The process for developing a new drug consists of the drug discovery phase which involves selecting drug candidates and the drug development phase which involves testing drug toxicity and effects by clinical trials. The overall process takes 10-15 years and costs 2.6 billion dollars on average, but the success rate of the drug development phase is less than 10%. The success of drug development is dependent on selecting promising drug candidates out of millions of chemical compounds, and lowering the failure rate using machine learning algorithms. To achieve these aims, I researched machine learning-based drug discovery methods such as personalized drug combination synergy prediction and response-based drug similarity prediction for drug candidate selection.
My first research topic is on predicting which drug combination will be the most effective for a cancer patient. When a single drug is used for treating a cancer, the cancer tries to find bypasses in the biological network, avoid the effects of the drug, and prevent apoptosis, which makes it difficult to treat. It is also difficult to develop anticancer drugs that are effective for several months. Hence, drug combination therapy, which can slow down resistance to relapse and have synergistic effects, is used as an alternative for cancer therapy. However, the synergy of drug combinations depend on the genetic characteristics of patients. Therefore, it is important to prescribe the most synergistic combination of drugs for a cancer patient based on their genomic information. I propose a regression model that predicts the synergy of cancer drug combinations for cancer cell lines using the genetic information of the cell lines and the pharmacological information of drugs. The model can find drug combinations as drug candidates for personalized cancer therapy.
My second research topic is on predicting response-based drug similarity. Ligand-based drug discovery is a traditional method for selecting drug candidates. Its principle is that structural analogs of molecules will bind to the same targets and produce the same effects. However, the structural similarity of two drugs does not always mean their effects will be similar. Even if the structures of the drugs are different, off targets of the drugs or the propagation of drug effects through biological pathways can produce the same effects. Therefore, it is necessary to find drug candidates based on response similarity, and not structural similarity. I propose ReSimNet which is a Siamese neural network model that predicts the differential gene expression similarity of two drugs. ReSimNet can recommend novel drug candidates that have response similarities to well-known drugs.