Improving chemical entity recognition through h-index based semantic similarity

Learning chemistry: exploring the suitability of machine learning for the task of structure-based chemical ontology classification

Chemical data is increasingly openly available in databases such as PubChem, which contains approximately 110 million compound entries as of February 2021. With the availability of data at such scale, the burden has shifted to organisation, analysis and interpretation. Chemical ontologies provide structured classifications of chemical entities that can be used for navigation and filtering of the large chemical space. ChEBI is a prominent example of a chemical ontology, widely used in life science contexts. However, ChEBI is manually maintained and as such cannot easily scale to the full scope of public chemical data. There is a need for tools that are able to automatically classify chemical data into chemical ontologies, which can be framed as a hierarchical multi-class classification problem. In this paper we evaluate machine learning approaches for this task, comparing different learning frameworks including logistic regression, decision trees and long short-term memory artificial neural networks, and different encoding approaches for the chemical structures, including cheminformatics fingerprints and character-based encoding from chemical line notation representations. We find that classical learning approaches such as logistic regression perform well with sets of relatively specific, disjoint chemical classes, while the neural network is able to handle larger sets of overlapping classes but needs more examples per class to learn from, and is not able to make a class prediction for every molecule. Future work will explore hybrid and ensemble approaches, as well as alternative network architectures including neuro-symbolic approaches.

Multi-domain semantic similarity in biomedical research

BACKGROUND

Given the increasing amount of biomedical resources that are being annotated with concepts from more than one ontology and covering multiple domains of knowledge, it is important to devise mechanisms to compare these resources that take into account the various domains of annotation. For example, metabolic pathways are annotated with their enzymes and their metabolites, and thus similarity measures should compare them with respect to both of those domains simultaneously.

RESULTS

In this paper, we propose two approaches to lift existing single-ontology semantic similarity measures into multi-domain measures. The aggregative approach compares domains independently and averages the various similarity values into a final score. The integrative approach integrates all the relevant ontologies into a single one, calculating similarity in the resulting multi-domain ontology using the single-ontology measure.

CONCLUSIONS

We evaluated the two approaches in a multidisciplinary epidemiology dataset by evaluating the capacity of the similarity measures to predict new annotations based on the existing ones. The results show a promising increase in performance of the multi-domain measures over the single-ontology ones in the vast majority of the cases. These results show that multi-domain measures outperform single-domain ones, and should be considered by the community as a starting point to study more efficient multi-domain semantic similarity measures.

BO-LSTM: classifying relations via long short-term memory networks along biomedical ontologies

BACKGROUND

Recent studies have proposed deep learning techniques, namely recurrent neural networks, to improve biomedical text mining tasks. However, these techniques rarely take advantage of existing domain-specific resources, such as ontologies. In Life and Health Sciences there is a vast and valuable set of such resources publicly available, which are continuously being updated. Biomedical ontologies are nowadays a mainstream approach to formalize existing knowledge about entities, such as genes, chemicals, phenotypes, and disorders. These resources contain supplementary information that may not be yet encoded in training data, particularly in domains with limited labeled data.

RESULTS

We propose a new model to detect and classify relations in text, BO-LSTM, that takes advantage of domain-specific ontologies, by representing each entity as the sequence of its ancestors in the ontology. We implemented BO-LSTM as a recurrent neural network with long short-term memory units and using open biomedical ontologies, specifically Chemical Entities of Biological Interest (ChEBI), Human Phenotype, and Gene Ontology. We assessed the performance of BO-LSTM with drug-drug interactions mentioned in a publicly available corpus from an international challenge, composed of 792 drug descriptions and 233 scientific abstracts. By using the domain-specific ontology in addition to word embeddings and WordNet, BO-LSTM improved the F1-score of both the detection and classification of drug-drug interactions, particularly in a document set with a limited number of annotations. We adapted an existing DDI extraction model with our ontology-based method, obtaining a higher F1 score than the original model. Furthermore, we developed and made available a corpus of 228 abstracts annotated with relations between genes and phenotypes, and demonstrated how BO-LSTM can be applied to other types of relations.

CONCLUSIONS

Our findings demonstrate that besides the high performance of current deep learning techniques, domain-specific ontologies can still be useful to mitigate the lack of labeled data.

Generating a Tolerogenic Cell Therapy Knowledge Graph from Literature

Tolerogenic cell therapies provide an alternative to conventional immunosuppressive treatments of autoimmune disease and address, among other goals, the rejection of organ or stem cell transplants. Since various methodologies can be followed to develop tolerogenic therapies, it is important to be aware and up to date on all available studies that may be relevant to their improvement. Recently, knowledge graphs have been proposed to link various sources of information, using text mining techniques. Knowledge graphs facilitate the automatic retrieval of information about the topics represented in the graph. The objective of this work was to automatically generate a knowledge graph for tolerogenic cell therapy from biomedical literature. We developed a system, ICRel, based on machine learning to extract relations between cells and cytokines from abstracts. Our system retrieves related documents from PubMed, annotates each abstract with cell and cytokine named entities, generates the possible combinations of cell–cytokine pairs cooccurring in the same sentence, and identifies meaningful relations between cells and cytokines. The extracted relations were used to generate a knowledge graph, where each edge was supported by one or more documents. We obtained a graph containing 647 cell–cytokine relations, based on 3,264 abstracts. The modules of ICRel were evaluated with cross-validation and manual evaluation of the relations extracted. The relation extraction module obtained an F-measure of 0.789 in a reference database, while the manual evaluation obtained an accuracy of 0.615. Even though the knowledge graph is based on information that was already published in other articles about immunology, the system we present is more efficient than the laborious task of manually reading all the literature to find indirect or implicit relations. The ICRel graph will help experts identify implicit relations that may not be evident in published studies.

Identifying Human Phenotype Terms by Combining Machine Learning and Validation Rules

Named-Entity Recognition is commonly used to identify biological entities such as proteins, genes, and chemical compounds found in scientific articles. The Human Phenotype Ontology (HPO) is an ontology that provides a standardized vocabulary for phenotypic abnormalities found in human diseases. This article presents the Identifying Human Phenotypes (IHP) system, tuned to recognize HPO entities in unstructured text. IHP uses Stanford CoreNLP for text processing and applies Conditional Random Fields trained with a rich feature set, which includes linguistic, orthographic, morphologic, lexical, and context features created for the machine learning-based classifier. However, the main novelty of IHP is its validation step based on a set of carefully crafted manual rules, such as the negative connotation analysis, that combined with a dictionary can filter incorrectly identified entities, find missed entities, and combine adjacent entities. The performance of IHP was evaluated using the recently published HPO Gold Standardized Corpora (GSC), where the system Bio-LarK CR obtained the best F-measure of 0.56. IHP achieved an F-measure of 0.65 on the GSC. Due to inconsistencies found in the GSC, an extended version of the GSC was created, adding 881 entities and modifying 4 entities. IHP achieved an F-measure of 0.863 on the new GSC.

Information Retrieval and Text Mining Technologies for Chemistry

Efficient access to chemical information contained in scientific literature, patents, technical reports, or the web is a pressing need shared by researchers and patent attorneys from different chemical disciplines. Retrieval of important chemical information in most cases starts with finding relevant documents for a particular chemical compound or family. Targeted retrieval of chemical documents is closely connected to the automatic recognition of chemical entities in the text, which commonly involves the extraction of the entire list of chemicals mentioned in a document, including any associated information. In this Review, we provide a comprehensive and in-depth description of fundamental concepts, technical implementations, and current technologies for meeting these information demands. A strong focus is placed on community challenges addressing systems performance, more particularly CHEMDNER and CHEMDNER patents tasks of BioCreative IV and V, respectively. Considering the growing interest in the construction of automatically annotated chemical knowledge bases that integrate chemical information and biological data, cheminformatics approaches for mapping the extracted chemical names into chemical structures and their subsequent annotation together with text mining applications for linking chemistry with biological information are also presented. Finally, future trends and current challenges are highlighted as a roadmap proposal for research in this emerging field.

Extracting microRNA-gene relations from biomedical literature using distant supervision

Many biomedical relation extraction approaches are based on supervised machine learning, requiring an annotated corpus. Distant supervision aims at training a classifier by combining a knowledge base with a corpus, reducing the amount of manual effort necessary. This is particularly useful for biomedicine because many databases and ontologies have been made available for many biological processes, while the availability of annotated corpora is still limited. We studied the extraction of microRNA-gene relations from text. MicroRNA regulation is an important biological process due to its close association with human diseases. The proposed method, IBRel, is based on distantly supervised multi-instance learning. We evaluated IBRel on three datasets, and the results were compared with a co-occurrence approach as well as a supervised machine learning algorithm. While supervised learning outperformed on two of those datasets, IBRel obtained an F-score 28.3 percentage points higher on the dataset for which there was no training set developed specifically. To demonstrate the applicability of IBRel, we used it to extract 27 miRNA-gene relations from recently published papers about cystic fibrosis. Our results demonstrate that our method can be successfully used to extract relations from literature about a biological process without an annotated corpus. The source code and data used in this study are available at https://github.com/AndreLamurias/IBRel.

Evaluating the effect of annotation size on measures of semantic similarity

Background

Ontologies are widely used as metadata in biological and biomedical datasets. Measures of semantic similarity utilize ontologies to determine how similar two entities annotated with classes from ontologies are, and semantic similarity is increasingly applied in applications ranging from diagnosis of disease to investigation in gene networks and functions of gene products.

Results

Here, we analyze a large number of semantic similarity measures and the sensitivity of similarity values to the number of annotations of entities, difference in annotation size and to the depth or specificity of annotation classes. We find that most similarity measures are sensitive to the number of annotations of entities, difference in annotation size as well as to the depth of annotation classes; well-studied and richly annotated entities will usually show higher similarity than entities with only few annotations even in the absence of any biological relation.

Conclusions

Our findings may have significant impact on the interpretation of results that rely on measures of semantic similarity, and we demonstrate how the sensitivity to annotation size can lead to a bias when using semantic similarity to predict protein-protein interactions.

Electronic supplementary material

The online version of this article (doi:10.1186/s13326-017-0119-z) contains supplementary material, which is available to authorized users.

Chemical named entity recognition in patents by domain knowledge and unsupervised feature learning

Medicinal chemistry patents contain rich information about chemical compounds. Although much effort has been devoted to extracting chemical entities from scientific literature, limited numbers of patent mining systems are publically available, probably due to the lack of large manually annotated corpora. To accelerate the development of information extraction systems for medicinal chemistry patents, the 2015 BioCreative V challenge organized a track on Chemical and Drug Named Entity Recognition from patent text (CHEMDNER patents). This track included three individual subtasks: (i) Chemical Entity Mention Recognition in Patents (CEMP), (ii) Chemical Passage Detection (CPD) and (iii) Gene and Protein Related Object task (GPRO). We participated in the two subtasks of CEMP and CPD using machine learning-based systems. Our machine learning-based systems employed the algorithms of conditional random fields (CRF) and structured support vector machines (SSVMs), respectively. To improve the performance of the NER systems, two strategies were proposed for feature engineering: (i) domain knowledge features of dictionaries, chemical structural patterns and semantic type information present in the context of the candidate chemical and (ii) unsupervised feature learning algorithms to generate word representation features by Brown clustering and a novel binarized Word embedding to enhance the generalizability of the system. Further, the system output for the CPD task was yielded based on the patent titles and abstracts with chemicals recognized in the CEMP task.The effects of the proposed feature strategies on both the machine learning-based systems were investigated. Our best system achieved the second best performance among 21 participating teams in CEMP with a precision of 87.18%, a recall of 90.78% and aF-measure of 88.94% and was the top performing system among nine participating teams in CPD with a sensitivity of 98.60%, a specificity of 87.21%, an accuracy of 94.75%, a Matthew's correlation coefficient (MCC) of 88.24%, a precision at full recall (P_full_R) of 66.57% and an area under the precision-recall curve (AUC_PR) of 0.9347. The SSVM-based CEMP systems outperformed the CRF-based CEMP systems when using the same features. Features generated from both the domain knowledge and unsupervised learning algorithms significantly improved the chemical NER task on patents.Database URL:http:// database. oxfordjournals. org/ content/ 2016/ baw049.

ChEBI in 2016: Improved services and an expanding collection of metabolites

ChEBI is a database and ontology containing information about chemical entities of biological interest. It currently includes over 46,000 entries, each of which is classified within the ontology and assigned multiple annotations including (where relevant) a chemical structure, database cross-references, synonyms and literature citations. All content is freely available and can be accessed online at http://www.ebi.ac.uk/chebi. In this update paper, we describe recent improvements and additions to the ChEBI offering. We have substantially extended our collection of endogenous metabolites for several organisms including human, mouse, Escherichia coli and yeast. Our front-end has also been reworked and updated, improving the user experience, removing our dependency on Java applets in favour of embedded JavaScript components and moving from a monthly release update to a 'live' website. Programmatic access has been improved by the introduction of a library, libChEBI, in Java, Python and Matlab. Furthermore, we have added two new tools, namely an analysis tool, BiNChE, and a query tool for the ontology, OntoQuery.

CHEMDNER: The drugs and chemical names extraction challenge

Natural language processing (NLP) and text mining technologies for the chemical domain (ChemNLP or chemical text mining) are key to improve the access and integration of information from unstructured data such as patents or the scientific literature. Therefore, the BioCreative organizers posed the CHEMDNER (chemical compound and drug name recognition) community challenge, which promoted the development of novel, competitive and accessible chemical text mining systems. This task allowed a comparative assessment of the performance of various methodologies using a carefully prepared collection of manually labeled text prepared by specially trained chemists as Gold Standard data. We evaluated two important aspects: one covered the indexing of documents with chemicals (chemical document indexing - CDI task), and the other was concerned with finding the exact mentions of chemicals in text (chemical entity mention recognition - CEM task). 27 teams (23 academic and 4 commercial, a total of 87 researchers) returned results for the CHEMDNER tasks: 26 teams for CEM and 23 for the CDI task. Top scoring teams obtained an F-score of 87.39% for the CEM task and 88.20% for the CDI task, a very promising result when compared to the agreement between human annotators (91%). The strategies used to detect chemicals included machine learning methods (e.g. conditional random fields) using a variety of features, chemistry and drug lexica, and domain-specific rules. We expect that the tools and resources resulting from this effort will have an impact in future developments of chemical text mining applications and will form the basis to find related chemical information for the detected entities, such as toxicological or pharmacogenomic properties.