Detecting miRNA Mentions and Relations in Biomedical Literature

Related work

Text-mining technologies are established for a variety of applications. For instance, the BioCreative competition^18,19 and BioNLP Shared Task^20–22 series have been conducted to benchmark text mining techniques for gene mention identification, protein-protein relation extraction and event extraction, among others.

To our knowledge, only limited work has been carried out in the area of miRNA-related text-mining. Murray et al. considered miRNA-gene associations from PubMed database using semantic search techniques²³. For their analysis, experimentally derived datasets were examined, combined with network analysis and ontological enrichment. Regular expressions were used to detect miRNA mentions. The authors claim to have optimized the approach to reach 100% accuracy and recall for detecting miRNAs mentions as in miRBase. Relations were identified based on a manually curated rule set. The authors extracted 1165 associations between 270 miRNAs and 581 genes from the whole MEDLINE.

The freely available miRSel¹² database integrates automatically extracted miRNA-target relationships from PubMed abstracts. A set of regular expressions is used for miRNA recognition that matches all miRBase synonyms and generic occurrences. The authors reach a recall of 0.96 and precision of 1.0 on 50 manually annotated abstracts for miRNA mention identification. Further, the relations between miRNA and genes were extracted at sentence level employing a rule-based approach. They evaluated on 89 sentences from 50 abstracts resulting in a recall of 0.90 and precision of 0.65. Currently, it hosts 3690 miRNA-gene interactions¹¹.

Since the miRNA naming convention has been formalized very early in comparison to other biological entities such as genes and proteins, applying text-mining approaches is relatively simple¹⁷. Thus, most of the previously applied text mining approaches for miRNA detection has been based on keywords. miRCancer uses keywords to obtain abstracts from PubMed, further miRNA entities have been identified using regular expressions based on prefix and suffix variations. Similarly, miRWalk database uses keyword search approach to download abstracts and applies a curated dictionary (compiled from six databases) for miRNA identification of human, rat, and mouse species²⁴. TarBase, miR2Disease, miRTarBase, and several others have followed related search strategies. However, several authors still tend to use naming variations for acronyms, abbreviations, nested representations, etc. for listing miRNAs. Additionally, in contrast to the previous text-mining approaches focusing purely on miRNA gene relations, we extend the information extraction approach additionally to retrieve miRNA-disease relations. Furthermore, we evaluate our approach using a larger corpus to achieve robustness. We differentiate between actual miRNA mentions (refered to as Specific miRNAs) and co-referencing miRNAs (Non-Specific miRNAs), which could in addition enhance keyword search. We evaluated three different relation extraction approaches, namely co-occurrence, tri-occurrence and machine learning based methods.

To support further research, our corpora are made publicly available in an established XML format as proposed by Pyysalo et al.²⁵, as well as the regular expressions used for miRNAs named entity recognition. In addition, our dictionary for trigger term detection and general miRNA mention identification are made available. To our knowledge, the annotated corpora as well as the information extraction resources are the most comprehensive developed so far.

Data curation and corpus selection

Named entities annotation. Mentions of miRNAs consisting of keywords (case-insensitive and not containing any suffixed numerical identifier) such as “Micro-RNAs” or “miRs” are annotated as Non-Specific miRNA. Names of particular miRNAs such as miRNA-101, suffixed with numerical identifiers are labeled as Specific miRNA. Numerical identifiers (separated by delimiters such as “,”, “/”, and “and”) occurring as part of specific miRNA mentions are annotated as a single entity. Box 1 depicts the annotation of specific miRNA mentions (including an example for part mentions). In addition, Disease, Gene/Protein, Species, and Relation Trigger are annotated. The detailed annotation guideline for annotating specific miRNA mentions is available as a supplementary file.

Mentions of disease names, disease abbreviations, signs, deficiencies, physiological dysfunction, disease symptoms, disorders, abnormalities, or organ damages are annotated as Disease. Only disease nouns were considered, adjective terms such as “Diabetic patients” are not marked; however, specific adjectives that can be treated as nouns were marked, e.g. “Parkinson’s disease patients”. Mentions referring to proteins/genes which are either single word (e.g. “trypsin”), multi-word, gene symbols (e.g. “SMN”), or complex names (including of hyphens, slashes, Greek letters, Roman or Arabic numerals) are annotated as Gene/Protein. Only those organisms that are having published miRNA sequences and annotations represented in miRBase database are labeled as Species. Any verb, noun, verb phrase, or noun phrase associating miRNA mention to either labeled disease or gene/protein term is annotated as Relation Trigger.

Relations annotation. We restrict the relationship extraction to sentence level and four different interacting entity pairs: Specific miRNA-Disease (SpMiR-D), Specific miRNA-Gene/Protein (SpMiR-GP), Non-Specific miRNA-Disease (NonSpMiR-D), and Non-Specific miRNA-Gene/Protein (NonSpMiR-GP). Relevant triples, an interacting pair (from one of the above-mentioned) co-occurring with a Relation Trigger in a sentence are defined to form a relation and can belong to one of the four above-mentioned Relation classes. On the contrary, if an interacting pair does not co-occur with any Relation Trigger then we do not tag such pair as a relation.

The annotation has been performed using Knowtator²⁶ integrated within the Protégé framework²⁷.

Corpus selection, annotation and properties. We develop a new corpus based on MEDLINE, annotated with miRNA mentions and relations. Shah et al.²⁸ showed that abstracts provide a comprehensive description of key results obtained from a study, whereas full text is a better source for biological relevant data. Thus, we choose to build the corpus for abstracts only. Out of 27001 abstracts retrieved using the keyword “miRNA”, 201 were randomly selected as training and 100 as test corpus. Two annotators performed the annotation. The first annotator annotated the training corpus iteratively to develop guidelines and built the consensus annotation. The second annotator followed these guidelines and annotated the same corpus. Disagreeing instances were harmonized by both the annotators through manual inspection for correctness and its adherence to the guidelines. Any changes to the guidelines were made if needed. During the harmonization process only the non-overlapping instances between the two annotators were investigated. Decisions were based on the rule that only noun forms were to be marked (specific adjectives that can be treated as nouns were also considered). In case of partial matches, where conflicting parts could be interpreted as an adjective were not resolved. For example, in “chronic inflammation”, marking either “chronic inflammation” or just “inflammation” were considered correct. Table 1 provides the inter-annotator agreement (measured as F₁, for both exact and boundary match, and Cohen’s κ) for the test corpus. Exact string match occurs only when both the annotators annotate identical strings, whereas in partial match fraction of the string has been annotated by either of the annotators. It is evident (cf. Table 1) that in almost all cases partial match performs better than exact string match, indicating variations in span of mentioned entities. An example annotation is shown in Box 1.

Table 2 shows the number of annotated concepts in the training and test corpora for each entity class and the count for manually extracted relations (triplets), categorized for different interacting entity pairs. Table 3 provides the overall statistics of the published corpora (additional information about the corpus is given in the README supplementary file).

Automated named entity recognition

For identification of specific miRNA mentions in text (cf. Table 4), we developed regular expression patterns using manual annotations of miRNA mentions as the basis. Similarly, a dictionary has been generated for general miRNA recognition. The regular expression patterns are represented in the format as defined by Oualline et al.²⁹. For simplicity and reusability, several aliases are defined (cf. Table 5) to be used in the final regular expression patterns for specific miRNA identification, given in Table 4. Detected entities are resolved to a unique miRNA name and disambiguated to adhere to standard naming conventions as authors use several morphological variants to report the same miRNA term. For example, miR-107 can be represented as miRNA-107, Micro RNA-107, MicroRNA 107, has-mir-107, mir-107/108, micro RNA 107 and 108, micro RNA (miR) 107 and so on. Thus, the identified miRNA entity has been resolved to its base form (e. g. hsa-microRNA-21 to hsa-mir-21 and microRNA 101 to mir-101) following the miRBase naming convention. Manual inspection of the test corpus for species distribution revealed that 71% of the documents belonged to human, followed by mouse (15%), rat (8%). Pig has 2 abstracts, zebrafish, HIV-1, HSV-1, and Caenorhabditis elegans 1 each (cf. Supplementary Figure A for the distribution). Thus, we assumed that most of the abstracts belonged to human and resolved the identified miRNA entities to human identifier in miRBase. Unique miRNA terms are mapped to human miRBase database identifiers through the mirMaid Restful web service. For those names where we do not retrieve any database identifiers, we fall back to another organism mention found in the abstract (if any), using the NCBI taxonomy dictionary (see below) (cf. Supplementary Figure B), otherwise we retain the unique normalized name (cf. Box 2).

We detect Species with a dictionary-based approach. The built dictionary consists of all the concepts from the NCBI taxonomy corresponding to only those organisms mentioned in miRBase.

Similarly, for identification of Disease and Gene/Protein mentions in text we adapted a dictionary-based approach. To detect Disease, we apply three dictionaries: MeSH, MedDRA³⁰ and Allie. For Gene/Protein, a dictionary³¹ based on SwissProt, EntrezGene, and HGNC is included. Gene synonyms which could be potentially tagged as miRNAs are removed to overcome redundancy. For example, genes encoding microRNA, hsa-mir-21 are named as miR-21, miRNA21 and hsa-mir-21, the gene symbol of MIR16 membrane interacting protein of RGS16 is MIR16, which can represent a miRNA mention.

The relation trigger dictionary comprises of all interaction terms from the training corpus. After reviewing the training corpus for relation trigger terms, we retrieved not one but many variants of the same relation trigger occurring in alternative verb-phrase groups. For example, “change in expression” can be represented as one of the following verb-phrases: Change MicroRNA-21 Expression, Expression of caveolin-1 was changed, Change in high levels of high-mobility group A2 expression, change of the let-7e and miR-23a/b expression, expression of miR-199b-5p in the non-metastatic cases was significantly changed, etc. To allow flexibility for capturing relation trigger along with its variants spanning over different phrase length, we first manually represented all the relations in its root form, such as “regulate expression” to “regulate” (cf. Relation_Dictionary.txt file in Dataset 1). The base form has been extended manually to different spelling variants, e.g. regulate to regulatory, regulation, etc., the detailed listing of variants is provided in Word_variations.txt in Dataset 1. Not all combinations of the root forms are logical; target and up-regulation terms cannot be combined to form a relation trigger. Thus, we additionally defined a set of relation combinations that are allowed (see Permutation_terms.txt in Dataset 1 for all combinations).

For all named entity recognition performed, the dictionary-based system ProMiner³¹ is used. Supplementary Table A (Dataset 1) provides a quantitative estimate of the entities available in the dictionaries used in this work.

Relation extraction

We consider three approaches for addressing automatic extraction of interacting entity pairs from free text, described in the following.

The co-occurrence approach serves as a baseline. Assuming all interactions to be present in isolated sentences, this approach is complete but may be limited in precision. Reducing the number of false positives can be achieved by filtering with the dictionary of relation triggers occurring in the same sentence. The rationale behind this filter is that the interaction is more likely to be described if such a term is present (we refer to this as tri-occurrence).

To increase the precision, we use a machine learning-based approach formulating the relation detection as a binary classification problem: each instance (consisting of a pair of entities) is classified either as not-containing a relation or belonging to one of the four-relation classes. Our system uses lexical and dependency parsing features. We evaluate linear support vector machines (SVM)³² as implemented in the LibSVM library, as well as LibLINEAR, a specialized implementation for processing large data sets³³, and naive Bayes classifiers³⁴. For more details, we refer to Bobić et al.³⁵.

Lexical features capture characteristics of tokens around the inspected pair of entities. The sentence text can roughly be divided into three parts: text between the entities, text before the entities, and text after the entities. Stemming³⁶ and entity blinding is performed to improve generalization. Features are bag-of-words and bi, tri, and quadri-gram based. This feature setting follows Yu et al. and Yang et al.^37,38. The presence of relation triggers is also taken into account, using the previously described manually generated list. Next to lexical features, dependency parsing (created using Stanford parser) provides an insight into the entire grammatical structure of the sentence³⁹ and was performed using the Stanford CoreNLP library (http://nlp.stanford.edu/software/corenlp.shtml). Deep parsing follows the shortest dependency path hypothesis⁴⁰. We analyzed the vertices v (tokens from the sentence) in the dependency tree from a lexical (text of the token) and syntactical (POS tag) perspective. Edges e in the tree correspond to the information about the grammatical relations between the vertices. Extracting relevant information from the dependency parse tree is usually done following the shortest dependency path hypothesis⁴⁰. Lexical and syntactical e-walks and v-walks on the shortest path are created by alternating sequence of vertices and edges, with the length of 3. We capture the information about the common ancestor vertex, in addition to checking whether the ancestor node represents a verb form (e.g. POS tag could be VB, VBZ, VBD, etc.). Finally, the length of the shortest path (number of edges) between the entities is considered as a numerical feature.

Performance evaluation of named entity recognition

Among the 201 abstracts present in the training corpus, 82% contained general miRNA mentions, in comparison to specific miRNAs with 45%. In Table 6, results for miRNA entity recognition are reported. Non-specific miRNA recognition is close to perfect. Specific miRNA mention recognition has an F₁ measure of 0.94.

For disease mention recognition, combined dictionaries, based on three established resources, resulted in 0.79 and 0.69 F₁ score for the training and test corpus respectively. The low score for disease identification could be due to the variation in disease mentions, such as multi-word, synonym combination, nested names, etc. However, the partial matches result for diseases reported 0.88 of F₁, providing the possibility for detection of similar text strings for better recall (cf. Supplementary Table B in Dataset 1). Genes/proteins dictionary showed a performance of 0.84 and 0.85 of F₁ in training and test corpus respectively.

The evaluation of the relation trigger dictionary (cf. Table 6) suggests that it covers a substantial part of the vocabulary with recall of 0.86 for the training and 0.79 for the test corpus.

Relation extraction

We queried MEDLINE for “miRNA and Epilepsy” documents, among which 16 documents containing miRNA-related relations were manually selected (cf. Supplementary Figure C for the detailed distribution statistics). To avoid any biased approach we choose Epilepsy disease domain. Manual inspection of these articles revealed 11.5% of miRNA-related associations occur outside the sentence level. Thus, our work focused on relations at sentence level. Sentences in which co-occurring entity pairs do not participate in any relation are tagged as false. A comparison of the different relation extraction approaches is shown in Figure 2. Supplementary Table D in Dataset 1 provides statistical details of the applied approaches given in Figure 2. If all the entities are correctly identified then co-occurrence based approach leads to 100% recall for relation extraction. The recall is not diminished using the tri-occurrence approach, as the true entity pairs remain constant, while the precision increases between 4pp (percentage points) and 17pp when compared to the co-occurrence based approach, reducing false positives (cf. Figure 2). However, overall the precision reaches less than 60%. In our work, we assume that all the entities have been identified giving a recall of 100% for both co-occurrence and tri-occurrence based approaches. Using the machine-learning based classification, precision is increased up to 76% for specific miRNA-gene relations for both LibLINEAR and LibSVM methods, although Naïve Bayes is not far behind. Similarly, these two methods performed nearly the same for specific miRNAs-disease relations, the F₁ measure is not substantially different but a trade-off between precision and recall can be observed. An increase in F₁ measure is observed for non-specific miRNA relations when Naïve Bayes method is applied, out performing other strategies. Nevertheless, preference of the method highly depends on the compromise one chooses, whether better recall or precision. Overall, better recall and acceptable precision can be achieved with tri-occurrence method.

Figure 2. Comparison of different relation extraction approaches.

On the x-axis, different entity pair relations are represented as SpMiR-D for Specific miRNA-Disease, SpMiR-GP for Specific miRNA-Gene/Protein, NonSpMiR-D for Non-Specific miRNA-Disease, and NonSpMiR-GP for Non-Specific miRNA-Gene/Protein.

Most relation extraction approaches are dependent on the performance of named entity recognition. The impact of error propagation coming from automated entity recognizers is evaluated by applying the tri-occurrence method on the automatically annotated training and test corpus, here termed as “NERTri”. Compared to the results on the gold standard entity annotation a drop of 13 pp for NonSpMiR-D, 7pp for NonSpMiR-GP, 22pp for SpMiR-D, and 30pp for SpMiR-GP in F₁ is observed for the test corpus. Overall performance of the NERTri approach on training and test corpus is detailed in Supplementary Table C in Dataset 1.

Extracting miRNA-Alzheimer’s disease relations from full MEDLINE

The database miR2Disease is queried to return all miRNA-disease relations occurring in Alzheimer’s disease. For comparison, we retrieved miRNA-disease relations from MEDLINE using NERTri approach, resulting in 41 abstracts containing 159 relations. Obtained triplets have been manually curated to remove 51 false positives. False negatives have not been accounted, which may result in loss of information (cf. Relation extraction section). Comparison between the relations obtained from miR2Disease and NERTri are summarized in Table 7. The miR2Disease database returns 28 evidential statements from 9 articles. Among these, only 14 evidences are present in abstracts. Moreover, 16 evidences are extracted from one full text document⁴⁴. Only two evidences are identified at abstract level among these 16 evidences. Overall, 26 miRNAs identified by miR2Disease refer to Alzheimer’s disease. Therefore, our text-based extraction proposes approximately three times more relations than the database provides.

The analysis of 17 false negative relations which are in the database but not found by our approach shows that most of the relations could be found only in full text and that the automatic system misses four miRNA-Alzheimer’s disease relations from abstracts. Manual inspection reveals that in three out of these missing four evidences the disease name is not mentioned in the sentence (relation occurred at co-reference level).

Extraction of miRNA-gene relations for Alzheimer’s diseases from full MEDLINE

Here we compare the performance of our relation detection NERTri with another text-mining database, miRSel. For comparison, 100 abstracts from PubMed were retrieved using the query “alzheimer disease”[MeSH Terms] OR (“alzheimer disease”[All Fields] OR “alzheimer”[All Fields]) AND (“micrornas”[MeSH Terms] OR “micrornas”[All Fields] OR “microrna”[All Fields]) AND (“2001/01/01”[PDAT]:“2013/7/4”[PDAT]). Manual inspection of these articles leads to 184 miRNA-gene relations, at sentence level, (Table 8) in 37 abstracts.

NERTri approach was able to identify 140 of these found relations in 28 abstracts. Among the 37 abstracts from the PubMed query, miRSel contained only 12 abstracts with 56 miRNA-gene relations (cf. Table 8). False negatives in our approach when compared with miRSel could not be directly identified as the database is not downloadable and searchable for disease specific relations. However, low intersection between miRSel and NERTri can be observed.

In summary, our approach provides AD related gene-microRNA relations from PubMed which have not been available in the database before.

Overall, the results are promising when compared with the miR2Disease and miRSel databases and indicate that we can extend the databases to a large extent with new relations. Such an approach makes it much easier to keep databases up to date. Nevertheless full text processing would most certainly increase the recall of automatic processing.