Abstract
High-throughput drug sensitivity screening has been utilized for facilitating the discovery of drug combinations in cancer. Many existing studies adopted a dose-response matrix design, aiming for the characterization of drug combination sensitivity and synergy. However, there is lack of consensus on the definition of sensitivity and synergy, leading to the use of different mathematical models that do not necessarily agree with each other. We proposed a cross design to enable a more cost-effective testing of sensitivity and synergy for a drug pair. We developed a drug combination sensitivity score (CSS) to summarize the drug combination dose-response curves. Using a high-throughput drug combination dataset, we showed that the CSS is highly reproducible among the replicates. With machine learning approaches such as Elastic Net, Random Forests and Support Vector Machines, the CSS can also be predicted with high accuracy. Furthermore, we defined a synergy score based on the difference between the drug combination and the single drug dose-response curves. We showed that the CSS-based synergy score is able to detect true synergistic and antagonistic drug combinations. The cross drug combination design coupled with the CSS scoring facilitated the evaluation of drug combination sensitivity and synergy using the same scale, with minimal experimental material that is required. Our approach could be utilized as an efficient pipeline for improving the discovery rate in high-throughput drug combination screening. The R scripts for calculating and predicting CSS are available at https://github.com/amalyutina/CSS.
Author summary Being a complex disease, cancer is one of the main death causes worldwide. Although new treatment strategies have been achieved with cancers, they still have limited efficacy. Even when there is an initial treatment response, cancer cells can develop drug resistance thus cause disease recurrence. To achieve more effective and safe therapies to treat cancer, patients critically need multi-targeted drug combinations that will kill cancer cells at reduced dosages and thereby avoid side effects that are often associated with the standard treatment. However, the increasing number of possible drug combinations makes a pure experimental approach unfeasible, even with automated drug screening instruments. Therefore, we have proposed a new experimental set up to get the drug combination sensitivity data cost-efficiently and developed a score to quantify the efficiency of the drug combination, called drug combination sensitivity score (CSS). Using public datasets, we have shown that the CSS robustness and its highly predictive nature with an accuracy comparable to the experimental replicates. We have also defined a CSS-based synergy score as a metric of drug interaction and justified its relevance. Thus, we expect the proposed computational techniques to be easily applicable and beneficial in the field of drug combination discovery.
