Heterogeneous Graph Learning
Knowledge graphs are visualization of information with multi-type relations (edges) among some multi-type entities (nodes) within an environment of interest. AThe heterogeneous aspect comes from the graph having two or more types of nodes and two or more types of edges.
Usage
In recent studies on Alzheimer's drug repurposing (Reference 1) Graph Neural Networks (GNN) were used to identify biological interactions from prior knowledge databases containing information on effective treatments and genes associated with high risk. The nodes of the graph included drugs, genes, pathways and gene ontology (GO) connected by interactions including drug-target interaction, drug-drug structural similarity, gene-gene interaction, gene-pathway association, gene-GO association and drug-GO association. A comprehensive graph can be created combining existing information from multiple sources on the biological interactions of complex drug-gene relationships, and from there we can use machine learning to train a model and address the incomplete knowledge in the graph. Note that in this paper the genes are encoded/embedded, simply meaning they are represented as numerical vectors/matrices to capture their function. To give a broad overview of the study's workflow:
- A knowledge graph is built to describe the interaction between drugs, genes, gene ontology and pathways
- Node
emeddingsembeddings are derived using a multi-relational Variational Graph Auto-Encoder (VGAE) - Machine learning model ranks drug-
canidatescandidates based on multi-level evidence - Drug combinations are searched for complementary exposure patterns, using previously ranked drug
canidatescandidates - Validate drug combinations with "oxidative stress responses"
By keeping the focus previously approved drugs, the graph can identify synergy between medications that treat complex diseases. The "Complementary Exposure Pattern" of analysis used suggests drug combinations are effective when the target of the drugs hit the disease module without overlap.
While this study breaks ground in applying knowledge graphs with multi-task learning to fragmented multi-modal data, it is subject to the limitations of each dataset it combines. The knowledge graph alone extracts data from:
- Universal protein-protein interactions from STRING
- Interaction between genes, drugs, GO, and pathways from Comparative
ToxicogenomicsToxicogenomic Database (CTD) - Drug-Drug associations based on structural similarities using scores from the RDkit package
- Gene IDs from National Center for Biotechnology information (NCBI)
- High confidence Alzheimer's associated genes from Agora's nominated gene list
If any of the data-driven drug efficacy is biased, then so is the model. Some of the datasets contained in vitro studies (lab based experiments,experiments usually in a test tube), which does not guarantee identical treatment outcomes. However, given the lack of clinical evidence from a huge list of compoundscompounds, this research could still provide insight on drug combinations in future research opportunities.