Graph neural networks are powerful computational tools that can predict relationships from graphs at the node, edge and graph level. They are especially useful for handling non-Euclidean data with lots of interdependencies – such as drug to drug interactions and drug predictions. A recent study has described a new graph method, GraphRepur, that predicts drugs that can be repurposed for the treatment of breast cancer.
Breast cancer is the most common type of invasive cancer in women. It accounts for 15% of all female cancer deaths. Current treatments are characterised by the hormone profile of the cancer tissue and more recently, on the gene-expression profile, using tests such as Oncotype DX or MammaPrint. However, due to the lengthy drug discovery process, drug repurposing opens another avenue for developing novel treatments for breast cancer.
Drug Repurposing – Network or Signature
Drug repurposing refers to identifying new uses for approved drugs or drugs that failed the development phases, rather than going through the normal discovery and clinical trial process. Although this process can still be lengthy, taking on average 6.5 years and costing $300 million – it is still shorter compared to traditional methods.
Previous drug repurposing computational methods are usually network-based or signature-based. Networks refer to interactions, such as between drugs, diseases or specific targets. Whereas, signature-based datasets use genomic data, such as gene expression profiles, to compare controls to ‘drug-altered’ profiles. Previously, these have been used to repurpose drugs for Alzheimer’s Disease.
The Graph Neural Network – GraphRepur
The study proposed the use of a graph neural network model, GraphRepur, which combines the two systems (as network-based methods ignore prior scientific knowledge and signature-based methods cannot compute drug-drug links information).
By taking the drug-exposure gene expression data from LINCS and using STITCH for the drug-drug links information, GraphRepur combines both datasets to predict new drugs for breast cancer.
GraphRepur was benchmarked against several existing machine-learning methods and outperformed them all according to a five-fold validation.
The most telling indication of GraphRepur being useful is that after being used on an external validation set, some of the predictions have now been confirmed by published studies.
GraphRepur identified 169 existing drugs that could be used in breast cancer and the researchers were able to validate these results by looking at current literature and trials. Of the 169 identified drugs, eight are currently in over 40 separate clinical trials for breast cancer. These include drugs previously used for multiple sclerosis, CLL and multiple myeloma.
However, there are limitations to GraphRepur. For example, it cannot be used for new drug discovery as it requires extensive datasets of known effective drugs. Also, breast cancer is a very heterogeneous disease and limited work was done on different cancer cell lines for this study (which GraphRepur performed worse at). Nonetheless, the researchers point out this could be due to the lack of cancer cell line datasets and the vast levels of tumour heterogeneity.
Overall, GraphRepur is an excellent tool for combining multiple ‘-omics’ for the use of drug repurposing, provided we can keep up with the tools’ demand for diverse datasets.