Predicting Cancer Drug Response Using an Adapted Deep Neural Network Model

Recent advancements in biotechnology have contributed to the concept of precision oncology through the application of machine learning algorithms. The proposed work focuses on the improvement of a novel Deep Learning model, known as Reference drug-based Deep Neural Network (RefDNN), applied to the prediction of cancer drug response. The model utilizes drug's structure similarity profiles (SSP) to describe the similarity between different reference cancer drugs and uses an SSP vector to weigh the pre-predicted drug response probability obtained by the use of Elastic Net (EN), with the weighted response to be the input of the Deep Neural Network. The prediction performance of RefDNN has been improved by adding a t- distributed stochastic neighbor embedding (t-SNE) based feature extraction estimator, through the integration of gene expression, copy number variants (CNV) and mutation data. This adaptation was used to characterise the model and customize the prediction procedure based on cell line data to provide more precise and time-efficient results. The performance of the proposed system was based on a 5-fold cross validation and was compared to the original RefDNN model, showing significant improvements in accuracy and reduction of the computational processing time.