Pattern recognition-based analysis of the material basis of five flavors of Chinese herbal medicines in Lamiaceae

Abstract

OBJECTIVE:

To study the correlation between five flavors (Wuwei) and the chemical substances of Chinese herbal medicines in Lamiaceae and to establish five flavors identification models.

METHODS:

A total of 245 herbs belonging to the Lamiaceae family were selected from the Pharmacopoeia of the People's Republic of China 2020 and Chinese Materia Medica. A database of the chemical substances of these herbs was constructed, with the chemical substances obtained from the professional literature and databases. A three-level classification system of the material components was established on the basis of the molecular structure and biosynthetic pathway of these substances. Apriori association rule analysis and feature selection were employed to obtain the material basis of the five flavors. A multiple logistic regression analysis method was employed to establish identification models for the five flavors.

RESULTS:

The association rule analysis revealed 34 high-value groups and 30 specific groups for the main flavors, and 39 high-value groups and 36 specific groups for the combined flavors. Sixteen groups of chemical components were the decisive groups for the main flavors, and 13 groups were the decisive groups for the combined flavors. Multiple logistic regression analysis was used to successfully establish identification models with an overall accuracy of 88.8% for the main flavors and 87% for the combined flavors.

CONCLUSIONS:

Five flavors are often characterized by the interaction of multiple classes of substances, and a single class of substances cannot be used to characterize flavors. The organic combination of multiple classes of substances is the material basis of the five flavors, both the main and combined flavors. Significant differences exist in the material basis of the main and combined flavors, suggesting that the “natural flavor” and “functional flavor” may have different material bases.

1. INTRODUCTION

The five flavors (Wuwei) theory is a unique medical theory of Traditional Chinese Medicine (TCM) that has guided the clinical practice of TCM for thousands of years and includes sour, bitter, sweet, pungent, and salty flavors. The concept of five flavors initially originated from people's perception and classification of the real taste and odor of Chinese herbs. It then gradually became associated with their effects, forming one of the unique properties of Chinese herbal medicines (CHM). Flavors are used to explain and summarize the effects of medicinal herbs and predict their actions. Huang Di Nei Jing¹ and Shen Nong Ben Cao Jing² were the first to incorporate flavors into the theory of medicinal properties and establish a relationship between flavors and efficacy. Su Wen Cang Qi Fa Shi Lun¹ in Huang Di Nei Jing summarizes the relationships between the five flavors and their effects. Shen Nong Ben Cao Jing² introduces the concept into herbal literature, making it one of the fundamental theories in TCM. Medical experts continuously supplemented and improved the theory, eventually forming a mature theory of the five flavors. According to this theory, a pungent flavor can disperse and promote; a sour flavor can shrink and astringe; a sweet flavor can nourish, relieve, and moderate; a bitter flavor can drain, dry, and strengthen; and a salty flavor can descend and soften.^3,4

With the continuous developments in science and technology, new hypotheses have been proposed regarding the properties of CHM. Zhao et al ⁵ proposed the thermodynamic theory of medicinal properties. Liu et al ^6,7put forward the five elements model of "drug-favor-substance-effect-function," which further developed into the research on the relationship between "substance-pharmacokinetics-effects". Kuang et al ⁸ proposed a new research model based on the splitability and combinativeness of the properties and flavors of Chinese medicine, that is, "one drug with X flavors and Y properties, where Y ≤ X".^9,10 Wang et al ¹¹ proposed the hypothesis of the three elements of "property-effect-substance" in Chinese herbs, which further developed into the study of the "property-structure" relationship in CHM.^12-14 Other studies have employed methods and ideas such as the traditional human taste panel method,^15-17 bionic technology,¹⁸ near-infrared spectroscopy technology,¹⁹ metabonomics,^20-22 network pharmacology,²³ bioin-formatics,^24,25 ingested chemicals,²⁶ gut microbiota,^27,28 ‘Xiang’s thinking,^29,30 and pharmaphylogeny.^31,32 These studies have provided methods that can be used for the modern interpretation of CHM property theories.^33-35 Although multiple methods are available for studying properties of CHM, the relationship between properties and chemical components has always been a research hotspot.^36,37 Exploring the material properties of CHM should return to Traditional Chinese Medicine theories to obtain the key points.³⁸ Therefore, the studying these five flavors should originate from systematic theory and return to systematic theory. ⁴ Additionally, reductionist methods should be used as the experimental means and specific entry points.

Phylogenetic studies on medicinal plants have confirmed a strong correlation between phylogenetic relationships, chemical components, and medicinal effects. Herbs belonging to the Lamiaceae family have a long history of use in traditional medicine. Studies have shown that the chemical components of Lamiaceae herbs are highly correlated with their cold-hot properties.^39-41 However, research on the relationship between chemical substances and these five flavors remains insufficient.

This study used Lamiaceae herbs as an example, classified their chemical substances into three-level categories, used data mining methods such as frequency analysis, association rule analysis, logistic regression, and feature selection to obtain the material basis of the five flavors of Lamiaceae herbs and established five flavor-recognition models.

The results showed significant differences in the material bases of pungent, bitter, and sweet flavors in the third-level classification. There were obvious differences in the material basis between the main and combined flavors. Association rule analysis identified 34 high-value groups and 30 specific groups for the main flavors and 39 high-value groups and 36 specific groups for the combined flavors. Feature selection revealed 16 decisive groups for the main flavors and 13 decisive groups for the combined flavors. These five flavors are often characterized by the interaction of multiple classes of substances, and a single class of substances cannot be used to characterize the five flavors. The organic combination of multiple classes of substances is the material basis of the five flavors, both the main and combined flavors.

Additionally, there are significant differences in the material basis of the main and combined flavors, suggesting that the natural flavor (the real taste and odor of CHM) and “functional flavor” (the effects of CHM) of the five flavors may have different material bases. Bicyclic monoterpenes, acyclic monoterpenes, acyclic sesquiterpenes, O-glycosides, ketones, and organic acids constitute the material basis for the pungent flavor (main); phenylpropylenes, and flavanones constitute the material basis for the sweet flavor (main); and bicyclic sesquiterpenes, bicyclic diterpenes, tricyclic diterpenes, and tetracyclic diterpenes constitute the material basis for the bitter flavor (main). Monocyclic monoterpenes, monocyclic sesquiterpenes, alcohols, and phenylethyl derivatives are shared by the pungent, sweet, and bitter flavors (main). Caffeetannins, bicyclic sesquiterpenes, and protocatechuic derivatives constitute the material basis for the pungent flavor (combined); tricyclic diterpenes, benzofurans, and esters constitute the material basis for the sweet flavor (combined); and flavones, monocyclic monoterpenes, other benzene derivatives, O-glycosides, and ketones constitute the material basis for the bitter flavor (combined). Bicyclic monoterpenes and alcohols are shared by the pungent, sweet, and bitter flavors (combined). Multiple logistic regression analysis was used to successfully establish five flavor identification models with an overall accuracy of 88.8% for the main flavors and 87% for the combined flavors.

2. MATERIALS AND METHODS

2.1. Establishment of the database of the chemical substances of herbs in Lamiaceae

Based on the Pharmacopoeia of the People's Republic of China 2020⁴² and Chinese Materia Medica,⁴³ 245 Lamiaceae medicinal plants were selected. The source of the chemical substances came from Pharmacopoeia of People's Republic of China 2020,⁴² Chinese Materia Medica,⁴³ Zhongyao Dacidian,⁴⁴ Quanguo Zhongcaoyao Huibian,⁴⁵ Modern Chinese Materia Medica,⁴⁶ Zhongyao Huoxing Chengfen Fenxi Shouce,⁴⁷ Changyong Zhongyao Jiqi Huoxing Chengfen Shouce,⁴⁸ Modern Study of Traditional Chines Medicine,⁴⁹ and Zhongyaocai Huoxing Chengfen Huaxue Jiegou Tuji,⁵⁰ as well as literature databases such as Chinese Biomedical Literature Database (CBM), China National Knowledge Infrastructure Database (CNKI), Wanfang Database, PubMed, and Web of Science.

The inclusion criteria for the literature selection were studies on the chemical components of the herbs with clear information on the original plants, medicinal parts, experimental methods, and research methods. The exclusion criteria included literature emphasizing extraction techniques with few descriptions of the chemical constituents; literature describing identification, content determination, quality control methods; literature discussing differences in chemical components of different medicinal parts; non-medicinal parts’ constituent research; studies on the relationship between harvesting time, processing, decoction, preparation, and component contents; and studies on the determination of chemical components of compound formulas.⁵¹

In addition, some chemical components were obtained from related databases, such as the Shanghai Institute of Organic Chemistry of CAS Chemistry Database, National Population Health Data Center (Pharmacy), Traditional Chinese Medicine Systems Pharmacology Database and Analysis Platform, Bioinformatics Analysis Tool for Molecular Mechanism of Traditional Chinese Medicine (BATMAN-TCM), and High-throughput Experiment- and Reference-guided Database of Traditional Chinese Medicine.

2.2. Treatment of medicinal properties

The five flavors of CHM generally refer to sour, bitter, sweet, pungent, and salty. In herbal literature, descriptions such as slightly bitter, slightly pungent, and slightly sweet distinguish the degree of medicinal flavors. These descriptions fall under the category of five flavors. Therefore, for the convenience of collation and statistics, the descriptions of slightly bitter, slightly pungent, and slightly sweet were uniformly classified as bitter, pungent, and sweet. In addition, some herbs often have multiple medicinal flavors, such as Jingjie (Herba Schizonepetae Tenuifoliae), Niuzhi (Herba Origani Vulgaris), and Xiazhicao (Herba Lagopsis Supinae), which have a "pungent and slightly bitter" flavor. In contrast Zhuchun (Herba Salviae Coccineae) has a "pungent, slightly bitter, and astringent" flavor. The terms "main flavor" and "combined flavor" are used to represent them, where "main flavor" indicates the first flavor of the drug, and "combined flavor" indicates the second and third flavors of the herb.^52,53

2.3. Confirmation and collection of chemical information

To ensure the uniqueness of the chemical substances, 12 characteristic parameters, including the Chinese and English chemical names, molecular formula, molecular weight, International Union of Pure and Applied Chemistry (IUPAC) name, simplified molecular input line entry system (SMILES), International Chemical Identifier (InChI), International Chemical Identifier Key (InChIKey),^54,55 structural formula, first-level classification, second-level classification, and third-level classification, were selected for each chemical component. Duplicate chemical components with different Chinese or English names were removed.

Chemical substance confirmation and information collection were based on recognized professional chemical databases, such as the Shanghai Institute of Organic Chemistry of CAS Chemistry Database, National Population Health Data Center (Pharmacy), PubChem and Reaxys. Other commercial chemical databases were used as supplements.

2.4. Establishment of a classification system for chemical substances

Based on Chinese Medicinal Chemistry and Natural Medicinal Chemistry, chemical components are divided into 12 major categories, including quinones, phenylpropanoids, flavonoids, terpenoids, steroids, alkaloids, tannins, cerebrosides, saccharides and glycosides, fatty acids, amino acids and proteins, and other compounds, which serve as the first-level classification. Furthermore, the second- and third-level classifications were carried out according to differences in the molecular structure and biosynthetic pathways of the chemical substances. The entire classification system can be found in our previous work.⁴¹

2.5. Data analysis tools and methods

Excel was used for database construction and frequency analysis, and IBM SPSS Statistics 26 (IBM Corp., Armonk‌, NY, USA) was used for general statistical analysis. IBM SPSS Modeler 18.0 (IBM Corp., Armonk, NY, USA) was used for Apriori association rule analysis and feature selection to study the chemical basis of the five flavors. Multiple logistic regression analysis was used to establish recognition models for the five flavors and evaluate the prediction accuracy.

3. RESULTS

A total of 193 herbs with known chemical substances were collected from the 245 herbs of the Lamiaceae family, of which four had unknown properties and were excluded. Additionally, 17 herbs with a first-level classification of < 2 were excluded, resulting in 172 herbs and 10 726 chemical substances. The botanical plant names of the herbs were checked at http://www.worldfloraonline.org on December 12, 2023. Because of the low proportion, the light and astringent flavors (Table 1) were excluded. Therefore, the light flavor was removed from the main flavor analysis and the light and astringent flavors were removed from the combined flavor analysis.

Table 1.

Five flavors and material components of herbs in Lamiaceae

Five flavors	CHM (main flavors)			Components (main flavors)			CHM (combined flavors)			Components (combined flavors)
	Count	Proportion (%)		Count	Proportion (%)		Count	Proportion (%)		Count	Proportion (%)
Light	2	1.16		29	0.27		2	2.13		217	3.47
Sweet	21	12.21		1020	9.51		13	13.83		749	11.98
Bitter	57	33.14		3298	30.75		47	50.00		3110	49.73
Pungent	92	53.49		6379	59.47		31	32.98		2127	34.01
Astringent	/	/		/	/		1	1.06		51	0.82
Total	172	100.00		10726	100.00		94	100.00		6254	100.00

Notes: CHM: Chinese herbal medicines.

3.1. Frequency analysis

3.1.1. Frequency analysis of main flavors

In the first-level classification of the main flavors, terpenoids accounted for the highest proportion (49.78%), followed by other compounds (17.30%) and flavonoids (11.70%). Differences were observed in the proportions of substances with pungent, sweet, and bitter flavors, with the most significant differences observed in terpenoids, other compounds, and quinones.

In the second-level classification, monoterpenoids accounted for the highest proportion (18.98%), followed by sesquiterpenoids (14.20%) and diterpenoids (10.62%). Differences were observed in the proportions of substances with pungent, sweet, and bitter flavors, with the most significant differences observed in monoter-penoids, sesquiterpenoids, and diterpenoids.

In the third-level classification, monocyclic mono-terpenes accounted for the highest proportion (8.78%), followed by flavones (6.82%) and bicyclic sesquiterpenes (6.10%). Differences were observed in the proportions of substances with pungent, sweet, and bitter flavors, with the most significant differences observed in monocyclic monoterpenes, bicyclic sesquiterpenes, pentacyclic triterpenes, tetracyclic diterpenes, and phenylpropionic acids.

Chi-square test results showed that the material composition ratios of pungent, bitter, and sweet flavors were significantly different in the first-, second-, and third-level classifications (P < 0.05), and the differences varied as the levels of substance classification changed. In terms of the percentage differences in the material composition ratios, the first level was smaller than the second level, and the second level was smaller than the third level, with average coefficients of variation of 47.56, 81.42, and 92.09%, respectively (Figure 1).

Figure 1. The material composition ratios for three levels of classification.

Only the top 15 items are shown in the figure. A: material composition ratios of pungent, bitter, and sweet flavors in the first-level classifications of main flavors; B: material composition ratios of pungent, bitter, and sweet flavors in the second-level classifications of main flavors; C: material composition ratios of pungent, bitter, and sweet flavors in the third-level classifications of main flavors; D: material composition ratios of pungent, bitter, and sweet flavors in the first-level classifications of combined flavors; E: material composition ratios of pungent, bitter, and sweet flavors in the second-level classifications of combined flavors; F: material composition ratios of pungent, bitter, and sweet flavors in the third-level classifications of combined flavors.

3.1.2. Frequency analysis of combined flavors

In the first-level classification of combined flavors, terpenoids accounted for the highest proportion (49.97%), followed by other compounds (18.99%) and flavonoids (10.71%). Differences were observed in the proportions of substances with pungent, sweet, and bitter flavors, with the most significant differences observed in other compounds, flavonoids, and phenylpropanoids.

In the second-level classification, monoterpenoids accounted for the highest proportion (17.06%), followed by sesquiterpenoids (14.07%) and diterpenoids (12.15%). Differences were observed in the proportions of substances with pungent, sweet, and bitter flavors, with the most significant differences observed in mono-terpenoids, diterpenoids, and O-containing compounds.

In the third-level classification, monocyclic mono-terpenes accounted for the highest proportion (7.53%), followed by pentacyclic triterpenes (6.31%) and bicyclic sesquiterpenes (6.18%). Differences were observed in the proportions of substances with pungent, sweet, and bitter flavors, with the most significant differences observed in monocyclic monoterpenes, pentacyclic triterpenes, flavones, bicyclic diterpenes, tetracyclic diterpenes, and phenylpropionic acids.

Chi-square test results showed that the material composition ratios of pungent, bitter, and sweet flavors were significantly different in the first-, second-, and third-level classifications (P < 0.05), and the differences varied as the levels of substance classification changed. In terms of the percentage differences in the material composition ratios, the first level was smaller than the second level, and the second level was smaller than the third level, with average coefficients of variation of 41.34, 86.22, and 95.87%, respectively (Figure 1).

3.2. Association rule analysis

According to a previous study, it was not possible to effectively distinguish the flavors at the first- and second-classifications.⁴¹ Therefore, this study only further analyzed the association rules for the third-level classification.

3.2.1. Selection of optimal parameters for association rule analysis

When using Modeler Apriori for association rule analysis, the maximum number of antecedents, minimum rule confidence, and minimum antecedent support are three important parameters. The selection of these parameters in previous studies was based on experience or limited exhaustiveness, without systematic analysis. ⁴¹ It is necessary to conduct in-depth research to enhance the quality of data analysis.

3.2.1.1. Exhaustive method

The exhaustive method enumerates the three parameters of the maximum number of antecedents, minimum rule confidence, and minimum antecedent support to obtain the best parameters. Considering that the properties of Chinese herbal medicines may be manifested through the interaction of multiple substances, the maximum number of antecedents was set from two to five with a step size of one. To ensure the effectiveness of the generated association rules, the minimum rule confidence was set from 60% to 80%, with a step size of 5%. The boundaries of the minimum antecedent support were defined based on the effective rules generated by the Apriori algorithm with a step size of one. Within the core groups range of 10-20, the parameters with the highest prediction accuracy were selected as optimalTable 2.

Table 2.

Parameter results of the exhaustive method

Consequent		Optimal parameter			No. of core groups	Logistic regression accuracy (%)	Single interaction accuracy (%)
		Maximum number of antecedents	Minimum rule confidence (%)	Minimum antecedent support (%)
Main flavors	Pungent	5	65	25	12	72	82
	Sweet	3	70	2	18	28	81
	Bitter	3	70	3	18	45	93
Combined flavors	Pungent	5	60	11	14	54	83
	Sweet	2	65	3	15	38	84
	Bitter	2	60	18	17	66	100

3.2.1.2. New parameter method

The exhaustive method selects parameters from a mathematical perspective, but does not consider the meaning of each parameter in the Apriori algorithm itself. Rule confidence is a conditional probability, with higher values indicating more reliable predictions of ante-cedents to consequences. Rule support is a measure of rule generalization, with higher values indicating greater practical value. The number of antecedents reflects the interactions between the substances. To obtain the core groups for a certain medicinal property, under the condition of equal confidence and support, fewer antecedents can be considered more important for predicting this property. Therefore, a new parameter, K, was constructed to measure the importance of substances in the core groups, and the optimal parameters were selected based on the highest prediction accuracy (Table 3).

Table 3.

Optimal conditions and values of the new parameter method

Consequent		The best value of new parameter	Constitute elements of new parameter method			No. of core groups	Logistic regression accuracy (%)	Single interaction accuracy (%)
			Coefficient of antecedents	Maximum support	Maximum confidence
Main flavors	Pungent	K7			√	15	76	90
	Sweet	K7			√	15	33	85
	Bitter	K1	√			15	43	93
Combined flavors	Pungent	K2	√	√		15	58	90
	Sweet	K1	√			15	46	76
	Bitter	K7			√	15	68	100

Comparison of the two parameter selection methods —overall, the new parameter method is superior to the exhaustive method (Figure 2).

Figure 2. Comparison of the results of the exhaustive and the new parameter methods.

3.2.2. High-value groups for main flavors

For the main flavor analysis, with pungent flavor as the consequent, a maximum number of antecedents of five, minimum rule confidence of 70%, minimum antecedent support of 14%, 5,147 valid rules, and 31 core groups were obtained. Using the optimal new parameters for sorting, the top 15 core groups were named "high-value groups." Substances, such as monocyclic monoterpenes, bicyclic monoterpenes, and monocyclic sesquiterpenes, had higher support.

With sweet flavor as the consequent, a maximum number of antecedents of five, minimum rule confidence of 70%, and minimum antecedent support of 1%, 15 702 valid rules, and 44 core groups were obtained. The top 15 core groups were named "high-value groups." Substances, such as phenylethyl derivatives, saturated fatty acids, and phenylpropionic acids, had higher support.

With bitter flavor as the consequent, a maximum number of antecedents of five, minimum rule confidence of 70%, and minimum antecedent support of 2%, 5865 valid rules, and 62 core groups were obtained. The top 15 core groups were named "high-value groups." Substances, such as o-phenanthrenequinones, other coumarins, and unsaturated fatty acids, had higher support.

3.2.3. High-value groups for combined flavors

For the combined flavor analysis, with pungent flavor as the consequent, a maximum number of antecedents of five, minimum rule confidence of 70%, and minimum antecedent support of 5%, 5527 valid rules and 36 core groups were obtained. The top 15 core groups were named "high-value groups." Substances, such as caffeetannins, protocatechuic derivatives, and monocyclic sesquiterpenes, had higher support.

With sweet flavor as the consequent, a maximum number of antecedents of five, minimum rule confidence of 70%, and minimum antecedent support of 2%, 1928 valid rules and 32 core groups were obtained. The top 15 core groups were named "high-value groups." Substances, such as tricyclic diterpenes, amino acids, and cyclo-lignans, had higher support.

With bitter flavor as the consequent, a maximum number of antecedents of five, minimum rule confidence of 70%, and minimum antecedent support of 9%, 1273 valid rules and 35 core groups were obtained. The top 15 core groups were named "high-value groups." Substances, such as alkanes, alcohols, and monocyclic monoterpenes, had higher support.

3.2.4. Specific groups

During the association rule analysis, some special rules had low support but extremely high confidence (100%). This indicates that these substances play unique roles in the recognition of flavors. Therefore, rules with low support but extremely high confidence (100%) and the maximum number of antecedents of 1 were defined as "specific rules," and the corresponding antecedents were defined as "specific groups."

For the main flavors, 170 herbs were included, and the minimum antecedent support was set to 1/170, which was 0.588%. With a rule confidence of 100%, and a maximum number of antecedents of 1, 18 specific groups were obtained with pungent flavor as the consequent, three specific groups with sweet flavor as the consequent, and nine specific groups with bitter flavor as the consequent.

For the combined flavors, 91 herbs were included, and the minimum antecedent support was set to 1/91, which was 1.09%. With a rule confidence of 100%, and a maximum number of antecedents of 1, 12 specific groups were obtained with pungent flavor as the consequent, three specific groups with sweet flavor as the consequent, and 21 specific groups with bitter flavor as the consequent.

3.2.5. Interaction among high-value groups

There were 34 high-value groups, 30 specific groups for the main flavors, 39 high-value groups, and 36 specific groups for the combined flavors. Monocyclic monoterpenes, monocyclic sesquiterpenes, alcohols, flavones, saturated fatty acids, phenylethyl derivatives, alkanes, pentacyclic triterpenes, flavonols, and unsaturated fatty acids were shared high-value groups for pungent, sweet, and bitter flavors in the main flavors. Other alkaloids, alkenes, bicyclic monoterpenes, alcohols, monocyclic sesquiterpenes, and benzyl derivatives were shared high-value groups for pungent, sweet, and bitter flavors in the combined flavors (Figure 3).

Figure 3. High-value groups and specific groups.

A: main flavors; B: combined flavors.

3.2.6. Differences in the material basis of main flavors and combined flavors

Further analysis of the high-value groups revealed significant substance differences between the main and combined flavors (Figure 4). For example, monocyclic monoterpenes, bicyclic monoterpenes, acyclic monoterpenes, and flavones showed significant differences in the pungent flavor between the main and combined flavors. It is speculated that the "natural flavor" and "functional flavor" attributes of the five flavors have different material bases. This is in line with the traditional understanding of the medicinal properties of CHM, which generally starts with tasting the taste and smelling the aroma and then continuously enriches its "flavor" as a therapeutic effect during clinical use.

Figure 4. Differences in the material basis of main and combined flavors.

3.3. Logistic regression modeling

3.3.1. Modeling high-value groups

We selected a multiple logistic regression model for modeling. The model inputs were high-value groups and the outputs were the main/combined flavors (pungent, sweet, and bitter). The multinomial procedure method used as the enter method. The basic category was sweet flavor. The quality of the models was evaluated using four parameters: likelihood ratio tests for equation significance, Wald tests for variable significance, Nagelkerke’s R2 and the accuracy of confusion matrices for the goodness of fit (Table 4).

Table 4.

Logistic regression modeling results of high-value groups

No.	Consequent	Independent variables	No. of variables	Sig. (P-values)	Nagelkerke R2 values	Percent correct (%)
						Pungent	Sweet	Bitter	Total
1	Main flavors	High-value groups	34	0.000	0.617	84.8	61.9	70.2	77.1
2	Main flavors	High-value groups+specific groups	64	0.000	0.785	87	81	77.2	82.9
3	Main flavors	High-value groups (single interaction)	76	0.628	0.991	100	95.2	96.5	98.2
4	Main flavors	High-value groups (all 2-way interactions)	159	0.650	0.990	100	95.2	96.5	98.2
5	Combined flavors	High-value groups	39	0.000	0.929	90.3	100	93.8	93.5
6	Combined flavors	High-value groups+specific groups	75	0.001	1.000	100	100	100	100
7	Combined flavors	High-value groups (single interaction)	60	0.428	1.000	100	100	100	100
8	Combined flavors	High-value groups (all 2-way interactions)	88	0.428	1.000	100	100	100	100

For the main flavor, when high-value groups were used as independent variables, the overall accuracy was 77.1%, indicating good model performance (Model 1). When both high-value and specific groups were used as independent variables, the overall accuracy increased to 82.9% (Model 2). However, the number of variables increases from 34 to 64, making the model less practical.

For the combined flavor, when high-value groups were used as independent variables, the overall accuracy was 93.5%, indicating good model performance (Model 5). When both high-value and specific groups were used as independent variables, the overall accuracy reached 100% (Model 6). However, the number of variables increases from 39 to 75, making the model less practical.

Further examination of the interaction between variables revealed that considering only the interaction among high-value groups, the prediction accuracy for the main and combined flavors was 98.2% (Models 3-4) and 100% (Models 7-8), respectively. This indicates that recognition of five flavor properties can be well achieved by examining high-value groups alone. This also suggests that different combinations of substances determine the flavor.

3.3.2. Decisive groups

Feature selection was performed for the 34 high-value groups for the main flavors to extract "decisive groups." Using the likelihood ratio chi-square test, variables with P-values of 1-probability > 0.8 were selected, resulting in 16 important variables (Table 5). According to the interaction relationships of high-value groups for the main flavors (Figure 3), bicyclic monoterpenes, acyclic monoterpenes, acyclic sesquiterpenes, O-glycosides, ketones, and organic acids were determined to be the basic substances of the pungent flavor. Phenylpropylenes and flavanones were determined to be the basic substances of the sweet flavor, whereas bicyclic sesquiterpenes, bicyclic diterpenes, tricyclic diterpenes, and tetracyclic diterpenes were identified as the basic substances of the bitter flavor. Monocyclic monoterpenes, monocyclic sesquiterpenes, alcohols, and phenylethyl derivatives were recognized as common substances for pungent, sweet, and bitter.

Table 5.

Results of feature selection based on high-value groups

No.	Main flavors			Combined flavors
	Variables	P-values of 1-probability		Variables	P-values of 1-probability
1	Monocyclic monoterpenes	1.00		Caffeetannins	0.98
2	Monocyclic sesquiterpenes	1.00		Bicyclic sesquiterpenes	0.98
3	Bicyclic monoterpenes	1.00		Flavones	0.96
4	Acyclic monoterpenes	1.00		Monocyclic monoterpenes	0.95
5	Bicyclic sesquiterpenes	0.99		Other benzene derivatives	0.93
6	Phenylpropylenes	0.99		Bicyclic monoterpenes	0.88
7	Acyclic sesquiterpenes	0.99		Protocatechuic derivatives	0.87
8	Alcohols	0.99		Alcohols	0.84
9	Bicyclic diterpenes	0.99		Tricyclic diterpenes	0.79
10	O-Glycosides	0.97		Benzofurans	0.71
11	Phenylethyl derivatives	0.94		O-Glycosides	0.71
12	Tricyclic diterpenes	0.93		Esters	0.62
13	Ketones	0.91		Ketones	0.61
14	Organic acids	0.90		/	/
15	Tetracyclic diterpenes	0.89		/	/
16	Flavanones	0.84		/	/

Similarly, feature selection was performed for the 39 high-value groups for the combined flavors to extract "decisive groups." Using the likelihood ratio chi-square test, variables with P-values of 1-probability > 0.6 were selected, resulting in 13 important variables (Table 5). According to the interaction relationships of high-value groups for the main flavors (Figure 3), caffeetannins, bicyclic sesquiterpenes, and protocatechuic derivatives were determined to be the basic substances of the pungent flavor. Tricyclic diterpenes, benzofurans, and esters were determined to be the basic substances of the sweet flavor, whereas flavones, monocyclic monoterpenes, other benzene derivatives, O-glycosides, and ketones were identified as the basic substances of the bitter flavor. Bicyclic monoterpenes and alcohols were recognized as common substances for pungent, sweet, and bitter.

3.3.3. Decisive groups modeling

A multiple logistic regression model was established, using the 16 identified variables above. As before, the model inputs were the 16 variables, with the outputs as the main flavors (pungent, sweet, and bitter), enter method as the multinomial procedure method, and the sweet flavor as the basic category. Models were evaluated using likelihood ratio tests for equation significance, Wald tests for variable significance, Nagelkerke’s R2 and the accuracy of the confusion matrices for the goodness of fit. We obtained an accuracy of 88.8%, with individual accuracies of 88% for pungent, 90.5% for sweet, and 89.5% for bitter flavors (Table 6, Model 2). It can be inferred that the 16 variables mentioned above can be regarded as "decisive groups," representing the decisive substances for the main flavors of Lamiaceae herbs.

Table 6.

Logistic regression results based on decisive groups

No.	Consequent	Independent variables	No. of variables	Sig. (P-values)	Nagelkerke R2 values	Percent correct (%)
						Pungent	Sweet	Bitter	Total
1	Main flavors	Decisive groups	16	0.001	0.354	78.3	38.1	54.4	65.3
2	Main flavors	Decisive groups (single interaction)	61	0.490	0.922	88	90.5	89.5	88.8
3	Main flavors	Decisive groups (all 2-way interactions)	124	0.261	0.922	87	95.2	89.5	88.8
4	Combined flavors	Decisive groups	13	0.011	0.453	54.8	38.5	83.3	67.4
5	Combined flavors	Decisive groups (single interaction)	48	0.468	0.894	77.4	84.6	93.8	87
6	Combined flavors	Decisive groups (all 2-way interactions)	66	0.419	0.894	80.6	84.6	91.7	87

For the combined flavors, the 13 identified variables were used to build a multiple logistic regression model using the enter method, resulting in an accuracy of 87%, with individual accuracies of 77.4% for pungent, 84.6% for sweet, and 93.8% for bitter flavors (Table 6, Model 5). It can be inferred that the 13 variables mentioned above can be regarded as "decisive groups" that represent the decisive substances for the combined flavors of Lamiaceae herbs.

4. DISCUSSION

The theory of CHM’s properties is an important part of the basic theory of TCM that elucidates the fundamental principles of disease prevention and treatment. Modern interpretation of this theory is crucial for the modernization of TCM theory,⁵⁶ quality markers,⁵⁷ overall evaluation of herb quality, processing and combination mechanisms,⁵⁸ analysis of compound formulas,⁵⁹ and development of new medicinal resources.⁶⁰ With the integration of multiple disciplines and the continuous application of new techniques and methods in TCM research, the intrinsic laws, material bases, and mechanisms of CHM’s properties are being constantly revealed, promoting the modern research progress of TCM.^61-64 However, current studies often focus on specific details, ignoring the holistic nature of TCM theory. Research results are difficult to explain from the basic theory of TCM, which still lacks support for the inheritance, innovation, and development of TCM, especially for the "re-clinical" of basic research.

This study selected Lamiaceae herbs as the entry point, combining macro and micro perspectives to investigate the correlation between the five flavors and their chemical substances, in alignment with the TCM theory. The results showed significant differences in the material bases of the pungent, bitter, and sweet flavors. These five flavors are often characterized by the interaction of multiple classes of substances, and a single class of substances cannot be used to characterize the flavors. The organic combination of multiple classes of substances is the material basis of the five flavors both the main and combined flavors. Among them, 16 classes of substances, such as monocyclic monoterpenes and monocyclic sesquiterpenes, were the decisive groups for the main flavors, and 13 classes of substances, such as coffeetannins and bicyclic sesquiterpenes, were the decisive groups for the combined flavors. There are significant differences in the material basis of the main and combined flavors, suggesting that the “natural flavor” and “functional flavor” of the five flavors may have different material bases. Additionally, the clinical significance of this study was prominent. On one hand, these models provide tools for recognizing the five flavors in new clinical drug development and new formulation development. On the other hand, it offers a practical five-flavor prediction tool for developing and utilizing new varieties of resources (CHM) based on the classical theories of TCM.

Considering the differences between the main and combined flavors, this study established several models for identifying flavors (Table 4, Table 6). The most practically significant models for the high-value groups were Model 1 in Table 4 (main flavors, accuracy 77.1%, 34 variables) and Model 5 in Table 4 (combined flavors, accuracy 93.5%, 39 variables). Regarding decisive groups, Model 1 in Table 6 (main flavors, accuracy 65.3%, 16 variables) and Model 4 in Table 6 (combined flavors, accuracy 67.4%, 13 variables) were more practical.

For both the high-value and decisive groups, the predictive accuracy was greatly improved when the interaction effects of independent variables were introduced, indicating that the interaction of substance components was crucial to the CHM’s properties. The predictive accuracy of combined flavors is higher than main flavors, suggesting that the combined flavors might reflect more of the "functional flavor" attribute, which is acquired, while the main flavors might combine both "natural flavor" and "functional flavor." Some herbs may have a main flavor of "natural flavor," some may have a main flavor of "functional flavor," and some may have both. To elucidate the "inconclusive" phenomenon of ancient records on CHM’s properties, future research on the five flavors must be conducted from a historical perspective, a dynamic perspective, a dialectical viewpoint, and a universally connected viewpoint. This will enable the elimination of falseness and the preservation of the essence, the restoration of "natural flavor" and "functional flavor", and the coordination of the relationship between the four natures, the five flavors, and other properties.

This study also explored the applicability of feature selection and association rule analysis to investigate CHM’s properties. By directly involving all 145 categories in the three-level classification in feature selection and selecting the top 15 groups as the decisive groups, their predictive accuracy was consistently lower than that of the results obtained using the new parameter methods. This finding is possible because feature selection primarily examines the importance of individual variables and does not account for variable interactions. This further proves that CHM’s properties are the external manifestations of the interactions of multiple constituents and that studying the medicinal properties by examining individual substance components alone cannot be conducted effectively.

In addition, this study found that although logistic regression modeling with decisive groups had good overall performance, the predictive accuracy for certain specific flavors was not high. For example, the overall accuracy of the main flavors was 65.3%, but for the sweet flavor, it was only 38.1%. The overall accuracy of the combined flavors was 67.4%, but for the sweet flavor, it was only 38.5%. Such discrepancies might be due to the inherently low proportions of specific flavors (sweet flavor accounted for 12% of the main flavors and 14% of the combined flavors) or data incompleteness, with many herbs having limited known constituents.

This study had some limitations. First, the logistic regression model did not include specific groups, such as pyrans and gallotannins, which had high support rates of 3.5% and 3.26%, respectively. These groups may play important roles in predicting medicinal properties and should be considered in future research. Second, although the new parameter method is superior to the exhaustive method, a retrospective analysis of the data showed that when analyzing the pungent of main flavors, certain important substances such as tricyclic sesquiterpenes (support rate 49.4%, confidence 72.6%) and cyclopentanes (support rate 15.3%, confidence 70.1%) should play important roles in predicting the pungent flavor but did not enter the high-value groups. Similarly, during the analysis of the bitter of main flavors, substances with relatively high support rates (≥ 4.7%), such as o-phenanthrenequinones, p-phenanthrene-quinones, and tricyclic diterpenes, entered the high-value groups but not the decisive groups. This indicates that there is still room for improvement and enhancement of the new parameter method for obtaining all valuable substances from the results of the association rule analysis. There are various methods for deriving decisive groups from high-value groups, but feature selection is the only one available. In future research, it will be necessary to further optimize data analysis methods to obtain more reasonable results. Third, this study was conducted based on a third-level classification of chemical substances. Each substances has complex biological effects. It is not yet possible to conduct an in-depth discussion of the material efficacy properties. As mentioned previously, the five flavors are regarded as a result of the interaction between the material components. With 34 high-value groups for the main flavors and 39 high-value groups for the combined flavors, it would be very complex to discuss their interactions. We will conduct further research and seek new methods to study the interrelationships between "material-efficacy-property" in the future.

In conclusion, this study explored the correlation between five flavors and chemical substances at the chemical skeleton level and established identification models, laying the foundation for further research on the CHM’s properties based on substance components.

Funding Statement

Supported by the National Natural Science Foundation of China: Research on the Identification of Cold-Hot Properties in Lamiaceae Herbs based on Infrared Spectroscopy Holistic Component Characteristic Markers (No. 81673622); Anhui Provincial Natural Science Foundation: Research on the Extraction and Identification of Holistic Compositional Characteristics of Warm-Hot Properties of Lamiaceae Herbs (No. 1508085MH202); Anhui Provincial Natural Science Research Project of Higher Education: Research on the Material Basis of Cold-Hot Properties of Lamiaceae Herbs based on Pattern Recognition and Energy Metabolism (No. 2023AH050773)