In total, 269 compounds from BLG were collected, which can be clustered into five groups according to their Tanimoto distance (Figure 2A). Cluster 2 contained the most compounds (n = 95), followed by cluster 4 (n = 54) and cluster 1 (n = 51). The structures of each cluster center are displayed in Figure 2B, which referred to the representative chemical scaffolds of the ingredients in BLG. Here, we evaluated the drug-likeness and ADME properties (PPB and human oral bioavailability) for initially excluding unsuitable candidates. Drug-likeness is utilized in drug design to describe how “druglike” a particular molecule is with respect to various molecular properties and structural features. As shown in Figure 2C, the classification model of drug-likeness eliminated 24 compounds with predicted positive probability lower than 50%, which left 245 compounds. Thereafter, oral bioavailability, one of the key pharmacokinetic indexes reflecting the efficiency of absorption, was further assessed. In this step, 88 compounds were predicted to have a bioavailability lower than or equal to 20%. PPB ability has significant effects on the pharmacodynamic action of a drug since one of the main mechanisms for drug uptake and distribution is through PPB. A compound with PPB <90% is considered suitable, while drugs with a high protein binding rate may have a low therapeutic index (Jie et al., 2018). During this process, three compounds were removed for inappropriate PPB property. Finally, 115 out of the 269 compounds in BLG were excluded, with the remaining 154 utilized for further study (Supplementary Table S1).

The C–T network was composed of interactions linking the screened compounds of BLG and their corresponding targets, which contained 7,845 CTIs connecting 154 compounds with 472 protein targets (Figure 3 and Supplementary Table S3). We further mapped the profiles of the influenza virus host proteins onto the C–T network and calculated the degree of each node. In total, there were 607 pairs of CTIs linked to 30 influenza-related targets. The average compound degree (D) of an influenza-related protein was 20.2, while the average influenza-related target degree (K) of a compound was 3.9. Among the top 20 targets with highest degree, ALDH1A1 (D = 150), MAPK1 (D = 149), and TP53 (D = 146) were influenza-related targets, indicating their important role in the anti-influenza effect of BLG. Interestingly, we found that each of the 154 compounds possessed at least one influenza virus host protein target. The top 5 compounds with the most influenza-related targets were acacetin (K = 11), hispidulin (K = 9), indigo red (K = 8), tryptanthrin (K = 8), and calycosin (K = 8). These compounds have been reported to exert potential as antiviral drug candidates against multiple viral infectious diseases, such as respiratory syncytial virus, severe acute respiratory syndrome (SARS), Middle East respiratory syndrome (MERS), COVID-19, human immunodeficiency virus-1 (HIV-1), coxsackie virus B3 (CVB3), and influenza–parainfluenza (Adhikari et al., 2021; Omrani et al., 2021; Mok et al., 2014; Chen et al., 2011; 2020; Mani et al., 2020). Overall, the C–T network revealed that the constituent components of BLG could exert synergistic anti-flu effects through targeting multiple influenza virus host proteins.

Based on the C–T network of the 154 compounds and the influenza virus host protein set, two network predictive models were built to identify novel anti-IAV agents from BLG. By setting the threshold of q < 0.05, models A and B identified 71 and 23 compounds as potent active compounds, respectively (Figure 4A and Supplementary Table S4). To improve the prediction accuracy, we merged the prediction results of the two models and obtained an intersection of the two sets of predicted positive compounds. The 23 compounds that had been predicted as potential anti-IAV agents simultaneously could serve as more promising candidates (Figure 4B). Indeed, some of them exhibited broad-spectrum antiviral effects in previously published experiments. For instance, indirubin was found to possess concentration-dependent inhibitory activity on Japanese encephalitis virus (JEV) replicated in vitro and exhibited good protective effects in a mouse model with lethal JEV challenge (Chang et al., 2012). It has also been reported to decrease the H1N1 susceptibility and alleviate lung damage in a restraint-stressed mouse model by regulating the mitochondrial antiviral signaling pathway (Chong et al., 2017). In addition, tryptanthrin was confirmed to effectively inhibit the CPE and virus yield (IC₅₀ = 1.52 μM) in HCoV-NL63-infected cells (Tsai et al., 2020). It also showed promise as a potential candidate against SARS-CoV-2 and other viruses (Mani et al., 2020; Xu et al., 2020). To sum up, evidence from literature preliminarily confirmed the reliability of the predicted results of network-based models, suggesting that the antiviral ability of these compounds against IAV deserves to be further experimentally validated, especially on resistant strains.

In vitro experimental evaluation was performed to further verify the in silico prediction of the potential anti-IAV candidates. By comprehensive considerion of the network-based predicted ranking, in terms of chemical structure, physicochemical property, availability, and accessibility, we finally selected the six most promising candidates from the 23 predicted compounds for evaluation of their anti-IAV efficacy: eupatorin, dinatin, linarin, tryptanthrin, indirubin, and acacetin (Figure 5). A cytotoxicity test was performed to determine whether the concentrations of the test samples used for the experiments would affect cell viability. The results showed that the TC₀ and TC₅₀ values of each compound were greater than 100 μg/ml.

CPE reduction assay was performed to evaluate the inhibition activity of the six compounds with two wild-type and one drug-resistant IAV strains: A/PR/8/34 (H1N1), A/Minfang/151/2000 (H3N2), and A/HebeiXinhua/SWL1106/2017 (oseltamivir- and amantadine-resistant H1N1). Our experiments included four different time points for drug administration: pretreatment, simultaneous treatment, posttreatment, and pre-incubation treatment assay. As listed in Table 2, all of the six compounds showed efficacy (selection index, SI > 1) against the wild-type strain A/PR/8/34 (H1N1). When the H1N1 virus was treated simultaneously or pre-incubated with the test compounds, these candidates tended to have stronger antiviral activity compared with other action modes, suggesting that they may directly suppress the H1N1 activity or impair viral adsorption. For the wild-type strain A/Minfang/151/2000 (H3N2), all the drug candidates except for dinatin also showed antiviral effects in the CPE reduction assay (Table 3). Among them, acacetin showed the best inhibitory activity with an IC₅₀ of 7.62 ± 1.45 µg/ml when it was added to the cells before virus infection, which indicates its potential anti-flu prophylactic effect.

We further evaluated whether the candidates are also effective against the drug-resistant strain. The results of the CPE reduction assay of the six predicted compounds toward A/HebeiXinhua/SWL1106/2017 (oseltamivir- and amantadine-resistant H1N1) are presented in Figure 6 and Table 4. As listed in Table 4, oseltamivir and ribavirin showed inefficacy in the four modes of drug administration tested for the resistant H1N1 strain, while all other tested candidates demonstrated inhibitory efficacy to some degree. Among them, acacetin and tryptanthrin exhibited stronger antiviral activity (Figure 6). Previous studies have shown their promising inhibitory effect against multiple viruses, such as coronavirus, respiratory syncytial virus, and influenza–parainfluenza viruses (Tsai et al., 2020; Adhikari et al., 2021). Additionally, according to the results of the drug administration modes (pretreatment, simultaneous treatment, and posttreatment), the earlier the samples were added to the cells, the better the protective effect they may possess on cells against the virus. Taking acacetin as an example, the average IC₅₀ values for the drug-resistant IAV strain with pretreatment, simultaneous treatment, and posttreatment were 8.44 ± 0.39, 24.33 ± 2.44, and 25.81 ± 1.85 μg/ml, respectively. Correspondingly, its SI (TC₅₀/IC₅₀) decreased. Besides, acacetin (IC₅₀ = 9.25 ± 0.87 μg/ml) and tryptanthrin (IC₅₀ = 7.02 ± 0.65 μg/ml) also exhibited good activities with pre-incubation treatment, indicating that they may have a direct inhibition function on the drug-resistant H1N1 strain. Overall, the six predicted anti-IAV candidates exhibited inhibitory efficacy in the CPE reduction assay both on wild-type and drug-resistant strains and tended to possess better activity with the pretreatment drug administration mode. Although the action intensities of the candidates were lower than that of the positive control drug zanamivir, the preliminary anti-IAV activities suggested their high potential as hit compounds that deserve to be further optimized.

As the CPE reduction experiments exhibited, all six tested compounds showed inhibitory activity against IAV. The high hit rates further proved the reliability of our prediction and demonstrated the high potential as anti-IAV agents of the remaining 17 compounds that have not been experimentally validated yet. Hence, it is plausible to hypothesize that these 23 compounds predicted by the network models are the main active anti-IAV constituents of BLG. We next explored the synergistic effects of these compounds to showcase the underlying MOAs of BLG against IAV. As shown in Figure 4B, the 23 compounds acted on 27 influenza virus host protein targets. These compounds shared a lot of common protein targets, while each target was also targeted by multiple compounds. Among the protein targets, TP53 interacted with most compounds (n = 22), followed by ALDH1A1 (n = 20) and MAPK1 (n = 16). TP53 was identified as the most extensive essential host factor-interacting virus-targeting protein (EHF-interacting VTP) for IAV, which holds great potential in the discovery of host-directed antiviral agents (Xie et al., 2015). In addition, it was reported that ALDH1A1 downregulates and translocates to the nucleus from the cytoplasm during H3N2 virus infection, which may be associated with lipid metabolism that plays multifaceted roles in the life cycle of influenza virus and in virus–host interactions of (Chan et al., 2010; Heaton and Randall, 2011; Wu et al., 2013). The host protein kinase MAPK1 was also validated as essential for the replication of influenza virus (Bakre et al., 2013). These findings suggest the key regulatory targets of BLG for exerting its anti-IAV effects.

Based on the 27 identified key targets, we further performed gene enrichment analysis, including GO term enrichment analysis and KEGG pathway annotation (Kanehisa and Goto, 2000), to explore the underlying molecular mechanisms of BLG against IAV. The top 20 molecular function terms in GO enrichment analysis are presented in Figure 7A. It is shown that these targets were involved in various basic molecular functions related to virus infection, such as the binding and activity of protein kinase, enzyme, and transcription factor. KEGG pathway annotation further elucidated the biological pathways that may be regulated by BLG (Supplementary Table S6). As shown in Figure 7B, most of these pathways were associated with virus infection. For instance, among the organismal system-related pathways, immune system is the largest group, with 13 pathways. It is universally acknowledged that respiratory virus infection usually induces effective immunity, including innate immune response and adaptive immune response, but overactive responses are correlated with pathophysiology (Stambas et al., 2020). In the category of human disease, 18 pathways are related to infectious diseases and 9 are correlated with drug resistance. Moreover, a lot of pathways are associated with signaling transduction, cell growth, and death. Figure 7C exhibits the top 20 enriched KEGG pathways with the highest q values. Accumulating pieces of evidence have demonstrated that these pathways are interconnected with the anti-influenza effect of drugs. For instance, the hypoxia inducible factor-1 (HIF-1) signaling pathway (ko04066) was one of the significantly enriched pathways with a q value of 6E−6. HIF-1α is an important factor in the development and repair of acute lung injury and could regulate glycolysis and AMPKα-ULK1-mediated autophagy, which ultimately affects IAV replication (Zhao et al., 2020). In addition, IAV infection is a complicated process, and its outcome is largely determined by the extensive release of chemokines and pro-inflammatory cytokines after its spread to the lung. Obviously, the chemokine signaling pathway (ko04062, q = 2E−6) plays an important role in this process (Zhang et al., 2020). A recent in vitro study has shown that the polysaccharides of BLG significantly decreased the expressions of chemokines stimulated by A/PR/8/34 (H1N1), including IP-10, MIG, and CCL-5 (Li et al., 2017). Taken together, the gene enrichment analysis highlighted the influenza-related molecular functions and biological pathways that merit attention, helping to understand the potential anti-IAV mechanisms of BLG.