Compound-level identification of sasang constitution type-specific personalized herbal medicine using data science approach

Abstract

Introduction

Sasang Constitutional Medicine (SCM) is a type of traditional Korean medicine where patients are classified as one of four Sasang constitution types (Sasang type) and medications consisting of medicinal herbs are prescribed according to the Sasang type. Despite the importance of personalized medicine, the operation mechanism is largely unknown. To gain a better understanding, we investigated the compound information that composes Sasang type-specific personalized herbal medicines on both multivariate and univariate levels.

Methods

Five machine learning classifiers including extremely randomized trees (ERT) were trained to investigate whether the Sasang type can be explained by compound information at the multivariate level. Hierarchical clustering was conducted to determine whether compounds are processed distributedly or specifically. Taxonomic and biosynthetic analyses were conducted on these compounds. A univariate level statistical test was conducted to provide more robust Sasang type-specific compound information.

Results

Using the trained ERT classifier, sixty important compounds were extracted. The sixty compounds were clustered into three groups, corresponding to each Sasang type-prominent compounds, suggesting that most compounds have specific preference for the Sasang type. Structural and biosynthetic characteristics of these Sasang type-prominent compounds were determined based on taxonomy and pathway analyses. Fourteen compounds showed statistically significant relevance with the Sasang type. Additionally, we predicted the Sasang type of unknown herbs, which were confirmed by their biological effects in functional assays.

Conclusion

This study investigated the personalized herbal medicines of the SCM using compound information. This study provided information on the chemical characteristics of the compounds that are essential for classifying the Sasang type of medicinal herbs, as well as predictions regarding the Sasang type of the commonly used but unidentified medicinal herbs.

1. Introduction

Sasang Constitutional Medicine (SCM) is a type of traditional Korean medicine, in which patients are classified into one of four Sasang constitution types (Sasang type): So-Eum (SE), So-Yang (SY), Tae-Eum (TE), and Tae-Yang (TY) [1]. The herbs are classified into four groups corresponding to the four Sasang types. For example, Panax Ginseng is used only for SE patients and Ephedra herba is used for TE patients. In addition, SCM is well-known for its personalized medicine characteristics, which means that despite the similar symptoms, the medications consisting of various medicinal herbs are prescribed differently depending on the patient's Sasang type.

Despite the importance of personalized medicine, however, it is only possible to assume the classification principle of SCM because the classification criteria for the herbs into four Sasang type groups or the meaning of each herb belonging to each Sasang type are largely unknown. Thus far, although various studies have been conducted to find the criteria and the meaning of herbal classification [[2], [3], [4]], most studies were conducted by applying the theoretical concepts of herbs used in conventional traditional medicine or by narratively reviewing the results for each Sasang type [3, 5]. Recently, chemical property-based various machine learning approaches have been applied to investigate natural products including herbal medicines [6, 7].

To gain a better understanding of the personalized medicine characteristics of SCM and its operation, we conducted a detailed investigation of the compound information that composes the Sasang type-specific personalized herbal medicines on both multivariate and univariate levels in this study. Using machine learning (ML) techniques and statistical analyses, compound patterns and specific compounds that enable Sasang type classification were found, and the chemical characteristics of the important compounds contributing to the classification were analyzed. The Sasang types of medicinal herbs, whose Sasang types were unidentified, were predicted based on compound information, and the prediction results were confirmed by a simple functional assay.

2. Methods

2.1. Identification of herbs

The list of herbs for each Sasang constitution type was obtained from 『Donguisusebowon Sinchukbon』 (『東醫壽世保元 辛丑本』; Longevity and Life Preservation of the Eastern Medicine), a book in which Jema Lee, the founder of SCM, presented the final result of his categorization of herbs by Sasang type. Previous studies [8, 9] that identified the medicinal herbs belonging to each Sasang type were also referenced. As a result, 144 herbs (47 for SE type, 37 for SY type, 44 for TE type, and 16 for TY type) were included in this study.

2.2. Construction of an herb-compound matrix dataset

Traditional Chinese Medicine Systems Pharmacology and Analysis Platform (TCMSP) (https://old.tcmsp-e.com/tcmsp.php), a database containing information on 499 herbs and 29,384 compounds, was used to extract compound information for each herb [10]. TCMSP was selected because the number of compounds included in the database has a greater advantage than other databases. When 144 herbs were analyzed using the database, 92 medicinal herbs (31 for the SE type, 30 for the SY type, 27 for the TE type, and 4 for the TY type) were included, allowing further analysis (Table 1).

Table 1.

List of SCM medicinal herbs.

	Sasang type	Chinese name	Pronunciation	English name
1	SE	白芍藥	Baishao	Paeoniae Radix Alba
2	SE	白朮	Baizhu	Atractylodes Macrocephala Koidz.
3	SE	半夏	Banxia	Arum Ternatum Thunb.
4	SE	陳皮	Chenpi	Citrus Reticulata
5	SE	川芎	Chuanxiong	Chuanxiong Rhizoma
6	SE	葱白	Congbai	Allii Fistulost Bulbus
7	SE	大腹皮	Dafupi	Areca Catechu L.
8	SE	當歸	Danggui	Angelicae Sinensis Radix
9	SE	大蒜	Dasuan	Allii Sativi Bulbus
10	SE	大棗	Dazao	Jujubae Fructus
11	SE	付子	Fuzi	Aconiti Lateralis Radix Praeparata
12	SE	甘草	Gancao	Licorice
13	SE	乾薑	Ganjiang	Zingiberis Rhizoma
14	SE	良薑	Gaoliangjiang	Alpiniae Officinarum Rhizome
15	SE	藿香	Guanghuoxiang	Pogostemon Cablin (Blanco) Benth.
16	SE	桂枝	Guizhi	Cinnamomi Ramulus
17	SE	厚朴	Houpu	Magnolia Officinalis Rehd Et Wils
18	SE	胡椒	Hujiao	Piperis Fructus
19	SE	靑皮	Qingpi	Citri Reticulatae Pericarpium Viride
20	SE	人蔘	Renshen	Panax Ginseng C. A. Mey.
21	SE	官桂	Rougui	Cinnanmomi Cortex
22	SE	生薑	Shengjiang	Zingiber Officinale Roscoe
23	SE	吳茱萸	Wuzhuyu	Evodiae Fructus
24	SE	香付子	Xiangfu	Cyperi Rhizoma
25	SE	茵蔯	Yinchen	Artemisiae Scopariae Herba
26	SE	鸎粟殼	Yingsuke	Papaveris Pericarpium
27	SE	益智仁	Yizhi	Alpiniae Oxyphyliae Fructus
28	SE	枳實	Zhishi	Aurantii Fructus Immaturus
29	SE	蒼朮	Cangzhu	Atractylodes Lancea (Thunb.)Dc.
30	SE	黃芪	Huangqi	Hedysarum Multijugum Maxim.
31	SE	砂仁	Sharen	Amomum Aurantiacum H. T. Tsai Et S. W. Zhao
32	SY	薄荷	Bohe	Menthae Herba
33	SY	柴胡	Chaihu	Radix Bupleuri
34	SY	車前子	Cheqianzi	Plantaginis Semen
35	SY	地骨皮	Digupi	Lycii Cortex
36	SY	獨活	Duhuo	Radix Angelicae Biseratae
37	SY	防風	Fangfeng	Saposhnikoviae Radix
38	SY	茯笭	Fuling	Poria Cocos (Schw.) Wolf.
39	SY	覆盆子	Fupenzi	Rubi Fructus
40	SY	甘遂	Gansui	Kansui Radix
41	SY	枸杞子	Gouqizi	Lycii Fructus
42	SY	荊芥	Jingjie	Schizonepetae Herba
43	SY	金銀花	Jinyinhua	Lonicerae Japonicae Flos
44	SY	苦蔘	Kushen	Sophorae Flavescentis Radix
45	SY	連翹	Lianqiao	Forsythiae Fructus
46	SY	牧丹皮	Mudanpi	Cortex Moutan
47	SY	木通	Mutong	Caulis Akebiae
48	SY	牛蒡子	Niubangzi	Fructus Arctii
49	SY	前胡	Qianhu	Peucedani Radix
50	SY	羌活	Qianghuo	Notopterygii Rhizoma Et Radix
51	SY	山茱萸	Shanzhuyu	Cornus Officinalis Sieb. Et Zucc.
52	SY	熟地黃	Shudihuang	Rehmanniae Radix Praeparata
53	SY	玄參	Xuanshen	Figwort Root
54	SY	澤瀉	Zexie	Alisma Orientale (Sam.) Juz.
55	SY	知母	Zhimu	Anemarrhenae Rhizoma
56	SY	山梔子	Zhizi	Gardeniae Fructus
57	SY	豬苓	Zhuling	Polyporus Umbellatus (Pers)Fr.
58	SY	黃栢	Huangbo	Phellodendri Chinrnsis Cortex
59	SY	川黃連	Huanglian	Coptidis Rhizoma
60	SY	沒藥	Moyao	Myrrha
61	SY	乳香	Ruxiang	Olibanun
62	TE	白果	Baiguo	Ginkgo Semen
63	TE	白芷	Baizhi	A. Dahurica (Fisch.) Benth. Et Hook
64	TE	栢子仁	Baiziren	Platycladi Semen
65	TE	五味子	Beiwuweizi	Schisandrae Chinensis Fructus
66	TE	大黃	Dahuang	Radix Rhei Et Rhizome
67	TE	浮萍	Fuping	Spirodelae Herba
68	TE	藁本	Gaoben	Ligustici Rhizoma Et Radix
69	TE	葛根	Gegen	Radix Puerariae
70	TE	苽蔕	Guadi	Calyx Cucumis
71	TE	黃芩	Huangqin	Scutellariae Radix
72	TE	桔梗	Jiegeng	Platycodon Grandiforus
73	TE	款冬花	Kuandonghua	Farfarae Flos
74	TE	蘿葍子	Laifuzi	Raphani Semen
75	TE	麻黃	Mahuang	Ephedra Herba
76	TE	牛黃	Niuhuang	Bovis Calculus
77	TE	蒲黄	Puhuang	Pollen Typhae
78	TE	桑白皮	Sangbaipi	Mori Cortex
79	TE	山藥	Shanyao	Rhizoma Dioscoreae
80	TE	升麻	Shengma	Cimicifugae Rhizoma
81	TE	石菖蒲	Shichangpu	Acoritataninowii Rhizoma
82	TE	使君子	Shijunzi	Quisqualis Indica
83	TE	酸枣仁	Suanzaoren	Ziziphi Spinosae Semen
84	TE	天門冬	Tiandong	Asparagi Radix
85	TE	烏梅	Wumei	Mume Fructus
86	TE	甘菊花	Yejuhua	Chrysanthemi Indici Flos
87	TE	薏苡仁	Yiyiren	Coicis Semen
88	TE	皂角	Zaojiaoci	Gleditsiae Spina
89	TY	蘆根	Lugen	Phragmitis Rhizoma
90	TY	木苽	Mugau	Chaenomeles Sinensis (Thouin) Koehne
91	TY	松花	Songhuafen	Pine Pollen
92	TY	松節	Songjie	Lignum Pini Nodi

SE, So-Eum type; SY, So-Yang type; TE, Tae-Eum type.

The compound information for each herb was extracted from the database and transformed into a simplified molecular-input line-entry system (SMILES) string [11]. A vector containing all compounds of 92 herbs was constructed using one-hot encoding (size of 1 × 4745). As a result, a herb-compound matrix (size of 92 × 4745) was constructed and used in analyses in this study. All data preprocessing and analysis were performed using Pandas, a Python library for data manipulation and analysis, and Scikit-learn, a Python module that integrates a broad range of machine learning algorithms [12].

2.3. Machine learning (ML) experimental details

2.3.1. ML model selection

Five well-known supervised machine-learning algorithms for classification were applied in this study. The models compared are as follows: extremely randomized trees (ERT), extreme gradient boosting (XGBoost), linear and nonlinear support vector machine (SVM), and multinomial logistic regression (Mlogit).

The ERT classifier is a decision tree-based ensemble method that is similar to random forests but uses randomly selected cut-off values rather than the optimal one. The strength of the ERT classifier is that it is robust to noise and can thus lead to a further decrease in overall variance while performing largely equal to or better than other tree-based classifiers [13]. Furthermore, the ensemble method can rank the importance of features used in a classification problem [14]. The XGBoost is a kind of Machine Learning algorithm belonging to a decision-tree-based ensemble and enrolls an advanced framework of gradient boosting [15]. The SVM classifier searches for the optimal hyperplane that maximizes the margin between classes in high-dimensional space [16]. The SVM classifier can be used as linear or nonlinear classifiers according to the applied kernel. The radial basis function (Gaussian kernel) was applied for the nonlinear SVM classifier. The one-vs.-rest scheme was used to apply the SVM, a binary classifier, into a multi-class problem. The Mlogit classifier is used to predict a nominal dependent variable with more than two categories [17]. The strength of the Mlogit model is that it measures how relevant a predictor (coefficient size) is and the direction of association (positive or negative) of the predictor.

2.3.2. Feature selection

The optimal number of features was selected by calculating the performance (accuracy) of the model while increasing the number of features from 10 to 4700 in increments of 10. The features were included from the feature (compound) with the highest feature importance score. It was calculated by a double nested cross-validation. This procedure was repeated ten times to avoid inconsistent results caused by randomness. Sixty features showed the best performance. The sixty features (compounds) showing the highest feature importance scores were selected by a nested cross-validation performance. All the feature selection procedures were conducted in the nested training set of each fold to avoid data leakage.

2.3.3. Hyperparameter optimization

For hyperparameter optimization, a randomized search on hyperparameters with nested cross-validation to avoid data leakage was conducted. The hyperparameter configuration can be varied across the folds because the hyperparameters were tuned for each fold. Supplementary Table S1summarizes the searched hyperparameters and their range for each ML model.

2.4. Model performance assessment

2.4.1. K-fold cross-validation

Each model was trained with stratified k-fold cross-validation (k = 4), in which the dataset was divided randomly into k disjoint subsets of approximately equal size according to the Sasang type.

2.4.2. AUROC, precision, recall, f1 score, and accuracy

The AUC was used to evaluate how well the ML model distinguishes the Sasang types with the compounds configuration of each herb. The area under the receiving operating characteristic (AUROC) curve was calculated using the implementation in the Scikit-learn Python package. The precision, recall, f1 score, and accuracy (equations (1), (2), (3), (4))) were used to evaluate the performance of the machine-learning model.

$$\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}=\frac{\mathrm{t}\mathrm{p}}{\mathrm{t}\mathrm{p}+\mathrm{f}\mathrm{p}}$$ (1)

$$\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}=\frac{\mathrm{t}\mathrm{p}}{\mathrm{t}\mathrm{p}+\mathrm{f}\mathrm{n}}$$ (2)

$$\mathrm{f}1\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}=\frac{2*\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}*\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}{\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}+\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}$$ (3)

$$\mathrm{a}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y}=\frac{\mathrm{t}\mathrm{p}+\mathrm{t}\mathrm{n}}{\mathrm{t}\mathrm{p}+\mathrm{t}\mathrm{n}+\mathrm{f}\mathrm{p}+\mathrm{f}\mathrm{n}}$$ (4)

In this equation, tp denotes the true positive; fp, false positive; tn, true negative; fn, false negative. Macro-average methods that treat all classes equally were used to calculate the average values in multi-class classification settings.

2.5. Cell-based analysis

2.5.1. Cell culture and reagents

AGS (derived from the stomach), HepG2 (derived from the liver) and NRK-52E (derived from the kidney) cells were obtained from American Type Culture Collection (ATCC, Rockville, MD). The cells were maintained in Dulbecco's modified Eagle's medium liquid (DMEM) with high glucose levels, 10% fetal bovine serum (FBS), 50 units/ml penicillin, and 50 μg/ml streptomycin at 37 °C in a humidified atmosphere containing 5% CO2. For all experiments, the cells were starved for 12 h in FBS-free media [18]. Curcuma longa Radix, Houttuynia cordata and Leonurus japonicus Houtt were extracted using the medicinal standard herbs, which are guaranteed by Korea FDA and produced by the pharmaceutical company (Daewon pharmacy, Korea) approved as the good manufacturing practice (GMP) system as previously described [19, 20].

2.5.2. MTT assay

The cells were plated at a density of 1 × 10⁵ cells per well in 48-well culture plates and incubated in an FBS-free medium for 12 h. AGS and NRK-52E cell were incubated with drugs for 24 h [18]. HepG2 cells were incubated with drugs for 1 h, followed by a treatment with AA (10 μM) for 12 h and then iron (5 μM) for 6 h. The cell viability was defined as relative to the untreated control [i.e., viability (% of control) = 100 × (absorbance of the treated sample)/(absorbance of control)] as previously described [19,20].

2.6. Statistical analysis

Analysis of variance (ANOVA) test was used to assess the differences in the number of compounds belonging to the four Sasang types. Fisher's exact test was used to determine if there are non-random associations between each compound and each Sasang type, at a univariate level. Multiple comparison correction was not conducted because the purpose of Fisher's exact test was to suggest relevant compound candidates.

3. Results

3.1. Basic characteristics of the compounds comprising the SCM medicinal herbs

Ninety-two medicinal herbs were analyzed (31 for the SE type, 30 for the SY type, 27 for the TE type, and four for the TY type) in the current study (Table 1). Before examining whether it was possible to discriminate the Sasang type using the compound information composing each herb, we first described the basic characteristics of the compound information. The average numbers of compounds were 99.3, 95.2, 65.0, and 34.5 for SE, SY, TE, and TY types, respectively (Fig. 1A). We examined the difference between the number of compounds among the different Sasang type groups. Although the difference of compound number among the Sasang types was marginally significant (F-stat = 2.68, p-val = 0.052), it seems that the TY type is the primary factor, as the compound number of the other types appears comparable. (Fig. 1B). The overlap of compound lists belonging to each Sasang type was also examined using a Venn diagram (Fig. 1C).

Fig. 1.

Basic characteristics of Sasang herb data (A) Box plot of the compounds data used in the analysis. The number of compounds for each Sasang type was analyzed using a box plot, and the average number of compounds for each type was calculated. SE, So-Eum type; SY, So-Yang type; TE, Tae-Eum type; TY, Tae-Yang type. (B) Result of one-way ANOVA. The data were plotted as a function of the means and 95% confidence intervals. Tukey was conducted as a post-hoc analysis. (C) Venn diagram of the compounds belonging to each Sasang type.

3.2. Classification of each herb to a specific sasang type being explained by the multivariate level compound information

ML classifiers were trained for Sasang type and their performance of the classifiers was assessed to find out whether the Sasang type information can be explained by compound combination at the multivariate level, i.e., whether it is possible to discriminate the corresponding Sasang type using the compound information composing each herb. Four TY-type herbs were excluded because of their small sample size [21,22]. Four samples were too small for the model to learn the generalizable pattern. As a result, 88 herbs were used for subsequent analyses.

Five well-known classification ML models were applied for this study. The ML models compared in this study were as follows: ERT, XGBoost, linear and nonlinear (RBF) SVM, and Mlogit (see Materials and Methods for more details). Since the ERT model outperformed the other models (Table 2), the ERT classifier was selected. The macro-averaged accuracy and f1 score of the ERT classifier were 0.511 ± 0.121 and 0.498 ± 0.129, respectively (mean ± SD).

Table 2.

Overall classification performance for each model.

Model	Precision	Recall	F1 score	Accuracy
ERT	0.505 $\pm$ 0.125	0.522 $\pm$ 0.139	0.498 $\pm$ 0.129	0.511 $\pm$ 0.121
XGBoost	0.394 $\pm$ 0.064	0.436 $\pm$ 0.078	0.374 $\pm$ 0.077	0.397 $\pm$ 0.067
SVM (RBF)	0.475 $\pm$ 0.066	0.482 $\pm$ 0.062	0.467 $\pm$ 0.070	0.489 $\pm$ 0.067
SVM(Linear)	0.441 $\pm$ 0.087	0.514 $\pm$ 0.109	0.420 $\pm$ 0.092	0.455 $\pm$ 0.085
Mlogit	0.464 $\pm$ 0.061	0.446 $\pm$ 0.061	0.422 $\pm$ 0.049	0.466 $\pm$ 0.067

Mean $\pm$ SD. ERT, extremely randomized trees; XGBoost, extreme gradient boosting; SVM, support vector machine; Mlogit, multinomial logistic regression.

The ERT classifier was investigated more thoroughly for a detailed analysis of the classifier performance. The macro-average AUROC of the classifier was 0.73 (Fig. 2). The classification performance of the ERT model for the individual Sasang type was as follows: the average precision, recall, and f1 score for the SE type were 0.612, 0.671, and 0.621, respectively; 0.464, 0.451, and 0.434, respectively, for the SY type; 0.440, 0.443, and 0.439, respectively, for TE type (Table 3, Supplementary Figure S1). The result showed statistically significant classification performance for the SE and SY type, suggesting that the configuration of the compound composing each herb has information to discriminate the Sasang type of each herb. In other words, the classification of each herb to a specific type can be explained by the compound configuration of the herbs.

Fig. 2.

Classification performance of the Sasang type decoder. Receiver operating characteristic (ROC) curve of 88 herbs with 60 compound features. SE, So-Eum type; SY, So-Yang type; TE, Tae-Eum type.

Table 3.

Classification performance of the extremely randomized trees classifier for individual Sasang type.

	Precision	Recall	F1 score	Accuracy
SE	0.612 $\pm$ 0.155	0.671 $\pm$ 0.132	0.621 $\pm$ 0.076	0.511 $\pm$ 0.121
SY	0.464 $\pm$ 0.271	0.451 $\pm$ 0.127	0.434 $\pm$ 0.189
TE	0.440 $\pm$ 0.132	0.443 $\pm$ 0.185	0.439 $\pm$ 0.159
Average	0.505	0.522	0.498	0.511

Mean $\pm$ SD. SE, So-Eum type; SY, So-Yang type; TE, Tae-Eum type.

3.3. Most compounds showing selectivity for particular sasang type

To avoid curse of dimensionality, feature selection was conducted. The optimal number of features was chosen by the double nested cross-validation performance (see the Methods 2.3.2). Sixty turned out to be the optimal number and the sixty features (compounds) that have the highest feature importance scores were selected by the nested cross-validation performance. The sixty features (compounds) were analyzed to understand which compounds are processed in a distributed manner and which are processed in a labeled-line manner for the purpose of classifying the Sasang type information. The sixty compounds were clustered based on the cosine similarity between each vector of the sixty compounds (size of 88 × 1) and representative Sasang type vector for each type (size of 88 × 1), representing the distribution of each compound and each Sasang type within 88 herbs, respectively. To define the clusters, the dendrogram was cut at the second level, resulting in three clusters and one compound (Fig. 3). We found that the three clusters corresponded to TE-prominent (n = 8), SY-prominent (n = 9), and SE-prominent compound groups (n = 42), which suggests that the majority of compounds have selectivity for particular Sasang type.

Fig. 3.

Clustered heatmap of the 60 compounds showing SC-prominent compounds. Each row represents each compound, and the column represents each SC type. The color represents the similarity (cosine similarity) between the distribution of each compound and that of each SC type. The data was clustered with respect to rows. The compounds were clustered into three groups at the second level dendrogram, resulting in TE-prominent, SY-prominent, and SE-prominent compounds, respectively. SE, So-Eum type; SY, So-Yang type; TE, Tae-Eum type.

3.4. Chemical characteristics of the sasasng type-prominent compounds

To identify the chemical characteristics of the Sasang type-prominent compounds, firstly, the chemical taxonomy of each cluster, i.e., the structural classification of chemical entities using ClassyFire [23], was identified (Supplementary Tables S2, S3, and S4). Many of the TE-prominent compounds showed a class of fatty acyls (7/8). Unlike TE-prominent compounds, SY-prominent compounds showed heterogeneous composition with various classes: prenol lipids (2/9), fatty acyls (2/9), coumarins and derivatives (2/9), carboxyl acids and derivatives (1/9), cinnamic acids and derivatives (1/9), and organooxygen compounds (1/9). The SE-prominent compounds were identified as prenol lipids (21/42), flavonoids (9/42), and benzene and substituted derivatives (4/42).

Furthermore, a review of the biosynthetic characteristics of the major secondary metabolites in sixty compounds that are crucial in the Sasang type classification could suggest an interesting point of view [24]. As shown in Supplementary Table S2, the majority of the components in the medicinal herbs classified as TE type were fatty acid-based substances synthesized via a polyketide biosynthetic pathway that requires various polyketide synthases. In the case of compounds from the medicinal herbs for the SY type, the shikimate, mevalonate, and polyketide pathways were involved in biosynthesizing these metabolites. As in the case of SY type, the compound list for the SE type showed that various biosynthetic pathways were involved for these compounds, but terpenoids appeared most commonly, basically biosynthesized through the mevalonic (isoprenoid) pathway [25].

3.5. Identification of sasasng type-specific compounds from a univariate level analysis

A univariate level statistical test (Fisher's exact test) was conducted on the sixty compounds showing a high feature importance score to find the compounds that are more relevant to the Sasang type, i.e., the Sasang type-specific compounds. Fourteen Sasang type-specific compounds were found accordingly, and the relative ratio of each compound for each Sasang type was analyzed (Fig. 4). Among them, ten, three, and one compound are relevant to the SE, SY, and TE types, respectively. Guaiene, -cis-.beta.-Elemene diastereomer, naringin, 3691-11-0, o-cymol, cadalin, cadinene, alpha-terpineol, germacrene, l-limonen are found to be SE-specific compounds (Fig. 4A); FER, marmesin, decyl acetate to be SY-specific compounds (Fig. 4B); and arachidonic acid to be TE-specific compounds (Fig. 4C).

Fig. 4.

SC type-specific compounds. Fourteen compounds showed significant relevance with the SC type (according to Fisher's exact test). The bar graph represents the relative frequency of each compound, indicating which SC type compound shows relevance. The relative frequency is $\frac{\#herbsincludingthecompoundforeachSCtype}{\#herbsbelongingtoeachSCtype}$. (A) SE-specific compounds. (B) SY-specific compounds. (C) TE-specific compound. SE, So-Eum type; SY, So-Yang type; TE, Tae-Eum type.

3.6. Sasang type prediction of unidentified medicinal herbs based on compound level information, and their investigations by functional assay

The trained ML classifier was applied to predict the Sasang type of medicinal herbs whose Sasang types are unidentified as a further application. Thirty medicinal herbs were selected according to the amount of the usage [26].

The trained ML classifier successfully predicted the Sasang type of the thirty herbs with the probability of which Sasang type each herb belongs. The prediction procedure was repeated 10 times for each herb, and the mean probability was presented to provide stability (Fig. 5 and Table 4). As a result, seven kinds of herbs (from number 1 to number 7), including Curcuma longa Radix and Eriobotrya japonica Lindley corresponded to SE type (Fig. 5A), and thirteen herbs (from number 8 to number 20), such as Houttuynia cordata, Cistanche deserticola, and Lindera strychnifolia Vill. are related to the SY type (Fig. 5B). Ten herbs (from number 21 to number 30) containing Leonurus japonicus Houtt. and Benincasa hispida Cogniaux were categorized as the TE type (Fig. 5C).

Fig. 5.

Predicted result of top 30 unidentified medicinal herbs using trained ERT model. The Sasang type of the 30 medicinal herbs was predicted using the trained ERT model. The bar represents the mean probability (ten trials repeated) of which constitution each medicinal herb belongs. The predicted Sasang type of each herb is visualized using background colors, and the herbs are clustered based on their predicted Sasang type. In each herbal group, the herbs were sorted according to their probability. The numbers indicate the name of herbs in Table 4 (A) Compounds 1–7 predicted as SE type. (B) Compounds 8–15 predicted as SY type. (C) Compounds 16–30 predicted as TE type. SE, So-Eum type; SY, So-Yang type; TE, Tae-Eum type.

Table 4.

Prediction probability for the top 30 unidentified medicinal herbs to belong to which constitution.

	Chinese name	English name	SE	SY	TE
1	鬱金	Curcuma longa Radix	0.766	0.125	0.110
2	枇杷葉	Eriobotrya japonica Lindley	0.519	0.231	0.250
3	防己	Sinomenium acutum Rehder et Wilson	0.459	0.298	0.243
4	丹蔘	Salvia miltiorrhiza Bunge	0.427	0.202	0.371
5	艾葉	Artemisia princeps Pampanini	0.426	0.239	0.335
6	丁香	Syzygium aromaticum Merrill et Perry	0.414	0.296	0.290
7	龍膽草	Gentiana scabra Bunge	0.354	0.304	0.342
8	魚腥草	Houttuynia cordata	0.129	0.665	0.206
9	肉蓯蓉	Cistanche deserticola	0.137	0.610	0.253
10	烏藥	Lindera strychnifolia Vill.	0.243	0.580	0.177
11	淫羊藿	Epimedium koreanum Nakai	0.282	0.569	0.149
12	肉荳蔲	Myristica fragrans	0.170	0.563	0.266
13	北沙參	Glehnia littoralis Fr. Schmidt ex Miquel	0.258	0.540	0.202
14	細⾟	Asarum sieboldii Miq.	0.268	0.532	0.199
15	辛夷	Magnolia denudata Desrousseaux	0.268	0.532	0.199
16	檳榔子	Areca catechu Linné	0.205	0.530	0.265
17	百合	Lilium lancifolium Thunberg	0.201	0.475	0.324
18	小茴香	Foeniculum vulgare	0.275	0.451	0.274
19	蛇床子	Torilis japonica	0.330	0.388	0.282
20	紫蘇葉	Perillae Folium	0.336	0.379	0.285
21	益母草	Leonurusjaponicus Houtt	0.101	0.220	0.679
22	冬瓜子	Benincasa hispida Cogniaux	0.153	0.210	0.637
23	決明子	Cassia tora Linné	0.301	0.143	0.556
24	杜仲	Eucommia ulmoides Oliver	0.294	0.162	0.543
25	白蒺藜	Tribulus terrestris	0.183	0.386	0.430
26	釣鉤藤	Uncariae Ramulus cum Uncus	0.352	0.239	0.409
27	麥芽	Hordeum vulgare Linné	0.257	0.337	0.405
28	川貝母	Fritillariae Cirrhosae Bulbus	0.293	0.310	0.397
29	牛膝	Twotoothed Achyranthes	0.366	0.266	0.368
30	威靈仙	Chinese Clematis	0.343	0.291	0.366

SE, So-Eum type; SY, So-Yang type; TE, Tae-Eum type.

Additionally, we wanted to know whether the predicted herbs with the highest rank in each type have a biological function in each organ corresponding to the Sasang type. According to SCM theory, each Sasang type is highly related to specific organs (i.e. SE, kidney; SY, stomach; TE, liver) [1, 2, 3]; Thus, we tested the anti-cancer effects of Curcuma longa Radix (SE) in the NRK-52E kidney cancer cell line and Houttuynia cordata (SY) in the AGS stomach cancer cell line as well as anti-oxidant effects of Leonurus japonicus Houtt (TE) in the HepG2 hepatocyte (Fig. 6A–C). The water extract of Curcuma longa Radix (the first ranking herb in SE) and Houttuynia cordata (the first ranking herb in SY) significantly inhibited the proliferation of cancer derived from kidney and stomach, respectively. Leonurus japonicus Houtt (the first ranking herb in TE) markedly inhibited the oxidative damage induced by AA + iron.

Fig. 6.

Cell viability in three types of cells. The cells were plated at a density of 1 × 10⁵ cells per well in 48-well culture plates and incubated in an FBS-free medium for 12 h. (A and B) AGS and NRK-52E cells were incubated with drugs for 24 h. (C) HepG2 cells were incubated with drugs for 1 h, followed by a treatment with AA (10 μM) for 12 h and then iron (5 μM) for 6 h. All data represent means ± SD of 4 independent experiments (*ρ < 0.05 and **ρ < 0.01 vs. control group; ##ρ < 0.01 vs. AA + iron-treated group). AA, arachidonic acid; con, vehicle-treated control.

4. Discussion

SCM is a unique form of personalized medicine in traditional Korean medicine, in which the patients are classified into one of four Sasang constitution types: SE, SY, TE, or TY. In SCM, herbal medicines (composed of various medicinal herbs) are prescribed based on the patients’ Sasang type, in addition to their symptoms, and the applied medicinal herbs were divided into four groups corresponding to the four Sasang types; the herbs themselves are inextricably linked to the concept of Sasang type. Furthermore, a study of the herbs could provide more objective insight than investigating the Sasang type-diagnosed patients by SCM experts to understand the intrinsic principle of SCM, in that the agreement rate for diagnosed Sasang type among three qualified SCM experts is between 52.5% and 68.4% [2,27]. Using a novel drug-centric approach, this study examined the compound patterns that enable Sasang type classification and the chemical characteristics of each herbal group that contributes to the medicinal effect of SCM via ML techniques.

A previous study examined the major botanical compounds, such as phenol, alkaloid, and terpenoid, contained in each herb from a biomedical point of view. Lim et al. reported that phenolics were dominant in the TY-type herbs, iridoids and triterpenes were in the SY-type herbs, saponins were in the TE-type herbs, and monoterpene and sesquiterpenes were in SE type herbs [3]. On the other hand, because this research was still limited as a review on the characteristics of each herb, more systematic and data-driven group-level characteristics of each constitutional herbal group consisting of corresponding medicinal herbs have not been identified.

This study aimed to identify the principle/criteria for classifying medicinal herbs into the Sasang types using a data-driven approach. The current study examined whether the principle of distribution is explainable at the multi-compound-level by multivariate analyses, such as ML [28]. The Sasang type classification could be explained by multi-compound configuration, being confirmed by statistical significance in the classification performance of the ML classifier on the SE and SY type herbs. Various patterns made by multiple compounds would help classify different type groups at the multivariate level. Although the classification performance for the TE type was statistically insignificant (p value = 0.25), the herbs were only investigated at the compound level in this study. Other factors, in addition to the compound factor, would help account for the Sasang type classification principles. It is important how much the Sasang type classification can be explained by other characteristics, such as traditional theory-based taste and action information, in addition to the compound information in further analyses.

This study investigated which compounds play important roles in type discrimination to interpret the multi-compound patterns derived from ML analysis. The 60 compounds deducted by the ML pattern analysis were examined using hierarchical clustering, and 14 Sasang type-specific compounds were found by statistical analysis. On the other hand, because this result reflects only univariate-level investigation, much more remains to be investigated for multivariate-level interpretation.

The characteristics of these compounds were analyzed using chemical information. The secondary metabolites in the plants were biosynthesized by the action of various enzymes and generally constituted the active ingredients of medicinal plants [29,30]. It was also reported that the building blocks that make up the drug target of the bodies and a line of enzymes normally involved in the biosynthesis of secondary metabolites [31,32]. Although it is difficult to assume that the compounds listed in Fig. 4 have the representative activity of each herbal medicine and there is little scientific evidence to directly connect the biosynthetic pathway of some of the representative compounds to SCM, the potential and novelty of these attempts themselves cannot be ignored. In the TE type, fatty acid-based materials are normally synthesized through the polyketide biosynthetic pathway. On the other hand, the shikimate, mevalonate, and polyketide pathways were involved in the SY constitution. In the case of SE, one of the most abundant components, terpenoids, was related to the mevalonic (isoprenoid) pathway. Although this analysis alone is not enough to explain the characteristics of herbal medicines in the SCM, at least there are some differences for each Sasasng type in terms of the biosynthetic pathway.

This study has some limitations. Owing to the small sample size, medicinal herbs of the TY type were excluded from the analyses. The TY type frequently has a small sample size in other studies regarding Sasang type diagnosis because of the rarity of the TY type in the population distribution in SCM. The majority of studies excluded the TY type from their analyses. The same situation occurred in this study although we adopted a drug-centric approach. In addition, in this study, the herb-compound matrix was constructed using one-hot encoding, not considering the chemical similarity between compounds. It would be better to consider vector embedding in future research. In addition, in this study, we provided information about the predictive type of the medicinal herbs using a trained ML classifier. However, readers should note that this prediction is only a result of nonlinear pattern computation based on compound information, not based on the whole information comprising the Sasang type, implying that this predictive information was not recommended for direct clinical application. If there exists a discrepancy between the prediction result and the clinical application, this indicates that additional information is needed to the dataset. Refining the model remains to conduct additional research in the future.

5. Conclusions

Here, we comprehensively investigated the compounds composing the Sasang type-specific personalized herbal medicines. We confirmed that most of the 60 compounds showing high feature importance determined by the ERT classifier have selectivity for particular Sasang type. We found that most compounds showed selectivity for particular Sasang type. We also investigated taxonomic and biosynthetic characteristics of the 59 Sasang-prominent compounds (8, 9, and 42 compounds for TE, SY, and SE, respectively). Furthermore, we identified 14 Sasang type-specific compounds showing statistically significance with the Sasang type at a univariate level. Lastly, using a trained ERT classifier, we predicted the Sasang type of commonly used but unidentified medicinal herbs, and indirectly confirmed the prediction result with a simple in vitro experiment examining the biological function of herbs.

Declarations

Author contribution statement

Ji-Hwan Kim: Conceived and designed the experiments; Contributed reagents, materials, analysis tools or data. Chang-Eop Kim: Conceived and designed the experiments; Contributed reagents, materials, analysis tools or data; Wrote the paper; Analyzed and interpreted the data.Sa-Yoon Park, Young Woo Kim: Performed the experiments; Analyzed and interpreted the data; Contributed reagents, materials, analysis tools or data; Wrote the paper. Yu Rim Song, Young Pyo Jang, Young Pyo Jang: Performed the experiments.

Funding statement

Prof. Ji-Hwan Kim was supported by National Research Foundation of Korea [NRF-2017R1C1B5017048 and NRF-2022R1I1A2066653]. Prof. Chang-Eop Kim was supported by National Research Foundation of Korea [2020M3A9E4103843]; Ministry of Health and Welfare [HF20C0087].

Data availability statement

Data will be made available on request.

Declaration of interest's statement

The authors declare no conflict of interest.

Additional information

Supplementary content related to this article has been published online at [URL].

Appendix A. Supplementary data

The following is the supplementary data related to this article: