Predicting Drug Response on Multi-Omics Data Using a Hybrid of Bayesian Ridge Regression with Deep Forest

An accurate drug response prediction for each patient is critical in personalized medicine. However, numerous studies that relied on single-omics datasets continue to have limitations. In addition, the curse of dimensionality considers a challenge to drug response prediction. Deep learning has remarkable prediction effectiveness compared to traditional machine learning, but it requires enormous amounts of training data which is a limitation because the nature of most biological data is small-scale. This paper presents an approach that combines Bayesian Ridge Regression with Deep Forest. BRR relies on the Bayesian approach, in which linear model estimation occurs based on probability distributions rather than point estimates. It was utilized to integrate multi-omics, a feature selection that calculates the coefficient as the feature importance. DF reduces the computational cost and hyper-parameter tuning cost. The Cancer Cell Line Encyclopedia CCLE was used as a dataset to integrate the gene expression, copy number variant, and single nucleotide variant. Root Mean Square Error, Pearson Correlation Coefficient, and the coefficient of determination were used as the evaluation metrics. The obtained findings show that the proposed model outperforms Random Forest and Convolutional Neural Network regarding regression performance; it achieved 0.175 for RMSE, 0.842 for PCC, and 0.708 for R2.