list of generalized IPF metrics is shown in Table 1.

The discovery of the efficacy of rapamycin was replicated via a pathway enrichment experiment. PubReminer was used to retrieve the research trend of rapamycin and breast cancer drugs. A total of 1502 abstracts were obtained, and the starting time was the year 2000. The experiment was designed to test if the gene interaction of rapamycin could be excavated by the text mining method from literature without reporting the relevance of breast cancer and rapamycin. All the gene pairs in the literature related to rapamycin from the years 1978 to 2000 were excavated, the active genes of rapamycin were obtained, and the enrichment analysis of the strategic gene pathway in this study was carried out. After applying the IPF_node, 1640 abstracts of rapamycin prior to the year 2000 were obtained and 243 genes were obtained. Afterwards, a standard pathway enrichment was obtained, and the top 0.5% of the pathways under each enrichment path index was statistically analyzed. As expected, the breast cancer pathway was listed in the enrichment results, and the results indicated that the potential activity of rapamycin can be obtained by enriching the gene pathway by text mining interaction genes.

The text mining system was investigated to bridge the drug, protein, and disease pathway in order to explore the pharmaceutical mechanism of rapamycin. Starting from the Literature Network application, the disease-related gene network was constructed, and 480 genes obtained by rapamycin-centric text mining were used to highlight the overlapping parts in the breast cancer gene network. All the breast cancer–related genes were collected from the STRING database. According to all the existing databases and text information, each gene was sorted for rapamycin correlation, and in this verification section, 100 breast cancer–related genes from STRING were selected. The breast cancer gene network was constructed according to the gene interaction mentioned in more than 40,000 papers, and the network was constructed using the literature network application program. After gene pathway enrichment analysis, the drug was associated with the pathway and Cytoscape was used for network visualization. In view of the relation between the pathway information and the disease, the drug was further associated with the disease. In order to further analyze the relationship between drugs and diseases, the distribution of the drug-active genes excavated in the disease gene network was analyzed.

In order to construct a disease-specific gene network, the genetic relationship of this network in nature was obtained from disease-related abstracts. Since Cytoscape is a high-quality visualization platform for network analysis, a literature network application program based on Cytoscape was applied to address the drug disease associations obtained after pathway enrichment. Figure 4 highlights 38 vital genes plotted as yellow circles, namely, STAT3, TP53, CDK4, CTLA4, AR, MYC, NOTCH1, IL6, ERBB2, CXCL12, BECN1, IGF1R, CDK2, EGF, ERBB4, MMP9, PIK3CA, CXCL8, ABCB1, EZH2, CDK6, SOX2, AKT1, CDH1, SRC, MTOR, ABCG2, KDR, CCND1, VEGFA, EGFR, ZEB1, ATM, PTEN, CXCR4, ERBB3, MDM2, and GATA3. These 38 genes are based on the intersection of the breast cancer text network and the drug rapamycin-active gene obtained in this strategy. The size of the point in the graph represents the degree of the point, the greater the degree, the larger the point, and the degree in this network is the number of proteins that interact with the protein. The edge thickness in the figure represents the number of sentences that support the protein-protein relationship. The edge color in the figure also represents the number of sentences that support the protein-protein relationship. It can be seen from the figure that the yellow bright spot covers the vast majority of breast cancer gene networks with moderately large spots. The 38 genes were enriched by the P value pathway, and 16 of them, that is, EGFR, IL6, TP53, CDK6, CDK4, PTEN, CDK2, KDR, AKT1, IGF1R, CCND1, VEGFA, PIK3CA, MDM2, MTOR, and the MYC signaling pathway belong to one of the MTOR signaling pathways. Among them, MTOR is an important gene targeted by rapamycin. The MTOR pathway plays an important role in multiple activities of rapamycin and is therefore linked to breast cancer. The reason that literature network is used to construct breast cancer–related network is that the protein interaction involved in constructing the network is obtained from the literature related to breast cancer, and it is the programmed realization of protein interaction based on sentence coexpression in this study. It is convenient for users to quickly construct interactive protein interaction networks based on text relationships. In this study, the breast cancer–related genes obtained from the STRING database were rearranged according to the text information, and the protein interaction information excavated from the text was reflected in the size of the protein gene points. Thus, breast cancer genes were given different weights. It is more convenient to give priority to the location of the active genes under the active conditions defined by the interaction. The overlap of disease and drug-active genes was observed and the possible mechanism of action was speculated.