Abstract Background and objective The multi-targets and multi-components of Traditional Chinese medicine (TCM) coincide with the complex pathogenesis of depression. Zhi-Zi-Hou-Pu Decoction (ZZHPD) has been approved in clinical medication with good antidepression effects for centuries, while the mechanisms under the iceberg haven't been addressed systematically. This study explored its inner active ingredients - potent pharmacological mechanism - DDI to explore more comprehensively and deeply understanding of the complicated TCM in treatment. Methods This research utilized network pharmacology combined with molecular docking to identify pharmacological targets and molecular interactions between ZZHPD and depression. Verification of major active compounds was conducted through UPLC-Q-TOF-MS/MS and assays on LPS-induced neuroblastoma cells. Additionally, the DDIMDL model, a deep learning-based approach, was used to predict DDIs, focusing on serum concentration, metabolism, effectiveness, and adverse reactions. Results The antidepressant mechanisms of ZZHPD involve the serotonergic synapse, neuroactive ligand-receptor interaction, and dopaminergic synapse signaling pathways. Eighteen active compounds were identified, with honokiol and eriocitrin significantly modulating neuronal inflammation and promoting differentiation of neuroimmune cells through genes like COMT, PI3KCA, PTPN11, and MAPK1. DDI predictions indicated that eriocitrin's serum concentration increases when combined with hesperidin, while hesperetin's metabolism decreases with certain flavonoids. These findings provide crucial insights into the nervous system's effectiveness and potential cardiovascular or nervous system adverse reactions from core compound combinations. Conclusions This study provides insights into the TCM interpretation, drug compatibility or combined medication for further clinical application or potential drug pairs with a cost-effective method of integrated network pharmacology and deep learning. Keywords: ZZHPD, Depression, Deep learning, Network pharmacology, Drug-drug interactions Highlights * • Integration of network pharmacology and deep learning uncovers the intricate interactions among diverse compounds in ZZHPD. * • The antidepressive mechanisms and key compounds of ZZHPD were identified and validated through in vitro analysis. * • This study proposes a cost-effective approach for developing synergistic drug combinations for clinical use. 1. Introduction Depression, the most common and severe mental disease, has been declared to be the principal cause of death by 2030 and imposes a heavy burden on human society [[45]1]. Single-target therapy and conventional medications prove inadequate for this multifactorial syndrome, owing to limited efficacy, numerous adverse reactions, and treatment resistance [[46]2]. There is an urgent and fundamental need to screen for new treatments with multigenetic effects and enhanced safety. Combinational therapy is becoming a powerful treatment strategy for complex diseases in clinical due to its advantage of synergistic or additive effect. Traditional Chinese medicine (TCM), historically used for depression healthcare, has gradually appear as a new and powerful treatment candidate with fewer side effects [[47]3,[48]4]. The medication philosophy of TCM is well consistent with the therapeutic idea of systems medicine for the complex depression treatment. It has the potential to decode the overall synergistic effects with multiple drugs on multiple targets, such as hippocampal neurons, neurotrophic factors, monoamine neurotransmitters, hypothalamic-pituitary-adrenal axis hyperactivity et al. [[49]5,[50]6]. Zhi-Zi-Hou-Po decoction (ZZHPD) has achieved reliable efficOacy with less side effects, ascribed to its composition of Gardenia jasminoides Ellis (ZZ), Citrus aurantium L. (ZS) and Magnolia officinalis Rehd. et Wils. (HP) [[51]7]. It could alleviate unpredictable, chronic, mild stress-induced depressant symptoms by improving the monoaminergic system, promoting hippocampal neurogenesis, restoring hypothalamic-pituitary-adrenal (HPA) axis function and increasing brain derived neurotrophic factor (BDNF) expression [[52]8]. ZZ showed quite fast antidepressant functions on CMS mice associated with BDNF signal transduction [[53]9]. The combination of magnolol and honokiol exhibits strong antidepressant-like effects, normalizing biochemical abnormalities of brain 5-HT and 5-HIAA in vivo [[54]10]. ZZHPD has demonstrated effectiveness in treating depression, benefiting significantly from the synergistic interaction beyond the individual effects of each herb. The myriad compounds in TCM may lead to drug-drug interactions (DDI), including pharmacological effects and unexpected adverse reactions. Multiple drugs or pairwise combinations enhance clinical outcomes for various complex diseases by synergistically targeting multiple disease pathways, either lowering the dosage for higher effectiveness or reducing side effects [[55]10]. Beneficial effects and adverse DDIs are closely linked to common biological targets/pathways or heterogeneous proteins across diverse diseases. Therefore, addressing fundamental issues at the systemic level, derived from the molecular level, is crucial to emphasize complex herbal formulas and devise novel therapeutic strategies for depressive patients [[56]10]. However, it is costly, infeasible, and challenging in practice to identify various DDIs and synergistic combinations through in vivo and in vitro biological tests. Significant computational approaches in the pharmacological and bioinformatics domains offer a promising tool for prioritizing pharmacotherapies [[57]11]. Network pharmacology based on computer science offers a systematic strategy for drug research to elucidate actions and interactions on multitargets, which is widely applied in the pharmacological research of TCM [[58]12,[59]13]. This system currently faces numerous challenges, particularly in effective data mining from massive heterogeneous datasets, including drug targets, pharmacological mechanisms, drug-organism/cell interactions, and multidrug treatments. Computational algorithms, especially deep learning (DL), can aid in analyzing vast amounts of information for decision-making and overcoming bottlenecks in complex DDI prediction or multitarget drug discovery, thus facilitating all stages of network pharmacology research [[60]14]. DeepDDI, a deep neural network, utilizes the names and structural information of drug-drug and drug-food constituent pairs as inputs to accurately generate 86 DDI types as the prediction output [[61]15]. Lee et al. introduced a deep feed-forward network with an autoencoder for predicting the pharmacological effects of DDIs, trained using the similarity profiles of structure, target gene, Gene Ontology term, and target gene of existing drug pairs [[62]16]. DDIMDL uses four drug features (chemical substructures, targets, enzymes, and pathways) in a separately or logically combined way to predict DDI-associated events with highly accurate and highly efficient performances [[63]17]. Identifying drug combinations with high synergistic efficacy and minimized adverse DDIs remains a critically important yet challenging task for clinical indications, drug discovery, and medication strategies. Therefore, this study aimed to integrate deep learning with network pharmacology to systematically elucidate the antidepressant mechanism and the internal interactions within ZZHPD ([64]Fig. 1). The primary contributions of this paper are outlined as follows: (1) Employing network pharmacology and UPLC-Q-TOF-MS/MS to screen and identify the core ingredients, respectively. (2) Conducting in vitro experimental research to analyze the gene expression of effective substances and key targets as predicted by network pharmacology. (3) DDIMDL is employed to predict six DDI types between 18 core antidepressant components in ZZHPD, including metabolism, serum concentration, therapeutic effect, nervous system effectiveness, cardiovascular adverse reactions and nervous system adverse reaction. Fig. 1. [65]Fig. 1 [66]Open in a new tab The integration framework of network pharmacology and deep learning. 2. Material and methods 2.1. Network construction-analysis and molecular docking 2.1.1. Active compound screening and potential targets prediction The active compounds in ZZHPD were collected from related literatures, and BATMAN-TCM [[67]18]. The ingredients in the BATMAN-TCM platform with two conditions (Score cutoff no less than 30, Adjusted P-value ≤0.05) were chosen as candidate components in ZZHPD for further analysis. The BATMAN-TCM database also provided corresponding target information for each active compound. Then, SEA [[68]19], PharmMapper [[69]20], and SwissTargetPrediction [[70]21] supplement the predicted targets based on the chemical structures of the bioactive constituents. The chemical structures of ZZHPD were obtained from PubChem. 2.1.2. Acquisition depression-associated targets DrugBank [[71]22], GeneCards [[72]23], TTD [[73]24], PharmGkb [[74]25], DisGeNET [[75]26], and OMIM [[76]27] were adopted to screen targets from the keywords of “depression”, “major depression”, “depressant”, “antidepressant”, “depressed”, “depressive”, “depressive disorder”, “major depressive disorder” and “depressing”. Aiming to standardize names, the protein names of all integrating targets from six different databases were turned into official gene symbols through the UniprotKB [[77]28] database with the organism limited to Homo sapiens. Afterwards, the latent targets of ZZHPD were obtained by overlapping the targets of active ingredients and antidepressant-related targets. 2.1.3. GO function enrichment and KEGG pathway enrichment analysis The intersecting antidepressant targets of ZZHPD were performed and conducted with R Bioconductor package to assess Gene Ontology (GO) [[78]29] functions and pathway enrichment with Kyoto Encyclopedia of Genes and Genomes (KEGG) [[79]30], in which the screening criteria is set as p-value cutoff = 0.05. 2.1.4. Construction of Chinese herbs-compounds-targets network The Chinese medicines-compounds-targets network had established with Cytoscape software (version 3.8.0) to elucidate the associated mechanisms between the bioactive components and the target protein [[80]31]. Calculation was carried out on the important network topology parameters of the compounds and related targets, such as the degree, closeness centrality, and betweenness centrality. The width of edges, the links between nodes, represents the strength of intermolecular interactions. 2.1.5. Protein-protein interaction network integration and the key genes identification Selected as Homo sapiens, 251 overlapping targets of ZZHPD for depression were input to the STRING [[81]32] platform to conduct the protein interaction network with the score greater than 0.95. The extracted protein-protein interaction (PPI) network in a CSV format was imported into Cytoscape to visualize the network. The molecular complex detection algorithm (MCODE) [[82]33], a small plugin of Cytoscape applications, was used to analyze the features of densely connected PPI network and obtain the network clusters of the module. 2.1.6. Molecular docking Prior to docking simulation, the two-dimensional (2D) structures of protein receptor were discovered in the protein crystal structure database PDB [[83]34] and the mol2 format of candidate components identified previously were obtained from PubChem. Then, the 2D protein receptor files were processed and converted to three-dimensional (3D) chemical structure using Chem3D. The processed target protein is hydrogenated and charged in the AutoDock Tools 1.5.6 software [[84]35], following the use of PyMOL software to remove the excess inactive ligands, such as water molecules and phosphate radicals in the target protein [[85]36]. All files were imported into AutoDock Vina 1.1.2 to calculate the binding free energy of the active compounds in the target protein structure [[86]37]. 2.2. Drug-drug interaction prediction 2.2.1. Deep learning model overview DDIMDL, a deep learning model, is proposed by Deng to predict five DDI types [[87]17]. The ECFP4 fingerprints were regarded as an input layer to the model to predict the binary classification activity of DDI. This classifier added batch normalization and dropout in every hidden layer and used a softmax activation function for binary activity as the last layer and the rectified linear unit (ReLU) to activate input and hidden layers. The EarlyStopping strategy was adopted in this paper to prevent overfitting and adaptive moment estimation (Adam) was used as an optimizer for all the experiments. 2.2.2. Molecular representation and data balance Extended Connectivity Fingerprints 4 (ECFP4), belonging to the extended connectivity fingerprint family, is also called Morgan fingerprints or a circular fingerprint. It is an approach that the SMILES strings of each drug were encoded into 2048-dimensional binary vector implemented by a Python package RDKit ([88]www.rdkit.org). As for the data imbalance problems, we applied the synthetic minority oversampling technique (SMOTE) to address them. SMOTE was conducted through random replication of samples of minority classes using the Python library imblearn. 2.2.3. Datasets Five DDI types need to be predicted: (1) Metabolism; (2) Serum Concentration; (3) Nervous System Effectiveness; (4) Cardiovascular Adverse Reactions; (5) Nervous System Adverse Reactions. Three datasets, namely DDI-sorted, DDI-antidp, and DDI-zzhpd, are categorized according to distinct prediction tasks as shown in [89]Table 1. Based on the type of label, the relevant data of the above five DDI types are sorted from the Shenggeng's dataset and termed DDI-sorted ([90]Supplementary Table S1). From the DDI-sorted dataset, 42 antidepressants and their relationships were selected as test sets for model performance verification tasks and named DDI-antidp ([91]Supplementary Table S2). Network pharmacology and UPLC-Q-TOF-MS/MS screened and identified 18 core compounds as predictive drugs and named DDI-zzhpd ([92]Supplementary Table S3). Table 1. The detailed information about three datasets. Dataset Relation types Drug number DDI number Label DDI- sorted __________________________________________________________________ Metabolism 1075 127234 YES Serum Concentration 1108 20795 Nervous System Effectiveness 686 14885 Cardiovascular Adverse Reactions 672 34945 Nervous System Adverse Reactions __________________________________________________________________ 593 __________________________________________________________________ 7693 __________________________________________________________________ DDI-antidp __________________________________________________________________ Metabolism 33 214 YES Serum Concentration 27 33 Nervous System Effectiveness 38 131 Cardiovascular Adverse Reactions 29 62 Nervous System Adverse Reactions __________________________________________________________________ 33 __________________________________________________________________ 66 __________________________________________________________________ DDI-zzhpd Metabolism 18 306 NO Serum Concentration 18 306 Nervous System Effectiveness Cardiovascular Adverse Reactions Nervous System Adverse Reactions [93]Open in a new tab 2.2.4. Prediction tasks description In this study, we have employed DDIMDL to the three-step DDI prediction of ZZHPD, which takes structural information encoded as two-dimensional vectors of two drugs as input, and predicts DDI types as an output that are human-readable sentences. A three-step procedure was carried out for DDI prediction task of ZZHPD by DDIMDL model as follows ([94]Fig. 2): (i) Step 1 was named as prediction interaction among the known drugs. Our dataset, consisting of drug pairs and their interaction types, was divided into training (80 %) and testing (20 %) sets. The DDIMDL model was then trained on the structural information of drug pairs, encoded as two-dimensional vectors. (ii) For this validation phase Step 2 of prediction interaction between the new drugs, we constructed a test set exclusively comprising 42 anti-depressive drugs, ensuring the training set did not include any interactions involving these drugs. (iii) During the Step 3 of prediction interaction among core components in ZZHPD, we focused on predicting DDIs among 18 core components of ZZHPD leveraging the trained DDIMDL model. The test set, named DDI-zzhpd, comprised 153 drug pairs across different DDI types without included labels. Fig. 2. [95]Fig. 2 [96]Open in a new tab Three steps for predicting DDI in ZZHPD. The orange, green and purple nodes represent the existing drugs, antidepressants, and core compounds in ZZHPD, respectively. Solid lines are known interactions and dashed lines are the relationships waiting for prediction. (For interpretation of the references to colour in this figure legend, the reader is