TL;DR
This paper introduces a novel method for causal feature selection using a knowledge graph (KG) that integrates biomedical literature and ontologies. By automatically extracting and structuring knowledge, the approach aims to identify potential confounders, colliders, and mediators for complex relationships like depression and Alzheimer's disease, augmenting human expertise and improving the rigor of observational studies.
What Problem Does This Paper Address?
Estimating causal effects from observational data is crucial but challenging due to uncontrolled confounding, which can distort findings. Traditionally, identifying confounders relies on expert knowledge and literature reviews, a process that is becoming increasingly infeasible as the volume of biomedical literature explodes. Furthermore, variables can play multiple causal roles (e.g., being both a confounder and a mediator), adding complexity. This paper addresses the need for a scalable, automated method to identify relevant causal variables by leveraging structured knowledge from both the vast biomedical literature and formal ontologies.
Methods
The authors developed a pipeline that combines machine reading systems with a knowledge graph (KG) and ontologies.
- Literature Scoping: They defined a search scope for relevant biomedical literature.
- Machine Reading: Three machine reading systems (SemRep, EIDOS, REACH) extracted structured knowledge (triples) from titles and abstracts.
- Knowledge Post-processing: This involved terminology mapping (e.g., UMLS to OBO), inferring missing links through logical closure (transitivity, symmetry), and filtering redundant or irrelevant information.
- KG Construction: The processed triples were integrated with biomedical ontologies using the PheKnowLator platform to create an ontology-grounded KG.
- KG Search: SPARQL queries, based on epidemiological definitions of confounders, colliders, and mediators, were used to search the KG. Dijkstra's shortest path algorithm was employed to find relationships.
- Causal Feature Selection: The KG search results were analyzed to identify variables and their potential causal roles.
Datasets
- PubMed: Used for scoping the literature.
- SemMedDB: A database of semantic predications extracted from PubMed, used for comparison.
- Ontologies: Various biomedical ontologies (e.g., BFO, GO, OBO, DO, HP, RO) were merged to construct the KG.
Key Insights and Ideas
- KG as a Bridge: Knowledge graphs are powerful tools for bridging the gap between unstructured literature and structured, computable knowledge, enabling automated reasoning for causal inference.
- Ontology Grounding is Key: Integrating ontologies provides a semantic backbone, helping to prune erroneous information from machine reading and ensure biological plausibility.
- Variables Play Multiple Roles: The study highlights that variables rarely fit neatly into a single causal role; many act as "chimeras" (playing multiple roles), which is critical for accurate causal modeling.
- Automated Discovery of Novel Confounders: The KG approach can uncover potential confounders that may not be explicitly reported in existing observational studies, offering new avenues for research and intervention.
Key Results
- The KG search identified 126 unique potential confounders for the depression-AD association, including phenotypes, drugs, and genes.
- Out of 58 identified confounder phenotypes, only 31 exclusively acted as confounders, while 27 played other roles.
- Obstructive sleep apnea emerged as a potential novel confounder for depression and AD.
- Anemia was identified as a variable playing combined roles (mediator/collider in KG, confounder in SemMedDB).
- The KG search identified 28 unique potential colliders and 18 unique potential mediators.
- The method produced interpretable mechanistic explanations in the form of reasoning paths.
Limitations and Future Work
- Incomplete Knowledge: Biomedical ontologies and literature are inherently incomplete, leading to potential information loss and erroneous inferences.
- Mapping Challenges: Imperfect terminology mapping between different knowledge sources (e.g., UMLS to OBO) resulted in concept loss and reduced granularity.
- Search Depth Constraints: The constrained search depth for efficiency might have limited recall for certain variable categories.
- KG vs. SemMedDB Discrepancies: The KG and SemMedDB search results did not always agree, highlighting the need for expert review to adjudicate conflicting assertions.
- Future Work:
- Develop tools for automated visualization of KG search outputs (causal graphs and reasoning paths).
- Improve mappings and expand ontology coverage.
- Gather meta-data for causal assertions to assess transportability.
- Explore causal feature selection for scenarios with multiple concurrent exposures.
- Address "M"-bias and other complex confounding scenarios.
Summary updated