Kuenzi et al discuss how machine learning models can be applied in understanding how or why a particular cancer responds to therapy. Previous models have been ‘black boxes’ and can predict outcomes without explaining the biological mechanisms underlying it. Thus, the researchers combined a visible neural network with conventional artificial neural networks to create a new model DrugCell that can predict and explain the response of a human cancer cell to new therapeutic drugs. This model was trained with the Cancer Therapeutics Response Portal (CTRP) v2 and the Genomics of Drug Sensitivity in Cancer (GDSC) database, representing various cell line-drug pairs. The DrugCell model outperformed other models in predicting a response using biological features only and had a total accuracy of rho = 0.80. DrugCell also learned information from somatic mutations beyond the tissue of origin unlike previous models that focused solely on the area provided. The model was also used to rank subsystems regulating sensitivity to 25 drugs. Researchers found that the top drug combinations nominated by DrugCell were more effective at cell killing, such as etoposide. Many prior machine learning models have not been used for modeling or validated experimentally for identifying drug combinations. Furthermore, DrugCell was also able to accurately identify subsystems that corresponded to successful drug combinations in patient-derived xenograft (PDX) models. Finally, DrugCell was able to predict the response of clinical patients who had estrogen receptor-positive metastatic breast cancer. DrugCell will serve as a foundation for future predictive modeling to represent the biological mechanisms associated with a drug response.