The antibacterial TCMs were counted according to the families and their properties using Excel, and then the data was imported into Rstudio1.3.959 and was further visualized using the ggplot2 package.

The information was imported into the R software platform, taking the families, properties, and efficacies of TCMs as the former item and the antibacterial effectiveness as the last item. Programs and the packages such as arules and arulesViz were used to mine and visualize the association rules (Hahsler, 2017). The correlation between the antibacterial activity of TCMs and the families, properties, and efficacies of TCMs was evaluated with the value of support, confidence, and lift of each rule.

The Taxonomy ID of TCMs original species was searched in the Taxonomy Database and uploaded to the Phylot website¹ to generate a text file of the phylogenetic tree. And then, the text file was uploaded to Interactive tree of life (ITOL)² for visual analysis (Ivica and Peer, 2011). In the phylogenetic tree, the antibacterial activity of the corresponding TCMs basic source species was marked. Finally, the distribution law of antibacterial TCMs in the phylogenetic tree of the basic source species was analyzed.

The data was imported into RStudio1.3.959 for preprocessing, which mainly included five parts: (1) Data integrity check: data with incomplete information was deleted. (2) Data encoding: the principle of “1 represents yes, 0 represents no” was used to describe the characteristics of TCMs properties, and the sequence 1,2,3.n was used to record the families’ characteristics. (3) Data type conversion: the dependent variables were transformed into the factor variables suitable for the classification model. (4) Data Splitting: the dataset was randomly split into 80% training data and 20% test data; (5) Data standardization and centralization: the data was standardized and centralized with the preProcess function.

This paper uses the recursive feature elimination method for feature selection, which combines the feature selection process with the training process. The model’s predictive ability was used to measure feature selection (Chen et al., 2009). Random Forest (RF), Naive Bayes (NB), Bagged Classification and Regression Tree (Bagged CART), and Linear Discriminant Analysis (LDA) were fitted using the rfe function in the caret package, and 10-fold cross-validation was used in each process.

After feature selection, six machine learning algorithms, namely Random Forest (RF), k-Nearest Neighbor (KNN), Classification and Regression Tree (CART), Support Vector Machine (SVM), Naive Bayes (NB), and Neural Networks (NN), were used to build models for different datasets. The train function implemented this process in the caret package, and five times 10-fold cross-validation was used in each model.

The receiver operator characteristic curve (ROC) of the model was established to predict the occurrence probability of the TCMs antibacterial activity, and calculating the area under the curve (AUC), comparing the performance and evaluating the predictive ability of each model (Baicus et al., 2011).

The parameters of the screened model were optimized to improve the performance further. To obtain the optimal model, the parameters combinations of all models were traversed through the combination of grid parameter adjustment method and random parameter adjustment method. Among them, the critical parameter that significantly affects the KNN algorithm was the neighbor number k. The NN algorithm’s parameters needed to adjust were the number of hidden units size and the weight attenuation parameter decay.

Seventy five percent ethanol was purchased from Sinopharm Chemical Reagent Co., Ltd. (Shanghai, China), and Luria-Bertani (LB) Culture was purchased from Shanghai Aladdin Biochemical Technology Co., Ltd. (Shanghai, China).

Dichroae Radix, Cirsii Japonici Herba Carbonisata, Swertiae Herba, Bubali Cornu, Callicarpae Formosanae Folium, Turpiniae Folium, Changii Radix, and Trachelospermi Caulis et Folium were all purchased from Qingdao Tiancheng Chinese Medicine Co., Ltd. (Shandong, China). The TCMs were identified by Professor Baoguo Li, Shandong University of Chinese Medicine.

Preparation of ethanol extract of TCMs: Firstly, 50 g of TCM samples were crushed and soaked 10 times in 75% ethanol for 1 h, followed by reflux extraction at 100°C for 2 h. Secondly, the extract was collected, and the remaining residue was extracted eight times using 75% ethanol at 80°C for 1 h. Finally, the two sections were mixed and vacuumed filtrated, and the filtrate was concentrated by rotary evaporation at 45°C to obtain 1 g/mL ethanol extract and stored in –20°C refrigerator until use.

Preparation of water extract of TCMs: 75% ethanol was replaced with pure water. According to the preparation method of alcohol extract, the water extract with 1 g/mL was obtained.

E. coli and S. aureus were certified and provided by the Ocean University of China and were stored at –20°C. Firstly, the bacteria were activated before use, and 15–20 mL LB culture was poured into the sterilized Petri dish to be spread evenly and solidified for use. Secondly, the inoculation loop was used to dip the bacterial solution, and a “Z” line was drawn on the solid medium. Then, after being rotated 90°, a second line was drawn in the same way, and the operated Petri dishes were sealed with sealing strips and incubated at 37°C for 24 h. Finally, the bacteria were picked up by an inoculation loop in sterile water, and the bacterial solution was shaken until even. Compared with the shaken McFarland Turbidity standard tube, the bacterial suspension containing about 1 × 106 ∼ 1 × 107 CFU⋅mL^–1 was prepared and placed for reserve.

A total of 100 μL of the prepared E. coli suspension were absorbed and evenly spread on the Petri dishes containing 15–20 mL LB culture. Four holes were drilled in the agar medium orderly of each dish. Different TCM extracts (100 μL) were added to the holes of each Petri dish. The water extract was added to the up and down spots, and the ethanol extract was added to the left and right holes. Finally, the inoculated plates were sealed with sealing strips, labeled corresponding information, and incubated for 24 h at 37°C in a constant temperature biochemical incubator. The production of a bacteriostatic ring was observed, and the diameter of the bacteriostatic ring (mm) was measured with a vernier caliper. The experiment was repeated three times.

The frequency analysis results on the properties of antibacterial TCMs showed that most medicines were cold nature, bitter, pungent and sweet tastes. Most of the meridians tropism were liver, lung, spleen, and stomach. The statistical results of the two kinds of antibacterial TCMs properties are the same (Figure 3A).

Association rules are an essential technology of data mining to mine the correlation between valuable data from a large amount of data, reflecting the interdependence and correlation between things (Tung et al., 2003). The association rules between TCMs properties and anti-gram-positive bacteria and anti-gram-positive bacteria were explored (support > 10%, confidence > 80%). The results showed a strong relationship between cold TCMs and the anti-gram-negative bacteria and anti-gram-positive bacteria activity, and the confidence degree was 81 and 85%, respectively. In terms of taste, sour TCMs had a higher degree of confidence (about 92%) with effective resistance to gram-negative bacteria, followed by bitter TCMs with a confidence of 81%. However, among the effective rules against gram-positive bacteria, bitter TCMs had the highest correlation degree with the confidence of 86%. Among the meridian tropism rules meeting the threshold, the most effective correlation with gram-negative bacteria was the gallbladder meridian (90%), followed by the large intestine meridian (89%), liver meridian (83%), heart meridian (83%), stomach meridian (83%), lung meridian (82%), and spleen meridian (82%), in order of confidence. In terms of gram-positive bacteria, the gallbladder meridian (94%), large intestine meridian (91%), heart meridian (87%), liver meridian (87%), stomach meridian (86%), lung meridian (86%), spleen meridian (85%) had the order of confidence from high to low. As for the efficacy of TCMs, the hemostatic efficacy showed a higher correlation with the confidence of 89 and 90%, respectively. Following purging fire efficacy with the confidence of effectiveness against gram negative bacteria and gram-positive bacteria were 89%, 93%, Respectively (Figure 3B).

According to statistics, 714 TCMs from 820 original species were effective against gram-negative bacteria, which were distributed in 170 families, 82 orders, 19 classes, and 9 phyla. Among them, the most significant number of species were found in Asteraceae (76), and Fabaceae (50), Rosaceae (41), Lamiaceae (34), and Ranunculaceae (32) with more than 30 species. The most numerous orders were Lamiales (84), followed by Asterales (80), Ranunculales (67), Rosales (57), and Fabales (54). A total of 730 TCMs were effective against gram-positive bacteria, from 832 original species, distributed in 179 families, 86 orders, 20 classes, and 9 phyla. Similar to the statistical results of TCMs against gram-negative bacteria, the top three families with the largest number of species were Asteraceae (76), Fabaceae (47), and Rosaceae (36), and the top three orders with the largest number of species were Lamiales (82), Asterales (77) and Ranunculales (61) (Figure 4A).

A total of 147 rules for the families of antibacterial TCMs and the activity against Gram-negative bacteria were screened (support > 4%, confidence > 60%). The results showed that the highest compelling correlation with anti-gram-negative bacteria activity was Ranunculaceae and Rosaceae, with the confidence of about 91%, followed by Polygonaceae with the confidence of about 90%. The rest were Lamiaceae, Oleaceae, and Asteraceae, which had more than 80% confidence. The orders with high effective associations with anti-gram-negative bacteria activity were Ranunculales, Rosales, Lamiales, Sapindales, and Caryophyllales (Figure 4B). With the same support and confidence as thresholds, the rules of TCMs families and anti-gram-positive bacteria activity were excavated. Polygonaceae, Ranunculaceae, Lamiaceae, Asteraceae, and Fabaceae were the top five families with the highest effective correlation against anti-gram-positive bacteria activity. The order with the highest confidence was Ranunculales (93%), followed by Lamiales, Sapindales, and Asterales (Figure 4C).

To further study the distribution of antimicrobial TCMs’ original species in the phylogenetic tree, all TCMs in Pharmacopoeia of the People’s Republic of China (Pharmacopoeia Commission China, 2020), Chinese Materia Medica (National Chinese Medicine Authority, 1999), Grand Dictionary of Chinese Traditional Medicine (Nanjing University of Chinese Medicine, 2006) and National Compendium of Chinese Herbal Medicine (Xie, 1976) were collected. The Chinese mineral medicines and some TCMs without Taxonomy ID were eliminated. There were 11,272 TCMs from 742 families. ITOL was used to construct the distribution of TCMs with antibacterial activity in the phylogenetic tree of the 11,272 TCMs original species. The end nodes of the branches represented families of different species, and species with similar genetic relationships will also be close in the tree. In the outer ring, the anti-gram-negative bacteria activity and anti-gram-positive bacteria activity were indicated in blue and red, respectively.

The results showed that 187 families had produced TCMs with antibacterial activity and showed specific aggregation. Most TCMs with antibacterial activity came from Tracheophyta (about 92% of all antibacterial TCMs). The rest were scattered in Metazoa, Fungi, and other kingdoms, consistent with the previous results of the mass of the number of plant TCMs (Figure 5A).

In the phylogenetic tree of the Viridiplantae kingdom, families with more than two species were identified in the figure due to the significant difference in the number of species in antibacterial TCMs families. The results demonstrated a specific aggregation of TCMs, mainly concentrated in the four branches of rosids, asterids, Liliopsida, and Ranunculales. The antibacterial TCMs in these four branches accounted for 74% of all plant antibacterial TCMs (Figure 5B).

Feature selection is an effective method to reduce the number of features in input data. Too much redundancy often increases the calculation overhead and hurts the model (Li et al., 2017). The recursive feature elimination method can remove features with low weight coefficients through continuous training to select high-quality variable subsets. Therefore, we applied the recursive feature elimination method to fit four RF, NB, LDA, and Bagged CART functions for feature selection. In the feature selection of the anti-E. coli model, 26 features were retained after RF selection, with an accuracy of 85.1% for cross-validation; 15 features were included after NB selection, with an accuracy of 83.8%; and 20 components were retained after LDA selection, with an accuracy of 83.7%. After Bagged CART selection, 18 features were kept, and the accuracy was 85.0% (Figures 7A–D and Supplementary Table 1). The feature selection results of the anti-S. aureus model were similar to those of the anti-E. coli model. The number of features retained after RF function feature selection was the largest, and the cross-validation accuracy was also high. After selection, 26 features were maintained with an accuracy of 85.3%. After selection by NB, LDA, and Bagged CART, 15, 20, and 18 components were retained, with accuracy of 83.5, 83.1, and 84.8%, respectively (Figures 7E–H and Supplementary Table 1). In this way, the number of features in the input data was reduced, and the model’s performance was improved.

This paper selected six machine learning algorithms, RF, KNN, CART, SVM, NB, and NN, in the prediction model of TCMs against E. coli and S. aureus. We evaluated the model performance by comparing the AUC value. According to the AUC results of the prediction model of anti-E. coli activity, the KNN and NN algorithms had better performance with the AUC value reaching more than 75%. The CART algorithm was slightly inferior to the above algorithms, the RF and NB algorithms had similar results, and the SVM algorithm had the worst performance (Figure 8A). In the prediction model of anti-S. aureus activity, RF, KNN, CART, and NN were the algorithms with AUC values above 75%, respectively, among which KNN and NN had the best performance. On the contrary, SVM and NB performed worst in this model (Figure 8B).

The model parameters were adjusted to achieve the optimal parameter combination to reduce the loss in the test set and improve the prediction accuracy. The optimization results were shown in the AUC value improvement. The top five algorithms of each model were selected from the AUC results of the antibacterial prediction models for optimization. In the prediction model of anti-E. coli activity, the NN algorithm was selected after LDA and Bagged CART, and KNN algorithms selected by NB, LDA, and KNN, were chosen for optimization. After optimization, the AUC value of each algorithm increased by varying degrees. Among them, the NN after Bagged CART selection had the best results, with the AUC value rising from 76.6 to 77.5% (Table 1 and Figure 9A). In the prediction model of anti-S. aureus activity, the NN algorithms selected by NB and LDA, and KNN algorithms, chosen by RF, Bagged CART, and NB, were optimized, respectively. After LDA selection, the results showed that NN performed best with the most significant increase, whose AUC value increased from 77.9 to 80.0% (Table 1 and Figure 9B).

The above models were applied to the screening of antibacterial TCMs. To predict the anti-E. coli anti-S. aureus activity, some TCMs were sorted out from Pharmacopoeia of the People’s Republic of China (Pharmacopoeia Commission China, 2020). The screening criteria were as follows:(1) TCMs with a bitter taste, cold nature, and return to the liver meridian; (2) excluding TCMs with antibacterial activity already reported in the literature; (3) excluding Chinese mineral medicines and TCMs with unclear source information. The prediction model predicted the antibacterial activity of the sorted TCMs, and the results were shown in the table (Supplementary Table 2). In the prediction results of the anti-E. coli activity, Cirsii Japonici Herba Carbonisata, and Swertiae Herba were effective, while six TCMs including Dichroae Radix and Turpiniae Folium were not. In the prediction results of anti-S. aureus activity, however, five TCMs, including Dichroae Radix and Cirsii Japonici Herba Carbonisata, were effective (Table 2).