Chinese herbal medicines have primarily originated from foods. Over long periods of practical living experience, the medicinal properties of many herbs were gradually established (Hou and Jiang, 2013; Gu and Pei, 2017). Subsequently, foods with therapeutic properties were progressively separated and designated for specialized use as medicines (Hou and Jiang, 2013; He et al., 2018; Long et al., 2022). Human foraging practices frequently entail the amalgamation of various food sources, a tendency that has played a significant role in the creation of Traditional Chinese Medicine (TCM) formulas (Hou and Jiang, 2013). From its inception, TCM may have involved the application of herbal combinations. Some of these herbal combinations were stable and clearly effective and were therefore documented and passed down as formulas through generations. This resulted in the creation of over 300,000 known formulas (Li et al., 2008), laying the foundation for clinical TCM treatments. However, pharmacological research of TCM formulas faces the significant challenge of analyzing combinations of 100 or more chemical compounds (which are also named components) per formula (Zhao et al., 2010). Statistics on the total amount of targets corresponding to each compounds in PubChem Bioassays database is 3.7 in average (Jalencas and Mestres, 2013; Hu et al., 2014). According to the number mentioned above, a given TCM formula could potentially regulate over 370 targets. Thus, the “one drug-one target” pharmacological research methodology is insufficient to explain the therapeutic effects and mechanisms of action associated with TCM formulas (Ding et al., 2020).

Deciphering the intricate pharmacological mechanisms associated with herbs and formulas is a monumental task for researchers in the field of TCM (Wang et al., 2021b; Li et al., 2022b). Due to the “black box” nature of complex biological systems, studies of formula efficacy would do well to take a more macroscopic approach (Yao et al., 2013; Huang et al., 2023), i.e., research needs be designed using a “system-to-system” framework for clinical and pharmacological investigations of entire formula instead of disassembling formulas and studying the components (Liang et al., 2012). This approach involves observing the relationships between formulas (input) and effects in biological systems (output). The research philosophy behind chemical drug development is fundamentally guided by reductionism, with antagonism serving as a primary principle (Saks et al., 2009). The key paradigm of drug discovery revolves around the creation of inhibitors or activators that specifically target particular molecular entities (Jendza et al., 2019; Gong et al., 2023). Over time, this approach has proven imperfect due to the discovery of off-target responses, which may have toxicological impacts or cause other side effects. Given the extensive range of enzymatic systems, classes, and isoforms that have been identified in biological systems, the development of many target-specific agents has relied on trial-and-error methodologies (Méndez-Lucio et al., 2016; Paydas, 2019). However, regulation of targets by formulas does not always require an extremely high level of specificity, and exceptionally high activity levels may not be necessary (Méndez-Lucio et al., 2016). Formulas themselves constitute complex systems, wherein synergistic interactions between components can lead to optimal effects to maximize impacts on the human biological systems (Chen et al., 2018). Research of the pharmacology associated with specific formulas therefore necessitates unveiling (or partially) of the “black box” that is synergistic interactions between components and their interactions with the human biological system. This requires accurate simulation of the alterations that occur in various nodes within the biological system due to regulation by a specific formula (Tan et al., 2019). The goal is to establish correlations (and ideally quantitative relationships) between changes in a formula and changes in clinical phenotypes.

Advances in computational biochemical analyses have ushered in a new age of TCM research (Barabási et al., 2011; Wang et al., 2022a). Cooperative regulation of multiple targets by multi-component medicine is an effective strategy for altering the output of complex systems (Csermely et al., 2005; Zhang et al., 2014; Ramsay et al., 2018). Mathematical models that reflect complex systems are exceptionally potent tools in systems biology research (Kitano, 2002; Liu and Barabási, 2016; Zhao et al., 2019). The advent of artificial intelligence (AI), particularly deep learning, has allowed the accumulation of TCM data with unprecedented depth and complexity (Chen et al., 2019). Studying a substantial number of effective formulas (rather than individual formulas) using phenotypes or clinical manifestations as outputs can allow elucidation of the intricate relationships among formulas, herbs, components, targets, phenotypes, and ZHENG. Computer science is a powerful tool that facilitates TCM research by allowing both establishment of relationships and large-scale collection of relevant data (Zhang et al., 2019a). However, it is crucial to exercise caution in utilizing such tools to ensure that results are grounded in reality. The establishment of trustworthy, accurate TCM databases will thus be a pivotal step in unraveling the complexities of herbs or formulas (Saks et al., 2009).

Research into the pharmacology underlying TCM necessitates the accumulation of extensive data for multiple parts of the TCM system: formulas, herbs, components, targets, phenotypes, and ZHENG (Han et al., 2017). Clear delineations of various relationships (e.g., formula-component, component-target, and target-phenotype relationships) are vital (Xu et al., 2021; Zhao et al., 2023b; Gan et al., 2023). Since 1960, databases have been developed and are now available for use in computational TCM research (Figure 1). In this review, we conduct a retrospective examination of the establishment of these databases, with a particular emphasis on comparison based on the inclusion of formula, herb, component, target, biological function, phenotype, and ZHENG data. We aim to consolidate and analyze the relationships between various entities within these databases, including formula-component, component-target, target-phenotype, and phenotype-ZHENG relationships. This review summarizes the trends, identifies gaps in the existing research, and suggests directions for future development of the databases related to systems pharmacology in TCM.

Initially, TCM formulas included only a small number of herbs. The herbs were consistently combined and administered in fixed proportions, which were documented and transmitted over time (Zha et al., 2015). With the evolution of medical practices, current formulas encompass not only ancient formulas but also modern empirical formulas and commercially prepared Chinese patent medicines. These formulas comprise the fundamental data within TCM formula databases, and there is a total of 21 such databases. These databases primarily contain fundamental information such as formula compositions, therapeutic functions, indications, and methods of use. Of the 21 databases, 18 are academic and 3 are commercial (Table 1). TCM-ID was one of the earliest TCM formula databases which brings the concept of formula-herb-component-target relationships in the form of databases (Chen et al., 2006). Chinese Medicine Think Tank (a big data analysis platform for TCM) houses the largest collection of TCM formula resources, including ∼300,000 formulas (Chinese Medicine Think Tank-a big data analysis platform for TCM, 2017). Over the Counter TCM Database and the Database of Standardized TCM Chinese Patent Drugs focus on marketed TCM formulas (Database of Standardized for Chinese patent drugs, 2017; OTC Chinese Herbal Medicine Database, 2017). These databases provide more comprehensive records of Chinese patent Drugs. DRUGDATAEXPY and Pharnexcloud are the major databases used in the pharmaceutical industry (DRUGDATAEXPY, 2009; Pharnexcloud, 2021). In addition to providing extensive formula resources, they also enable queries related to research, clinical trials, marketing, production inspections, and sales of specific formulas.

In 2016, with the emergence of databases like BATMAN-TCM that allow for customized predictive analysis of the pharmacological mechanisms of TCM formulas, databases for TCM were no longer limited to the functions of browsing and searching (Liu et al., 2016). The development of database functions became more geared towards analysis and prediction, and it was only then that TCM formula databases began to be utilized in a truly meaningful way. The principles of herbal compatibility are crucial for the effectiveness of TCM formulas. Formula databases not only allow for direct retrieval of herbal composition of formulas but also enable algorithmic-based analysis of the patterns governing herbal combinations. Through analysis of existing TCM formulas, CPMCP has summarized frequently used herbal combinations in TCM clinical practice. This functionality has helped to uncover the habitual pairings and contraindications between various herbs, shedding light on principles of herbal compatibility (Sun et al., 2022). Huabing data, an intelligent TCM big data platform, enables screening of disease-related formulas based on input herbal combinations. It dynamically calculates and analyzes the top 20 herbs, symptoms, diseases, and functions related to the treatment of certain diseases using formulas present in the database (Huabing data platform, 2023). TCMIP allows the selection of target TCM formula groups based on criteria such as formula composition and primary diseases treated. It calculates the frequency of herb usage, herb properties, inter-herb associations, and the frequency of formula targets within a selected formula group (Wang et al., 2021a). This approach facilitates innovative research of TCM formulas. For example, researchers have constructed a scoring system for the post-effects of drug combinations based on formula-herb relationships. The scoring system is utilized to recommend the most effective herbal combinations for certain diseases (Niu et al., 2023). The use of analytical algorithms on data from these databases can accelerate explorations of the intricate networks underlying TCM formula efficacy (Wang et al., 2021b).

One effective approach to simplifying the study of the pharmacological mechanisms of formulas is to make use of databases such as, ETCM, TCMIP, LTM-TCM, TCM-ID, which enable direct prediction of the pharmacological mechanisms of formulas (Chen et al., 2006; Xu et al., 2019; Wang et al., 2021a; Li et al., 2022a; Zhang et al., 2023). However, it's important to note that the quality of data and the dimensions covered vary among these databases, which can impact the accuracy of predictive results. Further evaluation is necessary for the data in these databases. Besides of that dosage and usage have varied significantly between regions and dynasties (Zha et al., 2015). Therefore, the results of the verification of the usage and dosage of individual herbs within formulas, as well as the sources of this information, should also be indispensable data within the database. The dosage of an herb significantly determines the concentrations of its components within the human body fluid and then may impacts the activity of component’s targets, which would be the key logic of both wet and dry experiments. However, most databases have not effectively cleaned dosage-related data during the inclusion of herbal formulas and data cleaning processes. Systematic research of ancient texts is essential to methodically organize and standardize ancient formulas. This highlights the needs of establishing uniform, rigorous standards, and quantitative dosage information for TCM data (Wang et al., 2021b).

Ancient Chinese people, through extensive medical practices over a long period of time, experimented with many medicinal substances derived from animals, plants, minerals, microorganisms, and other sources (Wang et al., 2017). These substances were documented, and their functions continuously verified over time. At present, TCM herbs encompass plant-based medicines, animal-derived medicines, and mineral-based medicines, with plant-based medicines being the predominant category. Herb databases are commonly used to compile fundamental information about herbs, such as properties, meridians, regions of usage, flavors, effects, and indications. This information came from various sources, including the Pharmacopoeia of the People’s Republic of China (ChP-TCM), the fourth national survey on Chinese Materia Medica Resources, books, literatures, and dictionaries. A total of 24 databases related to herbs have been identified, 21 of which are academic and 3 of which are commercial (Table 2). The Pharmacloud database contains the largest number of herb resources at −18,000 (Pharnexcloud, 2021). TCMID holds the distinction of being the academic database with the most extensive collection of herb resources, encompassing a total of 10,846 (Xue et al., 2012). China’s multi-ethnic composition means that various ethnic groups have discovered numerous herbs rooted in their own cultural practices. To enhance drug development, databases related to herbs from different ethnic groups have also been established, including databases for Tibetan, Mongolian, Uyghur, and Yao medicine.

The relationships between herbs and active components are currently key areas of focus in herb research (Fu et al., 2014; Zeng et al., 2022b). These linkages are included in herb databases. Active components are not only the primary materials that compose herbs but are also crucial for their therapeutic effects (Liu C. et al., 2018a). In 2018, following an update to the TCMID, there was a significant improvement in the coverage of herbs (Huang et al., 2018). Additionally, it introduced mass spectrometry (MS) data for these herbs, which served the purpose of distinguishing differences in the quality of herbs. Quantitative data for the characteristic components in each herb, as specified in the ChP-TCM, are available in the, ETCM and TCMIP databases (Wang et al., 2021a; Zhang et al., 2023). SymMap annotates components in four categories based on experimental MS data from ChP-TCM and from the literatures: quality control components, blood components, metabolite components, and other components (Wu et al., 2019). Utilizing herb-component relationship information from such a database, it is possible to construct more intricate features for herbs. This can be achieved, for example, by building heterogeneous herb-component-target networks. Such efforts enhance the accuracy of intelligent formula recommendation systems based on deep learning, such as FordNet (Zhou et al., 2021). Herb-component-target relationships in these databases also enable researchers to measure the effectiveness of specific herbs in treating diseases. This approach can then be used to identify herbs that are highly associated with specific diseases based on the importance of a particular target within a disease network (Wang et al., 2021d; Niu et al., 2023). For the identification of biological components in TCM, TCM-Suite gathered sequences and associated information for six marker genes: ITS2, matK, trnH-psbA, trnL, rpoC1, and ycf1 (Yang et al., 2022). Therapeutic efficacy of herbs is associated with the components and its content in the herb. A counterpart example in compound chemical drug is that there are fixed usage ratios for the synergistic effects of components (Ferrannini et al., 2022). For example, a fixed-ratio combination of insulin glargine and lixisenatide can better control the blood sugar levels in patients with diabetes (Aroda et al., 2016). Inappropriate ratios can lead to opposite effects (Létinier et al., 2023). In the context of components in herbs, the same principle holds true. Therefore, establishment of the herb-component relationships also requires the critical quantitative information-the content of components in herbs (Heinrich et al., 2022). Currently, there is a substantial accumulation of research on the identification and content measurements of components in herbs, including high-performance liquid chromatography, high-performance liquid chromatography-MS, and gas chromatography-MS, etc. (Arrizabalaga-Larrañaga et al., 2021; Papatheocharidou and Samanidou, 2023), but there is still a lack of databases for comprehensive aggregation and compilation of quantitative research data on components in herbs.

Regardless of whether a so-called “herb” of interest is a plant, animal product, mineral-based medicine, the active components of which are chemical substances. Herbal component databases include information about the chemical components that have been extracted or isolated from single herbs or formulas. Such databases source their data from the literature, experimental data, and/or preexisting databases, encompassing essential details such as chemical structure, and CAS registry number of component. A total of 28 databases related to chemical components in TCM herbs have been identified (Table 3).TCM-Suite has the largest number of TCM chemical components at 704321 components, but it only 54,868 herb-component relationships (Yang et al., 2022). Some of these databases are more specialized: TCMIO, CancerHSP, and NPACT primarily focus on collecting information about active components related to tumors, whereas CVDHD contains data about active components associated with cardiovascular diseases (Mangal et al., 2013; Tao et al., 2015; 2015; Liu et al., 2020). Databases for herb components specifically offer a wealth of resources for modern drug development (Fu et al., 2016; Zhang et al., 2023).

The relationships between components and targets represent a key link connecting two intricate systems: herbs and human biological systems (Stitziel and Kathiresan, 2017). Data mining and computational chemistry approaches are currently being used to collect and organize known component-target relationships and to predict and validate previously unknown component-target relationships (Chen et al., 2016). Several component databases TCM provide both information about known component-target relationships and functionalities for predicting such relationships. The HIT and HERB databases contain information about component-target relationships obtained through text mining of the literature (Fang et al., 2021; Yan et al., 2022). HIT categorizes component-target relationships component into three types: “Directly inhibit/activate,” “Indirectly inhibit/activate,” and “Enzyme substrate”. Users can refer to the associated literature to learn more about specific component-target relationships. More importantly, HIT facilitates automatic target mining and curation of “My-target” information from newly released PubMed literature (Yan et al., 2022).

For components lacking reported relationships with a target, several computational chemistry approaches have been significantly developed. These approaches include ligand-based methods, target-based methods, and target-ligand methods, all of which aim to predict relationships between components and proteins (Sadybekov and Katritch, 2023). The SwissTargetPrediction is a widely used web tool, available online since 2014, designed to predict the most probable protein targets of small molecules. Predictions are made using the similarity principle through reverse screening. In the latest updated version, the models have been recalculated, achieving a success rate of at least one correct human target in the top 15 predictions for more than 70% of external compounds (Daina et al., 2019). BindingDB is a database that focuses on relationships between small molecules and their corresponding targets. The BindingDB website provides specialized tools that leverage its extensive data collection, allowing researchers to generate hypotheses for protein targets of a given bioactive component or to predict components that are bound by a particular protein. Additionally, the website offers virtual component screening using methods like maximal chemical similarity, binary kernel discrimination, and support vector machines(Gilson et al., 2016). To meet the demand for predicting targets of components, component databases have also started incorporating target prediction functionality. SysDT is a model that was designed to predict potential targets of components within the TCMSP database (Ru et al., 2014). SysDT has demonstrated remarkable predictive performance for drug-target relationships (Yu et al., 2012). ETCM v2.0 uses a target identification method that is based on a two-dimensional ligand similarity search module within the D3CARP platform and utilizes data from Binding DB (Zhang et al., 2023). To enhance the accuracy of target prediction, TCMSID employs multiple target prediction methods, including similarity ensemble approach, SwissTargetPrediction, HitpickV2, PPB, PPB2, and CHEMBL (Zhang al., 2022a). LTM-TCM integrates component-target information from various sources, including the BATMAN-TCM, ChEMBL, and STITCH databases. LTM-TCM retains target scores from different sources to enable personalized target screening based on user-defined thresholds (Li et al., 2022a).

The systematic collection and organization of herb components in databases forms the foundation of target prediction to decipher the multiple pharmacological actions of a given compound. Target prediction methods have the potential to significantly shorten drug development timelines, but the accuracy of computational studies remains relatively low. In practice, even the most successful virtual screening campaigns typically result in only 10%–40% of candidate hits being confirmed through experimental validation (Sadybekov and Katritch, 2023). A multitude of virtual screening efforts produced predominantly discouraging outcomes. For instance, the antimalarial drug ebselen, which had been identified through an early virtual screening process, ultimately proved unsuccessful in clinical trials (Sadybekov and Katritch, 2023). Therefore, it is essential to conduct more comprehensive in vitro and in vivo studies and develop improved methods for evaluating the above results. These results recorded in online databases should also have clear indications of their sources, to aid researchers in assessing the reliability of the data.

Targets are the smallest functional units within an organism, serving as the internal nodes of complex systems (Turkarslan et al., 2014). They carry out various functions in numerous pathways and phenotypic responses, acting as bridges between medicines and the human biological system (Pfister and Ashworth, 2017; Santos et al., 2017). Drug mechanisms of action involve interactions between components and their targets. The initial paradigm in this area posited that a single component would act on a single target (Koeberle and Werz, 2014). However, further research revealed that nearly all natural and human-synthesized components interact with multiple targets (Plazas et al., 2022). Target databases primarily encompass genetic and protein-related information. Existing types of target databases include drug target databases, disease target databases, and specific target databases. These databases typically include basic information such as the target type, function, and origin, which are often sourced from the literature. UniProt, NCBI, and GeneCards are examples of target databases that provide comprehensive genetic and protein sequence information along with functional details (Table 4) (Safran et al., 2010; Brown et al., 2015; The UniProt Consortium, 2023). ETCM, TCMID, YaTCM, HIT, HERB, DisGeNET, and other databases also include information about targets, but these primarily focus on the relationships between targets and components or diseases (Piñero et al., 2017; Li et al., 2018a; Huang et al., 2018; Huang et al., 2018; Xu et al., 2019; Fang et al., 2021; Yan et al., 2022; Zhang et al., 2023). They often therefore have a decreased emphasis on the functional details of targets. Researchers have used target databases for purposes such as analysis of target-phenotype relationships (e.g., SymMap) (Wu et al., 2019; Lv et al., 2023). Target relationships in the TCMSP, TCMID, and TCM-ID databases have been used to map symptom-related genes and herb-related targets to human protein interaction networks (Chen et al., 2006; Xue et al., 2012; Ru et al., 2014; Huang et al., 2018). Through analysis of their topological relationships within a network, the distances between gene nodes can be calculated to infer distances between symptom modules, providing information about symptom co-occurrence and similarity. This approach has been employed to evaluate herb effectiveness for specific symptoms. It is a robust method for deciphering the mechanisms of herb and for predicting early-stage drug efficacy for diseases of interest (Gan et al., 2023).

Biological pathways can be considered as subsystems within complex systems. They serve as a framework for conducting pharmacological TCM research. These pathways can provide explanations for the complex mechanisms that link herbs to physiological changes. They often play significant roles in elucidation of interactions between drugs and biological functions (Wang et al., 2022b). Many target function databases integrate information about genes and genomes with higher-level functional annotations (Zeeshan et al., 2020). These data can then be used to systematically analyze gene functions based on known biological processes in an organism. Such databases are thus commonly utilized in conducting gene functional enrichment analyses, pathway-related analyses, and protein-protein interaction analyses. KEGG is a reference database for biological interpretation of genome sequences and other high-throughput data. The primary functionalities for biological process analyses are biochemical pathway mapping, metabolic network construction, genome comparison and merging, and enzyme database construction for target molecules (Kanehisa et al., 2017). BioCyc compiles and references genomes and metabolic pathways from thousands of sequenced organisms (Karp, 2005). Reactome systematically generates ordered molecular transformation networks, resulting in formation of classical metabolic maps. This database also associates human proteins with their molecular functions, offering a resource that serves as both a record of biological processes and a tool for discovering new functional relationships from data such as gene expression levels or mutations in tumor cells. Additionally, it can predict target biological processes of ion channels (Jassal et al., 2019). DAVID database consists of six tools: the functional annotation clustering, the functional annotation chart, the functional annotation table, gene functional classification, gene ID conversion, and gene name batch viewer (Sherman et al., 2022). The STRING database is used for analysis of protein-protein interactions. Individual protein queries generate a network composed of all proteins that interact with the queried protein (von Mering et al., 2003). This is particularly valuable for exploring interactions among input proteins; for example, it can be used to analyze the connections among differentially expressed proteins identified from proteomic data (Szklarczyk et al., 2021).

From the perspective of a complex system, the state of an organism corresponding to any abnormal phenotype is an abnormal steady state (Tyler et al., 2016). Such an abnormal steady state entails multiple nodes balance within the system. Likewise, interventions should target several nodes simultaneously to effectively restore the system to its normal steady state. Phenotype databases primarily focus on collecting data related to diseases, symptoms, and other phenotype-related entities. These databases provide robust datasets for those researching the mechanisms underlying TCM efficacy, primarily sourced from the literature and from other databases. Currently, a total of 13 databases have been compiled that provide detailed descriptions of diseases and symptoms (Table 5). TCMBank is the most comprehensive repository of disease-related resources, encompassing 32,529 data points (Lv et al., 2023).

The relationship between a target and the corresponding phenotype serves as a crucial bridge connecting a biological mechanism to the pathological manifestation in the human body. This connection was first established through the discovery of mutation-phenotype relationships. OMIM is a comprehensive repository that focuses on genetic and phenotypic data and interrelationships between the two. This database plays a pivotal role in naming and categorizing genetic phenotypes, thereby exerting a significant influence on the field of genetics (Funk et al., 2022). With the advent of the post-genomic era, the goal of deciphering the biological functions of target has evolved into the larger goal of delineating the intricate relationships between multiple genes and phenotypes. HPO and DisGeNET are comprehensive databases for analyzing and interpreting human gene-disease networks (Piñero et al., 2017; Köhler et al., 2021, 1). CPMCP and SymMap also include both TCM symptoms and modern medicine (MM) symptoms in an attempt to bridge TCM and modern medicine-based research through symptom associations (Wu et al., 2019; Sun et al., 2022).

TCM involves a unique, intuitive understanding of physiological states. ZHENG differentiation and treatment (辩证论治) is the fundamental approach guiding clinical practice in TCM. Diagnoses and treatments are made by taking into account the individual differences between patients (Zhou et al., 2014; Wang and Zhang, 2017). ZHENG is a summary of the pathological and physiological discrepancies at each stage of a disease. It is also determined by factors such as the disease site and the nature of the disease (Wang et al., 2022a). TCM practitioners prescribe different formulas based on the ZHENG to achieve therapeutic efficacy. Explorations of ZHENG-formula and ZHENG-phenotype relationships represent a challenging area of research in both clinical practice and foundational TCM studies. A ZHENG database, SoFDA, has been constructed to record and collect ZHENG data (Zhang et al., 2022b). It includes both macroscopic data, such as ZHENG, phenotypes, and TCM formulas, and microscopic data (molecular mechanisms). Such databases promote a deeper understanding of ancient systematic medicine, TCM, and modern medicine. SoFDA implements two common association measures (Jaccard and Cosine similarity) to quantify relationships between clinical entities (e.g., ZHENG, phenotypes, and formulas). This allows users to compute the degree of indirect associations between the three entities in terms of six shared features: symptoms, genes, enriched gene ontology (GO) terms, enriched pathways, network modules, and network density. However, databases specifically focused on ZHENG are currently limited in number, and there are few comprehensive phenotypes analyses related to ZHENG.