Deep Learning-based Drug Response Prediction Algorithm about Synergistic and Antagonistic Responses of Cell Lines and Drug Combinations

Cancer, a disease characterized by uncontrolled cell growth, has numerous types and often demands drug combinations to increase efficiency and reduce the side effects of the treatment. This combination needs advanced strategies to predict their effects: synergistic or antagonistic. The traditional approach relies on actual experiments; the huge cost of money and time makes the efficiency relatively low. Therefore, I integrated various deep-learning models to predict the synergy effect. GINConvNet combines a Graph Isomorphism Network (GIN) with a convolutional network, significantly outperforming other models in predicting the synergistic and antagonistic responses of cell lines to drug combinations. GIN identifies structural equivalences in graph-structured data (drugs and their molecular structures) regardless of label differences. This model processes drug and cell representations independently in dual tower architecture, employing the multiplication of drug concentrations to enhance interaction predictions. Its performance is evaluated using metrics such as root mean square error (RMSE), mean squared error (MSE), Pearson's r, and Spearman's ρ. The dataset is from NCI-ALMANAC, encompassing over 5,000 drug combinations across 60 cell lines, and was used for training and validation. GINConvNet demonstrated the lowest RMSE and MSE, which means higher accuracy of prediction, superior capability in capturing the complex interactions between drug combinations and cell lines, and strength in graph classification and similarity comparison tasks. This model might help clinicians tailor cancer treatments, potentially leading to more effective and personalized therapeutic strategies with relatively fewer side effects.