Rapid Knowledgebase Construction and Hypotheses Generation Using Extractive Literature Search

Shaked Launer-Wachs, Hillel Taub-Tabib, Yoav Goldberg, Yosi ShamaybioRxiv 2022 · DOI

TL;DR

This paper introduces SPIKE-KBC, a novel method and platform that empowers individual researchers to rapidly build personalized biomedical knowledgebases from scientific literature. This democratizes knowledge discovery, enabling faster hypothesis generation and exploration without requiring extensive bioinformatics expertise.

What Problem Does This Paper Address?

Building comprehensive biomedical knowledgebases is currently a monumental task. Traditional methods involve expensive, multi-year projects with teams of bio-curators and computer scientists. This limits existing knowledgebases to well-trodden research areas, leaving a vast "long tail" of specialized topics unexplored. This bottleneck hinders individual researchers and small labs from quickly synthesizing information and generating novel hypotheses relevant to their niche interests.

Methods

The core of SPIKE-KBC lies in extractive search. This technique combines standard keyword searching with "capture operators" that can extract specific pieces of information from selected documents. The authors extend this by developing a web-based platform that guides users through a structured process:

  1. Project Definition: Users define the classes of entities they are interested in (e.g., drugs, cancers, cell types).
  2. Entity List Generation: Extractive queries are used to mine the literature for mentions of these entities, creating initial lists.
  3. Relation Instance Extraction: Pairwise Boolean queries are then employed to find sentences where entities from different lists co-occur, suggesting potential relationships.
  4. Annotation: Users review these extracted instances, labeling them as positive or negative, and adding them to the knowledgebase with supporting evidence.
  5. Visualization and Exploration: The resulting knowledgebase can be visualized as graphs or trees, allowing users to explore connections and generate hypotheses.

This approach significantly lowers the barrier to entry, as it doesn't require deep NLP or computer science expertise. The system is designed to be iterative, allowing users to refine queries and expand their knowledgebases over time.

Datasets

The paper doesn't explicitly list pre-existing datasets used for training models. Instead, it leverages the vast corpus of publicly available biomedical literature (primarily from PubMed) as its data source. The authors demonstrate their method by constructing several knowledgebases, effectively creating new, domain-specific datasets for their use cases.

Key Insights and Ideas

  • Democratizing Knowledgebase Construction: The most significant insight is that building useful knowledgebases doesn't have to be an all-or-nothing, resource-intensive endeavor. SPIKE-KBC makes it accessible to individual researchers.
  • The Power of "Extractive Queries": This technique is the engine that drives the efficiency. By intelligently extracting specific information, it bypasses the need for full natural language understanding models in many cases.
  • Personalized Knowledgebases as a Scientific Tool: The paper highlights that these knowledgebases aren't just repositories of facts; they are active tools for hypothesis generation and exploration, directly integrated into the research workflow.
  • The "Long Tail" of Research: The authors emphasize that by enabling personalized knowledgebase construction, they are addressing the under-explored areas of scientific inquiry that are often overlooked by large-scale, community-focused projects.

Key Results

The authors demonstrate the effectiveness of SPIKE-KBC through five case studies:

  • Cancer Frontiers Knowledgebase: Built in a few hours, this knowledgebase mapped challenges and controversies in cancer research.
  • Tissue Engineering and Regeneration Knowledgebase: Constructed in about 4 hours, it generated over 3600 hypotheses.
  • Cancer Surgeries and Radiotherapeutics Knowledgebases: Built in approximately a day each, these yielded thousands of hypotheses.
  • Targeted Drug Delivery Knowledgebase: This more comprehensive knowledgebase took 4-5 days to construct and resulted in over 10^6 possible hypotheses.

Across these examples, the system consistently facilitated rapid hypothesis generation and exploration, with the resulting knowledgebases being directly linked to the literature for evidence.

Limitations and Future Work

  • Annotation Effort: While significantly reduced compared to traditional methods, manual annotation is still a necessary step, especially for more comprehensive knowledgebases. The authors note that the time investment for this can range from hours to weeks.
  • Noise Tolerance: For exploratory knowledgebases, some level of noise (inaccurate relations) is acceptable as hypotheses are linked to literature and can be reviewed by experts. However, for more public-facing resources, higher precision is required.
  • Future Work: The authors provide the SPIKE-KBC platform as open-source, encouraging further development and application. They also suggest that the potential for generating unpublished hypotheses can accelerate discovery across multiple fields of biomedicine. The paper also hints at the need for more sophisticated methods for handling synonyms and acronyms, particularly in complex domains like drug delivery.

Summary updated

Abstract

Abstract As knowledgebases become increasingly important for structuring vast amounts of scientific knowledge and making it accessible to researchers, their construction entails expensive multi-year projects involving teams of bio-curators, computer scientists, or both. This restricts the coverage of existing knowledgebases to a limited set of popular topics, leaving a long tail of more specialized interests uncovered. We present a methodology and a supporting tool to allow individual researchers or small teams, without background in bio-curation or computer science, to mine the scientific literature and construct ad-hoc, personalized, and literature-anchored knowledgebases, that are tailored around their specific research interests and support their scientific goals. The time investment involved in creating a knowledgebase ranges from a few hours to a few weeks, depending on the desired coverage and accuracy. We demonstrate the methodology by constructing knowledgebases for different purposes: a high-level overview of challenges and controversies in a field (the cancer frontiers knowledgebase); a mapping of main concepts and interactions in a field, to support lab-internal hypothesis generation (tissue engineering and regeneration, cancer surgery and radiotherapy knowledgebases); and a comprehensive and accurate knowledgebase designated as an online up-to-date resource for the wider research community (the cell specific drug delivery knowledgebase). In each case we show how the structured knowledgebase, coupled with effective visualizations, facilitates effective data exploration, hypothesis generation and meta-analysis. We implement the method as part of an open source web-based platform for knowledgebase construction, available publicly and freely at https://spike-kbc.apps.allenai.org .
BibTeX
@misc{launerwachs2022rapid,
  author = {Shaked Launer-Wachs and Hillel Taub-Tabib and Yoav Goldberg and Yosi Shamay},
  title = {{Rapid Knowledgebase Construction and Hypotheses Generation Using Extractive Literature Search}},
  year = {2022},
  doi = {10.1101/2022.02.13.480241},
  howpublished = {bioRxiv},
}
Figure