TL;DR
This paper argues that "knowledge-based" approaches are crucial for advancing biomedical data science, moving beyond simply analyzing large datasets. It surveys existing methods for representing and reasoning with biomedical knowledge, highlighting the potential for these systems to generate hypotheses, explain data, and ultimately accelerate biomedical discovery.
What Problem Does This Paper Address?
The paper addresses the challenge of effectively leveraging the vast and rapidly growing body of biomedical knowledge to drive data science research. While initiatives like the NIH's "Big Data to Knowledge" (BD2K) program have focused heavily on "big data," the "knowledge" component has often been an implicit assumption rather than an explicit focus. This leaves a gap in how to computationally represent, reason with, and apply existing biomedical knowledge to make sense of new data, generate hypotheses, and accelerate discovery. The authors contend that computational manipulation of knowledge is an under-appreciated but vital aspect of biomedical data science.
Methods
The paper doesn't introduce novel methods but rather surveys and categorizes existing approaches to knowledge representation and reasoning in biomedicine. Key areas discussed include:
- Knowledge Representation:
- Ontologies: Discusses the importance of community-curated ontologies (like those meeting OBO Foundry criteria) and terminological resources (UMLS, Snomed-CT, NCI thesaurus) for providing structured representations of biomedical entities and processes.
- Semantic Web Technologies: Explains the role of RDF, RDFS, OWL, and SPARQL in building knowledge graphs and enabling linked data, facilitating the integration of diverse biomedical information.
- Knowledge-based Inference:
- Logical Inference: Covers deductive reasoning and its restricted forms (description logics) for checking consistency and inferring relationships. It also mentions deductive retrieval (e.g., using Prolog or SWRL) for querying knowledge bases.
- Inference from Ontology Annotation: Highlights the use of gene ontology annotations for enrichment analysis, hypothesis generation about gene function, and mapping between phenotypes and diseases.
- Inference from Biomedical Literature: Discusses natural language processing (NLP) techniques for extracting knowledge from text, including tagging passages with ontological terms, identifying entities, and building more complex knowledge representations.
- Hypothesis Generation, Evaluation, and Modification: Explores Bayesian inference for quantitative reasoning, network-based inference (link prediction, community finding) for generating hypotheses, and the potential for systems to assist in explaining biomedical phenomena.
Datasets
Not discussed in the paper. The paper focuses on the conceptual framework and existing approaches rather than specific empirical evaluations on datasets.
Key Insights and Ideas
- Knowledge is not just data: The paper emphasizes that biomedical data science needs to go beyond simply processing large datasets. It requires the computational manipulation and reasoning with existing knowledge to extract deeper insights.
- Ontologies are foundational: Mature biomedical ontologies are crucial for providing a common language and structured representation of biomedical concepts, enabling interoperability and reproducibility.
- The Semantic Web as a framework: Technologies like RDF and OWL provide a powerful infrastructure for linking and integrating diverse biomedical knowledge sources, creating rich knowledge graphs.
- Inference is key to unlocking knowledge: Representing knowledge is only the first step; the ability to infer new information, generate hypotheses, and explain data is where the true value lies.
- The literature remains the largest knowledge repository: Despite the growth of databases, the published biomedical literature is still the most comprehensive source of knowledge, making text mining and NLP critical.
- The "knowledge acquisition bottleneck" is a major challenge: Manually curating and representing knowledge is time-consuming and difficult to scale, necessitating the development of automated methods.
- Representing ignorance is an open problem: Current systems struggle to represent what is not known, including gaps, ambiguities, and uncertainties, which are essential aspects of scientific knowledge.
- The ultimate vision: a system that synthesizes and reasons across disciplines: The paper posits that the biggest challenge and opportunity lie in building computational systems that can integrate disparate facts from across all biomedical disciplines to make sound judgments and generate novel hypotheses.
Key Results
This paper is a position paper and survey, not an empirical study. Therefore, it does not present quantitative results in terms of accuracy or performance metrics. Its "results" are conceptual:
- It establishes that knowledge-based approaches are a vital, yet under-appreciated, component of biomedical data science.
- It categorizes and surveys a range of existing methods for knowledge representation and inference in biomedicine.
- It highlights the significant potential of these approaches for hypothesis generation, data interpretation, and accelerating biomedical discovery.
Limitations and Future Work
- Limitations:
- Knowledge Acquisition Bottleneck: Human curation is difficult to scale, and automated methods are not yet sufficient.
- Representing Ignorance: Few computational methods exist to represent gaps, ambiguities, and uncertainties in knowledge.
- Scalability of Inference: Many promising inference methods are computationally intractable over large knowledge bases.
- Vision for Contribution: Developing a clear vision for how knowledge-based systems can effectively contribute to biomedical research remains a challenge.
- Future Work:
- Developing automated methods for knowledge acquisition and representation.
- Creating computational methods to represent and reason about ignorance.
- Improving the scalability and efficiency of inference methods.
- Building computational systems that can synthesize knowledge across disciplines, rank hypotheses, and generate novel mechanistic explanations.
- Exploring the synthesis of computational simulation with knowledge-based hypothesis generation and refinement.
Summary updated