Challenges in the construction of knowledge bases for human microbiome-disease associations

Varsha D. Badal, Dustin Wright, Yannis Katsis, Ho‐Cheol Kim, Austin D. Swafford, Rob Knight, Chun‐Nan HsuMicrobiome 2019 · DOI

Abstract

The last few years have seen tremendous growth in human microbiome research, with a particular focus on the links to both mental and physical health and disease. Medical and experimental settings provide initial sources of information about these links, but individual studies produce disconnected pieces of knowledge bounded in context by the perspective of expert researchers reading full-text publications. Building a knowledge base (KB) consolidating these disconnected pieces is an essential first step to democratize and accelerate the process of accessing the collective discoveries of human disease connections to the human microbiome. In this article, we survey the existing tools and development efforts that have been produced to capture portions of the information needed to construct a KB of all known human microbiome-disease associations and highlight the need for additional innovations in natural language processing (NLP), text mining, taxonomic representations, and field-wide vocabulary standardization in human microbiome research. Addressing these challenges will enable the construction of KBs that help identify new insights amenable to experimental validation and potentially clinical decision support.

Introduction: The Explosion of Microbiome Literature

The field of human microbiome research has experienced an unprecedented surge in activity over the last two decades. Driven by the decreasing costs of DNA sequencing, scientists are increasingly uncovering links between microbial communities and a wide array of physical and mental health conditions. However, this wealth of information remains largely trapped within the unstructured text of thousands of scientific publications.

The exponential growth of microbiome-related publications over the last two centuries.

As shown in the figure above, the volume of literature is now growing at a rate that makes manual curation by human experts impossible. To consolidate these scattered discoveries and enable multi-omics hypothesis generation, the field requires automated Knowledge Bases (KBs)—large-scale structured repositories that define entities like microbes and diseases and the complex relationships between them.

The Automated Knowledge Base Pipeline

Constructing a robust KB for the microbiome requires a multi-stage Natural Language Processing (NLP) pipeline. This process transforms raw literature into structured, actionable data.

The standard pipeline for extracting microbiome-disease associations from scientific literature.

The pipeline typically consists of several key steps:

  • Entity Extraction & Normalization: Identifying mentions of microbes and diseases within the text and mapping them to standardized identifiers. This involves moving from varied textual mentions (e.g., "E. coli" vs "Escherichia coli") to a single ground-truth entry.
  • Relationship Extraction: Determining whether a sentence describes a genuine biological association. This is often achieved through co-occurrence rules or advanced neural network models.
  • Knowledge Graph Completion: Using algorithmic methods to infer new, unstated relationships based on the existing network of extracted data.
  • Experimental Integration: Cross-referencing NLP-extracted claims with raw experimental metadata from repositories like Qiita or the European Nucleotide Archive (ENA) to validate findings.

Technical Challenges: Taxonomy and Dynamics

Despite the success of NLP in general biomedicine, the microbiome presents unique hurdles. One primary issue is "Taxonomy Chaos." There is no universally accepted taxonomy for microbes or diseases. A system might fail to recognize that P. gingivalis and Porphyromonas gingivalis refer to the same organism if the normalization step is not sufficiently robust. Furthermore, clinical dictionaries (like SNOMED-CT) often do not align perfectly with statistical ones (like ICD-10).

Additionally, microbiome-disease relationships are inherently dynamic. Unlike genetic associations, which are often treated as static, microbial populations fluctuate based on diet, medication, and environmental factors. Current KBs struggle to represent these conditional associations accurately.

Current Limitations and the Need for Scalability

Existing manual curation efforts, while high-quality, are fundamentally unscalable. For instance, the Disbiome database—one of the most comprehensive available—covers approximately 200 diseases but is based on the manual review of only 500 abstracts.

Furthermore, current NLP performance in this domain is lacking. State-of-the-art relation extraction models often achieve F1-scores below 0.8, which is insufficient for fully automated construction without human oversight. There is also a critical shortage of domain-specific training data; most existing biomedical NLP datasets focus on general species or diseases rather than the specific nuances of the human microbiota.

Conclusion and Future Directives

To bridge the gap between literature and structured knowledge, the authors propose three key directives for the research community:

  1. Standardization: Develop controlled vocabularies specifically for human microbiome-disease associations.
  2. Community Annotation: Build large, community-annotated text corpora to provide the "ground truth" needed to train more accurate deep learning models.
  3. Technological Translation: Adapt recent breakthroughs in NLP, such as pre-trained Transformer models (e.g., BERT), specifically for the microbiome domain.

By addressing these challenges, the scientific community can build KBs that not only consolidate existing knowledge but also help identify laboratory contamination biases. For example, KBs can expose contradictions where a microbe is reported as both "elevated" and "reduced" for the same disease, often flagging potential PCR reagent contaminants like Prevotella. Ultimately, these tools will generate new insights for clinical decision support.

Summary updated

Abstract

The last few years have seen tremendous growth in human microbiome research, with a particular focus on the links to both mental and physical health and disease. Medical and experimental settings provide initial sources of information about these links, but individual studies produce disconnected pieces of knowledge bounded in context by the perspective of expert researchers reading full-text publications. Building a knowledge base (KB) consolidating these disconnected pieces is an essential first step to democratize and accelerate the process of accessing the collective discoveries of human disease connections to the human microbiome. In this article, we survey the existing tools and development efforts that have been produced to capture portions of the information needed to construct a KB of all known human microbiome-disease associations and highlight the need for additional innovations in natural language processing (NLP), text mining, taxonomic representations, and field-wide vocabulary standardization in human microbiome research. Addressing these challenges will enable the construction of KBs that help identify new insights amenable to experimental validation and potentially clinical decision support.
BibTeX
@inproceedings{badal2019challenges,
  author = {Varsha D. Badal and Dustin Wright and Yannis Katsis and Ho‐Cheol Kim and Austin D. Swafford and Rob Knight and Chun‐Nan Hsu},
  title = {{Challenges in the construction of knowledge bases for human microbiome-disease associations}},
  booktitle = {Microbiome},
  year = {2019},
  doi = {10.1186/s40168-019-0742-2},
}
Figure