In this work, we pioneer a knowledgeable machine reading system, establishing connections between internal information from molecule structures and external information from biomedical text, as shown in Fig. However, it is non-trivial to jointly model the heterogeneous data in a unified framework, and challenging to learn the meta-knowledge without explicit human annotation. Inspired by human learners, it is desirable to build a knowledgeable machine reading system that versatilely learns from both information sources to better master molecule knowledge so as to assist biomedical research. Moreover, confined to single information sources, machine reading systems can hardly learn meta-knowledge beyond single information for deep molecule understanding. Specifically, information from molecule structure is concise but typically limited compared to information from wet-lab experiments, while information from biomedical text enjoys better abundance and flexibility but usually suffers from noisy extraction processes. This limits not only the generality of machine reading systems, but also the performance of knowledge acquisition due to the intrinsic nature of each information. To the best of our knowledge, all existing machine reading systems for biomedical knowledge acquisition are confined to either internal molecule structure information or external biomedical text information in isolation, and different models have to be developed to process each type of information. Moreover, humans are able to knowledgeably learn and leverage meta-knowledge within and across different information-establishing fine-grained mappings between semantic units from different information sources, e.g., functional groups and natural language phrases-for deep molecule understanding. Utilizing complementary information is typically crucial for human learners to achieve comprehensive molecule understanding. In comparison, biomedical text provides abundant flexible external information of molecule entities reported from wet-lab experiments.
Specifically, molecule structures provide concise standardized internal information, where functional groups and their positions are strong indicators of molecular properties and interactions. In the acquisition of biomedical molecule knowledge, humans are capable of versatilely reading different types of information that complementarily characterize molecule entities, including molecule structures and biomedical text. However, compared to human learners, machine reading systems still have a huge gap in terms of both versatile reading and knowledgeable learning 7. With the rapid progress of deep learning, machine reading systems are developed to automatically acquire biomedical knowledge by reading large-scale data, accelerating recent biomedical research in many cases 6. However, existing KBs are still far from complete due to the rapid growth of biomedical knowledge and the high cost of expert annotation.
To this end, people have built many biomedical knowledge bases (KBs), including PubChem 3, Gene Ontology 4, and DrugBank 5. For instance, experts study the structural properties of protein molecules to understand their mechanisms of action 1, and investigate the interactions between drugs and target molecules to prevent adverse reactions 2. Understanding molecule entities (i.e., their properties and interactions) is fundamental to most biomedical research areas. Experimental results show that our system even surpasses human professionals in the capability of molecular property comprehension, and also reveal its promising potential in facilitating automatic drug discovery and documentation in the future. By grasping meta-knowledge in an unsupervised fashion within and across different information sources, our system can facilitate various real-world biomedical applications, including molecular property prediction, biomedical relation extraction and so on. We solve the problem that existing machine reading models can only process different types of data separately, and thus achieve a comprehensive and thorough understanding of molecule entities. Inspired by humans that learn deep molecule knowledge from versatile reading on both molecule structure and biomedical text information, we propose a knowledgeable machine reading system that bridges both types of information in a unified deep-learning framework for comprehensive biomedical research assistance. To accelerate biomedical research process, deep-learning systems are developed to automatically acquire knowledge about molecule entities by reading large-scale biomedical data.
0 kommentar(er)
