Leveraging a pharmacogenomics knowledgebase to formulate a drug response phenotype terminology for genomic medicine

Abstract

Motivation

Despite the increasing evidence of utility of genomic medicine in clinical practice, systematically integrating genomic medicine information and knowledge into clinical systems with a high-level of consistency, scalability and computability remains challenging. A comprehensive terminology is required for relevant concepts and the associated knowledge model for representing relationships. In this study, we leveraged PharmGKB, a comprehensive pharmacogenomics (PGx) knowledgebase, to formulate a terminology for drug response phenotypes that can represent relationships between genetic variants and treatments. We evaluated coverage of the terminology through manual review of a randomly selected subset of 200 sentences extracted from genetic reports that contained concepts for ‘Genes and Gene Products’ and ‘Treatments’.

Results

Results showed that our proposed drug response phenotype terminology could cover 96% of the drug response phenotypes in genetic reports. Among 18 653 sentences that contained both ‘Genes and Gene Products’ and ‘Treatments’, 3011 sentences were able to be mapped to a drug response phenotype in our proposed terminology, among which the most discussed drug response phenotypes were response (994), sensitivity (829) and survival (332). In addition, we were able to re-analyze genetic report context incorporating the proposed terminology and enrich our previously proposed PGx knowledge model to reveal relationships between genetic variants and treatments. In conclusion, we proposed a drug response phenotype terminology that enhanced structured knowledge representation of genomic medicine.

Supplementary information

Supplementary data are available at Bioinformatics online.

1 Introduction

Genomic Medicine is an emerging medical discipline that involves using patients’ genomic information to support diagnostic or therapeutic decision-making (NHGRI, 2020). Genomic Medicine has the potential to enable medical practitioners to improve selection of targeted therapies, avoid side effects from ineffective therapy, and improve cost-effectiveness (Chantrill et al., 2015; Evans and Relling, 2004; Krynetskiy and McDonnell, 2007; Trusheim et al., 2007). Genetic testing (germline or somatic) involves examining a person’s DNA (Green and Botkin, 2003), and can reveal variants in genes that may contribute to an increased risk of disease [i.e. predictive genetic tests (Holtzman et al., 1997)] or a different response to drug therapies [i.e. pharmacogenomics (PGx) (Relling and Evans, 2015)]. The advancement of next-generation sequencing and clinical genetic testing has played an increasingly important role in the practice of Genomic Medicine (Kurian et al., 2017; Lerman et al., 1996; Reyna et al., 2001).

With the advancement in Genomic Medicine, specialized knowledgebases such as PharmGKB (Barbarino et al., 2018), ClinVar (Landrum et al., 2018) and CIViC (Griffith et al., 2017) were established to curate large bodies of scientific evidence on genes, variants and their effects on drug-related phenotypes. Knowledgebases like these are essential to the interpretation of genetic test results and thus influence the implementation of Genomic Medicine. Curators contributing to the development of such knowledgebases employ different approaches to curate evidence from publications. For example, PharmGKB combined natural language processing (NLP) as well as expert review to achieve high-quality data curation (Altman, 2007).

Despite the growing body of literature and increasing evidence of clinical validity and utility (Crews et al., 2012; Mrazek and Lerman, 2011) of genetic variants that are associated with drug dose, efficacy and safety (Johnson et al., 2011) or risk of diseases (Ratner et al., 2019), translating such information into clinical practice at the point of care remains a significant challenge (Kannry and Williams, 2013; Kochan et al., 2020; Roden et al., 2012; Tenenbaum et al., 2016). In order to systematically integrate genetic information into electronic health records (EHRs) and translate it into actionable treatment recommendations with a high level of consistency, scalability and computability, a formal knowledge model is needed to represent genetic and clinical information (Marsolo and Spooner, 2013). A knowledge model is a computer interpretable model or schema that organizes entities (data) and their relationships to one another within a database or knowledgebase. From the database perspective, knowledge modeling is useful for abstracting and decomposing complex concepts and can address issues related to data integration and data curation (Storey and Song, 2017). From a knowledgebase perspective, it is an essential component in order for stored data to be used by other systems and enable data to be linked with standardized, computable genomic knowledge, which might exist at different levels (variant, gene or phenotype) (Williams et al., 2019).

Previously, we analyzed the sublanguage patterns of unstructured text sections in 1565 oncology genetics reports and proposed a knowledge model that leveraged the Unified Medical Language System (UMLS) (Aronson, 2006; Bodenreider, 2004). That model represented PGx knowledge in genetic reports (Zhao et al., 2020) and was centered on genetic information and attempts to capture other key elements such as treatment, disorders and pathways. While genetic information and disorders were connected through cellular pathways that inform treatment choices, the connections between genetic information and individualized treatment were not captured by this model.

The reason for the lack of connections between genetic information and individualized treatment in the previous knowledge model was due to the lack of a terminology for drug response phenotype, an essential element in PGx studies (Xie and Frueh, 2005). Previously, when we constructed the knowledge model for genetic reports, we utilized the UMLS to identify concepts and classify them into semantic groups. While major elements in our model could be captured by the UMLS with high coverage (e.g. ‘Disease’, ‘Gene and Gene Product’, ‘Treatment’), we also found that some UMLS semantic groups (e.g. ‘Concept’, ‘Function’) were vaguely defined, heterogeneously classified (McCray et al., 2001) and not domain-specific.

Therefore, the aims of this study were to assemble a drug response phenotype terminology that could be used to enrich our existing PGx knowledge model (Zhao et al., 2020). We utilized a widely used PGx knowledgebase—PharmGKB, which contained high-quality human-curated drug response statements—to formulate a drug phenotype terminology and then used this terminology to enrich relationships between genetic variants and treatment. The drug response terminology was validated by comparison to the CIViC knowledgebase. The enriched PGx knowledge model resulting from this work will further facilitate integration of PGx information from genetic reports into EHRs.

2 Materials and methods

In this study, our previously proposed knowledge model (Zhao et al., 2020) for oncology genetic reports was enriched in four steps: (i) formulate a drug response phenotype terminology through data-driven analysis of PharmGKB variant annotation statements, (ii) evaluate the coverage of drug response phenotype terminology in genetic reports, (iii) compare the terminology to the CIViC knowledgebase and (iv) perform semantic analysis of unstructured text in genetic reports leveraging the UMLS and drug response phenotype terminology.

2.1 Formulation of drug response phenotype terminology

The PGx findings curated by PharmGKB were categorized into five categories of evidence and were summarized using standardized vocabularies of genes, drugs and diseases (Sangkuhl et al., 2008). PharmGKB curated two types of statements from the published literature (Barbarino et al., 2018): variant annotation and clinical annotation. Variant annotations were semi-structured summary sentences that summarized main findings from studies pertaining to a study population—but did not represent broader, knowledge-level assertions about the variation generally. In contrast, clinical annotation statements were broader knowledge-level assertions that linked to relevant variant annotation statements. Variant annotation statements served as evidence to support clinical annotation statements. For both types of annotations, PharmGKB utilized standardized vocabularies in the fields (genes, drugs and diseases) that overlapped with our previously proposed PGx knowledge model.

PharmGKB annotations included additional content that represented content missing from our knowledge model; therefore, in this study, we analyzed PharmGKB variant annotation statements because they represented a more detailed level of evidence that originated from research studies. Most importantly, the PharmGKB clinical annotation and variant annotation statements shared the same set of terms used to describe drug response phenotype because the former was a high-level summary based on the latter.

We combined three files (va_drug_ann.tsv, va_fa_ann.tsv, va_pheno_ann.tsv) that were available from https://www.pharmgkb.org/downloads. Each file contained multiple entries of variant annotations together with relevant details including Annotation ID, Variant, Gene, Chemical, PMID, Phenotype Category (assigned by PharmGKB and identical to the ClinVar drug response trait), Significance, Annotation Sentences, Study Parameters, Alleles and Chromosome. PharmGKB also provided mappings to vocabularies for Variant, Gene, Drug, Chemical and Disorders (from variants.zip, gene.zip, drugs.zip, chemicals.zip and phenotypes.zip, respectively) that were used in their statements. Those data were used to perform an iterative sentence structure analysis using an NLP pipeline, as follows (shown in Fig. 1):

Fig. 1.

Workflow to analyze semantic structures of variant annotation statements in PharmGKB. (1) PharmGKB source files, (2) example of a PharmGKB variant annotation statement, (3) structured annotation representations by existing terminologies to (4) formulation of Drug Response Phenotype terminology

1. We manually analyzed 50 randomly selected statements and developed an initial sentence structure with ‘language blocks’ that contained common phrases such as ‘is associated with’, ‘as compared to’ and ‘when treated with’.

2. We extracted all the identified ‘language blocks’ and concepts from PharmGKB vocabularies (Variant, Gene, Chemical and Disorder). For each extracted concept, a semantic label was assigned to the concept (labeled as ‘$Concept|Semantic$’), which enabled automated post-processing of the results. Those labeled concepts were considered ‘semantic blocks’. For example, in the statement ‘Allele A is associated with decreased dose of warfarin’, the phrases ‘Allele_A’, ‘decreased dose’ and ‘warfarin’ were identified as concepts and ‘is associated with’ was identified as a ‘language block’. For those extracted concepts, the semantic labels that were assigned were: ‘GENO’ for ‘Allele_A’, ‘PHENO’ for ‘decreased_dose’ and ‘CHEM’ for ‘warfarin’. Therefore, the semantic blocks representing those concepts were ‘$Allele_A|GENO$’, ‘$decreased_dose|PHENO$’ and ‘$warfarin|CHEM$’.

3. Phrases that could not be automatically mapped to PharmGKB vocabularies were analyzed using the MetaMap API (Aronson, 2006). Concepts from four key semantic groups (Genes and Gene Products, Treatment/Chemicals, Pathway and Disorders) that were included in the ‘semantic blocks’ were annotated using our previously proposed UMLS-derived knowledge model for genetic report interpretations (Zhao et al., 2020). Examples of sentences structures are listed in Supplementary Table S1.

4. An additional set of 50 randomly selected sentences that contained phrases that were not successfully annotated by the NLP pipeline were manually reviewed to develop new semantic categories for those phrases. The results of the manual curation were added to the NLP pipeline to improve its ability to completely annotate the corpus. This was an iterative process: each time a new semantic category was identified, phrases that belonged to the new semantic category were effectively removed from the unprocessed subset of data, and the remaining phrases were re-analyzed.

2.2 Evaluation of coverage of drug response phenotype terminology in genetic reports

To evaluate the utilization of the terms that were extracted from the PharmGKB knowledgebase and that were related to drug response phenotype, within real world clinical data sources, we reviewed Foundation Medicine genetic reports and clinical notes of a cohort of patients who received Foundation Medicine tests between 2016 and 2019. The Foundation Medicine genetic reports used for this study were from three panels: FoundationOne CDx, FoundationOne Heme and FoundationOne Liquid. Altogether, the three panels tested a total of 709 genes, which provided a diverse array of types of gene products about which phenotypic statements could be made. An automated analysis was performed using unstructured text in the ‘Genetic information’ and ‘Treatment options’ sections of the reports, which were the major sections that contained interpreted genetic information and provided evidence to support proposed treatment options (Fig. 2a). To avoid double-counting, duplicate sentences that interpreted the same genetic variant for different patients were removed (Fig. 2b). The MetaMap API (Aronson, 2006) was utilized to extract UMLS-identifiable concepts and assign those concepts to semantic groups.

Fig. 2.

Workflow for semantic analysis of unstructured text of genetic reports. (a) XML parsing to extract unstructured text from specified sections of the report, (b) extraction of unstructured text from genetic reports, (c) extraction of UMLS-identifiable concepts and drug response phenotypes, (d) grouping semantic annotations and co-occurrence analysis and (e) Network visualization. See text for details

A manual, expert-driven analysis was also performed. A sample of 200 sentences was randomly selected from the unstructured corpus of genetic reports that contained concepts in both the ‘Genes and Gene Products’ and ‘Treatments’ semantic groups. One annotator with a medical background was given a dictionary containing candidate drug response phenotype terms and definitions that were defined from the previous step (as shown in Supplementary Table S2). The annotator was asked to identify any drug response phenotypes mentioned in the selected sentences. The annotator was also instructed to identify all drug response phenotypes, without being limited to those provided in the dictionary, which would identify any missed terms. The semantic meaning of the terms was considered during this process. For terms listed in the dictionary, the annotator needed to agree with our definition in order to qualify as a match.

2.3 Comparison of drug response phenotype terminology to the CIViC knowledgebase

To evaluate the generalizability of our proposed phenotype terminology to other genomic knowledgebases, we compared our proposed drug response phenotypes with terminology used in the CIViC knowledgebase. CIViC was chosen as a comparator because it is not PGx-centric, and it contains structured evidence statements. In particular, CIViC defines Clinical Significance statements that describe how a variant is related to a specific, clinicially relevant property within the context of cancer (https://civic.readthedocs.io/en/latest/model/evidence/clinical_significance.html). Within CIViC, the scope of Clinical Significance is quite broad, including concepts related to therapeutic response but also other concepts including those related to diagnosis, prognosis and function. The terms and categories of evidence defined by CIViC were compared to the terms derived from PharmGKB.

2.4 Semantic analysis of UMLS-identifiable concepts and drug response phenotypes

To extend the knowledge model for genetic reports, the corpus developed in the previous steps (Fig. 2a and b) was used as the data source for a semantic analysis that leveraged the unstructured text from genetic reports, the UMLS, and the proposed drug response phenotype terminology (Fig. 2c; Terminology is listed in Supplementary Table S2 Column 1). We utilized a keyword-based information extraction system to extract drug response phenotypes from unstructured text, which streamlined the process of sentence detection and text normalization before matching to the drug response phenotype terminology. Drug response phenotype and UMLS-identifiable concepts were assigned to their respective semantic group (Fig. 2d) and co -occurrence relationships of semantic groups that occurred in the same sentence were calculated (Fig. 2e). For drugs belonging to the same parent category (e.g. olaparib, rucaparib, niraparib, which are all PARP inhibitors), only one ‘Treatment’ occurrence was counted in the semantic sentence. A semantic network (knowledge model) was built based on the co-occurrence of semantic groups.

2.5 Analysis of knowledge model on the diagnosis phase and the treatment phase

Our previously proposed knowledge model (Zhao et al., 2020) and the enriched knowledge model proposed in the previous section each modeled only static snapshots of knowledge about genomic medicine. Based on our previous work, ‘topic’ differences exist in clinical narratives (Zhao et al., 2021), which represent clinical context in different clinical phases. Therefore, text from the ‘Genetic information’ and ‘Treatment options’ sections of the clinical reports were analyzed using the same approaches as described above, with the goal of capturing temporal differences in the knowledge model that represent the early phases (e.g. prognosis, diagnosis, risk analysis) and later phases (e.g. treatment) of clinical care, respectively.

3 Results

3.1 Construction of drug response phenotype terminology

PharmGKB provided three data files that included statements related to variant drug annotation (9742 statements), variant functional assay annotation (1591 statements) and variant phenotype annotation (11 278 statements). The entire contents of all three files were analyzed, so in total 22 611 variant annotation statements were examined. PharmGKB framed its annotations as ternary relationships between a Genotype, a (Disease) Phenotype and a Chemical/Gene_product. In some cases, additional information about Population and a second Genotype to support comparative statements was included. In our manual review, two major semantic categories were identified for phrases that were not initially covered by the automated pipeline: ‘Drug Response Phenotype’ and ‘Population’. In addition, some of the ‘leftover’ phrases could not be categorized into either the ‘Drug Response Phenotype’ or ‘Population’ categories; examples include ‘fast metabolizer phenotype’ (when that phrase was used as an interpretation of a genotype rather than as a phrase capturing a drug response phenotype), ‘as determined luciferase reporter assays’ and ‘artificial constructs expressed’. It would be possible to represent those phrases as modifiers through extensions to the core statement model, but their subject matter was determined to be not specific to drug response so they were excluded from further analysis.

Table 1 contains a list of 35 ‘Drug Response Phenotype’ terms extracted from PharmGKB variant annotation statements. Supplementary Table S2 contains definitions for those terms from the National Cancer Institute Thesaurus (NCIT) and semantic categories from the UMLS; for the terms that definitions cannot be found in NCIT, definitions were derived from other academic publications. New definitions that more accurately reflected the usage of the terms within the PGx context were developed only when necessary and through a consensus-building process by the authors.

Table 1.

Sentence count summary of drug response phenotypes in unstructured text of genetic reports and PharmGKB (sorted by descending order of Genetic Reports sentence count)

Drug response phenotypes	Genetic reports	PharmGKB
Response	994	4575
Sensitivity	829	108
Survival	332	729
Resistance	252	136
Activity	184	692
Expression	163	246
Clinical benefit	101	5
Efficacy	97	10
Progression	69	7
Inhibition	42	22
Toxicity	27	243
Prognosis	26	8
Dose	25	1781
Reduction	17	115
Risk	16	4733
Likelihood	15	1610
Concentration	13	1155
Transcription	9	56
Uptake	6	16
Formation	5	21
Discontinuation	2	54
Exposure	2	224
Intolerance	2	17
Recurrence	2	16
Metabolism	1	113
AUC	0	68
Glucoruonidation	0	14
Steady-state level	0	29
Sulfation	0	67
Affinity	0	9
Clearance	0	957
Half life	0	2
Protein stability	0	6
Severity	0	628
Transport	0	23

Note: Example sentences from the unstructured text of genetic reports available in Supplementary Table S3).

Among 22 611 variant annotation statements, 18 495 (81.8%) statements contained a matching ‘Drug Response Phenotype’ term. The remainder of the statements (18.2%) were manually examined. We found that the genotype–drug relationships represented in those statements were related to either adverse drug events or changed physiology conditions, which were classified semantically by the UMLS as ‘Disease’ (e.g. major adverse cardiac events, ventricular arrhythmia, electrocardiogram QT prolonged). In the scope of this work, adverse drug events and changed physiology conditions were excluded from the core drug response phenotype terminology.

3.2 Coverage of drug response phenotype in genetic reports

There were 3481 sentences from the unstructured corpus of genetic reports where both ‘Genes and Gene Products’ and ‘Treatments’ were mentioned. A summary of drug response phenotypes in unstructured text of genetic reports were listed in Table 1. Note that a single sentence could contain multiple drug response phenotypes. The distribution of the number of drug response phenotypes in unstructured text of genetic reports was determined (Fig. 3). Most sentences contained 1 drug response phenotype while 31% sentences contained 0 drug response phenotype terms.

Fig. 3.

Distribution of number of drug response phenotypes in genetic reports sentences

Our manual evaluation using the randomly selected 200 sentences revealed that our proposed drug response phenotype terminology covered 96% of the terms that were identified in clinical genetic reports. Missed terms that were identified by the human annotator came from two majorsources. First, in some sentences, the keywords were not used in the same way as we defined. For example, ‘activity’ as defined in PharmGKB statements referred to ‘Function of a biological molecule (usually protein, RNA or DNA) that possesses catalytic activity.’ However, ‘activity’ had a broader use in genetic reports and was not restricted to catalytic activity (e.g. ‘A trial everolimus monotherapy for non-small cell lung cancer showed modest activity’). In another example, a genetic report contained the sentence ‘Ruxolitinib is a kinase inhibitor that targets JAK1 and JAK2 and is approved to treat intermediate or high-risk myelofibrosis’, where ‘risk’ referred to disease subtype rather than the result of a therapeutic action. Second, drug phenotypes were missed as a result of the use of abbreviations and variety in expression. For example, ‘PFS’ (progression-free survival) and ‘OS’ (overall survival) were both used as drug response phenotypes and also as survival measures, and ‘disease control rate’ was also used as a drug response phenotype in genetic reports. Therefore, in Table 1, we added counts from ‘PFS’, ‘OS’ and ‘disease control rate’. After iterative error analysis and corresponding improvements to the NLP pipeline, our system was able to identify 100% of the drug response phenotype terms in this corpus.

3.3 Comparison of drug response phenotype and terminology of the CIViC knowledgebase

We compared our proposed drug response phenotype terms with the terms used in the CIViC knowledgebase. We found our proposed phenotypes were similar to CIViC Evidence terms for ‘Clinical Significance’ in some categories, but we noted differences in others. CIViC describes six major Evidence categories: predictive, prognostic, diagnosis, predisposing, oncogenic and functional. Since two of these categories (‘predisposing’ and ‘oncogenic’) have no associated ‘Clinical Significance’ terms, the remaining four categories and their terms were compared to our proposed phenotypes. In the ‘predictive’ and ‘prognostic’ categories, the CIViC terms ‘sensitivity’, ‘response’ and ‘resistance’ were in common with our proposed drug response phenotypes, whereas terms in CIViC’s ‘diagnostic’ and ‘functional’ category shared no overlap with our proposed drug response phenotype. The ‘prognostic’ category has a similar concept to our ‘outcomes’ drug response phenotype but has an important distinction in that the CIViC concept is not associated with a particular therapeutic and is therefore not a drug response phenotype. This reflects the semantic distinction between ‘prognostic’ and ‘predictive’ evidence in CIViC and similar cancer variant interpretation knowledgebases. Regarding terms in the ‘functional’ category, our phenotypes and CIViC’s terms focused on different aspects of functional change: our phenotypes emphasized changes in the gene/protein functions will further affect drug metabolism while CIViC’s terms focused on the broader gain or loss of non-descript functional properties.

3.4 Semantic analysis of unstructured text of genetic reports

We retrieved 16 744 non-duplicate sentences from the ‘Genetic information’ (11 451 sentences) and ‘Treatment options’ (5293 sentences) sections from 1565 patients’ genetic reports. From these sentences, we utilized the MetaMap API and extracted 9721 UMLS-identifiable concepts that were mapped to 115 of the 127 semantic types in the UMLS. Those concepts were grouped into 20 semantic categories. On average, each sentence had 10.7 words.

Subgraphs of the top 50 most frequently co-occurred relationships (edges) and the associated semantic categories (nodes) from previous and current knowledge models were plotted in Figure 4a1 and a2. Figure 4b1 and b2 shows examples of expanded subgraphs for the term ‘PIK3CA’, highlighting functional and pathway roles of PIK3CA. As part of the PI3K signaling pathway (Di Nicolantonio et al., 2010), PIK3CA is linked to the available targeted treatment ‘everolimus’ (Loi et al., 2013) and clinical trials related to ‘temsirolimus’ (Fleming et al., 2012; Janku et al., 2012). With the addition of ‘Drug response phenotype’ to the knowledge model, additional information can be represented regarding PIK3CA’s association with response to mTOR inhibitors (Janku et al., 2012) and the usage of those drugs in treating patients who develop resistance to HER2 inhibitors (Nahta and O'Regan, 2010).

Fig. 4.

(a1) Knowledge model subgraph (top 50 edges); Key elements for a patient’s genetic profile are Disorders, Pathway, Treatments, and Genes and Gene Products. (a2) Knowledge model subgraph (top 50 edges); the thickness of the edges in the network represents frequencies of co-occurrence associations between two groups. Size of the nodes represent node degree and color gradient represents the distance (by edge weight) between each node and ‘Drug Phenotype’; darker color represents closer associations, as for Drug Response Phenotype and Procedures. (b1 and b2) Expanded subgraph for term ‘PIK3CA’ showing the addition of the new node ‘Drug Response Phenotype’. (a1) and (b1) are from a previous publication (Zhao et al., 2020)

3.5 Analysis of knowledge model on the diagnosis phase and the treatment phase

Leveraging our previously proposed ‘topics’, or ‘clinical phases’, text from the ‘Genetic information’ section was analyzed to identify and represent information essential for phases from prognosis to test result discussion (Phase 1–3 in Fig. 5) and text from the ‘Treatment options’ section was analyzed to identify and represent information essential to medication prescription and clinical outcome (Phase 4–5 in Fig. 5).

Fig. 5.

Dynamic knowledge model for genomic medicine (nodes highlighted in red are significant in the knowledge model)

We found that in model (a), knowledge was centered (highlighted in red) around genetic variants, risk of cancer/metastasis and pathways that were linked to responsiveness or resistant to specific treatment. In model (b), drug phenotypes and procedures were more significant (illustrated in a comparable size to ‘Gene and Gene Product’). There were also new nodes in the knowledge model: ‘Organization’—which represented guideline and regulatory agencies' role in the treatment process; and ‘Disease Model’—which referred to the source of evidence for drug response phenotypes.

4 Discussion

Previously, we proposed a knowledge model to represent PGx knowledge (Zhao et al., 2020), which captured key elements such as treatment, disorders and pathways. However, the connections between genetic information and individualized treatment were not captured by this model. In this study, we formulated a drug response phenotype terminology to enrich our existing PGx knowledge model. There have been several attempts to generate PGx ontologies, including the Suggested Ontology for Pharmacogenomics (SO-Pharm) (Coulet et al., 2006) and Pharmacogenomics Ontology (PO) (Dumontier and Villanueva-Rosales, 2009). Both SO-Pharm and PO provided web standards for integrating and representing PGx knowledge through the web ontology language (OWL) (Antoniou and Van Harmelen, 2004). Samwald et al. (2015) also proposed a PGx ontology and already made progress in integrating genomic clinical decision support systems (CDSS) and OWL reasoning into a clinical application, and offered clinical recommendations based on patient information. However, none of those studies provided a terminology for drug response phenotypes. While the existing systems could be used to generate treatment recommendations, further refinement of recommendations would be needed to prioritize them and provide actionable recommendations (Monnin et al., 2019). This required incorporation of drug response information in CDSS. Another study by Monnin et al. proposed a simple ontology that represents PGx knowledge—PGxO (Monnin et al., 2019). While the PGxO contained terms related to drug response phenotypes, the four terms that were included (Efficacy, Toxicity/ADR, Metabolism/PK, Dosage, Other) did not represent the variety of concepts and level of detail of terms that were used in PharmGKB statements and genetic reports.

We utilized PharmGKB variant annotation statements for this study, but we did not import existing PharmGKB vocabularies directly. Rather, we analyzed PharmGKB variant annotation statements using an NLP pipeline and formulated a drug response phenotype terminology through sublanguage patterns mining. From our analysis of PharmGKB variant annotation statements, we were able to identify 35 drug response phenotypes, 29 of which had a matching definition in NCIT. For some of the phenotypes that were modeled, more detailed variations of the phenotype exist. For example, ‘survival’ can include sub-concepts such as ‘overall survival’ and ‘progression-free survival’. Some NCIT definitions were defined from an individual patient perspective, which did not match our PGx context completely. Since our proposed drug response phenotype terminology aimed to serve knowledge models within the PGx context, which was more focused on describing drug response on a population level, we developed our own definitions to meet this need. NCIT definitions, however, were still useful in describing an individual-level understanding of those terms, and thus they were provided in Supplementary Table S2.

In this first version of the drug response phenotypes terminology, we focused on establishing the main concepts; the potential hierarchical relationships of terms will be explored in future work. In addition, this study examined genetic reports that were focused on oncology treatment. While we anticipate that a significant proportion of the statements that were analyzed would be generalizable beyond oncology, future studies that examined other types of reports would be needed to confirm that hypothesis.

The PGx research community is well aware of the challenges in our ability to integrate molecular test results with patients’ EHRs, which prevent the realization of the full potential of PGx knowledge for both clinical practice and translational research (Boyce et al., 2013; Samwald et al., 2012). This is due in part to the gap between knowledgebases and clinical decision-making information needs. Therefore, in this study we performed sublanguage analysis to understand content differences between two data sources: genetic reports and knowledgebases. In Table 1, we observed there is a difference in distribution of drug response phenotypes discussed between the two data sources. In genetic reports, the most frequently discussed drug response phenotypes were related to outcomes, such as ‘response’, ‘sensitivity’ and ‘survival’. Most pharmacokinetics-related drug response phenotypes were not discussed as frequently in reports as they were in the knowledgebase; there were approximately 1000 statements related to ‘concentration’ and ‘clearance’ in the knowledgebase but those concepts were discussed very rarely (<5 times) in genetic reports. ‘Dose’ and ‘likelihood’ are also discussed less frequently in genetic reports than in the knowledgebase. Since ‘likelihood’ and ‘risk’ have overlap in meaning, this may just reflect sublanguage differences between data sources. Another possible example of term preference between data sources includes ‘efficacy’ and ‘clinical benefit’, which were discussed more frequently in genetic reports than in knowledgebases.

Our work has several limitations: (i) variants located in non-coding regions can have indirect but significant effects on drug response. For example, variants present in one gene can affect the phenotype of the product of another gene. Our knowledge model does not currently capture those complicated relationships because most PGx statements documented variants within coding regions using a simple one-to-one (gene-phenotype) relationship. (ii) In many clinical studies, drugs are given in combination and drug response phenotype can be affected by a drug-drug interactions that were not represented in statements in knowledgebases or addressed in the present manuscript. (iii) The analysis of drug response phenotypes described here focused on the characterization of terms currently in use and does not disambiguate the semantics of the statements themselves, which would be best addressed by the authors of drug response statements. (iv) Each genetic knowledgebase adopts a different data and terminology model. While we performed an in-depth analysis of one and used another for validation, additional knowledgebases could provide further support for the proposed terminology.

Currently, we are validating our terminology using other databases, including, CIViC and ClinVar. From the comparison of our knowledge model and CIViC’s evidence model, we found our knowledge model to share similar components with CIViC’s evidence model, which includes ‘Gene/Variants’, ‘Treatment’ and ‘Disorders’. In addition, our model includes one new key element: ‘Pathway’, which extends the proposed knowledge model beyond implications for treatment effectiveness to include interpretations for the mechanism of gene-treatment relationship. From the comparison of our proposed drug response phenotypes and CIViC’s Clinical Significance terms, we found both similarities and differences. While this may be due to the difference in scope between CIViC’s Clinical Significance terms and our drug response phenotype, we believe that over time ongoing efforts within the genomic medicine community will continue to improve harmonization of terminology in use, which will be reflected both in our drug response phenotype terminology and in our corresponding knowledge model.

In addition, we note that there is an important distinction between evidence for pharmacogenomic knowledge (as modeled in this work) and the clinical classification of variants under professional society guidelines and informed by this evidence. Both CIViC therapeutic response evidence, and some of the evidence used in this model, may be used to inform clinical classifications. However, much work remains to be done to formalize the application of the evidence statements from this study under clinical classification guidelines.

This work is not intended to assert or constrain the terminology used by experts that author drug response statements. Instead, this project analyzed a large corpus of statements to identify the terminology currently in use and created a computational model for capturing the content of drug response statements, which also contributed to a drug response knowledge model. The terminology used in drug response statements could be evaluated and harmonized through consensus to help improve consistency among drug response statements, similar to a terminology harmonization effort conducted by the Clinical Pharmacogenetics Implementation Consortium (CPIC) (Caudle et al., 2017), a process that could be informed by the analysis and results described here.

We anticipate that our proposed drug response phenotype terminology and the development of a more complete PGx knowledge model represents an important step in the harmonization of evidence and knowledge statements. Creating structured statements using common terminology will improve the exchange of knowledge between genomic medicine organizations, knowledgebases and clinical practices, facilitate the delivery of that knowledge to consumers, and increase the utilization of genomic knowledge in clinical decision making.

5 Conclusions

In conclusion, we leveraged PharmGKB, a comprehensive PGx knowledgebase, to formulate terminologies for drug response phenotypes. The terminology achieved a coverage of 96% drug response phenotype concepts in genetic reports. Utilizing the terminology, we were able to observe a sublanguage difference in distribution of drug response phenotypes between genetic reports and knowledgebases, which signified a need to bridge information provided in the knowledgebase and information needs in a clinical setting. Moreover, we were able to enrich our existing PGx knowledge model using the new terminology. Therefore, our proposed drug response phenotype terminology is an important step toward filling the gaps in current Genomic Medicine knowledge representation.

Data Availability

The data underlying this article are available in Zenodo, at https://doi.org/10.5281/zenodo.7186984. The source data from PharmGKB are licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.

Funding

This study was supported by the Mayo Clinic Center for Individualized Medicine, Mayo Clinic Division of Digital Health Sciences, Subcontract 5000821-5500001377 under Prime 75N97019P00280 [R.R.F.], and Subcontract 776-01280-000-107 under Prime GS35F0565T/D15PD00739 [R.R.F.]. Research reported in this publication was supported by the National Human Genome Research Institute of the National Institutes of Health under Award Number R35HG011899. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health, or the Departments of Pediatrics and Biomedical Informatics of the Ohio State University College of Medicine.

Conflict of Interest: none declared.