High-level screening technologies have generated a vast volume of drug-sensitivity data for a panel of cancer cell lines and hundreds of chemicals. By identifying molecular genetic factors of drug sensitivity and developing novel anticancer medicines, computational approaches to analysing these data can assist in the development of anticancer therapies. Conventional deep learning models lack the ability to select the best imputation strategy or to handle missing values. This may compromise the dataset’s originality and introduce data sensitivity issues. To address these issues, we introduce the Explainable Drug Graph Attention Transformer (EDrGAT), which proposes task-specific integration of feature-level graph attention and engineered temporal pharmacodynamics features to predict drug sensitivity. To make the proposed model more data-sensitive, lag, rolling mean, and Exponential mean are computed. The cat boost regressor model performs best for imputation and for data with genomic features such as cell line, drug name, and drug concentration (IC50), and is trained using EDrGAT. We demonstrate the model’s performance using metrics such as RMSE, R2, and training and validation loss/accuracy. The proposed model achieves an R2 of 93% by outperforming previous state-of-the-art models.
Anushya et al. (Tue,) studied this question.