EHR data.

The EHR data included approximately 1.2 million narrative patient notes from 2004 to 2010. Admission notes, discharge summaries and outpatient visits for a total of approximately 178,000 patients at New York Presbyterian Hospital were analyzed after obtaining IRB approval.

Reference standard dataset.

A reference standard of 253 drugs reported to cause the ADE pancreatitis was collected. Information in the reference standard was compiled from Micromedex/Drugdex, review articles, case reports, and reliable websites (the complete dataset and the references are given in Table S1). Drugs following two inclusion criteria were taken into account in the reference standard: 1) well-established by Micromedex (a trusted medical database), or by literature reviews or reports where the pancreatitis follows a reasonable temporal sequence from administration of the drug and the ADE is confirmed by de-challenge/re-challenge (cessation of the drug and the symptoms and new exposure to the drug and reappearance of the symptoms); 2) probable/possible where the drug is considered be responsible for the adverse effect but other possible causes are not totally excluded (more information is provided in Table S1).

Drug structure.

The DrugBank database [12] was used to obtain the structures of drugs included in this study. Some drugs whose structures were not available in DrugBank were manually represented using the Molecular Operating Environment (MOE) software [13].

EHR data mining.

The natural language processing (NLP) system MedLEE [14] was used to process the unstructured narrative data, and to extract and normalize relevant clinical entities, such as medications, diseases, symptoms, and associated temporal information. In addition, structured data was also obtained, consisting of abnormal laboratory test results that were associated with pancreatitis. Determining the pancreatitis outcome was based on laboratory test values [15], [16], which were extracted from structured EHR data (amylase ≥300 U/L or lipase ≥120 U/L), as these values are standardly used to determine potential occurrences of pancreatitis. Drug names were standardized and mapped to UMLS codes using MedLEE [14], and then generic names were obtained using RXNORM [17]. As an example, the brand name Videx was extracted from a note and then mapped to the UMLS code C0592249 by MedLEE, and subsequently RXNORM was used to map it to the generic name didanosine (UMLS code C0012133). A detailed process has been described in previous publications [18], [19]. The standard DPA method using the Odds Ratio (OR) measure was used to generate signals consisting of statistical associations between drugs and the event pancreatitis [9], [20]. DPA is based on frequency analysis consisting of 2×2 contingency tables containing drug-event pairs. OR is a measure of association between drugs and adverse events that can be calculated from the contingency table [9], [20]. Associations were quantified by the lower 5^th percentile of the Odds Ratio measure (OR05) [9]. To qualify as signals, associations had to meet two criteria: (1) OR05>1.25; (2) associations must pass a statistical test of independence P-value <0.05, based on the one-sided Fisher exact test and a Bonferroni correction, a conservative adjustment for multiple comparisons.

Molecular fingerprint-based modeling (MFBM).

The development process of this type of MFBM has been described in more detail in our previous publication [10]. The structures of the drugs involved in this study (i.e., the drugs in the reference standard and the candidate drugs generated by DPA method) were downloaded from the DrugBank database [12] and subjected to different preprocessing steps with the module Wash in the Molecular Operating Environment (MOE) software [13]. In this module, simple metal salts were disconnected and only the active ingredient was retained (i.e. the largest molecular fragment), the protonation state was considered neutral and explicit hydrogen atoms were added.

In the next step, four different molecular fingerprints were calculated for the drugs using the software MOE to carry out a comparative study of their performances: a) MACCS (MACCS structural keys), b) TGD (Typed Graph Distances), c) TGT (Typed Graph Triangles) and d) GpiDAPH3 (Graph 3-Point Pharmacophore) [13]. The basic idea is to represent a molecule using a bit vector that codifies the existence or absence of structural features, functional groups, pharmacophore features or molecular properties [21]–[25]. All the molecular fingerprints were calculated from the 2D molecular graph. MACCS fingerprints codify 166 structural keys. As an example, using MACCS fingerprints, some substructures represented in the molecule C₇H₁₃-NH₂ are: bit 19-seven membered ring, bit 84-NH₂ (amine group). TGD is a 2-point pharmacophore fingerprint codifying pair of atoms using graph distance and two atom types (the atom type could be donor, acceptor, polar, anion, cation, hydrophobe). TGT is a similar fingerprint but codifies triplets of atoms (3-point pharmacophore). GpiDAPH3 also codifies 3 pharmacophoric features calculated from the 2D molecular graph. All triplets of atoms are coded using graph distances and atom types (there are 8 possible atom types).

The final step consisted of similarity assessments between fingerprints of pairs of drugs using the Tanimoto coefficient (TC) measure of similarity. TC, also known as the Jaccard index, is one of the measurements most widely applied in the scientific literature for measuring similarity [24]. The range for the TC covers values from 0 to 1, where 0 means “minimum similarity” and 1 means “maximum similarity”. The TC between two fingerprints A and B is defined as:where N_A is the number of features present in fingerprint A, N_B is the number of features present in fingerprint B, and N_AB is the number of features present in both fingerprints A and B.

To compare a test set of drugs against the reference standard set of drugs known to cause pancreatitis, a similarity matrix file based on TC for all the drugs was calculated using the Fingerprint Cluster module and the sim_matrix2txt.svl script in MOE [13]. Therefore Sim_ij is the similarity score (TC values) between the ith drug and the jth drug in the matrix. The final similarity scoring provided by the model for a drug in the test is defined as the maximum pairwise TC obtained against each drug in the pancreatitis reference standard set. As an example, the drug entecavir in the test set was compared through TC to all the drugs in the reference standard and the maximum TC was considered the final score pointing out that the most similar drug in the reference standard was ganciclovir.

Evaluation.

Performance evaluation of the method using DPA by itself [9] was compared to the method that combined DPA with MFBM. The evaluation centered on the proportion of true signals identified by each of the approaches based on the reference standard.

DPA by itself.

The evaluation of DPA by itself was based on comparing the drugs that were selected with OR05 value greater than the previously established cutoff of 1.25 with the drugs in the pancreatitis reference standard dataset. Since only associations above the estimated cut-off were taken into account, the ROC curve cannot be plotted with all the drugs included in the EHR. For this reason the precision of the method (TP/TP+FP) was calculated as a standard comparative measurement of the performance. The method was compared to random results through one-sided Fisher’s exact test using the DrugBank database [12] as a resource, as explained below in Results.

Combination of DPA and MFBM.

All the drug candidates highlighted using DPA were subjected to MBFM-based analysis. As described previously, a comparison between molecular fingerprints from candidates selected by OR05 and the reference standard set of drugs was carried out through the calculation of the Tanimoto coefficient (TC). The maximum pairwise TC obtained against each drug in the pancreatitis reference standard set was considered to be the final similarity score.

The performance of the MFBM was assessed through a leave-one-out cross validation method. Each drug included in both sets (candidates selected through OR05 and already included in the reference standard set) was taken out and evaluated by the model in order to compare the performance with the rest of the candidate drugs selected by OR05. Precision-Recall and Receiver Operating Characteristic (ROC) curves were plotted, considering as true positives the drugs included in our reference standard and false positives the rest of candidate drugs.

When evaluating performance, the false positives according to our system are drugs not included in our reference standard. Nevertheless, it is possible that some of the drugs selected and not included in the reference standard are causally related to pancreatitis. For this reason, a bibliographic search using Micromedex/Drugdex database [26], [27] and case reports from the literature was carried out to confirm whether the drugs found by the method but not included in the reference standard dataset could be the cause of the adverse event under study (see Table S2). A final evaluation of the combined model has been completed for this test set through Precision-Recall and ROC curves.

Examples of drugs belonging to the same pharmacological category.

An example of a drug not included in the initial pancreatitis reference standard set is megestrol, a progesterone derivative with antineoplastic properties used in the treatment of carcinoma. According to our MFBM, megestrol is structurally similar to norethindrone, another progestogen drug used as oral contraceptive included in the reference standard (see Table 2 and Table S2). There are some publications that relate the drug norethindrone as a possible cause of pancreatitis with positive rechallenge [28]. Although not many reports were found relating megestrol as the cause of pancreatitis, some information is available establishing pancreatitis as a possible adverse event due to the use of contraceptive pills containing megestrol [29], [30]. However, further studies will be necessary to explain the potential adverse event in the drug.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Examples of candidates selected through the combination of DPA and MFBM (MACCS fingerprints) and similar molecules in the pancreatitis reference standard, along with OR05 (lower 5^th percentile of the Odds Ratio measure of association in DPA analysis) and TC (Tanimoto coefficient) values.

https://doi.org/10.1371/journal.pone.0041471.t002

Entecavir is a nucleoside analog used in clinic as an antiviral for the treatment of hepatitis B. A possible role of entecavir in the development of pancreatitis has been discussed previously [31]. Elevations in serum amylase were also reported in patients receiving entecavir although no clinical pancreatitis was diagnosed [32]. Our system detected a structural similarity between entecavir and ganciclovir, another antiviral drug used to treat cytomegalovirus infections which was included in the reference standard because it was identified in the literature and in Micromedex causally related to pancreatitis [26], [33].

Another case found in the EHR associated with pancreatitis is the polyene antifungal antibiotic drug amphotericin B. This drug was also detected by MFBM as a drug similar to doxorubicin, an anthracycline antibiotic used in clinic in the treatment of different types of cancer (see Table 2). Although amphotericin B could be the treatment for fungal infections associated to pancreatitis, there are some reports in the literature that confirm the potential risk of this drug. A report describes an increased serum lipase levels with clinical signs of pancreatitis in some patients treated with liposomal amphotericin B therapy [34]. Another report describes a case of pancreatitis in an HIV patient possibly due to amphotericin B [35].

Examples of drugs belonging to different pharmacological categories.

The combination of the DPA and MFBM can also detect drugs that belong to different pharmacological classes (see Table 2). An interesting example of drug pointed out by MFBM is loperamide, a piperidine derivative opioid-receptor agonist that is very effective for the treatment of diarrhea. The structure of loperamide is similar to haloperidol, a drug in the reference standard with different pharmacological category since haloperidol is a butyrophenone belonging to antipsychotic medications and used in the treatment of schizophrenia. There are some reports in the literature that confirm the potential adverse effect of lopiramide probably due to a spasm at the sphincter of Oddi or inhibition of the release of pancreatic polypeptide [36]–[38]. On the other hand, haloperidol was reported to cause pancreatitis in a study of antipsychotic drugs inducing pancreatitis [27], [39]. The temporal relationship between the adverse event and the beginning of the therapy could indicate a causal relationship.

Another example of similar drugs belonging to different pharmacological classes is pyrimethamine and lamotrigine (drug included in the reference standard). Pyrimethamine, an antiparasitic drug used in the treatment of malaria and toxoplasmosis, is selected by MFBM as a drug structurally related to lamotrigine (both drugs present a chloro substituted phenyl with a diamine azine ring), an anticonvulsant used to treat epilepsy, bipolar disorders and depression. Pancreatitis is described as a possible ADE for the combination pyrimethamine/sulfadoxine [40]. However, more studies will be necessary to confirm the ADE.

An interesting case to study further is micafungin, an antifungal drug indicated for the treatment of candidiasis by inhibiting the production of 1,3-beta-D-glucan, an important component of the fungal cell walls. Although fungi infections can occur to patients with pancreatitis, and therefore the prescription of an antifungal drug is necessary in some cases [41], there is also a report in the literature describing a case of acute pancreatitis probably caused by micafungin [42]. Our model detected that micafungin resembles the peptidic structure of ceruletide (see Table 2), a drug included in our reference standard and used in experimental animal models to induce pancreatitis [43]. Further studies will be necessary to shed some light about the relationship between micafungin and pancreatitis.

Cosyntropin, a derivative of adrenocorticotropic hormone used to diagnose cortisol disorders, is another case selected by DPA and MFBM as a drug related to pancreatitis. In this case, the information gathered in Micromedex database clearly indicates that pancreatitis is an ADE associated to the drug [26]. Cosyntropin presents some structural similarity regardless to caspofungin, an antifungal drug included in the reference standard set.