A carregar...
Publicação
Extracting phenotype-gene relations from biomedical literature using distant supervision and deep learning
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 1.06 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
As relações entre fenótipos humanos e genes são fundamentais para entender completamente a origem de algumas abnormalidades fenotípicas e as suas doenças associadas. A literatura biomédica é a fonte mais abrangente dessas relações. Diversas ferramentas de extração de relações têm sido propostas para identificar relações entre conceitos em texto muito heterogéneo ou não estruturado, utilizando algoritmos de supervisão distante e aprendizagem profunda. Porém, a maioria dessas ferramentas requer um corpus anotado e não há nenhum corpus disponível anotado com relações entre fenótipos humanos e genes. Este trabalho apresenta o corpus Phenotype-Gene Relations (PGR), um corpus padrão-prata de anotações
de fenótipos humanos e genes e as suas relações (gerado de forma automática) e dois módulos de extração de relações usando um algoritmo de distantly supervised multi-instance learning e um algoritmo de aprendizagem profunda com ontologias biomédicas. O corpus PGR consiste em 1712 resumos de artigos, 5676 anotações de fenótipos humanos, 13835 anotações de genes e 4283 relações. Os resultados do corpus foram parcialmente avaliados por oito curadores, todos investigadores nas áreas de Biologia e Bioquímica, obtendo uma precisão de 87,01%, com um valor de concordância inter-curadores de 87,58%. As abordagens de supervisão distante (ou supervisão fraca) combinam um corpus não anotado com uma base de dados para identificar e extrair entidades do texto, reduzindo a quantidade de esforço necessário para realizar anotações manuais. A distantly supervised multi-instance learning aproveita a supervisão distante e um sparse multi-instance learning algorithm para treinar um classificador de extracção de relações, usando uma base de dados padrão-ouro de relações entre fenótipos humanos e genes. As ferramentas de aprendizagem profunda de extração de relações, para tarefas de prospeção de textos biomédicos, raramente tiram proveito dos recursos específicos existentes para cada domínio, como as ontologias biomédicas. As ontologias biomédicas desempenham um papel fundamental, fornecendo informações semânticas e de ancestralidade sobre uma entidade. Este trabalho utilizou a Human Phenotype
Ontology e a Gene Ontology, para representar cada par candidato como a sequência de relações entre os seus ancestrais para cada ontologia. O corpus de teste PGR foi aplicado aos módulos de extração de relações desenvolvidos, obtendo resultados promissores, nomeadamente 55,00% (módulo de aprendizagem profunda) e 73,48% (módulo de distantly supervised multi-instance learning) na medida-F. Este corpus de teste também foi aplicado ao BioBERT, um modelo de representação de linguagem biomédica pré-treinada para prospeção de texto biomédico, obtendo 67,16% em medida-F.
Human phenotype-gene relations are fundamental to fully understand the origin of some phenotypic abnormalities and their associated diseases. Biomedical literature is the most comprehensive source of these relations. Several relation extraction tools have been proposed to identify relations between concepts in highly heterogeneous or unstructured text, namely using distant supervision and deep learning algorithms. However, most of these tools require an annotated corpus, and there is no corpus available annotated with human phenotype-gene relations. This work presents the Phenotype-Gene Relations (PGR) corpus, a silver standard corpus of human phenotype and gene annotations and their relations (generated in a fully automated manner), and two relation extraction modules using a distantly supervised multi-instance learning algorithm, and an ontology based deep learning algorithm. The PGR corpus consists of 1712 abstracts, 5676 human phenotype annotations, 13835 gene annotations, and 4283 relations. The corpus results were partially evaluated by eight curators, all working in the fields of Biology and Biochemistry, obtaining a precision of 87.01%, with an inter-curator agreement score of 87.58%. Distant supervision (or weak supervision) approaches combine an unlabeled corpus with a knowledge base to identify and extract entities from text, reducing the amount of manual effort necessary. Distantly supervised multi-instance learning takes advantage of distant supervision and a sparse multi-instance learning algorithm to train a relation extraction classifier, using a gold standard knowledge base of human phenotype-gene relations. Deep learning relation extraction tools, for biomedical text mining tasks, rarely take advantage of existing domain-specific resources, such as biomedical ontologies. Biomedical ontologies play a fundamental role by providing semantic and ancestry information about an entity. This work used the Human Phenotype Ontology and the Gene Ontology, to represent each candidate pair as the sequence of relations between its ancestors for each ontology. The PGR test-set was applied to the developed relation extraction modules, obtaining promising results, namely 55.00% (deep learning module), and 73.48% (distantly supervised multi-instance learning module) in F-measure. This test-set was also applied to BioBERT, a pre-trained biomedical language representation model for biomedical text mining, obtaining 67.16% in F-measure.
Human phenotype-gene relations are fundamental to fully understand the origin of some phenotypic abnormalities and their associated diseases. Biomedical literature is the most comprehensive source of these relations. Several relation extraction tools have been proposed to identify relations between concepts in highly heterogeneous or unstructured text, namely using distant supervision and deep learning algorithms. However, most of these tools require an annotated corpus, and there is no corpus available annotated with human phenotype-gene relations. This work presents the Phenotype-Gene Relations (PGR) corpus, a silver standard corpus of human phenotype and gene annotations and their relations (generated in a fully automated manner), and two relation extraction modules using a distantly supervised multi-instance learning algorithm, and an ontology based deep learning algorithm. The PGR corpus consists of 1712 abstracts, 5676 human phenotype annotations, 13835 gene annotations, and 4283 relations. The corpus results were partially evaluated by eight curators, all working in the fields of Biology and Biochemistry, obtaining a precision of 87.01%, with an inter-curator agreement score of 87.58%. Distant supervision (or weak supervision) approaches combine an unlabeled corpus with a knowledge base to identify and extract entities from text, reducing the amount of manual effort necessary. Distantly supervised multi-instance learning takes advantage of distant supervision and a sparse multi-instance learning algorithm to train a relation extraction classifier, using a gold standard knowledge base of human phenotype-gene relations. Deep learning relation extraction tools, for biomedical text mining tasks, rarely take advantage of existing domain-specific resources, such as biomedical ontologies. Biomedical ontologies play a fundamental role by providing semantic and ancestry information about an entity. This work used the Human Phenotype Ontology and the Gene Ontology, to represent each candidate pair as the sequence of relations between its ancestors for each ontology. The PGR test-set was applied to the developed relation extraction modules, obtaining promising results, namely 55.00% (deep learning module), and 73.48% (distantly supervised multi-instance learning module) in F-measure. This test-set was also applied to BioBERT, a pre-trained biomedical language representation model for biomedical text mining, obtaining 67.16% in F-measure.
Descrição
Tese de mestrado em Bioinformática e Biologia Computacional, Universidade de Lisboa, Faculdade de Ciências, 2019
Palavras-chave
Literatura biomédica Extração de relações Corpus Padrão-Prata Supervisão distante Aprendizagem profunda Teses de mestrado - 2019