TL;DR
Biological pathways are crucial for understanding life, but integrating information from diverse sources reveals many contradictory statements. This paper introduces a method to resolve these "semantic conflicts" by considering the context (like cell type or disease) in which an interaction occurs, significantly reducing the number of conflicts and improving pathway data quality.
What Problem Does This Paper Address?
Biological pathways, which describe how molecules interact, are fundamental to understanding life's mechanisms. As more pathway information is generated from both curated databases and literature mining, a significant challenge arises: inconsistencies and contradictions. For instance, one source might state that protein A increases the activity of gene B, while another claims it decreases it. These "semantic conflicts" can render pathway simulations unreliable and hinder downstream analyses like gene annotation or network dynamics. The paper aims to quantify these conflicts, understand their causes, and propose a method for their resolution.
Methods
The authors developed a systematic approach to identify and resolve semantic conflicts.
- Integrated Pathway Dataset Construction: They combined data from the KEGG pathway database (a curated source) and a large corpus of PubMed abstracts (literature mining).
- Meta-Relation Mapping: They mapped various interaction terms (e.g., "activates," "inhibits," "phosphorylates") into two general "meta-relations": 'increase' and 'decrease'.
- Context Definition: They defined four types of context: organ (OG), cell-type (CT), disease (DS), and drug (DR). These contexts specify the conditions under which a pathway interaction is valid.
- Conflict Identification: A conflict was defined as an interaction where both 'increase' and 'decrease' rules exist.
- Context-Based Conflict Resolution: The core innovation is using context. If two conflicting rules have different, non-overlapping contexts (e.g., "A increases B in liver" vs. "A decreases B in brain"), they are no longer considered contradictory. They developed a method to group rules by identical context combinations.
- Balance Analysis: For remaining conflicts, they introduced a "minor rule frequency" (MRF) metric to assess the balance of evidence supporting each side of a conflict. A low MRF suggests one side might be erroneous.
Datasets
- KEGG Pathway Database: Used as a source of curated human pathways.
- PubMed Abstracts: A large corpus (13,214,710 abstracts) was mined for pathway information.
- Ontologies for Contexts:
- MeSH (Medical Subject Heading) database for organ and disease terms.
- DrugBank database for drug terms.
Key Insights and Ideas
- Literature Mining is a Major Source of Conflict: The study found that conflicts are far more prevalent in pathways extracted from literature than in curated databases, highlighting the need for robust validation of LM-derived data.
- Context is King (for Resolution): The most significant finding is that considering biological context (cell type, organ, disease, drug) can resolve a substantial majority of apparent conflicts. This aligns with the biological reality that molecules can have different roles in different cellular environments.
- Dual Functions are Real: Some conflicts that persist even after context consideration likely represent genuine "dual functions" of molecules, where they can indeed have opposing effects under different (perhaps unannotated) conditions.
- MRF as a Diagnostic Tool: The MRF metric provides a quantitative way to distinguish between potentially erroneous information (dominating conflicts) and genuine dual functions (balanced conflicts).
Key Results
- High Conflict Rate: Approximately 19.4% of integrated pathway interactions were found to contain conflicting information.
- Contextual Resolution Power: Considering context reduced the percentage of conflicting rule groups dramatically, down to 13.6% when all four context types were applied. Even single contexts showed significant conflict reduction.
- Literature vs. Database: KEGG alone showed a very low conflict rate (0.86%), reinforcing its reliability compared to literature mining.
- Conflict Balance: For remaining conflicts, the majority (94.3%) were "balanced" (MRF > 0.25), suggesting dual functions rather than simple errors. Only a small percentage (5.7%) were "dominating" (MRF < 0.25), indicating potential text mining errors.
Limitations and Future Work
- Limitations:
- The study acknowledges that not all context information was available for all literature-derived pathways.
- The MRF analysis suggests that some conflicts might be due to unconsidered contexts beyond the four studied (e.g., measurement time, ethnic origin).
- The authors note that "react" or "translocate" relations were not directly used for conflict analysis due to their less direct directional interpretation.
- Future Work:
- The authors suggest exploring additional context types to further resolve conflicts.
- They propose using the MRF metric to identify novel dual-role information and improve literature mining techniques by correcting identified errors.
- The ultimate goal is to clean pathway databases for more accurate secondary analyses.
Summary updated