Developing a Knowledge Graph for Pharmacokinetic Natural Product-Drug Interactions

Abstract

Background:

Pharmacokinetic natural product-drug interactions (NPDIs) occur when botanical or other natural products are co-consumed with pharmaceutical drugs. With the growing use of natural products, the risk for potential NPDIs and consequent adverse events has increased. Understanding mechanisms of NPDIs is key to preventing or minimizing adverse events. Although biomedical knowledge graphs (KGs) have been widely used for drug-drug interaction applications, computational investigation of NPDIs is novel. We constructed NP-KG as a first step toward computational discovery of plausible mechanistic explanations for pharmacokinetic NPDIs that can be used to guide scientific research.

Methods:

We developed a large-scale, heterogeneous KG with biomedical ontologies, linked data, and full texts of the scientific literature. To construct the KG, biomedical ontologies and drug databases were integrated with the Phenotype Knowledge Translator framework. The semantic relation extraction systems, SemRep and Integrated Network and Dynamic Reasoning Assembler, were used to extract semantic predications (subject-relation-object triples) from full texts of the scientific literature related to the exemplar natural products green tea and kratom. A literature-based graph constructed from the predications was integrated into the ontology-grounded KG to create NP-KG. NP-KG was evaluated with case studies of pharmacokinetic green tea- and kratom-drug interactions through KG path searches and meta-path discovery to determine congruent and contradictory information in NP-KG compared to ground truth data. We also conducted an error analysis to identify knowledge gaps and incorrect predications in the KG.

Results:

The fully integrated NP-KG consisted of 745,512 nodes and 7,249,576 edges. Evaluation of NP-KG resulted in congruent (38.98% for green tea, 50% for kratom), contradictory (15.25% for green tea, 21.43% for kratom), and both congruent and contradictory (15.25% for green tea, 21.43% for kratom) information compared to ground truth data. Potential pharmacokinetic mechanisms for several purported NPDIs, including the green tea-raloxifene, green tea-nadolol, kratom-midazolam, kratom-quetiapine, and kratom-venlafaxine interactions were congruent with the published literature.

Conclusion:

NP-KG is the first KG to integrate biomedical ontologies with full texts of the scientific literature focused on natural products. We demonstrate the application of NP-KG to identify known pharmacokinetic interactions between natural products and pharmaceutical drugs mediated by drug metabolizing enzymes and transporters. Future work will incorporate context, contradiction analysis, and embedding-based methods to enrich NP-KG. NP-KG is publicly available at https://doi.org/10.5281/zenodo.6814507. The code for relation extraction, KG construction, and hypothesis generation is available at https://github.com/sanyabt/np-kg.

Graphical Abstract

1. Introduction

Complementary health approaches involving the use of botanical and other natural products have increased substantively since passage of the Dietary Supplement Health and Education Act in 1994. Up to 18% of adults reported regular use of natural products in the United States (US) [1], and annual sales of herbal dietary supplements in the US increased again, by 9.7% in 2021 [2]. Although not intended to replace pharmaceutical drugs, co-consuming natural products with drugs is common [3,4]. Older adults are the largest consumers of both pharmaceutical drugs and natural products, with up to 88% in the US reporting concomitant use [5]. However, such concomitant use may result in pharmacokinetic natural product-drug interactions (NPDIs) and potentially, unexpected drug responses [6]. Although substantial research has been devoted to understanding mechanisms and clinical effects of drug-drug interactions, corresponding information about NPDIs remains lacking. With the safety concerns related to the growing use of natural products, understanding the mechanisms underlying their interactions with other xenobiotics is imperative to prevent or minimize potential adverse events.

Pharmacokinetic NPDIs occur when a natural product alters the absorption, distribution, metabolism, and/or excretion of a co-consumed drug, potentially resulting in reduced treatment efficacy or adverse events [7]. For example, the popular botanical product green tea (Camellia sinensis), which is available as a beverage and dietary supplement, precipitated an interaction with the beta-blocker and anti-hypertensive agent nadolol, possibly by inhibiting intestinal human organic anion transporting polypeptide(s) (OATPs), uptake transporters that facilitate absorption of the substrate from the gut lumen into the systemic circulation [8]. This pharmacokinetic interaction in turn led to a decrease in the blood pressure lowering effect of nadolol. Although this study focused on potential adverse events resulting from an NPDI, beneficial effects are also possible. For example, piperine, a constituent of black pepper (Piper nigrum L.), can increase the absorption of co-administered xenobiotics by inhibiting intestinal cytochrome P450 (CYP) 3A4, leading to an increase in xenobiotic oral bioavailability [9].

Understanding biochemical mechanisms underlying clinically significant pharmacokinetic NPDIs can help prevent or minimize adverse events [4]. Similar to drug-drug interactions, computational approaches to elucidate potential mechanisms can help answer the ‘why’ question; that is, why the NPDI occurs and through what mechanism(s) [10]. Prior work in computational discovery of potential NPDIs has focused on mining scientific abstracts to classify NPDIs involving dietary supplements [11] and developing a knowledge graph (KG) to identify plausible drug-supplement interactions [12]. Using scientific abstracts to classify supplement-drug interactions has shown promise; however, the model developed by Wang et. al. [11] was trained using a drug-drug interaction dataset due to lack of labeled NPDI datasets. Schutte et. al. [12] extracted knowledge from scientific abstracts to develop a literature-based KG and evaluated the graph with discovery patterns followed by a clinical review. Computational prediction has also utilized homogeneous graph structure constructed from existing resources for chemicals and foods [13]. Dietary supplements sold in the US represent a broad group of products, including botanicals, which are regulated differently than pharmaceutical drugs [14]. The US Food and Drug Administration (FDA) can prohibit the sale of these products only if they are found to be unsafe; they do not approve them prior to marketing.

Following a similar workflow used during drug development to identify pharmacokinetic drug-drug interactions [15], potential pharmacokinetic NPDIs can be evaluated by first conducting in vitro studies to determine if a natural product extract (e.g., some quantity of green tea) phytoconstituent (e.g., catechin) inhibits or induces the function of a drug metabolizing enzyme or transporter, which may or may not have unforeseen negative consequences. Results are then used to quantitatively predict the clinical impact of the interaction using in vitro to in vivo extrapolation (IVIVE). Potential clinically relevant pharmacokinetic NPDIs suggested by IVIVE are then evaluated via dynamic (e.g., physiologically based pharmacokinetic) modeling and simulation and/or human clinical studies [4].

Although effective, in vitro experiments followed by IVIVE often predict NPDIs that are not confirmed by more time-consuming and costly clinical studies. For example, a mechanistic static model predicted green tea to increase the systemic exposure to the selective estrogen receptor modulator raloxifene by inhibiting intestinal uridine 5’-diphospho-glucuronosyltransferase (UDP-glucuronosyltransferase, UGT), but the opposite was observed when tested in healthy volunteers [16,17]. Although green tea is promoted for weight loss, mental alertness, relieving headaches and digestive symptoms, and cardio protection, the popularity and widespread availability of green tea products, as well as the likelihood of co-consumption with pharmaceutical drugs increases the risk of potential clinically significant green tea-drug interactions [16].

Developing novel methods that help scientists make accurate and timely NPDI predictions is crucial to reduce the number of patients who experience adverse events from NPDIs. We hypothesized that a KG built using rigorous biomedical ontologies and enriched with domain-specific information from the scientific literature will generate plausible mechanistic explanations for pharmacokinetic NPDIs that can be used to guide robust NPDI research. In this study, we designed and built a novel KG, termed NP-KG, as a first step towards testing this hypothesis. We evaluated whether NP-KG accurately represented known pharmacokinetic mechanisms for NPDIs involving two exemplar natural products: green tea and kratom. Green tea was selected due to its worldwide use and the published results related to NPDIs discussed above. Kratom (Mitragyna speciosa) is an emerging botanical product commonly used to self-treat pain, anxiety, and opioid withdrawal symptoms. Kratom leaves contain numerous alkaloids that produce stimulant effects at low doses and opioid-like effects at higher doses. However, safety concerns have been raised by various federal agencies related to kratom toxicity and potential NPDIs [18–21]. We further demonstrated the potential of NP-KG to generate hypotheses for mechanisms underlying pharmacokinetic NPDIs related to kratom reported in published case reports.

2. Materials and Methods

A biomedical KG combines concepts from different domains to represent entities such as molecules, genes, diseases, and chemical substances as nodes in the graph. Edges in the KG represent relations between the entities. We describe the construction of the ontology-grounded KG, the process of semantic relation extraction from full texts of the literature, and integration into NP-KG (Figure 1). Semantic relation extraction produces predications (subject-relation-object triples) from texts, where the subjects and objects can form the nodes in the KG, and relations form edges between them. Lastly, we discuss the strategies used to evaluate NP-KG.

Figure 1.

Overview of NP-KG that combines biomedical ontologies, drug databases (FDA Drug Interaction Data, Drug Central, Drug Interaction Knowledge Base), natural products, and literature-based graph constructed from green tea and kratom publications.

2.1. Data Integration

An ontology is a formal representation of knowledge in a domain of discourse. Ontologies are used to formally structure the classes and relations within a domain, where classes refer to sets of entities in the world and relations are properties that hold for two or more entities [22]. For example, the Chemical Entities of Biological Interest (ChEBI) is a biomedical ontology representing molecular entities, focusing on chemical compounds and relations between the entities [23]. ‘Catechin’ (ChEBI:23053) is an example of a class in ChEBI ontology and has a unique identifier, name, and definition. Biomedical ontologies represent our current understanding of biological reality and can be used to represent knowledge as knowledge statements. The Open Biological and Biomedical Ontology (OBO) Foundry is a family of interoperable ontologies in the biomedical domain following a set of principles that include open use, non-overlapping content, and common syntax, as well as relations [24]. As ontologies contain entities in a domain and relations between the entities, they can be represented as graphs, where nodes of the graph represent classes or instances and edges represent an axiom or relation involving the connected classes or instances [22].

We used the Phenotype Knowledge Translator (PheKnowLator) workflow to construct an ontology-grounded KG consisting of the OBO Foundry ontologies and linked data sources. PheKnowLator (v3.0.0) is a Python library that constructs large-scale, biomedical KGs by semantically integrating biomedical ontologies and complex heterogeneous data [25]. Entities from the OBO Foundry ontologies (diseases from Mondo Disease Ontology [26]; phenotypes from Human Phenotype Ontology [27]; anatomical entities from Uber Anatomy Ontology [28]; biological processes, cellular components, and molecular functions from Gene Ontology [29]; proteins from Human Protein Ontology [30]; pathways from Pathway Ontology [31]; chemicals from ChEBI ontology [23]; genes and variants from Sequence Ontology [32]; and cells from Cell Ontology [33] and Cell Line Ontology [34]) and linked data sources were integrated in the KG with the PheKnowLator workflow. The workflow was implemented in three steps: Ontology Merging, Data Preparation, and KG construction. The Ontology Merging step identified potential errors in the ontology files (such as identifier, entity normalization, and punning errors), handled duplicate terms, removed deprecated classes, and merged the ontologies. The Data Preparation step downloaded and processed all data sources to build the edges for the KG as well as mapped identifiers between data sources for each edge type. Finally, the KG construction step created node and edge lists from the ontology and data sources in the prior steps, added node and edge data and attributes to the KG, and output the KG files with triples.

In addition to the ontologies and data sources originally included in the PheKnowLator workflow [35], we extended the workflow to include the Ontology of Adverse Events [36] and data from the following drug data sources in the ontology-grounded KG:

• Drug Interaction Knowledge Base (DIKB) (v2017) [37,38]: evidence of enzyme substrates and inhibition, including in vitro information, drug label statements, and results from randomized clinical trials. To simplify the knowledge representation, only the positive evidence present in DIKB was included in the KG.

• Drug Central database (v2017) [39]: in vitro evidence of enzyme inhibition, drug-transporter interactions, drug-bacteria interactions, and enzyme and transporter substrates.

• FDA Drug Interaction database (v2017) [40]: in vitro and clinical evidence of enzyme and transporter substrates and inhibitors. We included all data where the reported fold change in area under the receiver operating characteristic curve of the drug substrate for an enzyme or transporter is at least 2-fold in the presence of a purported inhibitor drug. This cut-off was chosen because it represents strong positive evidence of a clinically measurable pharmacokinetic mechanism (i.e., the primary drug clearance pathway involves a specific enzyme or transporter that can be inhibited by a drug to a clinically measurable extent).

The 2017 versions of the above databases were included to enable time-slicing in the KG. The time-slicing approach for KG evaluation splits the graph to predict chronologically later links. All entities in the above data sources were mapped to the OBO Foundry ontologies before integration into the KG. We then implemented the Network Transformation for Statistical Learning (OWL-NETS) method in the PheKnowLator workflow with a goal to build a semantically rich KG to facilitate hypothesis generation [41]. The OWL-NETS method decoded all Web Ontology Language (OWL)-encoded classes and axioms into clinically and biologically relevant edges in the graph. To extend the PheKnowLator workflow with the above data sources, we applied the instance-based approach for KG construction, where knowledge statements from the ontologies are added to the KG in the form of subject-relation-object triples without the associated logical definitions. The resulting KG was saved as subject-relation-object triples and as a Python NetworkX multidigraph [42], which is a directed graph containing nodes and edges of multiple types.

2.2. Representation of Natural Products

Inclusion of natural products and their chemical constituents in the OBO Foundry is limited. Natural products and their constituents are also limited in drug databases such as Drug Bank and Drug Central, which are commonly used to predict drug-drug interactions or drug-related adverse events. The Drug Ontology [43], FoodOn ontology [44], and the NCBI Taxonomy [45] contain natural products but do not provide information about the pharmacokinetic interactions precipitated by natural products or their constituents. The ChEBI ontology contains some natural product constituents, including catechin and the kratom alkaloid mitragynine, but the coverage of constituents and their metabolites is incomplete. We addressed this gap by integrating extensions designed in our prior work in the ChEBI ontology to include information about natural products and link them to their constituents [46]. In brief, we created semantic representations for natural products using data from the Global Substance Registration System [47], including information about natural product constituents and their metabolites, translated them to logical statements in OWL, integrated the statements into the ChEBI (Lite) ontology, and finally into the KG.

2.3. Literature-based Graph

The literature-based graph (Figure 2) was created from full texts from the scientific literature related to green tea and kratom.

Figure 2.

Workflow for semantic relation extraction and literature-based graph construction.

2.3.1. Semantic Relation Extraction

Semantic relation extraction produces predications from texts. We extracted predications from two biomedical relation extraction systems to add to the ontology-grounded KG. The scope of the literature for the literature-based graph included all PubMed-indexed articles related to green tea and kratom (including keywords of scientific names, synonyms, and their constituents) and pharmacokinetic interactions. The search was restricted to English language articles only and placed no restrictions on animal or human studies. Articles included mechanistic NPDI studies, case reports, and clinical studies from January 1982 to March 2022 for green tea, and from January 1988 to March 2022 for kratom. The search strategies are available in Appendix A.

2.3.1.1. Relation Extraction Systems

SemRep [48] and the Integrated Network and Dynamic Reasoning Assembler (INDRA) [49] with the Reading and Assembling Contextual and Holistic Mechanisms from Text (REACH) biological reader [50] were used to extract semantic relations. SemRep (v1.8) is a natural language processing system that infers relations between entities in biomedical texts and is informed by syntactic and semantic constraints combined with biomedical domain knowledge [48]. SemRep uses MetaMap (v2018) to identify biomedical entities in texts and maps to Unified Medical Language System (UMLS) concepts [51]. SemRep achieved 0.69 precision and 0.42 recall for relation extraction from PubMed titles and abstracts in a test collection [48]. Although SemRep has mainly been used for titles and abstracts, our pipeline extended SemRep to process full texts.

REACH (v1.6.3) is an information extraction system designed to robustly parse full texts of biomedical literature and extract cancer signaling pathways. REACH focuses on extracting information about biochemical interactions of proteins and has demonstrated high precision and throughput in uncovering mechanistic knowledge from the literature compared to manually curated databases and other biological relation extraction systems [50]. Evaluation of REACH for extracting mechanistic information from the biomedical literature showed a precision of 0.62 and estimated recall of 0.486 based on a throughput evaluation [50]. INDRA (v1.19) provides a framework to extract predications using REACH. As the performance metrics of SemRep and REACH were computed from different datasets, and SemRep has not yet been comprehensively evaluated on full texts, we performed a baseline evaluation on a subset of green tea articles to evaluate the relation extraction systems. Full texts of 13 articles were processed with both SemRep and REACH to extract predications. The predications were then compared with human extracted data from the Center of Excellence for Natural Product Drug Interaction Research (NaPDI Center) database [6] to calculate the recall for SemRep and INDRA/REACH.

We extracted full texts from the literature in our search strategy from PubMed Central and from text mining of PDFs using the PDFMiner library [52] and Python script adapted from Hoang et. al. [53]. The available full texts were processed separately by SemRep and INDRA/REACH to extract predications.

2.3.2. Literature-based Graph Construction

To create the literature-based graph from the extracted predications, we linked all subjects, objects, and relations from the SemRep and INDRA/REACH predications to the OBO Foundry ontologies to avoid ambiguity and standardize the graph. This process, termed entity linking, can be achieved with automated methods (such as named entity recognition tools) or curated mappings. We used a combination of both approaches to standardize subjects, objects, and relations in the predications to OBO Foundry ontologies. First, relations extracted by SemRep and INDRA/REACH were standardized to Relation Ontology terms based on domain, range, and definitions of the relations after consulting with experts in the field. SemRep also extracted negated predications involving negative relations such as neg_inhibits and neg_interacts_with. These were stored separately as there were no corresponding negative relations in the OBO Foundry ontologies. Next, gene functions and phenotype concepts identified by SemRep were mapped to concepts in the Gene Ontology and the Human Phenotype Ontology using database cross references in UMLS. Subjects and objects extracted by the INDRA/REACH system were mapped to the OBO Foundry ontologies using the INDRA BioOntology module.

We next applied the OntoRunNER OGER++ wrapper [54] with custom term lists containing OBO Foundry ontologies (from the ontology-grounded KG) to identify candidate mappings for the remaining unmapped subjects and objects. To reduce noise from the relation extraction systems, all candidate mappings were manually reviewed to determine the correct identifiers in the OBO Foundry, and predications with unmapped subjects and objects were excluded. We further filtered the predications to exclude relations that were not useful for hypothesis generation, such as converts_to, diagnoses, method_of, and process_of. Certain semantic types (such as activities and behaviors, concepts and ideas, organizations) and generic concepts (such as Animals, Disease, Persons, Organism, Syndrome, Patients) provided by Zhang et. al. [55] were also excluded. We used the INDRA assembly module for deduplicating, normalizing namespaces, and generating confidence scores for the predications extracted from INDRA/REACH.

After entity linking and processing, predications from SemRep and INDRA/REACH were combined into a single file with OBO identifiers for all subjects, relations, and objects. We then applied symmetric and transitive closure over the combined predications using the CLIPSPy library [56]. The closure process used entailment properties to infer new predications (termed inferred predications) based on the rules for symmetry and transitivity of the relations. We included the relations ‘interacts_with’ and ‘molecularly_interacts_with’ in the symmetric closure rules and the relations ‘part of’, ‘precedes’, and ‘positively_regulates’ in the transitive closure rules based on the relation properties in the Relation Ontology. As an example, applying closure over a predication involving a symmetric relation, such as ‘catechin interacts_with nadolol’, results in the inferred predication ‘nadolol interacts_with catechin’. The inferred predications were combined with those extracted from the literature. We then created the literature-based graph as a NetworkX graph (separate from the ontology-grounded KG) from the combined predications, where the nodes in the graph represented the subjects and objects from the predications, and relations formed the edges between the nodes. The metadata related to each predication, including the PubMed ID of the article, year of publication of the article, confidence score (if available), and the sentence from which the predication was extracted, was also added in the graph.

2.4. Evaluation

We combined the ontology-grounded KG and the literature-based graph to create NP-KG and store as another NetworkX multidigraph. This combined KG is referred to as NP-KG in the following sections. We evaluated NP-KG with case studies of green tea and kratom. We also conducted an error analysis to identify errors from relation extraction and contradictory information in NP-KG.

2.4.1. Ground Truth

Potential pharmacokinetic interactions that have been identified through in vitro experiments and clinical studies were considered as ground truth for NPDIs. If a clinical study had not yet been conducted, results from in vitro experiments and mechanistic static model predictions sufficed as ground truth. In case of contradictory results between in vitro and clinical studies, the clinical studies were considered the ground truth. We obtained ground truth for KG evaluation from the NaPDI Center, which maintains a human curated database that contains results from in vitro experiments and clinical pharmacokinetic NPDI studies [6].

2.4.2. Evaluation Strategies

I. Knowledge Recapturing

We evaluated the potential of NP-KG to capture existing knowledge about green tea and kratom interactions with enzymes and transporters. We first collected OBO Foundry ontology identifiers for the enzymes and transporters in the NaPDI Center database. Next, for all nodes related to green tea (green tea and the green tea catechins (−)-epicatechin-3 gallate (ECG), (−)-epigallocatechin gallate (EGCG), epicatechin, catechin, and gallocatechin) and kratom (kratom and the kratom alkaloid mitragynine), we extracted direct edges (if any) and shortest paths (if any) from the nodes to enzymes and transporters in NP-KG. Edges and paths extracted from NP-KG were compared to the ground truth information in the database. Results from in vitro experiments and the published literature in the database involving inhibition, induction, and no interactions for the enzymes and transporters were included in the comparison.

We calculated the number of congruences, contradictions, and missing knowledge from NP-KG while comparing to data in the database. Congruence was established when the direct edge or shortest path in NP-KG concurred with the ground truth information about the interaction. Contradiction was established if the direct edge or shortest path in NP-KG contradicted the results of the interaction in the database. An example of congruence and contradiction between nodes X and Y is shown (Figure 3). In case of both congruence and contradiction, we manually reviewed the evidence related to the information from metadata present in NP-KG. For edges or paths where congruence or contradiction could not be established, such as edges or paths involving the relations interacts_with, molecularly_interacts_with, and directly_regulates_activity_of, the edges and paths were not included in the calculation of congruences or contradictions. Shortest paths were obtained using the bidirectional shortest path algorithm in the NetworkX library.

Figure 3.

Congruence and contradiction between node X and an enzyme or transporter Y in NP-KG, where ground truth information between the nodes states that X inhibits Y.

II. Meta-path Discovery

Meta-paths are defined as sequences of connected relations between nodes, node types, or sequences of node types in a KG. They are commonly used in KG construction and hypothesis generation [55,57]. We applied direct edge searches and meta-paths in NP-KG to generate hypotheses for the pharmacokinetic NPDIs involving green tea and kratom. Figure 4 shows the direct edges and meta-path searches applied for discovering the interactions between a natural product or its constituent and a drug, with an interacting enzyme or transporter. The meta-path searches were implemented on the NetworkX multidigraph by finding all nodes and edges in the neighborhood of green tea and kratom nodes (including the green tea and kratom constituent nodes) in NP-KG and applying filters to the neighboring nodes and edges for relevant relations, enzymes, transporters, and drugs. The filtered nodes and edges that matched the sequences shown in Figure 4 were then extracted for visualization and review.

Figure 4.

Direct edges and meta-path searches implemented in NP-KG for natural products or natural product constituents and drugs with an interacting enzyme or transporter.

Relations A and B in the direct edges and meta-path searches (Figure 4) were filtered to include the relations interacts_with, molecularly_interacts with, associated_with, directly_regulates_activity_of, positively_regulates, inhibits, capable_of_regulating, capable_of_positively_regulating, correlated_with, is_substrate_of, transports, and regulates_activity_of. We then evaluated the results based on the available ground truth information for the interactions.

We selected five natural product-drug pairs for the meta-path discovery, namely green tea-raloxifene, green tea-nadolol, kratom-midazolam, kratom-quetiapine, and kratom-venlafaxine. These interactions were chosen due to the published clinical results and case reports that provided evidence and potential mechanisms for the interactions. Green tea has been shown to interact with the anti-osteoporosis drug raloxifene by decreasing raloxifene systemic exposure, likely due to green tea-mediated alterations in processes involved in the intestinal absorption of raloxifene [16,17]. As mentioned above, the interaction between green tea and nadolol led to reduced systemic exposure and blood pressure-lowering effects of nadolol [8]. Potential pharmacokinetic interactions between kratom and quetiapine and/or venlafaxine were published in a recent case report on kratom use disorder [58]. The kratom-quetiapine interaction was also reported in a case report of a kratom-related death [59]. Proposed mechanisms for these interactions include inhibition of CYP3A activity by the kratom alkaloid mitragynine [60,61].

3. Results

3.1. Data Integration

The ontology-grounded KG created from the PheKnowLator workflow included entities from ontologies representing diseases, genes, phenotypes, chemicals, anatomy, cell lines, proteins, and pathways from 10 OBO Foundry ontologies and 13 publicly available data sources [35]. After integrating green tea, kratom, and their constituents and metabolites, 15 classes, 96 axioms, and 13 individuals were added to ChEBI Lite ontology, bringing the total to 156,113 classes, 1,201,077 logical axioms, and 13 individuals. The merged ChEBI ontology, individual semantic representations, and logical extensions are publicly available [46].

The ontology-grounded KG contained 745,250 nodes and 7,224,186 edges. Appendix A Table A.1 shows the number of edges in the ontology-grounded KG with the edge type, edge source, and edge label. The edges chemical-transporter (N=90), chemical-molecule (N=391), chemical-inhibitor (N=272), and chemical-substrate (N=514) were included in the extended workflow from the drug data sources.

3.2. Literature-based Graph

We obtained full texts of 735 scientific articles for green tea and 59 articles for kratom from PubMed. INDRA/REACH extracted 12,270 predications from 716 full-text articles (89.05%). SemRep extracted 250,360 predications from 637 full-text articles (79.23%). There were 13,676 predications from SemRep for green tea and kratom after deduplication, concept reduction, and filtering. There were 7,999 predications for green tea and kratom from INDRA/REACH after processing with the INDRA assembly module. Implementation of the transitive and symmetric closures over the predications resulted in an additional 13,637 inferred predications. Appendix Table A.2 describes the predications extracted from the literature with their counts, source, and the OBO Foundry ontology mapped relation.

The baseline evaluation of SemRep and INDRA/REACH with 13 green tea articles resulted in 522 predications extracted by SemRep and 123 predications extracted by INDRA/REACH compared to 179 predications from the NaPDI Center database. The recall values for SemRep and INDRA/REACH were 0.31 and 0.20, respectively. The combined recall for the systems was 0.42.

The literature-based graph constructed from the combined and deduplicated predications for green tea and kratom contained 3,510 nodes and 25,421 edges. Figure 5 presents an overview of the types of nodes and edges in the literature-based graph. Figure 6 shows the distribution of the edges in the literature-based graph by edge type.

Figure 5.

Node types and edge types in the literature-based graph (with edge counts greater than 100). Rectangles represent edges with edge labels from the graph.

Figure 6.

Distribution of edges in the literature-based graph, including edges extracted from the literature and inferred edges. The x-axis presents the log of the number of edges and y-axis presents the edge labels mapped to OBO Foundry ontologies. Edges included in the symmetric closure are marked with *, and edges included in the transitive closure are marked with **.

3.3. Evaluation

After integration of the ontology-grounded KG and the literature-based graph, the heterogeneous NP-KG contained 745,512 nodes and 7,249,576 edges. The literature-based graph added 262 unique nodes and 25,390 unique edges to NP-KG. Table 1 describes the graph counts for the ontology-grounded KG, the literature-based graph, and NP-KG.

Table 1.

Graph counts for ontology-grounded KG, literature-based graph, and NP-KG

Parameter	Ontology-grounded KG	Literature-based Graph	NP-KG (% change from ontology-grounded KG)
Nodes	745,250	4,157	745,512 (0.035%)
Edges	7,224,186	27,784	7,249,576 (0.35%)
Average Degree	9.69	6.68	9.72 (0.31%)
Node Density	1.301 × 10−5	0.002	1.304 × 10−5 (0.23%)

The integration of the literature-based graph resulted in a 0.035% and 0.35% increase in nodes and edges, respectively, in NP-KG. As the literature search strategy was focused on green tea- and kratom-drug interactions, the impact of including the literature-based graph was highest for the green tea and kratom nodes. Figures 7(a) and 7(b) compare the in- and out-degrees of green tea and kratom nodes in the ontology-grounded KG and NP-KG, where in-degree is the number of edges into the node and out-degree is the number of edges exiting the node. The in- and out-degrees of every node, except kratom, increased after the integration of the literature-based graph in NP-KG.

Figure 7.

Comparison of green tea and kratom node degrees between ontology-grounded KG and NP-KG. 7a. In-degrees of green tea and kratom nodes. 7b. Out-degrees of green tea and kratom nodes. The x-axis presents the log of the in- and out-degrees and y-axis presents the node labels.

I. Knowledge Recapturing

Table 2 summarizes the congruent and contradictory knowledge from NP-KG related to pharmacokinetic interactions between the natural products-related nodes and enzymes and transporters when compared to the ground truth information. For the green tea-related nodes, we performed 59 searches for direct edges or shortest paths in NP-KG involving 19 enzymes and 8 transporters. For the kratom-related nodes, we performed 14 searches for direct edges or shortest paths involving 10 enzymes and 1 transporter. Figure 8 shows examples of direct edges and shortest paths in NP-KG. Some nodes contained multiple edges between them, such as catechin and UGT1A1 (Figure 8a), mitragynine and CYP3A (Figure 8d), and mitragynine and CYP2D6 (Figure 8e). Results with both congruent and contradictory edges between the nodes were manually reviewed to verify congruence and/or contradiction.

Table 2.

Summary of congruences and contradictions for direct edges and shortest paths in NP-KG compared to ground truth information.

	Green Tea (%)	Kratom (%)
Congruence	23 (38.98)	7 (50.0)
Contradiction	9 (15.25)	3 (21.43)
Edges/paths exist but no congruence or contradiction	25 (42.37)	3 (21.43)
Both congruence and contradiction	2 (3.39)	1 (7.14)
Total searches	59	14

Figure 8.

(a-c) Edges and/or shortest path(s) between green tea and UGT enzymes (UGT, UGT1A1, UGT1A8, and UGT1A10) in NP-KG. (d-e) Shortest path(s) between kratom and cytochrome P450 enzymes (CYP3A4, CYP2D6) in NP-KG. If an edge in the path is derived from the literature, the year of publication is noted with the edge label. The rounded rectangles represent nodes and rectangles represent edges in NP-KG.

Examples of manually reviewed searches include the following:

• The edges inhibits, positively_regulates, and directly_positively_regulates_quantity_of between EGCG and CYP1A2 and EGCG and UGT1A1 suggest both congruence and contradiction. Manual review showed that contradictions in the literature were responsible for the evidence of both congruence and contradictions in NP-KG.

• Mitragynine, the major alkaloid in kratom leaves, is a time-dependent inhibitor of CYP3A and reversible inhibitor of CYP2D6 [60]. The edges inhibits and directly_positively_regulates_quantity_of between mitragynine and CYP3A4 (Figure 8d) imply both congruence and contradiction with the ground truth. Manual review of the edges shows evidence of statements in the literature suggesting mitragynine inhibits (PubMed ID: 24174816) and induces (PubMed ID: 23274770) CYP3A4 in vitro.

• The inhibits edges and positively_regulates edge between mitragynine and CYP2D6 (Figure 8e) imply both congruence and contradiction with ground truth information. Manual review in this case revealed errors from relation extraction for the positively_regulates edge between mitragynine and CYP2D6.

Congruence or contradiction could not be established for 25 (42.37%) searches for green tea nodes and 3 (21.43%) searches for kratom nodes as the edges included general interaction predicates such as interacts_with and regulates_activity_of. For example, NP-KG found that mitragynine interacts_with CYP2C19, although the ground truth is that mitragynine inhibits CYP2C19. The search was thus insufficient to conclude congruence or contradiction in this case. Supplementary File 1 presents edges and paths between nodes for the knowledge recapturing evaluation strategy for all kratom and green tea-related nodes.

II. Meta-path Discovery

We applied direct edge and meta-path searches for discovery of pharmacokinetic NPDIs involving green tea and kratom in NP-KG (Figure 4). Table 3 summarizes the results from the direct edges and meta-path discovery for green tea-raloxifene, green tea-nadolol, kratom-midazolam, kratom-quetiapine, and kratom-venlafaxine.

Table 3.

Summary of results of meta-path discovery in NP-KG for the natural product-drug pairs.

Natural product-drug pair	Enzyme(s)	Transporter(s)
Green tea-raloxifene	CYP3A4, UGT	ND
Green tea-nadolol	See Appendix Table A.4	See Appendix Table A.5
Kratom-midazolam	CYP2D6, CYP3A4	P-glycoprotein
Kratom-quetiapine	CYP3A4	P-glycoprotein
Kratom-venlafaxine	CYP2D6, CYP3A4	ND

(ND=None Detected)

Green tea-raloxifene:

Raloxifene is an anti-osteoporosis agent and an intestinal UGT substrate. Results from in vitro experiments showed concentration-dependent inhibition of intestinal UGT activity by green tea and green tea catechins [16]. In the subsequent clinical study conducted to determine the effects of green tea on raloxifene pharmacokinetics in 16 healthy adult participants, an interaction was observed. Specifically, relative to water, green tea decreased the area under the plasma concentration versus time curve (AUC) of raloxifene by ~30%. However, the change in raloxifene AUC was opposite to the in vitro-in vivo prediction (increase in raloxifene AUC). These observations, combined with the lack of effect of green tea on raloxifene half-life and the raloxifene-to-glucuronide AUC ratios, suggested that inhibition of intestinal UGT activity was not responsible for the interaction [17]. Meta-path discovery in NP-KG identified CYP3A4 and UGTs for the pharmacokinetic interaction between green tea and raloxifene. NP-KG also identified raloxifene to be a substrate for CYP2C9 but with no connection to green tea. Figure 9 shows a graphical snapshot of the paths among green tea, green tea constituents, and raloxifene nodes in NP-KG focusing on pharmacokinetic interactions.

Figure 9.

Nodes and edges in NP-KG related to pharmacokinetic interactions between green tea and raloxifene with enzymes (CYP3A4, CYP2C9, UGT). Rounded rectangles represent nodes and rectangles represent edges in NP-KG. If an edge is derived from literature, the year of publication is noted with the edge label. Note that nodes can contain multiple edges between them.

Green tea-nadolol:

Nadolol is a non-selective beta-adrenoceptor blocker used to treat hypertension. A clinical pharmacokinetic interaction study involving a green tea beverage and nadolol showed green tea to decrease nadolol AUC by 80% and to decrease the systolic blood pressure lowering effects of nadolol in ten healthy adults [8]. In vitro experiments showed that nadolol is a substrate for OATP1A2 and that EGCG inhibited OATP1A2-mediated nadolol uptake using OATP1A2-expressing human embryonic kidney cells. However, the existence of OATP1A2 remains controversial [62,63], and the interaction may involve other mechanisms, such as upregulation of efflux transporters [8]. Meta-path discovery in NP-KG identified 8 drug metabolizing enzymes and 12 transporters for the pharmacokinetic interaction between green tea and nadolol. The list of drug metabolizing enzymes and transporters is available in Appendix Tables A.4 and A.5, respectively.

Kratom-midazolam:

Midazolam is a short-acting benzodiazepine and a CYP3A substrate. A kratom extract and the kratom alkaloid mitragynine have been shown to inhibit CYP2D6, CYP2C9, and CYP3A activities in vitro [60,61]. A mechanistic static model predicted a pharmacokinetic interaction between kratom and midazolam based on time-dependent inhibition of CYP3A by mitragynine [60]. Meta-path discovery in NP-KG identified 2 interacting enzymes, CYP3A4 and CYP2D6, and one transporter, P-glycoprotein, for the pharmacokinetic interaction between kratom and midazolam. However, midazolam is not a substrate for CYP2D6 nor P-glycoprotein

Kratom-quetiapine and kratom-venlafaxine:

Venlafaxine is an antidepressant and a dual CYP3A/CYP2D6 substrate. Quetiapine is an antipsychotic and a CYP3A substrate. A recent case report described potential pharmacokinetic kratom-drug interactions involving these two drugs [58]. Proposed mechanisms for the interactions include inhibition of CYP3A and CYP2D6 [58]. Meta-path discovery in NP-KG identified CYP3A4 and P-glycoprotein for the interaction between kratom and quetiapine and CYP3A4 and CYP2D6 for the interaction between kratom and venlafaxine. Figures 10a and 10b show graphical snapshots for paths between kratom and interacting drugs.

Figure 10a.

Nodes and edges in NP-KG related to pharmacokinetic interactions between mitragynine and midazolam with interacting enzymes (CYP3A4, CYP2D6) and a transporter (P-glycoprotein). 10b. Nodes and edges in NP-KG related to pharmacokinetic interactions between mitragynine and drugs (quetiapine and venlafaxine) with interacting enzymes (CYP3A4, CYP2D6) and a transporter (P-glycoprotein). Rounded rectangles represent nodes and rectangles represent edges in NP-KG. If an edge is derived from the literature, the year of publication is noted with the edge label.

Appendix Table A.3 presents the results from the meta-path discovery searches for the natural product-drug pairs examined.

Error Analysis

Error analysis of predications related to green tea-drug interactions (Figures 8–9) and kratom-drug interactions (Figures 8, 10a-b) showed that multiple incorrect predications exist in NP-KG generated by the relation extraction systems. Examples of incorrect predications included midazolam inhibits CYP2D6 (2014), mitragynine inhibits UGT2B7 (2014), mitragynine stimulates midazolam (2021), 7-hydroxymitragynine inhibits mitragynine (2014).

Manual verification of the predications identified broad errors from the relation extraction systems from the literature. First, sentences in texts containing ‘increase or decrease in area under the plasma concentration-time curve’ were commonly misread by the systems and not represented accurately by the extracted predication. Second, in some cases, negation in sentences was not accurately extracted, leading to inaccurate predications. Third, substrates of enzymes and transporters mentioned in the literature are not extracted by either of the relation extraction systems as they do not contain rules to identify the ‘substrate of’ relation in text. Improving relation extraction from the literature and applying constraints to NP-KG during construction and hypothesis generation can help resolve some of the errors. Other errors from the ontology-grounded KG also exist, such as midazolam interacts_with CYP2E1 (source: Comparative Toxicogenomics Database), UGT1A4 moleculary_interacts_with UGT1A6 (source: STRING Database), and CYP1A2 molecularly_interacts_with CYP2C8 (source: STRING Database).

There are several contradictory edge pairs between nodes in NP-KG that are not included in the errors described above. Contradictory edges may be defined as edges that contradict each other semantically, such as inhibits and positively_regulates. For example, the pairs mitragynine inhibits CYP2D6 (2013, 2017) and mitragynine positively_regulates CYP2D6 (2013), mitragynine inhibits CYP3A4 (2013) and mitragynine directly_positively_regulates_quantity_of CYP3A4 (2013), EGCG inhibits UGT (2018) and EGCG positively_regulates UGT seem contradictory. These contradictions can be reasonable in cases where the evolution of knowledge leads to multiple edges between nodes, as the evidence related to the nodes is updated (in the literature or elsewhere). Although we can track the year of publications for the sources of the predications, manual verification and fact-checking of the contradictory edges are required to improve the hypotheses generated by NP-KG. These procedures can be aided by automated filtering strategies for incorrect predications due to relation extraction [64].

5. Discussion

To our knowledge, this study is the first to combine existing curated information from biomedical ontologies with computable knowledge extracted from full texts of scientific articles to construct a KG focused on NPDIs. NP-KG is novel in that it combines biomedical ontologies, publicly available databases, and domain-specific information from the scientific literature to generate mechanistic hypotheses for potential pharmacokinetic NPDIs. We demonstrated that NP-KG can be used to recapture knowledge about pharmacokinetic NPDIs, as well as provide a template for predicting NPDIs that have not been studied extensively to guide researchers in the design and conduct of new studies.

We applied two evaluation strategies to validate the hypotheses generated from NP-KG for green tea- and kratom-drug interactions. The first provided direct edges and shortest paths between nodes related to the natural products in NP-KG and potential interacting enzymes and transporters. When compared to curated data from the NaPDI Center database, we identified both congruent and contradictory information in NP-KG. Further analysis showed that contradictory information was the result of contradictions present in the literature or errors from the relation extraction systems. Manual review was required in multiple cases where both congruent and contradictory edges existed between the nodes (15.25% searches for green tea, 21.43% searches for kratom). Although our targeted searches could identify some contradictions, a complete contradiction analysis of NP-KG was out of scope for this study.

Resolving contradictions in literature-based discovery and biomedical KGs is an active area of research. An investigation of clinically relevant, contradictory predications in SemMedDB [65] (semantic predications database for PubMed abstracts generated using SemRep) revealed multiple reasons for the contradictions, including characteristics related to the patients, known controversies, contradictions present in the literature, and the progression of knowledge [66]. Although the analysis was restricted to abstracts in the clinical domain, both the immediate context (source sentence) and full abstract were required to detect and resolve the contradictions [66]. Similarly, Sosa and Altman [67] demonstrated that literature-based KGs suffer from several challenges that give rise to contradictory information in the drug repurposing domain, namely, errors in relation extraction, errors in entity normalization leading to false positive contradiction pairs, contradictions in the biomedical literature due to progression of science and variation in research methods, and inconclusive evidence. Enhancing NP-KG with contextual information, including more granular information, quantitative facts (e.g., dosage), negative results, confidence scores, and source sections of the predications can help resolve contradictions and improve knowledge inference.

The second evaluation strategy for NP-KG applied direct edges and meta-path searches for select natural product-drug pairs, specifically green tea-raloxifene, green tea-nadolol, kratom-midazolam, kratom-quetiapine, and kratom-venlafaxine. Meta-paths are also known as discovery patterns or pathways in literature-based discovery and can be of varying lengths. Prior work suggests that they can correspond to mechanisms of pharmacological efficacy and thus may be used to generate plausible hypotheses from KGs [57]. In related work, Schutte et. al. [12] applied meta-paths (Dietary_Supplement-Gene-Drug and Dietary_supplement-Gene1-Biological_Function-Gene2-Drug) to identify NPDIs involving supplements from a literature-based KG. Meta-path discovery for the green tea-raloxifene interaction identified CYP3A4 and UGTs involved in the mechanism. Although inhibition of intestinal UGT by green tea catechins is supported by in vitro data, the clinical study did not confirm this mechanism [16,17]. This in vitro-in vivo disconnect suggests that NP-KG may need to be expanded to more data domains, including physicochemical properties of drugs and natural products, and the gut microbiome domain such as with the gutMEGA database [68] and microbe-drug associations from the microbiota-active substance interactions database [69], to generate mechanistic hypotheses.

For the green tea-nadolol interaction, although the meta-path searches identified OATP1A2 as one of the transporters involved in the interaction, other enzymes and transporters may or may not be involved in the interaction and require further review. There is no evidence to support that nadolol is a substrate of the CYP enzymes identified by NP-KG for the interaction between green tea and nadolol (Appendix Table A.4). NP-KG meta-path discovery results suggested that mechanisms of putative pharmacokinetic interactions between kratom and midazolam, quetiapine, and venlafaxine primarily involve CYP3A4, CYP2D6, and P-glycoprotein. Some of these results are supported by the published literature [60,70–72] while some are yet unknown. The evaluation results suggest that NP-KG provides a major advance in support of our hypothesis for computational discovery of pharmacokinetic NPDI mechanisms by combining biomedical ontologies and databases with the scientific literature.

Constructing a KG by integrating full texts of the scientific literature with biomedical ontologies is challenging due to the limited availability of full texts in computable format and lack of interoperability within vocabularies used in ontologies, databases, and relation extraction systems. Large-scale KGs developed for drug-related applications, such as drug-drug interaction, drug-target prediction, and drug repurposing [55,57,73,74], are typically constructed either from curated data sources (such as ontologies, drug databases) or the scientific literature. KGs constructed from the scientific literature typically rely on manual annotations or automated relation extraction systems (e.g., SemRep). The literature-based graph constructed in this study combined predications from two relation extraction systems for more than 700 full-text articles. Node and edge types in the literature-based graph (Figures 5 and 6) showed that the relation extraction systems can capture relevant information from the scientific literature for different types of biomedical entities and supplement the ontology-grounded KG with domain-specific knowledge.

Comparison of the node degrees between the ontology-grounded KG and NP-KG showed greater connectivity of the green tea and kratom nodes in NP-KG. Although the number of nodes and edges added from the literature-based graph were significantly less than the nodes and edges in the ontology-grounded KG, integration of the literature-based graph contributed to filling the gaps in knowledge about green tea- and kratom-drug interactions in the ontology-grounded KG. As green tea is a popular natural product and is used worldwide, there were substantially more articles and predications related to green tea in NP-KG relative to kratom. There were also more ground truth interactions available for green tea in the knowledge recapturing evaluation strategy. We expect that this bias will remain when scaling NP-KG to include more natural products; however, the integration of different knowledge domains and KG completion methods can help discover unknown results about the interactions and mechanisms of less researched natural products.

Automated extraction and integration of the predications from the literature enabled us to draw inferences from NP-KG suggested in the published literature, and metadata included in the graph were used to verify the accuracy of the edges. Data integration and KG inference were further supported by the choice of instance-based representation for data integration, where subject-relation-object triples were retained in the ontology-grounded KG and the literature-based graph. The standardization of relations extracted by SemRep and INDRA/REACH to OBO Foundry ontologies presented in this study (Appendix Table A.2) will further facilitate the integration of literature-based knowledge with biomedical ontologies. Analysis of the predications also revealed errors in the relation extraction systems that, if improved, can lead to more accurate results and fewer contradictions.

Our overall goal is to use NP-KG to understand the knowledge gaps pertaining to NPDIs and generate mechanistic hypotheses for novel NPDIs with supporting evidence. This goal can be accomplished in several ways following the inclusion of more natural products in NP-KG. First, symbolic approaches using meta-paths and literature-based discovery methods similar to those described in this study can be used to discover unknown NPDIs. Second, NP-KG can be used to generate hypotheses and verify biological plausibility of pharmacovigilance signals from spontaneous reporting systems. Recent studies to detect adverse events related to natural products and NPDIs from spontaneous reporting systems, including the FDA Adverse Event Reporting System and CFSAN Adverse Event Reporting System, have shown promise [75–77]. The third approach involves graph representation learning methods for KG completion and hypothesis generation through link prediction and semantic similarity. KG completion methods can be used to overcome the gaps in knowledge about natural products in NP-KG. Recent advances in graph representation learning methods that generate embeddings from the KG have shown promise in discovering new edges in KGs and ranking mechanisms based on the similarity of nodes and edges in the embedding space [10,55,78]. The generated hypotheses can then be validated through manual review by domain experts, in vitro experiments, physiologically based pharmacokinetic modeling and simulation, and clinical studies [79,80]. To generalize the above approaches for all NPDIs, we need broader inclusion of knowledge domains to represent the mechanisms, broader processing of published literature, improved capture of context to address contradictions and assign confidence to edges, and rigorous evaluation of inference relative to specific use cases.

Limitations and Challenges

There are several limitations to this study. First, no manual filtering was used for the search strategies, which could have added non-relevant predications to the literature-based graph. Second, the literature-based graph contains known relation extraction errors due to incorrect entity recognition, missing predications (due to low recall), and complex sentence syntax [50]. The reported recall for relation extraction from abstracts with SemRep is 42% [48]; however, our baseline evaluation showed that the recall is lower for full texts from the pharmacokinetic literature compared to human curated data. SemRep has not been formally evaluated on full texts. Third, the relation extraction systems extracted predications from the full texts without targeting specific sections or information in tables. Fourth, our current approach for entity linking in the literature-based predications mapped only 60–70% of all subjects and objects in the predications, leading to potential missing information in NP-KG. Extensions of named entity recognition systems with custom data sources and relation extraction systems for natural products, such as in SemRepDS [12], can help improve entity linking in the KG. Potential data sources include the Integrated Dietary Supplements Knowledge Base [81] and FoodKG [82] to increase representation of natural products in existing resources. Fifth, this study applies meta-paths and shortest path searches in the KG for evaluation and although this approach is computationally efficient, the results are limited. For example, meta-paths applied in this study do not use information about metabolites of the natural product or the drug that could improve the understanding of the mechanism(s) underlying a given NPDI. Graph representation learning methods can also be applied to generate embeddings from the KG and broaden the scope of the results beyond the symbolic approaches used in this study. Finally, validation of NPDI mechanisms is challenging due to the limited availability of ground truth information. Drug Bank and Drug Central do not include data on NPDIs that can be used to evaluate predictions, as is common for drug-drug interactions. The Natural Medicine Interaction Checker database and Natural Medicines Comprehensive Database provide limited information for validating new hypotheses for drug-supplement interaction mechanisms [11,12]. Validation of NPDIs is thus either accomplished through the published literature or adaptation of drug-drug interaction datasets for food-drug interactions [11–13,83]. Our evaluation focused on green tea and kratom with ground truth information extracted from the NaPDI Center database. We only searched for enzymes or transporters with known interactions in the knowledge recapturing evaluation strategy. Although not shown in this study, utilizing NP-KG to discover beneficial natural product-drug interactions is possible. A broader search for other pharmacokinetic targets (e.g., nuclear receptors) may lead to more insights into these complex interactions.

6. Conclusions

NP-KG combined existing curated information in the form of ontologies and databases with domain-specific full texts of scientific articles to construct a KG focused on NPDIs. We demonstrated the potential of NP-KG to generate hypotheses that are grounded in both existing knowledge and suggested mechanisms from the scientific literature. NP-KG can be used to narrow the scope of potential hypotheses for pharmacokinetic NPDIs and help scientists in the preclinical assessment of potential clinically significant NPDIs. Future work will seek to address limitations of the relation extraction systems, extend NP-KG to other natural products, and apply improved hypothesis generation methods and evaluation. We are adding more natural product information in the KG, including additional natural products of interest and other knowledge domains. NP-KG is publicly available at https://doi.org/10.5281/zenodo.6814507. The code for relation extraction, KG construction, and hypothesis generation is available at https://github.com/sanyabt/np-kg.

Statement of Significance.

Problem of Issue	Understanding the mechanisms of pharmacokinetic natural product-drug interactions (NPDIs) is key to preventing adverse events.
What is Already Known	Pharmacokinetic NPDIs occur when a natural product alters the absorption, distribution, metabolism, and/or excretion of a co-consumed drug, potentially resulting in adverse events. NPDI mechanisms and clinical effects are less studied than drug-drug interactions, and computational discovery of NPDIs can guide robust NPDI research.
What this Paper Adds	A knowledge graph (KG) was developed that combined biomedical ontologies, scientific literature, and linked open data to generate mechanistic hypotheses for pharmacokinetic NPDIs. The KG was evaluated using the exemplar natural products green tea and kratom.

Abbreviations

ChEBI

Chemical Entities of Biological Interest

CYP

Cytochrome P450

DIKB

Drug Interaction Knowledge Base

ECG

(−)-epicatechin-3 gallate

EGCG

(−)-epigallocatechin gallate

FDA

Food and Drug Administration

INDRA

Integrated Network and Dynamical Reasoning Assembler

IVIVE

In vitro to in vivo extrapolation

KG

Knowledge Graph

NPDI

Natural Product-Drug Interaction

OATP

Human Organic Anion Transporting Polypeptide

OBO

Open Biological and Biomedical Ontology

OWL

Web Ontology Language

OWLNETS

Network Transformation for Statistical Learning

PheKnowLator

Phenotype Knowledge Translator

REACH

Reading and Assembling Contextual and Holistic Mechanisms from Text

UGT

uridine 5’-diphospho-glucuronosyltransferase (UDP-glucuronosyltransferase)

UMLS

Unified Medical Language System

US

United States

Appendix A

Search Strategies

Search strategies for green tea and kratom were adapted from the following:

Birer-Williams C, Gufford BT, Chou E, Alilio M, VanAlstine S, Morley RE, McCune JS, Paine MF, Boyce RD. A new data repository for pharmacokinetic natural product-drug interactions: from chemical characterization to clinical studies. Drug metabolism and disposition. 2020 Oct 1;48(10):1104–12.

Search Strategies for Green tea

Clinical Studies

(Clinical Trial [PT] AND (“green tea”[All Fields] OR “green teas”[All Fields] OR “green tea extract”[All Fields] OR “Camellia sinensis”[All Fields] OR “Cha ye”[All Fields] OR “Cha-yeh”[All Fields] OR “Camelliae folium”[All Fields] OR “Camelliae sinensis folium “[All Fields] OR “Matcha tea”[All Fields] OR “Sencha tea”[All Fields] OR “Thea sinensis”[All Fields] OR “Catechin”[All Fields] OR “Catechins”[All Fields] OR “epicatechin”[All Fields] OR “gallocatechin”[All Fields] OR “epigallocatechin”[All Fields] OR “epigallocatechin gallate”[All Fields] OR “EGCG”[All Fields] OR “Epicatechin-3-gallate”[All Fields] OR “(−)-epigallocatechin gallate”[All Fields] OR “(−)-epicatechin”[All Fields] OR “(−)-epigallocatechin”[All Fields] OR “(−)-epicatechin-3-gallate”[All Fields]) AND “drug interactions”[All Fields]) NOT Review [PT]

Case Reports

Case Reports [PT] AND (“green tea”[All Fields] OR “green teas”[All Fields] OR “green tea extract”[All Fields] OR “Camellia sinensis”[All Fields] OR “Cha ye”[All Fields] OR “Cha-yeh”[All Fields] OR “Camelliae folium”[All Fields] OR “Camelliae sinensis folium “[All Fields] OR “Matcha tea”[All Fields] OR “Sencha tea”[All Fields] OR “Thea sinensis”[All Fields] OR “Catechin”[All Fields] OR “Catechins”[All Fields] OR “epicatechin”[All Fields] OR “gallocatechin”[All Fields] OR “epigallocatechin”[All Fields] OR “epigallocatechin gallate”[All Fields] OR “EGCG”[All Fields] OR “Epicatechin-3-gallate”[All Fields] OR “(−)-epigallocatechin gallate”[All Fields] OR “(−)-epicatechin”[All Fields] OR “(−)-epigallocatechin”[All Fields] OR “(−)-epicatechin-3-gallate”[All Fields]) AND “drug interactions”[All Fields]

Mechanistic Studies

Step 1) Log into My NCBI and go to Pubmed: https://www.ncbi.nlm.nih.gov/pubmed/.

Step 2) In the advanced search form, clear the search history.

Step 3) Paste in this query into builder (using “edit”) and click “add to history”—this is referred to as “#1” in the rest of this search strategy:

“Cytochrome P-450 Enzyme System”[MeSH Terms] OR “Cyto- chrome P450 Family 1”[MeSH Terms] OR “Cytochrome P450 Family 2”[MeSH Terms] OR “Cytochrome P450 Family 3”[MeSH Terms] OR CYP1A1[All Fields] OR CYP1A2[All Fields] OR CYP1A3[All Fields] OR CYP1A4[All Fields] OR CYP1A5[All Fields] OR CYP2D6[All Fields] OR CYP2C9[All Fields] OR CYP2A6[All Fields] OR CYP2C8 [All Fields] OR CYP2C19[All Fields] OR CYP2B6[All Fields] OR CYP2B1[All Fields] OR CYP2E1[All Fields] OR CYP3A4[All Fields] OR CYP3A5[All Fields] OR UGT1[All Fields] OR UGT1A1[All Fields] OR UGT1A3[All Fields] OR UGT1A4[All Fields] OR UGT1A5[All Fields] OR UGT1A6[All Fields] OR UGT1A7[All Fields] OR UGT1A8[All Fields] OR UGT1A9[All Fields] OR UGT1A10[All Fields] OR UGT2[All Fields] OR UGT2A1[All Fields] OR UGT2A2 [All Fields] OR UGT2A3[All Fields] OR UGT2B4[All Fields] OR UGT2B7[All Fields] OR UGT2B10[All Fields] OR UGT2B11[All Fields] OR UGT2B15[All Fields] OR UGT2B17[All Fields] OR UGT2B28[All Fields] OR B3GAT1[All Fields] OR B3GAT2[All Fields] OR B3GAT3[All Fields].

Step 4) Paste in this query into builder (using “edit”) and click “add to history”—this is referred to as “#2” in the rest of this search strategy.

“Solute Carrier Proteins”[MeSH Terms] OR “Membrane Transport Proteins”[MeSH Terms] OR “P-gp”[All Fields] OR “p-glycoprotei- n”[All Fields] OR BCRP[All Fields] OR OCT2[All Fields] OR “Organic Cation Transporter 2”[MeSH Terms] OR “Organic Cation Transport Proteins”[MeSH Terms] OR MATE1[All Fields] OR “SLC4A Proteins”[MeSH Terms] OR MATE-2K[All Fields] OR “SLC4A Proteins”[MeSH Terms] OR OATP[All Fields] OR OAT1 [All Fields] OR “Organic Anion Transport Protein 1”[MeSH Terms] OR OAT3[All Fields] OR UGT1[All Fields] OR “Glucuronosyltransfer- ase”[MeSH Terms] OR ABC[All Fields] OR “ATP-Binding Cassette Transporters”[MeSH Terms].

Step 5) Paste in this query into builder (using “edit”) and click “add to history”—this is referred to as “#3” in the rest of this search strategy.

(“green tea”[All Fields] OR “green teas”[All Fields] OR “green tea extract”[All Fields] OR “Camellia sinensis”[All Fields] OR “Cha ye”[All Fields] OR “Cha-yeh”[All Fields] OR “Camelliae folium”[All Fields] OR “Camelliae sinensis folium “[All Fields] OR “Matcha tea”[All Fields] OR “Sencha tea”[All Fields] OR “Thea sinensis”[All Fields] OR “Catechin”[All Fields] OR “Catechins”[All Fields] OR “epicatechin”[All Fields] OR “gallocatechin”[All Fields] OR “epigallocatechin”[All Fields] OR “epigallocatechin gallate”[All Fields] OR “EGCG”[All Fields] OR “Epicatechin-3-gallate”[All Fields] OR “(−)-epigallocatechin gallate”[All Fields] OR “(−)-epicatechin”[All Fields] OR “(−)-epigallocatechin”[All Fields] OR “(−)-epicatechin-3-gallate”[All Fields]) Step 6) Paste in this query into builder (using “edit”) and click “add to history”—this is referred to as “#4” in the rest of this search strategy.

(Pharmacokinetics[MeSH Terms] OR pharmacokinetic[All Fields]) or (inhibit[All Fields] or inhibition[All Fields]) OR substrate[All Fields]. Step 7) Paste in this query into builder (using “edit”) and click “add to history”—this is referred to as “#5” in the rest of this search strategy.

#3 AND #4 AND (#1 OR #2)

Search Strategies for Kratom

Clinical Studies

(Clinical Trial [PT] AND (“Kratom”[All Fields] OR “mitragynine”[All Fields] OR “mitragynine ethanedisulfonate”[All Fields] OR “SK and F 12711”[All Fields] OR “SKF 12711”[All Fields] OR “SK and F-12711”[All Fields] OR “mitragynine, (16E)-isomer”[All Fields] OR “mitragynine, (3beta,16E)- isomer”[All Fields] OR “mmitragynine, (3beta,16E,20beta)-isomer”[- All Fields] OR “kratom alkaloids”[All Fields] OR “kmitragynine monohydrochloride”[All Fields] OR “Mitragyna speciosa”[All Fields] OR “Nauclea speciose”[All Fields] OR “Biak-biak”[All Fields] OR “Cratom”[All Fields] OR “Gratom”[All Fields] OR “Ithang”[All Fields] OR “Kakuam”[All Fields] OR “Katawn”[All Fields] OR “Kedem- ba”[All Fields] OR “Ketum”[All Fields] OR “Krathom”[All Fields] OR “Kraton”[All Fields] OR “Kratum”[All Fields] OR “Madat”[All Fields] OR “Mambog”[All Fields] OR “Mitragynine”[All Fields] OR “Mitra- gynine extract”[All Fields] OR “Thang”[All Fields] OR “Thom”[All Fields] OR “7-hydroxymitragynine”[All Fields] OR “7-hydroxy-mitra- gynine”[All Fields] OR “mitragynine pseudoindoxyl”[All Fields] OR “Paynantheine”[All Fields]) AND “drug interactions”[All Fields]) NOT Review [PT]

Case Reports

Case Reports [PT] AND (“Kratom”[All Fields] OR “mitragynine”[All Fields] OR “mitragynine ethanedisulfonate”[All Fields] OR “SK and F 12711”[All Fields] OR “SKF 12711”[All Fields] OR “SK and F-12711”[All Fields] OR “mitragynine, (16E)-isomer”[All Fields] OR “mitragynine, (3beta,16E)-isomer”[All Fields] OR “mmitragynine, (3beta,16E,20beta)-isomer”[All Fields] OR “kratom alkaloids”[All Fields] OR “kmitragynine monohydrochloride”[All Fields] OR “Mitragyna speciosa”[All Fields] OR “Nauclea speciose”[All Fields] OR “Biak-biak”[All Fields] OR “Cratom”[All Fields] OR “Gratom”[All Fields] OR “Ithang”[All Fields] OR “Kakuam”[All Fields] OR “Katawn”[All Fields] OR “Kedemba”[All Fields] OR “Ketum”[All Fields] OR “Krathom”[All Fields] OR “Kraton”[All Fields] OR “Kratum”[All Fields] OR “Madat”[All Fields] OR “Mambog”[All Fields] OR “Mitragynine”[All Fields] OR “Mitragynine extract”[All Fields] OR “Thang”[All Fields] OR “Thom”[All Fields] OR “7-hydroxymitragynine”[All Fields] OR “7-hydroxy-mitragynine”[All Fields] OR “mitragynine pseudoindoxyl”[All Fields] OR “Paynantheine”[All Fields]) AND “drug interactions”[All Fields]

Mechanistic Studies

Step 1) Log into My NCBI and go to Pubmed: https://www.ncbi.nlm.nih.gov/pubmed/.

Step 2) In the advanced search form, clear the search history.

Step 3) Paste in this query into builder (using “edit”) and click “add to history”—this is referred to as “#1” in the rest of this search strategy:

“Cytochrome P-450 Enzyme System”[MeSH Terms] OR “Cyto- chrome P450 Family 1”[MeSH Terms] OR “Cytochrome P450 Family 2”[MeSH Terms] OR “Cytochrome P450 Family 3”[MeSH Terms] OR CYP1A1[All Fields] OR CYP1A2[All Fields] OR CYP1A3[All Fields] OR CYP1A4[All Fields] OR CYP1A5[All Fields] OR CYP2D6[All Fields] OR CYP2C9[All Fields] OR CYP2A6[All Fields] OR CYP2C8 [All Fields] OR CYP2C19[All Fields] OR CYP2B6[All Fields] OR CYP2B1[All Fields] OR CYP2E1[All Fields] OR CYP3A4[All Fields] OR CYP3A5[All Fields] OR UGT1[All Fields] OR UGT1A1[All Fields] OR UGT1A3[All Fields] OR UGT1A4[All Fields] OR UGT1A5[All Fields] OR UGT1A6[All Fields] OR UGT1A7[All Fields] OR UGT1A8[All Fields] OR UGT1A9[All Fields] OR UGT1A10[All Fields] OR UGT2[All Fields] OR UGT2A1[All Fields] OR UGT2A2 [All Fields] OR UGT2A3[All Fields] OR UGT2B4[All Fields] OR UGT2B7[All Fields] OR UGT2B10[All Fields] OR UGT2B11[All Fields] OR UGT2B15[All Fields] OR UGT2B17[All Fields] OR UGT2B28[All Fields] OR B3GAT1[All Fields] OR B3GAT2[All Fields] OR B3GAT3[All Fields].

Step 4) Paste in this query into builder (using “edit”) and click “add to history”—this is referred to as “#2” in the rest of this search strategy.

“Solute Carrier Proteins”[MeSH Terms] OR “Membrane Transport Proteins”[MeSH Terms] OR “P-gp”[All Fields] OR “p-glycoprotei- n”[All Fields] OR BCRP[All Fields] OR OCT2[All Fields] OR “Organic Cation Transporter 2”[MeSH Terms] OR “Organic Cation Transport Proteins”[MeSH Terms] OR MATE1[All Fields] OR “SLC4A Proteins”[MeSH Terms] OR MATE-2K[All Fields] OR “SLC4A Proteins”[MeSH Terms] OR OATP[All Fields] OR OAT1 [All Fields] OR “Organic Anion Transport Protein 1”[MeSH Terms] OR OAT3[All Fields] OR UGT1[All Fields] OR “Glucuronosyltransfer- ase”[MeSH Terms] OR ABC[All Fields] OR “ATP-Binding Cassette Transporters”[MeSH Terms].

Step 5) Paste in this query into builder (using “edit”) and click “add to history”—this is referred to as “#3” in the rest of this search strategy.

(“Kratom”[All Fields] OR “mitragynine”[All Fields] OR “mitragynine ethanedisulfonate”[All Fields] OR “SK and F 12711”[All Fields] OR “SKF 12711”[All Fields] OR “SK and F-12711”[All Fields] OR “mitragynine, (16E)-isomer”[All Fields] OR “mitragynine, (3beta,16E)-isomer”[All Fields] OR “mmi- tragynine, (3beta,16E,20beta)-isomer”[All Fields] OR “kratom alka- loids”[All Fields] OR “kmitragynine monohydrochloride”[All Fields] OR “Mitragyna speciosa”[All Fields] OR “Nauclea speciose”[All Fields] OR “Biak-biak”[All Fields] OR “Cratom”[All Fields] OR “Gratom”[All Fields] OR “Ithang”[All Fields] OR “Kakuam”[All Fields] OR “Kataw- n”[All Fields] OR “Kedemba”[All Fields] OR “Ketum”[All Fields] OR “Krathom”[All Fields] OR “Kraton”[All Fields] OR “Kratum”[All Fields] OR “Madat”[All Fields] OR “Mambog”[All Fields] OR “Mitragynine”[All Fields] OR “Mitragynine extract”[All Fields] OR “Than- g”[All Fields] OR “Thom”[All Fields] OR “7-hydroxymitragynine”[All Fields] OR “7-hydroxy-mitragynine”[All Fields] OR “mitragynine pseudoindoxyl”[All Fields] OR “Paynantheine”[All Fields]).

Step 6) Paste in this query into builder (using “edit”) and click “add to history”—this is referred to as “#4” in the rest of this search strategy.

(Pharmacokinetics[MeSH Terms] OR pharmacokinetic[All Fields]) or (inhibit[All Fields] or inhibition[All Fields]) OR substrate[All Fields]. Step 7) Paste in this query into builder (using “edit”) and click “add to history”—this is referred to as “#5” in the rest of this search strategy.

#3 AND #4 AND (#1 OR #2)

Table A.1.

Edges in the ontology-grounded KG with edge counts, edge labels, and edge data source(s).

Edge Type	Edge Count	Edge Label	Edge Source(s)
chemical-disease	171,506	substance that treats	BioPortal, CTD
chemical-gene	16,701	interacts with	BioPortal, CTD
chemical-biological process	304,686	molecularly interacts with	BioPortal, CTD
chemical-molecular function	26,788	molecularly interacts with	BioPortal, CTD
chemical-cellular component	46,372	molecularly interacts with	BioPortal, CTD
chemical-pathway	29,248	participates in	Reactome Pathway Database
chemical-protein	66,828	interacts with	BioPortal, CTD
chemical-transporter	90	transports	FDA Drug Interaction database
chemical-molecule	391	molecularly interacts with	FDA Drug Interaction database, DIKB, Drug Central
chemical-inhibitor	272	inhibits	FDA Drug Interaction database, DIKB
chemical-substrate	514	substrate of	FDA Drug Interaction database, DIKB, Drug Central
gene-pathway	107,025	participates in	CTD
gene-phenotype	23,525	causes or contributes to	DisGeNET
gene-protein	19,523	has gene product	UniProt
pathway-cellular component	15,982	has component	Reactome Pathway Database
pathway-molecular function	2,424	has function	Reactome Pathway Database
protein-anatomy	30,682	located in	GTEx, Human Protein Atlas
protein-cell	75,318	located in	GTEx, Human Protein Atlas
protein-biological process	129,408	participates in	UniProt
protein-molecular function	70,085	has function	UniProt
protein-cellular component	82,366	located in	UniProt
protein-pathway	118,158	participates in	Reactome Pathway Database
protein-protein	618,069	molecularly interacts with	STRING database

Table A.2.

Edges in the literature-based graph with edge counts, edge source(s), OBO label, and OBO identifier (N=25,421).

Edge Name	Edge Count (%)	Edge Source(s)	OBO Label	OBO Identifier
acetylation	3 (0.012)	INDRA/REACH	protein acetylation	GO_0006473
activation	2515 (9.893)	INDRA/REACH	directly regulates activity of	RO_0002448
affects	353 (1.389)	SemRep	capable of regulating	RO_0002596
associated_with	110 (0.433)	SemRep	correlated_with	RO_0002610
augments	236 (0.928)	SemRep	capable of positively regulating	RO_0002598
causes	126 (0.496)	SemRep	causally influences	RO_0002566
coexists_with	1434 (5.641)	SemRep	existence overlaps	RO_0002490
complicates	1 (0.004)	SemRep	exacerbates condition	RO_0003309
decrease_amount	354 (1.393)	INDRA/REACH	directly negatively regulates quantity of	RO_0011010
dehydroxylation	5 (0.02)	INDRA/REACH	molecularly interacts with	RO_0002436
demethylation	3 (0.012)	INDRA/REACH	protein demethylation	GO_0006482
dephosphorylation	25 (0.098)	INDRA/REACH	protein dephosphorylation	GO_0006470
deubiquitination	1 (0.004)	INDRA/REACH	protein deubiquitination	GO_0016579
disrupts	153 (0.602)	SemRep	negatively regulates	RO_0002212
glycosylation	6 (0.024)	INDRA/REACH	protein glycosylation	GO_0006486
hydroxylation	8 (0.031)	INDRA/REACH	protein hydroxylation	GO_0018126
increase_amount	437 (1.719)	INDRA/REACH	directly positively regulates quantity of	RO_0011009
inhibits	3326 (13.084)	SemRep, INDRA/REACH	inhibits	RO_0002449
interacts_with	2421 (9.524)	SemRep	interacts with	RO_0002434
methylation	7 (0.028)	INDRA/REACH	protein methylation	GO_0006479
part_of	1927 (7.58)	SemRep	part of	BFO_0000050
phosphorylation	64 (0.252)	INDRA/REACH	phosphorylates	RO_0002447
precedes	1 (0.004)	SemRep	precedes	BFO_0000063
predisposes	27 (0.106)	SemRep	causes or contributes to condition	RO_0003302
prevents	53 (0.208)	SemRep	capable of inhibiting or preventing pathological process	RO_0002599
produces	248 (0.976)	SemRep	produces	RO_0003000
stimulates	11071 (43.551)	SemRep	positively regulates	RO_0002213
treats	505 (1.987)	SemRep	is substance that treats	RO_0002606
ubiquitination	1 (0.004)	INDRA/REACH	ubiquitinates	RO_0002480

Table A.3.

Results from meta-path discovery in NP-KG for natural product-drug pairs. Meta-path descriptions are available in Figure 4.

Meta-path	Results
Green Tea – Raloxifene
Green tea | Green tea constituent – Relation – Raloxifene	1. EGCG inhibits raloxifene 2. EGCG coexists with raloxifene 3. EGCG interacts with raloxifene
Raloxifene – Relation – Green tea | Green tea constituent	No results
Green tea | Green tea constituent – Relation A – Enzyme or Transporter – Relation B – Raloxifene	No results
Green tea | Green tea constituent – Relation A – Enzyme or Transporter, Raloxifene – Relation B – Enzyme or Transporter	1. Green tea inhibits UGT, Raloxifene inhibits/interacts with UGT 2. ECG inhibits UGT, Raloxifene inhibits/interacts with UGT 3. Catechin positively regulates/interacts with/directly regulates activity of UGT, Raloxifene inhibits/interacts with UGT 4. Catechin positively regulates CYP3A4, Raloxifene inhibits CYP3A4 5. EGCG inhibits/positively regulates UGT, Raloxifene inhibits/interacts with UGT 6. EGCG positively regulates CYP3A4, Raloxifene inhibits/interacts with UGT
Green Tea – Nadolol
Green tea | Green tea constituent – Relation – Nadolol	1. Green tea inhibits nadolol 2. EGCG inhibits nadolol 3. EGCG directly regulates activity of nadolol 4. Catechin interacts with nadolol
Nadolol – Relation – Green tea | Green tea constituent	1. Nadolol positively regulates EGCG 2. Nadolol interacts with EGCG 3. Nadolol positively regulates catechin 4. Nadolol positively regulates green tea leaf
Green tea | Green tea constituent – Relation A – Enzyme or Transporter – Relation B – Nadolol	SLCO1A2 gene (OATP1A2)
Green tea | Green tea constituent – Relation A – Enzyme or Transporter, Nadolol – Relation B – Enzyme or Transporter	See Tables A.4 and A.5.
Kratom – Midazolam
Kratom | Kratom constituent – Relation – Midazolam	No results
Midazolam – Relation – Kratom | Kratom constituent	1. Mitragynine inhibits midazolam 2. Mitragynine positively regulates midazolam
Kratom | Kratom constituent – Relation A – Enzyme or Transporter – Relation B – Midazolam	No results
Kratom | Kratom constituent – Relation A – Enzyme or Transporter, Midazolam – Relation B – Enzyme or Transporter	1. Mitragynine interacts with/inhibits/directly regulates activity of/directly positively regulates quantity of CYP3A4, Midazolam is substrate of/interacts with CYP3A4 2. Mitragynine inhibits/interacts with/positively regulates CYP2D6, Midazolam inhibits CYP2D6 3. Mitragynine inhibits/directly negatively regulates quantity of ABCB1, Midazolam interacts with ABCB1
Kratom – Quetiapine
Kratom | Kratom constituent – Relation – Quetiapine	No results
Quetiapine – Relation – Kratom | Kratom constituent	No results
Kratom | Kratom constituent – Relation A – Enzyme or Transporter – Relation B – Quetiapine	No results
Kratom | Kratom constituent – Relation A – Enzyme or Transporter, Quetiapine – Relation B – Enzyme or Transporter	1. Mitragynine interacts with/inhibits/directly regulates activity of/directly positively regulates quantity of CYP3A4, Quetiapine is substrate of CYP3A4 2. Mitragynine inhibits/directly negatively regulates quantity of P-glycoprotein, Quetiapine transports P-glycoprotein
Kratom – Venlafaxine
Kratom | Kratom constituent – Relation – Venlafaxine	No results
Venlafaxine – Relation – Kratom | Kratom constituent	No results
Kratom | Kratom constituent – Relation A – Enzyme or Transporter – Relation B – Venlafaxine	No results
Kratom | Kratom constituent – Relation A – Enzyme or Transporter, Venlafaxine – Relation B – Enzyme or Transporter	1. Mitragynine interacts with/inhibits/directly regulates activity of/directly positively regulates quantity of CYP3A4, Venlafaxine is substrate of CYP3A4 2. Mitragynine inhibits/interacts with/positively regulates CYP2D6, Venlafaxine is substrate of/inhibits CYP2D6

Table A.4.

Enzymes identified from meta-path discovery in NP-KG for green tea-nadolol.

Enzyme	OBO Identifier
cytochrome P450 1A1 (CYP1A1)	PR_000006101
cytochrome P450 1A2 (CYP1A2)	PR_000006102
cytochrome P450 2D6 (CYP2D6)	PR_000006121
cytochrome P450 2E1 (CYP2E1)	PR_000006122
cytochrome P450 3A4 (CYP3A4)	PR_000006130
cytochrome P450 7A1 (CYP7A1)	PR_000006148
sulfotransferase 1A1 (SULT1A1)	PR_000015818
UDP-glucuronosyltransferase 1A1 (UGT1A1)	PR_000017048

Table A.5.

Transporters identified from meta-path discovery in NP-KG for green tea-nadolol.

Transporter	OBO Identifier
high affinity copper uptake protein 1 (SLC31A1)	PR_000015083
ileal sodium/bile acid cotransporter (IBAT / SLC10A2)	PR_000014918
long-chain fatty acid transport protein 4 (FATP-4 / SLC27A4)
multidrug and toxin extrusion protein 2 (SLC47A2 / MATE2)	PR_000015153
multidrug resistance-associated protein 4 (MOAT-B / ABCC4)	PR_000003560
phospholipid-transporting ATPase ABCA1	PR_000003537
phospholipid-transporting ATPase IB (ATP8A2)	PR_000029291
sodium-dependent dopamine transporter (DA transporter / SLC6A3)	PR_Q01959
solute carrier family 22 member 6 (SLC22A6 / OAT1)	PR_000014993
solute carrier organic anion transporter family member 1A2 (SLCO1A2 / OATP1A2)	PR_000015222
solute carrier organic anion transporter family member 1B3 (SLCO1B3 / OATP1B3)	PR_000015224
zinc transporter ZIP4 (SLC39A4)	PR_000015131