Abstract

   Parkinson’s disease (PD) is the second most common neurodegenerative
   disease with a fast-growing prevalence. Developing disease-modifying
   therapies for PD remains an enormous challenge. Current drug treatment
   will lose efficacy and bring about severe side effects as the disease
   progresses. Extracts from Ginkgo biloba folium (GBE) have been shown
   neuroprotective in PD models. However, the complex GBE extracts
   intertwingled with complicated PD targets hinder further drug
   development. In this study, we have pioneered using single-nuclei RNA
   sequencing data in network pharmacology analysis. Furthermore,
   high-throughput screening for potent drug-target interaction (DTI) was
   conducted with a deep learning algorithm, DeepPurpose. The strongest
   DTIs between ginkgolides and MAPK14 were further validated by molecular
   docking. This work should help advance the network pharmacology
   analysis procedure to tackle the limitation of conventional research.
   Meanwhile, these results should contribute to a better understanding of
   the complicated mechanisms of GBE in treating PD and lay the
   theoretical ground for future drug development in PD.

   Keywords: Parkinson’s disease, Ginkgo biloba folium, network
   pharmacology, deep learning, single-nuclei RNA sequencing

Introduction

   Parkinson’s disease (PD) is mainly manifested by progressive motor
   impairment ([33]Armstrong and Okun, 2020), leading to severe damage to
   the everyday lifestyle of 6.1 million patients worldwide ([34]Shandilya
   et al., 2022). Prevalence and disability-adjusted life years (DALYs) of
   PD have been increasing in recent decades ([35]Collaborators, 2019),
   causing an enormous burden on the medical system and economy
   ([36]Rocca, 2018).

   The development of novel PD therapeutics is in urgent demand. Without
   available disease-modifying therapy, current treatments for PD are only
   symptomatic ([37]Vijiaratnam et al., 2021). Long-term symptomatic
   therapy brings about adverse events, such as dyskinesia and impulse
   control disorders ([38]Voon et al., 2017).

   Herbal medicines, including extracts from Ginkgo biloba extract (GBE),
   have gradually come to attention as novel therapies for PD. Herbal
   medicines generally share the advantages of multilevel functions with
   fewer adverse effects ([39]Yin et al., 2021). Among the most frequently
   applied herbal medicines, GBE has been used in clinical therapies since
   the early 1970s ([40]Saponaro et al., 1971; [41]Bartolo, 1973). One of
   the most investigated applications of GBE is in treating
   neurodegenerations, represented by Alzheimer’s disease and mild
   cognitive impairment ([42]Singh et al., 2019; [43]Nowak et al., 2021;
   [44]Tomino et al., 2021). Various studies have also validated that a
   mixture of GBE exerts neuroprotective function on both in vivo and in
   vitro PD models, including toxin-induced PD models on rats ([45]Yu et
   al., 2021), toxin-induced PD mice ([46]Rojas et al., 2009; [47]Rojas et
   al., 2012), transgenic PD mice ([48]Kuang et al., 2018), and in vitro
   cultured cell models ([49]Yang et al., 2001; [50]Kang et al., 2007).
   Subsequent research is hindered by the mixture nature of GBE and its
   multi-target effects. A complicated extraction procedure is required to
   obtain bioactive components from GBE ([51]Liu et al., 2022; [52]Ma et
   al., 2022), and the procedure is still ongoing improvement
   ([53]Boateng, 2022). The variability of extracts results in difficulty
   in repeating results across studies, thereby hampering the exploration
   of molecular mechanisms. Delineating the effects of a single active
   component in GBE would contribute to proposing feasible targets and
   aiding future drug development for PD.

   Herein, advances in analytical pharmacy and bioinformatics would help
   to detangle the complex molecular mechanisms underlying the therapeutic
   efficacy of GBE in PD. The single-nuclei RNA sequencing (snRNA-seq) has
   emerged as a powerful tool for identifying and characterizing cell
   types, states, and lineages ([54]Slyper et al., 2020). Recently, the
   snRNA-seq approach was conducted to analyze the transcriptome in
   midbrains of PD patients ([55]Smajić et al., 2022). Therefore, we took
   the unprecedented chance to investigate GBE effects in a
   cell-type-specific manner. We intended to focus on microglia and
   astrocytes in addition to neurons when studying the effects of GBE.
   Since PD is attributed to a selective loss of dopaminergic neurons in
   the substantia nigra. Meanwhile, mounting clinical and experimental
   evidence illuminated that glial cells, especially microglia and
   astrocytes, were not only responders but also significant mediators in
   PD pathogenesis ([56]Sorrentino et al., 2019; [57]Bartels et al.,
   2020). Thus, modulating microglia and astrocytes functions is a
   promising pharmacological strategy for treating PD ([58]Grotemeyer et
   al., 2022; [59]Lee et al., 2022).

   The deep learning approach can be another handy tool to guide
   pharmacological studies, including drug-target prediction, drug
   repurposing, and novel drug discovery ([60]Zhavoronkov et al., 2019;
   [61]Issa et al., 2021; [62]Zhu et al., 2021). Experimental measurement
   of the compound–protein binding affinity remains the most accurate
   method for studying drug-target interactions. However, conventional
   methods are costly, time-consuming, and laborious, which are infeasible
   for investigating the multifarious drug-target interactions (DTI)
   between complex GBE ingredients and numerous PD targets. Therefore,
   deep learning has been used to conduct high throughput DTI analyses,
   which could help to screen out potent DTI between GBE ingredients and
   PD-related bio-targets.

   In this study, we tended to identify active components in GBE for PD
   along with its cell-type-specific targets. Network pharmacology
   analysis was conducted, integrating data from snRNA-seq and existing
   drug datasets. A cell-type-specific compound-target-pathway network was
   established, and DTI was subsequently investigated with a deep learning
   algorithm. Then, we validated the results by molecular docking. This
   research will contribute to a better understanding of the molecular
   mechanisms of treating PD with GBE.

Methods

Collecting and selecting compounds in GBE

   Firstly, components of GBE were collected via searching the terms:
   “ginkgo folium,” “folium ginkgo,” and “Yinxingye” in databases. TCMSP
   ([63]Ru et al., 2014) ([64]https://old.tcmsp-e.com/index.php, version
   2.3), TCMID ([65]Huang et al., 2018) ([66]http://bidd.group/TCMID/,
   version 2.0) and SymMap ([67]Wu et al., 2019)
   ([68]http://www.symmap.org/, version 2.0) databases rendered 307, 94
   and 319 ingredients of GBE, respectively. All data were collected on 18
   May 2022.

   Secondly, PubChem CID was retrieved from PubChem
   ([69]https://pubchem.ncbi.nlm.nih.gov/) to identify each component.

   We also searched the PubMed database with the following terms " (Ginkgo
   biloba leaf OR Ginkgo biloba folium) AND (components OR ingredients OR
   metabolite)" and added ginkgolide K, which was not timely updated in
   databases ([70]Li et al., 2018a).

   Thirdly, chemical properties and pharmacokinetic profiles of components
   were retrieved. The chemical properties of components were annotated
   via SwissADME ([71]http://www.swissadme.ch) ([72]Daina et al., 2017),
   which provided information on molecular weight, lipophilicity (log P
   [o/w]), number of H-bond acceptors, number of H-bond acceptors, number
   of rotatable bonds, and topological polar surface area (TPSA). ADMETlab
   2.0 ([73]Xiong et al., 2021) ([74]https://admetmesh.scbdd.com/) was
   employed to evaluate compound pharmacokinetics and toxicity. To assess
   components’ oral bioactivity, HobPre ([75]www.icdrug.com/ICDrug/ADMET)
   ([76]Wei et al., 2022), a classification model, was exploited. The
   ADMET profiles of components were obtained from pkCSM ([77]Pires et
   al., 2015) ([78]http://structure.bioc.cam.ac.uk/pkcsm).

   With all the above data collected, Lipinski’s Rule ([79]Lipinski et
   al., 2001) was subjected to assess the draggability of collected
   compounds. A total of 25 selected compounds were selected and listed in
   [80]Supplementary Table S1. These compounds met the following criteria:
   molecular weight of fewer than 500 Da; log P [o/w] lower than five and
   higher than −2; five or fewer hydrogen bond donor sites and tenor fewer
   hydrogen bond acceptor sites; the number of rotatable bonds less than
   10.

Acquiring potential molecular targets of GBE components

   Every selected component has been searched in SymMap ([81]Wu et al.,
   2019) ([82]http://www.symmap.org/, version 2.0) database, and 272
   potential molecular targets were retrieved. SymMap database integrates
   target information from HIT ([83]Ye et al., 2011)
   ([84]http://lifecenter.sgst.cn/hit/), TCMSP, HPO ([85]Köhler et al.,
   2021) ([86]https://hpo.jax.org/app/), DrugBank ([87]Wishart et al.,
   2018) ([88]https://go.drugbank.com/),
   NCBI([89]https://www.ncbi.nlm.nih.gov/) and HERB ([90]Fang et al.,
   2020a) ([91]http://herb.ac.cn/) databases. Additional pharmacoproteomic
   and pharmaco-transcriptomic data were obtained manually. Additional
   ginkgolide J, ginkgolide M, and ginkgolide K targets data, which is not
   included in the above databases, was retrieved from the Comparative
   Toxicogenomics Database (CTD) ([92]Davis et al., 2017) (URL:
   [93]http://ctdbase.org/). All data were collected on 18 May 2022. After
   removing duplicates, 283 genes were identified as putative GBE targets
   for PD.

Acquiring PD-related-targets in different cell types from single-nuclei RNA
sequencing data

   Gene expression profile of different cell types from the idiopathic
   Parkinson’s disease patient’s brain snRNA-seq ([94]Smajić et al., 2022)
   ([95]GSE157783) was used to identify the disease-related targets in
   this study. Cell-type-specific genes were identified using the
   Quasi-Poisson generalized linear model implemented in the fit models
   function of the R package monocle3 (version 1.0.0) ([96]Trapnell et
   al., 2014). The cutoff q coefficient was set at 0.05 to obtain
   differentially expressed genes (DEGs) in each cell type. The potential
   targets were identified by overlapping genes of GBE targets and DEGs in
   different cell types of PD. Intersections were visualized with R
   package VennDiagram (version 1.7.3) ([97]Chen and Boutros, 2011).

PPI networks construction

   Protein-protein interaction (PPI) network of all targets was
   constructed using Cytoscape software (version 3.9.1) with data from
   STRING ([98]Szklarczyk et al., 2015) (version 10.0) database. The
   confidence score cutoff was set at 0.4.

GO and KEGG pathway enrichment analysis

   R package topGO (version 2.46.0) and cluster profile (version 4.2.2)
   was employed to conduct Gene Ontology (GO) and KEGG pathway analysis.
   Reference gene data were retrieved using R package, org. Hs.eg.db
   (version 3.14.0). The p-value cutoff was set at 0.05, and the q-value
   cutoff was set at 0.01 for all analyses. Top clusters from GO and KEGG
   enrichment were visualized using R package ggplot2 (version 3.3.5) and
   enrichplot (version 1.14.2). All mentioned analysis was conducted on R
   version 4.1.2.

Drug-target interaction (DTI) prediction with DeepPurpose

   Pre-trained model CNN_CNN_BindingDB provided by DeepPurpose ([99]Huang
   et al., 2020) ([100]https://github.com/kexinhuang12345/DeepPurpose) was
   used to calculate the binding score between selected targets and their
   proven ligands. In this pre-trained model, Convolutional Neural Network
   (CNN) was chosen to encode SMILES of components and the amino acid
   sequence. The Binding Database (BindingDB), a public drug-target
   binding benchmark dataset, was employed to provide measured binding
   affinities. DeepPurpose generates predictions via a Multi-Layer
   Perceptron (MLP), one of the most common artificial neural networks.
   All amino acid sequences of the selected targets were collected from
   UniProt ([101]Consortium, 2020) ([102]https://www.uniprot.org/). The
   SMILES of each component were obtained from the PubChem database
   ([103]https://pubchem.ncbi.nlm.nih.gov/).

Molecular docking

   Molecular docking was performed using the SwissDock ([104]Grosdidier et
   al., 2011) server ([105]http://www.swissdock.ch/). 3D structure of
   MAPK14 protein was obtained from RCSB PDB ([106]https://www.rcsb.org/)
   with PDB ID: 1WBS. The chemical structure of ginkgolide J and
   ginkgolide A was obtained from the PubChem database
   ([107]https://pubchem.ncbi.nlm.nih.gov/). The DockPrep plugin of
   Chimera (version 1.16, build 42,360) was employed to prepare the
   structures before docking. Docking results were analyzed and visualized
   using UCSF Chimera (version 1.16, build 42,360) and LigPlot
   ([108]Laskowski and Swindells, 2011) (version 2.2.5).

Results

Potential active components and related targets of GBE

   Chemical components of GBE were searched and collected from TCMSP,
   TCMID, and SymMap databases and manually checked references from