A Novel Method for Predicting Drug Synergy Based on Matrix Factorization

Combination therapy has proven to be highly effective in the treatment of complex diseases such as cancerand infectious diseases. When compared to monotherapy, drug combination therapy can improve cancertreatment efficacy, reduce drug dose-dependent toxicity, and prevent drug resistance. In this study, we presenta novel regression method based on matrix factorization. Using K-fold nested cross-validation, we comparethe results of the presented method to the results of two novel regression methods, PRODeepSyn andDeepSynergy, on the DrugComb database. DrugComb collects data from four studies: I the O'Neil dataset,(ii) the Forcina dataset, (iii) the NCI Almanac dataset, and (iv) the CLOUD dataset. DeepSynergy is a feedforward neural network that converts sample input vectors into a single output value known as the synergyscore. DeepSynergy used chemical information derived from drugs as well as genomic data pertaining todisease biology. PRODeepSyn is a deep neural network that predicts synergy scores based on cell lineembeddings and drug features using Batch Normalization. PRODeepSyn constructs the feature vector foreach drug using the molecular fingerprint and descriptors to represent the structural and physicochemicalproperties of drugs, and for cell line features they integrate three types of heterogeneous cell line featurescontaining gene expression data, gene mutation data, and interactions between gene expression products toconstruct cell line embeddings. To compare these methods, the mean square error (MSE), root mean squareerror (RMSE), and Pearson correlation coefficient (PCC) between predictions and ground truth are used asprimary evaluation metrics. The results show that the presented matrix factorization method outperformedthe PRODeepSyn and DeepSynergy methods across the platform.

Keywords:

drug synergy , deep learning , graph convolutional network , feed forward neural network