Optimizing text mining methods for improving biomedical natural language processing
Mehryary, Farrokh (2022-02-04)
Optimizing text mining methods for improving biomedical natural language processing
(04.02.2022)
Turun yliopisto
Julkaisun pysyvä osoite on:
https://urn.fi/URN:ISBN:978-951-29-8744-3
https://urn.fi/URN:ISBN:978-951-29-8744-3
Tiivistelmä
The overwhelming amount and the increasing rate of publication in the biomedical domain make it difficult for life sciences researchers to acquire and maintain all information that is necessary for their research. Pubmed (the primary citation database for the biomedical literature) currently contains over 21 million article abstracts and more than one million of them were published in 2020 alone.
Even though existing article databases provide capable keyword search services, typical everyday-life queries usually return thousands of relevant articles. For instance, a cancer research scientist may need to acquire a complete list of genes that interact with BRCA1 (breast cancer 1) gene. The PubMed keyword search for BRCA1 returns over 16,500 article abstracts, making manual inspection of the retrieved documents impractical. Missing even one of the interacting gene partners in this scenario may jeopardize successful development of a potential new drug or vaccine. Although manually curated databases of biomolecular interactions exist, they are usually not up-to-date and they require notable human effort to maintain. To summarize, new discoveries are constantly being shared within the community via scientific publishing, but unfortunately the probability of missing vital information for research in life sciences is increasing.
In response to this problem, the biomedical natural language processing (BioNLP) community of researchers has emerged and strives to assist life sciences researchers by building modern language processing and text mining tools that can be applied at large-scale and scan the whole publicly available literature and extract, classify, and aggregate the information found within, thus keeping life sciences researchers always up-to-date with the recent relevant discoveries and facilitating their research in numerous fields such as molecular biology, biomedical engineering, bioinformatics, genetics engineering and biochemistry.
My research has almost exclusively focused on biomedical relation and event extraction tasks. These foundational information extraction tasks deal with automatic detection of biological processes, interactions and relations described in the biomedical literature. Precisely speaking, biomedical relation and event extraction systems can scan through a vast amount of biomedical texts and automatically detect and extract the semantic relations of biomedical named entities (e.g. genes, proteins, chemical compounds, and diseases). The structured outputs of such systems (i.e., the extracted relations or events) can be stored as relational databases or molecular interaction networks which can easily be queried, filtered, analyzed, visualized and integrated with other structured data sources. Extracting biomolecular interactions has always been the primary interest of BioNLP researcher because having knowledge about such interactions is crucially important in various research areas including precision medicine, drug discovery, drug repurposing, hypothesis generation, construction and curation of signaling pathways, and protein function and structure prediction.
State-of-the-art relation and event extraction methods are based on supervised machine learning, requiring manually annotated data for training. Manual annotation for the biomedical domain requires domain expertise and it is time-consuming.
Hence, having minimal training data for building information extraction systems is a common case in the biomedical domain. This demands development of methods that can make the most out of available training data and this thesis gathers all my research efforts and contributions in that direction.
It is worth mentioning that biomedical natural language processing has undergone a revolution since I started my research in this field almost ten years ago. As a member of the BioNLP community, I have witnessed the emergence, improvement– and in some cases, the disappearance–of many methods, each pushing the performance of the best previous method one step further. I can broadly divide the last ten years into three periods. Once I started my research, feature-based methods that relied on heavy feature engineering were dominant and popular. Then, significant advancements in the hardware technology, as well as several breakthroughs in the algorithms and methods enabled machine learning practitioners to seriously utilize artificial neural networks for real-world applications. In this period, convolutional, recurrent, and attention-based neural network models became dominant and superior. Finally, the introduction of transformer-based language representation models such as BERT and GPT impacted the field and resulted in unprecedented performance improvements on many data sets. When reading this thesis, I demand the reader to take into account the course of history and judge the methods and results based on what could have been done in that particular period of the history.
Even though existing article databases provide capable keyword search services, typical everyday-life queries usually return thousands of relevant articles. For instance, a cancer research scientist may need to acquire a complete list of genes that interact with BRCA1 (breast cancer 1) gene. The PubMed keyword search for BRCA1 returns over 16,500 article abstracts, making manual inspection of the retrieved documents impractical. Missing even one of the interacting gene partners in this scenario may jeopardize successful development of a potential new drug or vaccine. Although manually curated databases of biomolecular interactions exist, they are usually not up-to-date and they require notable human effort to maintain. To summarize, new discoveries are constantly being shared within the community via scientific publishing, but unfortunately the probability of missing vital information for research in life sciences is increasing.
In response to this problem, the biomedical natural language processing (BioNLP) community of researchers has emerged and strives to assist life sciences researchers by building modern language processing and text mining tools that can be applied at large-scale and scan the whole publicly available literature and extract, classify, and aggregate the information found within, thus keeping life sciences researchers always up-to-date with the recent relevant discoveries and facilitating their research in numerous fields such as molecular biology, biomedical engineering, bioinformatics, genetics engineering and biochemistry.
My research has almost exclusively focused on biomedical relation and event extraction tasks. These foundational information extraction tasks deal with automatic detection of biological processes, interactions and relations described in the biomedical literature. Precisely speaking, biomedical relation and event extraction systems can scan through a vast amount of biomedical texts and automatically detect and extract the semantic relations of biomedical named entities (e.g. genes, proteins, chemical compounds, and diseases). The structured outputs of such systems (i.e., the extracted relations or events) can be stored as relational databases or molecular interaction networks which can easily be queried, filtered, analyzed, visualized and integrated with other structured data sources. Extracting biomolecular interactions has always been the primary interest of BioNLP researcher because having knowledge about such interactions is crucially important in various research areas including precision medicine, drug discovery, drug repurposing, hypothesis generation, construction and curation of signaling pathways, and protein function and structure prediction.
State-of-the-art relation and event extraction methods are based on supervised machine learning, requiring manually annotated data for training. Manual annotation for the biomedical domain requires domain expertise and it is time-consuming.
Hence, having minimal training data for building information extraction systems is a common case in the biomedical domain. This demands development of methods that can make the most out of available training data and this thesis gathers all my research efforts and contributions in that direction.
It is worth mentioning that biomedical natural language processing has undergone a revolution since I started my research in this field almost ten years ago. As a member of the BioNLP community, I have witnessed the emergence, improvement– and in some cases, the disappearance–of many methods, each pushing the performance of the best previous method one step further. I can broadly divide the last ten years into three periods. Once I started my research, feature-based methods that relied on heavy feature engineering were dominant and popular. Then, significant advancements in the hardware technology, as well as several breakthroughs in the algorithms and methods enabled machine learning practitioners to seriously utilize artificial neural networks for real-world applications. In this period, convolutional, recurrent, and attention-based neural network models became dominant and superior. Finally, the introduction of transformer-based language representation models such as BERT and GPT impacted the field and resulted in unprecedented performance improvements on many data sets. When reading this thesis, I demand the reader to take into account the course of history and judge the methods and results based on what could have been done in that particular period of the history.
Kokoelmat
- Väitöskirjat [2824]