The data used in this study were derived from our previous research, which systematically analyzed TCM prescriptions to identify potential drug candidates for the treatment of pox (Wu and Guo, 2024). The prescriptions were manually collected from online databases and local libraries to ensure comprehensive coverage. All data are publicly available and accessible through our prior publication. Given that this study focused on the dosage effect in TCM analysis, two types of prescriptions were excluded: 1. Prescriptions containing botanical drugs with non-quantifiable dosage units, such as “50 grains of glutinous rice,” “two dates,” or “a handful of bamboo leaves.” 2. Botanical drugs for which relevant information—specifically bioactive metabolites and targets—could not be retrieved using the TCM-LTM database (Li et al., 2022).

The network components for analyzing TCM include botanical drugs, their bioactive metabolites, the targets of these metabolites, and the pox viruses along with their associated targets. To retrieve the metabolites and their corresponding targets, the LTM-TCM database was utilized. The LTM-TCM database specifically designed for network pharmacology research in TCM, is comprised of data from 14 authoritative databases, encompassing information on 9,122 botanical drugs, 34,967 metabolites, and 13,109 targets (Li et al., 2022). The bioactive metabolites within the database are defined based on Lipinski’s rule. Using this resource, we collected the bioactive metabolites and their corresponding targets for each of the 94 prescriptions.

Although pox viruses are diverse, they exhibit a degree of homology. Four prevalent species including vaccinia, smallpox, chickenpox, and mpox were selected for this study. A total of 686 disease-associated target genes were identified from the GeneCards database (Stelzer et al., 2016) using these four pox viruses as search keywords.

In this study, the network was constructed based on the “botanical drugs–bioactive metabolites–target–pox virus” framework. In the unweighted network, each relationship was treated equally, and a weight of one was assigned to the edges connecting the components of the network. For the dosage-weighted network, the dosage of each botanical drug was used to assign weights to the corresponding edges. The key step in this process involved constructing a standardized dosage vector, which was derived by standardizing the dosages of botanical drugs within the prescriptions. For example, in the prescription called Chong He powder, the components included one Fen of Bai Yao Zi, one Fen of Gan Cao, and one Qian of Xiong Huang. Using the unit conversion relationships (Qiu, 1996) summarized in Table 1, these traditional units were converted into metric units, resulting in a dosage vector of [0.4 g, 0.4 g, 4 g], which was then normalized by dividing each value by the minimum weight (0.4 g in this case) to yield a standardized dosage vector of [1, 1, 10]. The standardized dosage vector was subsequently applied as the weight to the edges connected to the corresponding botanical drugs. In contrast, in the non-dosage network, the standardized dosage vector for this prescription was [1, 1, 1].

For both the non-dosage and dosage-weighted networks, nodes with a degree higher than the average were identified as key targets of the prescription. These key targets were subsequently compared to pox-related targets, and the overlapping targets were defined as drug targets for the prescription. To explore the potential working mechanisms of the prescriptions, the version of 4.14.4 clusterProfiler R package (Wu et al., 2021) was employed for Gene Ontology (GO) annotation based on the filtered drug targets. A significance threshold of p < 0.05 and a minimum of 10 enriched genes associated with each pathway were applied to identify biologically relevant pathways, which provides insights into the potential mechanisms underlying the therapeutic effects of the prescriptions.

The output of the network analysis included the predicted drug targets and their corresponding pathways. To quantify the differences in output between the non-dosage and dosage-weighted networks, two indicators were created: 1. Difference in Drug Targets (DeDT) to measure the disparity between the predicted drug targets from the two networks; 2. Difference in Pathways (DePy) to quantify the variation in the enriched pathways associated with the predicted drug targets. D e D T   o r   D e P y = 1 − S 1 ∩ S 2 S 1 ∪ S 2

The 1-Jaccard Similarity Index (Kosub, 2019) was employed to quantify the difference. The set of drug targets or pathways predicted by the non-dosage network was denoted as S₁, while S₂ represents the corresponding predictions from the dosage-weighted network.

The input to the network consisted of botanical drugs and their associated metabolites.

Wilcoxon rank-sum test was utilized to assess the statistical significance of relationships between network input and output variables, as well as between botanical drug counts and their corresponding standard deviations. For evaluating linear associations, Pearson correlation coefficients were applied to examine two key relationships: 1. Between the Dedis and DeSD indicators; 2. Between botanical drug counts and standard deviations.

The 94 filtered prescriptions from a prior study were analyzed (Supplementary Table S1). These prescriptions were derived from historical medical texts spanning the Song, Yuan, Ming, and Qing dynasties in China. After comparing outputs from non-dosage and dosage-weighted networks for each prescription, we observed that the DeSD and Dedis values - used to quantify the differences in inputs between the two types of networks - ranged from 0.12 to 34.37 and 0.35 to 99, respectively. Additionally, the DeDT and DePy indicators, which were employed to access the difference in network analysis outputs, ranged from 0.0% to 68.9% and 0.0%–74.6%, respectively, suggesting the impact of dosage across prescriptions (Supplementary Table S2).

DeSD and Dedis were created to delineate the differences in network input attributable to dosage from two distinct perspectives. DeSD quantified the deviation of dosages from the mean value, while Dedis measured the distance between evenly distributed dosages and specific dosage values. A significant linear relationship was observed between these two indices (cor = 0.49, p = 4e-7). To evaluate the impact of incorporating dosage into the network, outcomes from networks that considered dosage were compared with those that did not, resulting in the generation of DeDT and DePy indicators. Based on the median values of DeDT and DePy, the 94 results were categorized into two groups: high output change and low output change. Upon comparing the DeSD and Dedis values between these groups, it was observed that only Dedis exhibited a statistically significant difference (Figure 1).

Botanical drugs and their metabolites serve as inputs to the network at two distinct levels. Botanical drugs typically comprise a complex mixture of metabolites, which act as the direct agents interacting with target genes, thereby determining the efficacy of the botanical drug. To further investigate the relationship between DeSD/Dedis and DeDT/DePy, a comparative analysis found that Dedis did not exhibit statistically significant difference at the metabolite level. In light of this observation, we proposed that the therapeutic effects of botanical drugs were driven by the synergistic interactions of multiple bioactive metabolites, rather than the mere summation of their individual functions. This conclusion was supported by the well-documented phenomenon that isolated so-called key metabolites from botanical drugs often failed to replicate the therapeutic effects observed in whole extracts (Gilbert and Alves, 2003; Rasoanaivo et al., 2011; Caesar and Cech, 2019).

To better illustrate these relationships, a clock model was constructed (Figure 2A). As shown, the function pointer shifted when the two hands formed an arc, but it still pointed to the same function when the arc was not long (Figure 2A2). The arc represented the Dedis and the function pointer represented the output of the dosage-weighted network. The clock model illustrated that Dedis was a necessary and insufficient condition to determine the output change (Figure 2A1,A2,A3). In addition, the standard deviation (SD) was the intrinsic property abstracted as the length of the hands. The difference in length of the two hands cannot change the direction of the function pointer. Therefore, the DeSD should not be used as the indicator to predict whether the dosage-weighted network might yield altered output or not. When the angle between the two hands remained unchanged, the arc increased with the increased length of the hands, indicating a linear relationship between DeSD and Dedis (Figure 2A4).

From this clock model illustration, we can easily deduce that the efficacy of a drug combination will be enhanced when the dosage of each component multiplies proportionally (when the hand is extended, the original hand overlaps with the new one without changing the direction of the function pointer, meaning the function remains the same) (Figure 2A1).

To demonstrate the impact of dosage on TCM network analysis, we selected Chong He Powder from our dataset of 94 prescriptions (selection criteria: Dedis >0). This prescription, historically indicated for pox-infected patients with dryness and heat symptoms, comprises three botanical drugs with distinct dosages: realgar for 1 Liang, Bai Yao zi for 1 Qian, and Gan Cao for 1 Qian. This case study examined how dosage weight redistributed influence among the botanical drugs and refined the potential pathways, aligning with the documented clinical efficacy of the prescription.

The comparison between non-dosage and dosage-weighted network analyses in our work focused on the same prescriptions. This approach has demonstrated how integrating dosage information provides a distinct perspective for deciphering prescription mechanisms compared to the traditional non-dosage method. Importantly, the dosage-weighted network analysis can also be applied to compare different prescriptions containing identical botanical drugs but distinct dosage proportions. For example, “Gan Jie Soup” and “Ren Shen Gan Jie Soup” share the same botanical drug components but with different dosage ratios. Historical records indicated that these prescriptions from different eras were used to treat similar clinical symptoms. When applying dosage-weighted network analysis to these two formulations, identical drug targets were predicted. This finding explained how these prescriptions with varying dosages could have comparable clinical efficacy for nearly identical symptom profiles. In contrast, non-dosage network analysis would fail to reveal this dosage-efficacy relationship, as it would invariably produce identical predictions regardless of dosage variations. Therefore, dosage-weighted network pharmacology offers a valuable alternative approach to understanding combination drug principles through the lens of dosage optimization.

Secondary analysis revealed a weak inverse correlation between prescription dosage variability (standard deviation) and the number of botanical drugs (r = −0.3, p = 0.001; Figure 4A). Notably, total prescription dosage remained stable across various numbers of botanical drugs (Figure 4B). This observation may reflect established combination drug safety practices, where individual drug dosages are typically minimized to maintain overall dosage within a safe therapeutic range. Such an approach may potentially reduce the adverse effects while preserving treatment efficacy (Ali et al., 2017).