Enhancing Synergistic Drug Combination Model Through Dimension Reduction in Cancer Cell Lines

This paper addresses the challenge of dealing with varying dimensions in drug combination models designed to predict synergistic effects across cancer cell lines. Cancer cell lines features vector may exceed 50,000. Such high-dimensional data is unsuitable for learning modeling approaches. To address this challenge, a range of dimension reduction techniques are employed, including feature selection methods, autoencoders, and the LINCS project. The paper proposes a comprehensive comparative analysis on O’Neil drug combination dataset of various dimension reduction techniques to identify the top 1,000 most influential cancer cell line features and compare the optimal model derived from these techniques with recent works in the field to assess its performance and advancements over existing approaches. Initially, a variance-based selection approach is used to reduce the gene set, followed by the application of entropy and Gini-based methods to further refine gene expression. Additionally, Gaussian feature selection is applied to pinpoint the 1,000 most significant genes. Autoencoders and variational autoencoders are also applied to reduce the dimensionality of gene expression data. Furthermore, the paper leverages the LINCS project to select the 1,000 genes according to its landmark gene set. These selection techniques are implemented across two distinct models, each focused on different drug features, including Modered drug features and graph convolution network features derived from drug molecular data. The results demonstrate that the variational autoencoder is the most appropriate feature selection technique for both models especially for graph features. Finally, the optimal model achieves a mean square error (MSE) of 227.73, mean absolute error (MAE) of 9.44, coefficient of determination ( \({R}^{2}\) ) of 0.55, Pearson correlation coefficient ( \({CC}_{p}\) ) of 0.75, and Spearman correlation coefficient ( \({CC}_{s}\) ) of 0.74. In terms of the synergy class label, the optimal model obtains a 0.96 area under the ROC curve (ROC-AUC), a 0.86 area under the precision-recall curve (PR-AUC), Precision of 0.83, Kappa of 0.74, and Accuracy of 0.92.