Source-linked AI summary
Neuro-symbolic representation learning on biological knowledge graphs
Mona Alshahrani, Mohammed Asif Khan, Omar Maddouri, Akira R Kinjo, Núria Queralt-Rosinach, Robert Hoehndorf
TL;DR
Biological knowledge graphs integrate heterogeneous data and ontology-based knowledge, but existing graph feature-learning methods have not been widely evaluated on structured biological knowledge. The paper combines automated reasoning with neural embedding learning to represent explicit and inferred graph information, achieving high-accuracy edge prediction that often matches or outperforms state-of-the-art methods. The approach supports using Semantic Web knowledge graphs for biological data analysis and discovery.
Problem
Feature-learning methods for graph-structured data have not been widely applied and evaluated on structured biological knowledge, despite the growing use of Semantic Web technologies and knowledge graphs in biology.
Method
The method combines automated reasoning over ontologies with neural-network feature learning to generate node embeddings from biological knowledge graphs.
Results
The embeddings support biological edge prediction with high accuracy, often matching or outperforming state-of-the-art methods across disease-gene, drug-target, drug-indication, gene-function, and related tasks.
Takeaways & Limitations
Semantic Web knowledge graphs can support biological data analysis and discovery in addition to storing, retrieving, and querying biological information.
Takeaways & Limitations
The approach relies on qualitative data and ignores quantitative information such as edge strength or confidence values.
Abstract
from arXiv · showhide
Motivation: Biological data and knowledge bases increasingly rely on Semantic Web technologies and the use of knowledge graphs for data integration, retrieval and federated queries. In the past years, feature learning methods that are applicable to graph-structured data are becoming available, but have not yet widely been applied and evaluated on structured biological knowledge. Results: We develop a novel method for feature learning on biological knowledge graphs. Our method combines symbolic methods, in particular knowledge representation using symbolic logic and automated reasoning, with neural networks to generate embeddings of nodes that encode for related information within knowledge graphs. Through the use of symbolic logic, these embeddings contain both explicit and implicit information. We apply these embeddings to the prediction of edges in the knowledge graph representing problems of function prediction, finding candidate genes of diseases, protein-protein interactions, or drug target relations, and demonstrate performance that matches and sometimes outperforms traditional approaches based on manually crafted features. Our method can be applied to any biological knowledge graph, and will thereby open up the increasing amount of Semantic Web based knowledge bases in biology to use in machine learning and data analytics. Availability and Implementation: https://github.com/bio-ontology-research-group/walking-rdf-and-owl Contact: robert.hoehndorf@kaust.edu.sa
1 Introduction
Biological databases increasingly use Semantic Web standards, Linked Data, and ontologies, but machine learning must integrate explicit data with inferred semantic knowledge. The paper develops a neuro-symbolic embedding method and applies it to biological edge prediction, matching or surpassing state-of-the-art approaches in many cases.
- Motivation: Semantic Web technologies and Linked Data have become successful in biology, supporting interoperable data integration through shared identifiers and standardized relations.Biological databases increasingly publish data as Linked Data, while FAIR principles promote interoperability through linking and federated queries.
- Motivation: Ontologies formally represent domain concepts and their interconnections, but ontology inferences are not always used to infer new relations between biomedical data.Ontologies are widely used for biological annotation and metadata, while reasoning is increasingly applied for ontology development and quality assurance.
- Motivation: Graph feature-learning methods exist, but biological knowledge requires interoperability between Linked Data and ontology-based semantic knowledge, including automated reasoning over inferred information.This integration also supports handling data consistency and incompatible conceptualizations.
- Contribution: The proposed method combines automated ontology reasoning with neural-network feature learning to generate node embeddings for Semantic Web and Wikidata knowledge graphs.The embeddings encode information from the semantic layer and are designed for biological knowledge graphs.
- Contribution: The embeddings predict disease genes, drug targets, drug indications, gene functions, and other biological associations, often matching or outperforming state-of-the-art methods.The paper applies the representations to edge prediction tasks with biological meaning.
- Contribution: Linked Data and ontologies can form biological knowledge graphs that support not only retrieval and search but also data analysis and discovery of novel biological knowledge.
2 Methods
The method constructs integrated biological knowledge graphs from ontology annotations and heterogeneous databases, computes OWL 2 EL inferences, and learns node embeddings from edge-labeled random walks. These embeddings are generated through a modified DeepWalk and skip-gram pipeline with configurable walk and representation parameters.
- 2.1 Data Description: The knowledge graph integrates three biomedical ontologies with annotations and relations from protein, chemical-protein, gene-disease, drug, side-effect, indication, and phenotype databases.The graph includes GO, HPO, and Disease Ontology resources alongside biological interaction and annotation datasets.
- 2.1 Data Description: Biological identifiers are mapped across source databases, while ontology annotations are represented as relations between biological instances and ontology-class instances.The construction distinguishes biological entities from ontology classes and assigns new IRIs to annotation instances.
- 2.2 Reasoning: OWL 2 EL provides polynomial-time reasoning over class hierarchies, object properties, and instance classification for the generated knowledge graphs.The profile supports constructs including intersections, existential quantification, subclass relations, property chains, transitivity, and reflexivity.
- 2.2 Reasoning: The knowledge graph is deductively closed with respect to OWL 2 EL by adding explicit inferences between named individuals and classes.The inferred graph is generated with the OWL API and ELK reasoner, then used as the basis for node-embedding generation.
- 2.3 Walking RDF and OWL: A modified DeepWalk represents each random walk as an alternating sequence of vertices and edge labels, preserving relation types during graph traversal.This edge-labeled walk is the corpus used for learning node representations.
3 Results
The workflow integrates heterogeneous biological data and ontologies into a reasoned knowledge graph, then learns node embeddings from both asserted and inferred relations. These embeddings support edge prediction and drug-repurposing tasks, with performance varying by relation type and matching or exceeding reported alternatives in several settings.
- 3.1 Neuro-symbolic feature learning using Semantic Web technologies: The workflow builds a two-layer biological knowledge graph, deductively adds inferred triples, and learns node embeddings from random walks over the closed graph.The graph combines instance-level biological data with ontology classes and annotations; automated reasoning adds inferred relations before Word2Vec skip-gram embedding learning.
- 3.2 Edge Prediction: Edge-prediction performance varies by relation type, reaching F-measure 0.94 and ROCAUC 0.98 for has target edges but only ROCAUC 0.77 for has disease phenotype.Reasoning slightly improves predictions between instances but mostly decreases performance for instance-to-ontology-class edges; hierarchical outputs also impose consistency constraints.
- 3.2 Edge Prediction: The authors conclude that ontology-informed embeddings contain useful biological information, while application-specific knowledge graphs may still be needed for particular predictive tasks.The method is presented as supporting data analysis and discovery from integrated Linked Data and ontologies, but the paper notes a need for more specific graphs in some applications.
- 3.3 Drug repurposing on biological knowledge graphs: Drug-pair embeddings predict shared indications or targets with ROCAUC up to 0.87, compared with side-effect similarity baselines of 0.75 for shared targets and 0.83 for shared indications.The evaluation removes direct indication and target edges from the graph before training, testing whether embeddings encode related information indirectly.
- 3.3 Drug repurposing on biological knowledge graphs: Drug embeddings also assign significantly greater cosine similarity to drugs sharing a top-level ATC category than to drugs in different categories (p < 3 · 10^-4).This similarity-based evaluation uses the full knowledge graph and tests whether embedding proximity reflects shared therapeutic classification.
4 Discussion
The approach extends biological knowledge-graph analysis by combining symbolic representations with predictive modeling, while remaining constrained by its treatment of qualitative data. It depends on separating biological content from metadata and supports using Semantic Web graphs for biological discovery.
- Ignoring quantitative information limits the approach because edge weights can represent either relation confidence or biological association strength.The authors note that incorporating weights would require distinguishing these different semantics.
- The method relies on data models that separate biological content from metadata such as provenance and authorship.Such metadata remains important for retrieval and data quality, but is not the primary biological content modeled by the approach.
- Semantic Web knowledge graphs can represent biological phenomena, support data analysis, and build predictive models beyond storage and querying.The discussion highlights heterogeneous relations, incremental information addition, inference rules, and Web standards as advantages.