Comprehensive survey of target prediction web servers for Traditional Chinese Medicine

Abstract

Traditional Chinese medicine (TCM) is characterized by multi-components, multiple targets, and complex mechanisms of action and therefore has significant advantages in treating diseases. However, the clinical application of TCM prescriptions is limited due to the difficulty in elucidating the effective substances and the lack of current scientific evidence on the mechanisms of action. In recent years, the development of network pharmacology based on drug systems research has provided a new approach for understanding the complex systems represented by TCM. The determination of drug targets is the core of TCM network pharmacology research. Over the past years, many web tools for drug targets with various features have been developed to facilitate target prediction, significantly promoting drug discovery. Therefore, this review introduces the widely used web tools for compound-target interaction prediction databases and web resources in TCM pharmacology research, and it compares and analyzes each web tool based on their basic properties, including the underlying theory, algorithms, datasets, and search results. Finally, we present the remaining challenges for the promising future of compound-target interaction prediction in TCM pharmacology research. This work may guide researchers in choosing web tools for target prediction and may also help develop more TCM tools based on these existing resources.

Graphical abstract

1. Introduction

New drug discovery is a complex, expensive, time-consuming, and challenging undertaking [1]. Drugs play a role in treating diseases by acting on specific targets and producing corresponding biological outcomes [2,3]. Drug targets generally refer to molecules in the human body and pathogens that can interact with drugs, including enzymes, receptors, ion channels, transporters, nucleic acids, and other molecules [4]. In addition, there are other therapeutic targets, such as protein-protein interactions and nucleic acid (DNA/RNA)-protein interactions. Traditional drug discovery has mainly focused on highly specific inhibitors with a single target, in which the most specific drugs acting on one particular target of individual illnesses are sought [5]. It is challenging to treat complex diseases effectively using drugs with a single target. Using monomer compounds to treat complex diseases makes obtaining an excellent curative effect difficult (see Fig. 1, Fig. 2, Fig. 3).

Fig. 1.

Number of citations of each web serves.

Fig. 2.

Application of SEA web servers: (A) journal publications; (B) Top 10 research areas; (C) Top 10 countries/regions where authors belong.

Fig. 3.

Application of TCMSP web servers: (A) journal publications; (B) Top 10 research areas; (C) Top 10 countries/regions where authors belong.

Traditional Chinese medicine (TCM) has a long history in China and has played an essential role in the struggle between the Chinese nation and disease. It is the accumulation of much experience to form classic prescriptions with definitive clinical efficacy. TCM provides a rich material basis for discovering and developing modern drugs. TCM-based new drug development, especially for treating complex diseases, has gradually attracted increasingly more attention. However, due to the complex chemical compositions and modes of action of TCM, the understanding of TCM is still limited; the theoretical research of modern TCM lags, which hinders the development of modern TCM [6]. TCM includes various active compounds, and drug-drug interactions may occur among each combination, affecting the pharmacokinetics of drugs in vivo [7]. TCM's essence, concepts, and methods in treating diseases differ from those of modern medicine, which are discovered by targeting a specific protein. TCM plays its role through multiple targets and pathways. Therefore, correct identification and validation of drug-target interactions is the first step in activity prediction and mechanism elucidation of TCM. This highlights the critical role of identifying compound target interactions for the accuracy of specific TCMs.

There are many methods for target prediction, which are mainly divided into experimental and computational categories. The most direct method experimentally is the biochemical test, which directly evaluates the binding ability of small molecule compounds and corresponding proteins after incubation and elution to judge whether the small molecule binds to the target [8]. In addition, there are genetic and genomic approaches; for example, after locking the target of small molecules in a specific range of gene clusters, a series of point mutations or knockouts are performed on the gene clusters [9]. If the mutated or knock-out samples show the same effects as the small molecules, the protein encoded by the mutated or knock-out genes will likely be a target of the small molecules. The accuracy of the experimental method is relatively high; however, these experiments are costly and time-consuming, and a thorough search is not feasible because there are millions of drug-like compounds and hundreds of potential targets [10]. In recent years, computational methods that can quickly predict compound targets have emerged. Their strong development momentum could provide new potential drug-target interaction candidates for experimental biological validation and reduce the time and cost of biological experiments [11,12]. These methods are often called target identification, “in silico target prediction,” or “target fishing” [13,14]. Recently, increasingly more evidence indicates that a combination of target prediction and network pharmacology analyses is a feasible and powerful way to analyze the molecular mechanism of a drug and TCM [15,16].

In recent years, researchers have developed many machine learning methods to predict the targets of small molecules, many of which provide web services. The advantage of a web-based target prediction tool is that it is easy to operate and can be used even if you do not understand the underlying methods, which pharmaceutical chemists and biologists favor. In recent years, web-based target prediction tools have also been applied to TCM, which has played an important role in elucidating TCM's active substances and mechanisms of action. In this article, we first summarized the currently widely used drug-target interaction prediction databases and web resources for TCM pharmacology research, an area that is under tremendous development. Subsequently, we compared different web resources based on their basic properties, including fundamental theory, algorithms, and datasets. Third, we compared different web resources based on their basic properties and search results.

2. Public databases make drug-target data available

With the rapid development of bioinformatics, the number of drug-target interaction databases has increased in recent years, and they provide crucial support for TCM pharmacology research. These biological activity databases contain many chemical compositions, structure information, and activity data; provide free data retrieval for research; and deepen our understanding of the targets [14]. An overview of many of the significant public databases used in TCM pharmacology is presented in Table 1. These databases contain biological activity data verified through experiments of many compounds for many proteins. Their common goal is to integrate different data types to facilitate drug design and new drug development, which is also of great interest to pharmaceutical companies [17].

Table 1.

Overview of many public drug-target data resources available today. Due to the fast development of the area, this list does not claim to be exhaustive.

Databases	Latest release	Data	Web address
ChEMBL [18]	ChEMBL 32, released 03-01-2023	ChEMBL contains >2 million small molecules with >1.2 million bioactivity data points.	https://www.ebi.ac.uk/chembl/g/
DrugBank [19]	version 5.1.10, released 01-04-2023	DrugBank contains 15,441 drug entries, including 2739 approved small molecule drugs, 1575 approved biologics (proteins, peptides, vaccines, and allergenics), 134 nutraceuticals, and over 6716 experimental (discovery-phase) drugs. Additionally, 5294 non-redundant protein (i.e. drug target/enzyme/transporter/carrier) sequences are linked to these drug entries.	http://www.drugbank.ca
Comparative Toxico- genomics Database [20]	version 6.1, released 03-01-2023	CTD provide 51 million toxicogenomic relationships for over 17,100 chemicals, 54,300 genes, 6100 phenotypes, 7270 diseases and 202,000 exposure events, from 600 comparative species	http://ctdbase.org/
Therapeutic Target Database [21]	2022 version, released 09-29-2021	TTD contains 5059 patented drugs (collected from 3145 patents of WIPO, USA, Europe, Japan, etc.) targeting 215 successful, 236 clinical trial and 207 patent-recorded targets (with the structures of 4774 patented drugs drawn and provided in.mol files, and with the target activities of 3388 patented drugs collected from literature).	http://db.idrblab.net/ttd/
BindingDB [22]	2023 version, released 03-01-2023	BindingDB contains 41,328 Entries, each with a DOI, containing 2,680,191 binding data for 9000 protein targets and 1,153,558 small molecules.	http://www.bindingdb.org/bind/index.jsp
TDR Targets Database [23]	version 6.1, released 03-01-2021	TDR Targets Database contained 14,315 drug entries and 5239 targets.	https://tdrtargets.org/
The IUPHAR/BPS Guide to PHARMACOLOGY [24]	version 04-2022	The IUPHAR/BPS Guide to PHARMACOLOGY contains 11,596 ligands, 6219 targets proteins, and 49558 binding constants.	http://www.guidetopharmacology.org/

Public drug-target databases vary in data size, type, and how they are implemented. Public databases such as ChEMBL [18], DrugBank [19], and TTD [25] continuously increase in size, with the ChEMBL latest release comprising approximately 2 million distinct small molecules and approximately 1.2 million bioactivity data points. In recent years, a remarkable development is that the size of databases has increased, and different types of data have begun to be integrated [26]. DrugBank and TTD link approved small molecules with their known and explored therapeutic proteins and pathways. CTD provides manually curated chemical-gene/protein interactions, and chemical-disease and gene-disease relationships [20]. Combined with functional and pathway data, these data help develop hypotheses about the underlying mechanisms of environmental effects on diseases. In addition, the update speed of these data is relatively fast. Except for SuperTarget, all the other databases had versions released after 2020, ensuring the timeliness of this database information. These databases provide a convenient method for researchers to find identified targets of compounds.

3. Web servers for target prediction

With the massive growth of drug target databases in recent years, the calculation tools for predicting small molecule protein targets have become more critical. The user inputs the molecular structure into the computational bioactivity model and generates the “prediction” target of the compound based on the “informed guess” of the database without experiments [27]. However, their use is limited by complex algorithms and specific operating software systems, even if the algorithm can be obtained as open-source material. It is encouraging that web-based technology provides an opportunity to solve the problem of using computer simulation target prediction methods. Some free web resources for assisted target prediction analysis have recently emerged. These web services bring great convenience to the research of TCM pharmacology. By consulting the literature, we summarized the target prediction databases commonly used for the pharmacology of TCMs. Table 2 lists websites that apply to the target prediction of small molecules. It is worth noting that many platforms, such as TCMID [28], TCMSP [29], and BATMAN-TCM [30], which combine TCM ingredients, properties, targets, and diseases based on TCM information, have been established in recent years. However, we have also noticed that some web services are inaccessible (see Table 3).

Table 2.

Public web services that can be used for the target prediction of small molecules.

Type	Database	Web address	Year
Target prediction web servers	SEA [31]	http://sea.bkslab.org/	2007
	STITCH [32]	http://stitch.embl.de/	2008
	PharmMapper [33]	http://www.lilab-ecust.cn/pharmmapper/	2010
	idTarget [34]	http://idtarget.rcas.sinica.edu.tw	2012
	HitPick [35]	http://mips.helmholtz-muenchen.de/hitpick/	2013
	ChemMapper [36]	http://lilab-ecust.cn/chemmapper/	2013
	SwissTargetPrediction [37]	http://www.swisstargetprediction.ch/index.php	2014
	Superpred [38]	http://prediction.charite.de/	2014
	Polypharmacology Browser 2 [39]	http://ppb2.gdb.tools	2018
TCM related databases	TCMID [28]	http://www.megabionet.org/tcmid/	2012
	TCMSP [29]	http://tcmspw.com/tcmsp.php	2014
	BATMAN-TCM [30]	http://bionet.ncpsb.org/batman-tcm/	2016
	ETCM [40]	http://www.tcmip.cn/ETCM/index.php/Home/Index/index.html	2018
	SymMap [41]	https://www.symmap.org	2019
	HERB [42]	http://herb.ac.cn/	2021
	HIT2.0 [43]	http://www.badd-cao.net:2345/	2021
Not available	TarFisDock [44]	http://www.dddc.ac.cn/tarfisdock	2008
	DRAR-CP [45]	https://cpi.bio-x.cn/drar/?page=home	2011
	ChemProt [46]	http://www.cbs.dtu.dk/services/ChemProt/	2011
	TCMAnalyzer [47]	http://www.rcdd.org.cn/tcmanalyzer	2016
	TCM-Mesh [48]	http://mesh.tcm.microbioinformatics.org/	2017

Table 3.

Initial publishing journal of the web server articles and impact factors.

Type	Database	Publish	IF
Target prediction web servers	SEA	Nature Biotechnology	68.16
	STITCH	Nucleic Acids Research	19.16
	PharmMapper	Nucleic Acids Research	19.16
	idTarget	Nucleic Acids Research	19.16
	HitPick	Bioinformatics	6.931
	ChemMapper	Bioinformatics	6.931
	SwissTargetPrediction	Nucleic Acids Research	19.16
	Superpred	Nucleic Acids Research	19.16
	Polypharmacology Browser 2	Journal of Chemical Information and Modeling	6.162
TCM related databases	TCMID	Nucleic Acids Research	19.16
	TCMSP	Journal of Cheminformatics	8.489
	BATMAN-TCM	Scientific Reports	4.996
	ETCM	Nucleic Acids Research	19.16
	SymMap	Nucleic Acids Research	19.16
	HERB	Nucleic Acids Research	19.16
	HIT 2.0	Nucleic Acids Research	19.16

3.1. Impact of the web servers

One way to measure the impact of these databases is to consider the impact factors of the journals in which these databases were published. The journal impact factors are from the Journal Citation Reports (Clarivate Analytics, 2022). Journals that published these databases had an impact factor of >5.998, and the median impact factor of these journals was 19.16 (Table 5). We also collected and compared their citation counts from the Web of Science on March 15, 2023. Most of the 13 databases included in the analysis enjoyed a high number of citations (median = 112). The top three published databases, TCMSP, STITCH, and SwissTargetPrediction, have accumulated 2650 citations, more significant than the combined citations of the remaining 10 predictors. These results suggested that target prediction tools are based on solid interest of the scientific community.

Table 5.

Comparisons among the Web servers for TCM-related web servers.

Index	Function	Prediction methods	Dataset					Database	Input type
			Formulas	Herb	Ingredients	Target	Dieases
TCMID	TCMID provides information on six different areas (Fomula, Herb, Compound, Disease, Drug and Target) and their links.	NA	99582	10846	43413	17521	2679	Encyclopedia of Traditional Chinese Medicines/TCM@Taiwan/DrugBank/OMIM	TCM formula name Herb name in Chinese name or English Compound name STITCH database ID CAS number
TCMSP	Compounds, targets, and diseases; Herbal information ingredients with their ADME-related properties; Compounds-Targets/Targets-Diseases relationships;	SysDT model	NA	499	29384	3311	837	Drugbank/TTD/PharmGKB	Herbal name ingredient's chemical name InChIKey CAS number target name disease name
BATMAN-TCM	TCM ingredients' target prediction; Functional analyses of targets; The visualization of ingredient-Target-pathway/disease association network; KEGG biological pathway with highlighted targets; Comparison analysis of multiple TCMs	Similarity-based target prediction	46914	8159	25210	NA	NA	TCMID/Drugbank/TTD/KEGG/OMIM	TCM formula name Herb name PubChem_CID chemical structure of InChI format
ETCM	ETCM includes multiple aspects of clinical and functional essential information on TCM herb species, TCM formulas, herbal ingredients, validated or predicted drug targets, as well as related diseases	Similarity-based target prediction	3959	402	7284	2266	4323	The Fourth National Survey on Chinese/Materia Medica Resources Pharmacopoeia of the People's Republic of China (2015 version)/China Food and Drug Administration/Pubchem/OMIM DisGeNET/ORPHANET/HPO	TCM formula name Herb name in Chinese name or English Herbal name ingredient's chemical name
SymMap	SymMap integrates traditional Chinese medicine (TCM) with modern medicine (MM) through both internal molecular mechanism and external symptom mapping, thus provides massive information on herbs/ingredients, targets, as well as the clinical symptoms and diseases they are used to treat for drug screening efforts.	Similarity-based target prediction	NA	698	25975	20965	14082	HERB/ETCM/TCMSP/BATMAN-TCM/GeneCards MalaCards	Herb name in Chinese name or English or latin name or pinyin name Compound name CAS number PubChem ID
HERB	HERB database is a natural medicine database platform that integrates high-throughput experimental data and reference mining data. This database provides functions such as browsing, searching, viewing, and downloading for TCM herbs, TCM herbal active ingredients, target gene, disease, high-throughput experiment and reference mining data.	NA	NA	7263	49258	12933	28212	SymMap/TCMID/TCMSP/TCM-ID/NCBI GEO database/PubMed database	Herb name in Chinese name or English or latin name or pinyin name Compound name CAS number Gene name Disease name
HIT 2.0	HIT2.0 is a comprehensive searching and curation platform for herbal ingredients and target information based on literature evidences.Alternatively, uses can enter My-target curation system to check the detailed compound-target relationship and create the latest personal targeting profiles	Text-mining technology	NA	1250	1237	2208	NA	PubMed literatures/TTD/Drugbank	Herbs:Chinese name, Chinese characters or English name. Herbal ingredient: name, CAS number or Pubchem CID

3.2. Applications of the web servers

With the rapid growth of web servers for target prediction, their use has steadily increased over the past years. We selected two widely used databases, SEA and TCMSP, and analyzed the publication year, research area, and country or region of the publication that was cited. The number of publications using both databases increased with time, showing similar rapidly growing trends. For a deeper understanding of the related research areas of the journals using these databases, we further counted the number of research areas of these journal papers. The results showed that among the types of literature using the SEA database, the number of published types of literature in pharmacology pharmacy was the highest, followed by biochemistry, molecular biology, chemistry, and computer science. Unlike the SEA database, the types of literature using the TCMSP database were complementary, and the number of integrative medicine documents was the highest. Pharmacology pharmacy ranked second regarding the proportion of research volume in journal publications. This indicates that researchers are more inclined to use the database dedicated to TCM for integrative, complementary medicine. This can also be shown in the countries and regions where the authors of these journals belong. Among the researchers who use TCMSP data, 93.8% are from China. However, countries and regions are widely distributed in the SEA database, and Chinese researchers still make up a considerable proportion. Overall, applications of web servers in pharmacology and complementary integrative medicine have become increasingly popular in recent years. Notably, the research in integrative, complementary medicine accounts for a large proportion of all research.

3.3. Overview of the target prediction web servers

3.3.1. Prediction method

Several methods for predicting targets have been developed. The existing computational target prediction methods are generally divided into ligand-based methods centered on small molecules and structure-based processes based on protein three-dimensional structure information [49]. The ligand-based method predicts the target according to the principle of chemical similarity; that is, similar chemical structures show similar biological activities [50,51]. These methods first collect known ligands acting on the target and then determine the similarity between the chemical under evaluation and the known ligand set to infer the drug's interaction target [52]. The majority of web tools for target prediction are based on the principle of ligand-based methods.

The model based on the chemical structure is premised on similar chemical structures showing similar biological activities and depends on prior knowledge of bioactive ligands and protein structures [53,54]. Therefore, describing molecular structures to the computer is a crucial problem of this model. When a user enters a compound, a more complex binary vector, such as a molecular fingerprint, is usually obtained in the intermediate conversion step from a 1D sequential textual format (e.g., SMILES) to a 2D structure. The 3D structure requires additional information about the 3D conformation of the compounds. There are many molecular descriptors, and they show great differences in evaluating the similarity between molecules [55]. As shown in Table 4, compounds in the chemical structure-based model are divided into 2D and 3D fingerprints. The 2D molecular fingerprinting method focuses on the similarity of molecular chemical structure, while the 3D focuses more on pharmacophore similarity. Similarity ensemble approach (SEA) [31], SuperPred [38], HitPick [35], and The Polypharmacology Browser (PPB2) [39] use 2D molecular fingerprinting. Still, some specific methods exist; the difference lies in defining substructures and coding methods. SEA, SuperPred, and PPB2 adopted circular fingerprint-extended connectivity fingerprints, which were the most widely used 2D molecular fingerprints for building quantitative structure-activity relationship models of compounds [56]. PPB also represents two other molecular fingerprints: molecular fingerprint encoding composition, molecular shape, and pharmacophores. HitPick generated Morgan fingerprints with feature invariants similar to functional-class fingerprints [57]. In addition, numerous online tools using 3D similarity methods are emerging, such as ChemMapper, which explores chemical and target pharmacology relationships against any given small molecule by SHAFTS. The 3D similarity calculation is driven by molecular shape and chemotype features [58]. SwissTargetPrediction searches for similar molecules in 2D and 3D; the 2D measure uses path-based binary fingerprints, while the 3D measure is based on Electroshape 5D (ES5D) [59,60].

Table 4.

Comparisons among the Web servers for Target Prediction. The comparison contains the search method, molecular fingerprints, algorithm, different input types, general statistics, and limitations of several TCM databases.

Type of Method	Web servers	Searching Method	Molecular fingerprints	Algorithm	Dataset			Source	Input type
					Compounds	Targets	interactions
Chemical structure-based	SEA	2D Similarity	ECFP	BLAST	3665	N	N	ChEMBL MDDR	SMILES
	STITCH	NA	NA	Random Forests	430,000	N	N	DrugBank GLIDA TTD CTD ChEMBL PDSP PDB	SMILES
	Superpred	2D Similarity	ECFP4	Tanimoto coefficients	341000	1800	665000	SuperTarget ChEMBL BindingDB	PubChem ID SMILES MOL file Chemical structure
	SwissTargetPrediction	2D and 3D Similarity	Openbabel FP2/ES5D	Tanimoto coefficients/Manhattan distance	376,342	3068	580496	ChEMBL	Chemical structure SMILES SDF MOL2
	HitPick	2D Similarity	FCFP	NN searches with NB machine learning	99,572	1375	145549	STITCH	SMILES
	PPB 2	2D Similarity searching	MQN Xfp ECFP	NN searches with NB machine learning/NN searches with NB machine learning	344163	1720	555346	ChEMBL	SMILES Chemical structure
	ChemMapper	3D Similarity searching	SHAFTS	Random walk algorithm	800,000	NA	NA	ChEMBL	SMILES Chemical structure
Structure-based methods	PharmMapper	Pharmacophore based	NA	Triangle hashing (TriHash) Genetic algorithm (GA) optimization	NA	23236	16159	TargetBank DrugBank BindingDB PDTD	SMILES Chemical structure

Note: ECFP: Extended Connectivity Fingerprints; FCFP: Functional-Class Fingerprints; ES5D: Electroshape vectors; MQN: Molecular Quantum Number; Xfp: atom category extended atom-pair fingerprint SHAFTS: SHApe-FeaTure Similarity; NN: nearest neighbor; NB: Naıve Bayes; DNN: Deep Neural Network.

Another widely used target recognition method is the structure-based target prediction, which refers to the protein-ligand interaction mode corresponding to the desired pharmacological action, and can be regarded as the greatest common divisor shared by a group of active molecules [61,62]. Therefore, the interaction between protein targets and any ligand can be characterized by matching the 3D chemical structures with the model characteristics [63]. PharmMapper is an online web server based on a pharmacophore model that identifies potential target candidates for small molecules [64].

3.3.2. Searching method

The target prediction method of each web server is different. The simplest and fastest target prediction method relies on querying the structural similarity of compounds, such as in SuperPred and SwissTargetPrediction. SuperPred suggests protein targets according to the molecular similarity between the query compound and compounds with known biological activities using the Tanimoto coefficient and sorts the targets accordingly. SwissTargetPrediction predicts the target of a compound based on its similarity to the 2D and 3D structures of known compounds. The 2D similarity is also quantified by the Tanimoto coefficient of the query molecule and the screening molecule, and 3D similarity is quantified by the Manhattan distance between the electric shape vectors (ES5D) of 20 conformational isomers of the query molecule and the screening molecule. The final score corresponds to the combination of similarity measures based on a logistic regression of similarity values, in which the most similar ligands use 2D and 3D similarity measures [65]. Unlike single compound similarity, the SEA server uses the similarity ensemble method based on the sequence alignment algorithm BLAST to group proteins based on ligand topology. A set of web servers, such as HitPick and PPB2, use machine learning algorithms to generate predictions. HitPick applies the B-score method for hit identification combining one-nearest-neighbor similarity searching [66] and Laplacian-modified naïve Bayesian target models [67] to predict targets of the identified hits. PPB2 combines nearest neighbor searches with naïve Bayes and deep neural network (DNN) machine learning to enhance performance. Nearest neighbor and naïve Bayes combinations perform best in terms of precision statistics in PPB2.

3.3.3. Data set

Target prediction methods focus on ligand-based methods to identify structures similar to those from a bioactivity database. Therefore, the number of internal datasets, usually from several databases, is a critical factor affecting the web server quality. These web servers cover between 3665 and 800,000 compounds, including drugs and drug-like compounds, while the number of protein targets ranges between 1720 and 23,236 (Table 4). ChemMapper, PharmMapper, and STITCH contain human proteins and proteins of other species. Another important indicator by which to evaluate the web servers is the number of compound/ligand-target interactions. This number varies between 16,159 and 665,000. The SuperPred database contains the largest number of compound/ligand-target interactions. These compounds and target data derived from source databases include MDDR [68], PDSP [69], GLIDA [70], Binding DB [22], TTD [71], CTD [72], KEGG [73], DrugBank [19], SuperTarget Matador [74], STITCH [32], ChEMBL [18], and PDTD [75]. Typically, the ChEMBL database is used as a public source for chemical structures, while DrugBank, TTD, and BindingDB are for bioactivity. Generally, SuperPred and SwissTargetPrediction contain the largest quantity of data.

3.4. TCM-related web servers

As databases for herbal ingredients emerge, various web servers are used for TCM analysis to bridge the gap between TCM and modern Western medicine [47]. Unlike the commonly used target prediction websites, these web servers often record information related to TCM collected from different resources and using text-mining methods, including information on herbs, chemicals, targets, drug-target networks, and related drug target-disease networks. Recent studies compared the currently available web servers for TCM analysis in more detail [76]. Our paper focuses on the TCMID, TCMSP, BATMAN-TCM, ETCM, SymMap, and HERB databases, which store herbal compound-target information (Table 5). These databases all provide herbal compound-target-disease networks, which support studies on the mechanism of TCM. These databases also have their unique functions.

TCMID has the greatest number of medicines and TCMs, with 99,582 and 8,159, respectively. HERB has the most number of components collected, with 49,258. The greatest number of targets is in SymMap, with 20,965 targets. In addition to TCM and its components, SymMap includes 1717 corresponding TCM syndromes, corresponding to 961 Western medicine symptoms. In this way, SymMap relates TCM and modern medicine from phenotype to the molecular level. The TCMSP and ETMC databases include many herbal entries with ADME properties for drug screening and evaluation.

In addition, TCMSP, BATMAN-TCM, ETCM, SymMap, and HERB are intelligent, integrated pharmacological research platforms of TCM that can be used for target prediction, functional analysis, and reverse searches of TCM and compound prescriptions. It is worth noting that TCMSP not only contains the known targets of compounds but also establishes a target prediction method based on random forest and support vector machine (SVM), called the SysDT model, which integrates chemical, genomic, and pharmacological information [77]. The target prediction method of BATMAN-TCM is based on the similarity of chemical structure [30]. The core idea is to rank possible drug-target interactions based on their similarity to known drug-target interactions obtained from the DrugBank, KEGG, and TTD databases [16].

3.5. Comparative analysis of target prediction web servers in TCM pharmacology research

To further compare the web servers, we collected 54 indexed components of different TCMs from Chinese Pharmacopoeia and compared the ability of other web servers to predict known targets (Supplementary Table S1). Compound structures (SMILE) were collected from PubChem by querying their compound name. Next, the known targets were retrieved from DrugBank and TTD, and 54 known targets of chemical constituents were collected (Supplementary Table S2) (see Fig. 4).

Fig. 4.

Web servers based on chemical structure-based model workflow.

We counted the number of targets predicted by each web server, and the results showed that the number varied greatly, ranging from several to hundreds. The number of target outputs by SwissTargetPrediction and PPB2 database was certain. The number of targets predicted by most web servers was less than 100. Next, we performed statistical tests to identify the accuracy of each web server in predicting known targets. As shown in Fig. 5, considering that at least one target was predicted, SwissTargetPrediction demonstrated the best performance in accuracy (53.7%), followed by PPB2 (51.9%), SuperPred (44.4%), and SEA (37.0%), while the others were below 25%. It is worth noting that the prediction accuracy of the unique databases of TCM, such as TCMID, TCMSP, and BATMAN-TCM, was not high.

Fig. 5.

The search results of different web servers for 54 compounds: (A) the number of targets predicted by each web server; (B) the predicted results of different web servers.

4. Discussion and conclusion

Drug target identification is one of drug discovery's most critical and complex steps [78]. The construction and development of computational methods and data-driven technological approaches have brought new ideas for studying the material basis and action mechanisms of TCM. These systems provide new ideas for researching complex TCM systems and offer further scientific and technological support for rational clinical drug use and new drug research and development. In this review, we witnessed the efforts of current computing technology in drug target prediction and the modernization of TCM in the form of online tools. We also note that all developers have made commendable efforts to provide non-computing users with tools to enable a broad audience to analyze drug targets quickly. Each web server has its advantages regarding the specificity of its data and algorithm. Although significant progress has been made in its methodology and applications, current TCM target research has several problems and limitations.

Unfortunately, the target prediction databases applied to TCM are based on Western medicine compositions and target databases. Unlike Western drugs, most herbal composite compound targets are unknown. For example, even by comprehensive literature curation, only 6% of 9862 herbal ingredients in HIT were found to have direct or indirect targets in PubMed abstracts [79]. TCM-related information has typical big data characteristics. The existing web tools have their emphases, and the information in a single database is limited and biased. Therefore, we should build high-quality databases and strengthen the association of major databases. The prediction methods used in the target prediction database mainly include target discovery based on ligand structure and target prediction based on receptor structure characteristics. Ligand-based drug target prediction is performed according to the similarity between molecules to those in the database. However, the affinity between the compounds and potential targets is not reflected. Moreover, the mode of interaction between the compound and the target cannot be predicted. Therefore, it is usually necessary to comprehensively use different prediction tools and make a selective experimental verification of the target through manual judgment.

Many scholars have proposed AI algorithm models. With the rapid accumulation and digital transformation of drug research and development data, as well as the accelerated development of AI technology, machine learning models such as decision trees, random forest, and SVM, as well as deep learning algorithms such as DNNs, convolutional neural networks, and cyclic neural networks, are gradually applied to the field of drug discovery, such as peptide synthesis, virtual screening, toxicity prediction, drug monitoring and release, pharmacophore modeling, quantitative structure-activity relationships, drug repositioning, multiple pharmacological and physiological activities, and other drug discovery processes [80]. For example, a multi-target drug prediction platform for anti-AD was established based on a machine learning algorithm. The global drug database was predicted and mined, from which 13 drugs were selected, which acted on at least one anti-AD drug target that could resist AD through various actions [81]. AI technology based on computational learning is also suitable for predicting and screening the activity of TCM with complex components. AI technology has been successfully applied to the analysis of complex functions of TCM and the screening of new drugs in TCM, providing a new direction and ideas for speeding up the modernization of TCM.

By summarizing and sorting out the multi-dimensional big data of TCM from different sources, an extensive dataset of molecular-target interactions according to the characteristics of TCM is constructed. Using deep learning to establish an intelligent prediction model of "molecular-target" can improve predictability, reliability, and repeatability. Moreover, efficient algorithms that follow the complex characteristics of TCM are lacking. Therefore, the development of TCM needs high-quality data and efficient algorithms. This is conducive to the modernization and internationalization of TCM and may provide significant technical support for drug development, clinical diagnosis, and personalized medicine.

Author contribution statement

All authors listed have significantly contributed to the development and the writing of this article. ⟨/p⟩

5. Data availability statement

Data included in article/supplementary material/referenced in article.

Funding

This work was supported by the National Natural Science Foundation of China (Nos. 41806191), China's National Key Research and Development Program (2017YFC1702703), Shandong Taishan Scholar Climbing Project, China, the youth innovation team of Shandong University of Traditional Chinese Medicine, China.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary data

The following is the Supplementary data to this article.