NeuMF: Predicting Anti-cancer Drug Response Through a Neural Matrix Factorization Model

Abstract

Background: Anti-cancer drug response is urgently required for individualized therapy. Measurements with wet experiments are costly and time-consuming. Artificial intelligence-based models are currently available for predicting drug response but still have challenges in prediction accuracy Objective: Construct a model to predict drug response values for unknown cell lines and analyze drug potential association properties in sparse data. Methods: Propose a Neural Matrix Factorization (NeuMF) framework to help predict the unknown responses of cell lines to drugs. The model uses a deep neural network to figure out drug and cell lines' latent variables. In NeuMF, the inputs and the parameters of the multi-layer neural network are simultaneously optimized by gradient descent to minimize the reconstruction errors between the predicted and natural values of the observed entries. Then the unknown entries can be readily recovered by propagating the latent variables to the output layer. Results: Experiments on the Cancer Cell Line Encyclopedia (CCLE) dataset and Genomics of Drug Sensitivity in Cancer (GDSC) dataset compare NeuMF with the other three state-of-the-art methods. NeuMF reduces constructing drug or cell line similarity and mines the response matrix itself for correlations in the network, avoiding the inclusion of redundant noise. NeuMF obtained drug averaged PCC_sr of 0.83 and 0.84 on both datasets. It demonstrates that NeuMF substantially improves the prediction. Some essential parameters in NeuMF, such as the strategy of global effect removal and the scales of the input layer, are also discussed. Finally, case studies have shown that NeuMF can better learn the latent characteristics of drugs, e.g., Irinotecan and Topotecan are found to act on the same pathway TOP1. The conclusions are in line with some existing biological findings. Results: Experiments on the Cancer Cell Line Encyclopedia (CCLE) dataset and Genomics of Drug Sensitivity in Cancer (GDSC) dataset compare NeuMF with the other three state-of-the-art methods. NeuMF reduces constructing drug or cell line similarity and mines the response matrix itself for correlations in the network, avoiding the inclusion of redundant noise. NeuMF obtained drug averaged PCC_sr of 0.83 and 0.84 on both datasets. It demonstrates that NeuMF substantially improves the prediction. Some essential parameters in NeuMF, such as the global effect removal strategy and the input layer scales, are also discussed. Finally, case studies have shown that NeuMF can better learn the latent characteristics of drugs, e.g., Irinotecan and Topotecan are found to act on the same pathway TOP1. The conclusions are in line with some existing biological findings. Conclusion: NeuMF achieves better prediction accuracy than existing models, and its output is biologically interpretable. NeuMF also helps analyze the correlations between drugs.