Abstract
Cardiotoxity is a severe adverse effect associated with anti-cancer drug treatment. Many research focus on myocardial excitation, the cytotoxicity to cardiomyocyte is also important. This toxicity seems to be assessed by simple in vitro experiments, however, such challenges are limited. A reason for this may be that cellular sensitivities differ by culture environment, which affect gene expression profile. Therefore, we searched gene features essential to reproduce cytotoxicity. To consider as many toxicity mechanisms as possible, we utilized data in the Genomics of Drug Sensitivity in Cancer project. We can access drug sensitivity data for about 1000 cancer cells to about 300 anti-cancer drugs together with gene expression profiles for each cell. Using these data, supervised machine learning was performed to construct a neural network model, in which sensitivities such as EC50 are returned when gene expression profile and drug property are input. Using this model, we extracted genes having high impacts in the model calculation when cardiomyocyte gene expression profile is input in combination with various drugs. The impact was evaluated by calculating partial derivatives to the model output, thus one can obtain a gene number-length vector. We clustered the drugs using the impact vectors and extracted high impact genes for each cluster. By gene ontology analyses, the extracted genes were confirmed to be associated with cytotoxicity. Further validation is required, but this result suggests that the extracted genes are key features to assess various types of toxicities.
