While drugs have been widely and successfully used to treat various diseases, most drugs cause different adverse events (AEs), commonly called adverse drug reactions (ADRs). These ADRs are sometimes severe and significantly affect public health. Indeed, ADRs are listed as the fourth killer after heart disease, cancer, and stroke [1]. Therefore, it is critical to carefully study the ADRs and underlying mechanisms.

Multiple studies have been conducted to automatically identify ADRs in text using Natural Language Processing (NLP) techniques. Different types of data sources such as electronic health records [2], scientific publications, and social media data have been used to extract ADRs. A lexicon of ADR-related terms and concepts was compiled from different sources such as the Unified Medical Language System (UMLS) [3] and the side effect resource (SIDER) [4] and was used to match the ADR mentions in user comments retrieved from DailyStrength (http://www.dailystrength.org) by Leaman et al. [5]. Nikfarjam and Gonzalez used the same user comment data set and developed an association rule mining approach to tag ADR mentions [6]. Similarly to Leaman et al., Gurulingappa et al. [7] also developed a lexicon-based matching approach to identify ADRs in text using the lexicon created based on the Medical Dictionary for Regulatory Activities (MedDRA) [8] and DrugBank [9]. However, rather than using user comments from social media, Gurulingappa et al. used the abstracts of case reports as their data source. Product labels have also been used as data sources to extract ADRs and create knowledge bases of known ADRs [10, 11]. A review of recent techniques on ADR extraction from text from various data sources is available in [12].

An important group of ADRs is neuropathy. Using FDA reported package insert documents and drug safety records, our previous studies identified 242 neuropathy-inducing drugs (NIDs) through mining various public resources and drug labels [13, 14]. We have previously developed an Ontology of Drug Neuropathy Adverse Events (ODNAE) that ontologically represents 214 NIDs, corresponding chemicals of these drugs, chemical function, adverse events associated with these drugs, and various other chemical characteristics [14]. Our study also showed that ODNAE provides an ideal platform to systematically represent and analyze AEs associated with neuropathy-inducing drugs and generate new scientific insights and hypotheses [14]. One weakness of the ODNAE study is that ODNAE only collects neuropathy-related ADRs commonly found in drug package insert documents and misses the collection of non-neuropathy ADRs from different sources.

In addition to enhanced literature mining, ontology can also be used for advanced class effect analysis. Specifically, an AE-specific drug class effect is defined to exist when all the drugs in a specific drug class (or drug group) are associated with an AE. In a recent study on cardiovascular drug-associated AEs, a proportional class-level ratio (PCR) value was defined and used to identify drug class effect on different AEs [15]. Specifically, when the PCR value equals to 1, it means that a class effect of a group of drugs on a specific AE exists. Previous PCR-based heatmap analyses identified many important drug class effects on different AEs [15].

In addition to the official FDA drug package insert documents, FDA also collects large amounts of spontaneous ADR case reports. To better understand these case report data, it is critical to use standardized terminologies or ontologies to identify drugs, ADRs, and associated data from the text reports. Therefore, ontology-based literature mining becomes critical. Previously, we applied the Vaccine Ontology (VO) [16] to enhance our literature mining of interferon-gamma related [17], Brucella-related [18], and fever-related [19] gene interaction networks in the context of vaccines and vaccinations. In these studies, we used and expanded SciMiner [20], a literature mining program with a focus on scientific article mining. SciMiner uses both dictionary- and rule-based strategies for literature mining [20].

To better study biological interaction networks, we have also developed a literature mining strategy CONDL, or Centrality and Ontology-based Network Discovery using Literature data [19]. The centrality analysis here refers to the application of different centrality measures to calculate the most important genes (i.e., hub genes) of the resulting gene-gene interaction network out of biomedical literature mining. Centrality measures, including degree, eigenvector, closeness, and betweenness, have been studied [19, 21]. The CONDL strategy was applied to extract and analyze IFN-γ and vaccine-related gene interaction network [21] and vaccine and fever-related gene interaction network [19], and our results showed that centrality analyses could identify important genes and raise novel hypotheses based on literature mined gene interaction networks.

The main purpose of this study was to develop a CONDL method for literature mining of all ADRs associated with neuropathy inducing drugs (NIDs) and used the mined results for systematic network and class effect analyses. Using MedDRA [8], ODNAE [14], Chemical Entities of Biological Interest (ChEBI) [22], and Ontology of Adverse Events (OAE) [23], we developed an ontology-based ADR-SciMiner tool for identifying ADRs from drug labels and applied it to NIDs to ontologically model their ADR-associated characteristics. The literature mined results were then used for ontology-based class effect analysis, leading to new scientific discoveries.

The contributions of this study are multiple fold. First, we developed and applied an ontology-based SciMiner literature mining approach, which was then used to mine the FDA TAC 2017 dataset. It is a huge challenge to identify all ADRs using textual description of ADR case reports. Our MedDRA/OAE-based SciMiner literature mining approach was successfully used to mine the FDA TAC 2017 dataset with a special focus on 53 neuropathy-inducing drugs (NIDs). Our study demonstrates the important role of the MedDRA controlled terminology and ontologies (e.g., ChEBI, OAE, and ODNAE) in the literature mining and further ADR analysis. Second, we constructed an ADR-ADR network and applied centrality analysis to identify the hub ADRs in the network. Third, among the 53 NIDs, our ChEBI-based analysis found three benzimidazole drugs, which formed a drug class effect on 43 ADRs. An OAE analysis of these ADRs further identified many enriched ADR categories. Based on the results, we can hypothesize that the proton-pump inhibition role, common to all the three benzimidazole drugs, might participate in different pathways leading to these ADRs. To our knowledge, our study represents the first of such literature mining-derived ontology-based drug class effect analysis.

The present study is based on a subset of US FDA drug labels, which was included in the 2017 Text Analysis Conference (TAC) Adverse Drug Reaction Extraction from Drug Labels track. We used this data set as a proof of concept as well as to develop a prototype version of ADR-SciMiner. We assumed that if an ADR is mentioned in the file of a drug, it is associated with the drug. However, it is likely that the ADR occurs within a negation or speculation statement such as ‘depression was not observed as an ADR of the drug’ or ‘depression might be an ADR of the drug’. Therefore, more semantic oriented NLP analysis techniques may be developed to identify whether an ADR is really associated with a drug or not.

To identify the most salient ADRs associated with NIDs, we created ADR-ADR networks both specific to NIDs and non-NIDs using a threshold of 50% for association. In other words, two ADRs were connected by an edge, if they co-occur in at least 50% of the NIDs or non-NIDs. Six of the central ADRs in the NID specific network were also included in the non-NID specific network, showing that these are prevalent and commonly occur together both in NID and non-NID cases. The other ADRs in Table 2 are central only in the NID associated network, which might reveal that they are more NID specific. As future work, we plan to extend the network analysis by including the specific drugs to the network as well and creating bipartite drug-ADR networks. The types of relations between drugs and ADRs can be identified by using the Interaction Network Ontology (INO) [24].

Our study identified three benzimidazole drugs (i.e. lansoprazole, pantoprazole, and omeprazole) that induce similar profiles of ADRs. Overall these three drugs have been found safe in terms of their associated ADR reports [33–35]. For example, a previous study with 10,008 users of lansoprazole in daily practice indicated that the most frequently reported lansoprazole ADRs were diarrhoea, headache, nausea, skin disorders, dizziness, and generalized abdominal pain/cramps, but no evidence of rare ADRs were found [33]. Current study found many ADRs associated with each of these three drugs, and all these three drugs are associated with 43 ADRs, commonly behavioral and neurological, digestive, muscular, and skin ADRs. A common reason for stopping pantoprazole usage was found to be the diarrhea ADR [34], which is also listed as one of the 43 ADRs.

A previous study suggested that these three drugs have similar profiles to interact with other drugs (most commonly vitamin K antagonist), suggesting a class effect [36]. According to the ODNAE records [14], lansoprazole, omeprazole, and pantoprazole are all associated with neuropathy adverse events. Our study found 43 AEs commonly shared with these three benzimidazole drugs. Interestingly, many of these AEs are also found to be the hubs of the highly enriched NID network from our literature mining data centrality analysis. It is likely that these three benzimidazole drugs, which function as proton-pump inhibitors, use the same or similar pathways to induce neuropathy adverse events.

It is noted that the ontology-based drug class effect study is novel in many aspects compared to its original report [15]. First, compared to the previous report using the drug package insert information, our study uses the data generated from literature mining of FDA provided case report data. Second, given the large size of AE data for each vaccine, we were able to identify many AEs commonly used by a class of drugs, in our case, 43 AEs associated with the three benzimidazole drugs. Our OAE-based analysis was able to further identify the common patterns among these AEs. Such a high throughput study was not reported in the previous package insert document-based studies.

The ADR identification performance is not yet optimal and there is still much room for improvement. The majority of falsely identified ADR terms by SciMiner could be grouped into three types: (1) incorrect mapping of acronyms to ADRs (e.g., ‘all’, as in ‘all patients’, mapped to ‘acute lymphocytic leukaemia’); (2) ADR that may not be caused by the current drug (e.g., ‘caution is needed in patients with diabetes’); and (3) ADRs that occur as discontinuous entities in text (e.g., ‘corneal ulceration’ is an ADR, but does not occur as a continuous text fragment in ‘corneal exposure and ulceration’). Integration of other dictionaries such as SNOMED CT [37] into ADR-SciMiner will be explored to possibly expand the ADR dictionary thus to improve the recall. Identifying whether a term is an acronym for an ADR or not, determining whether an ADR that occurs in a drug label is really caused by that drug, and detecting ADRs that occur as discontinuous text fragments in text require deeper semantic understanding of the sentences by considering the context information (i.e., the surrounding words) of an ADR in text. Our current method is a dictionary and rule-based method, which does not consider the context of an ADR occurrence in text. These challenges can be tackled by using machine learning methods with features that capture context information and utilize the syntactic analysis of the sentences such as their dependency parses.

As future work, we plan to develop machine learning based methods to improve the accuracy of ADR tagging as well as the detection of the associations between ADRs and drugs. We will also extend our approach to include all available structured drug labels in the DailyMed database, maintained by National Institute of Health. DailyMed currently contains listings of 95,513 drugs submitted to the US FDA, about 28,000 of which are prescription drugs for human. Our ontological study of NIDs will be extended using this larger drug label dataset.

In this study we developed an MedDRA and ontology-based SciMiner literature mining pipeline, applied the pipeline to mine a FDA text set for ADRs associated with neuropathy-inducing drugs, performed centrality network analysis, and drug class effect studies. Our approach identified scientific insights regarding these drug-specific ADRs. Our study demonstrates the feasibility of using ontology-based literature mining, network analysis, and drug class effect classification to efficiently identify and study specific drugs and their associated ADRs.