Machine learning decreases experimental costs of drug combination screening for translational applications
FIMM researchers have developed an efficient machine learning model that requires only a limited set of dose-response measurements for accurate prediction of drug combination synergy in a given patient sample. The minimal-input web-implementation, named DECREASE, supports cost-effective combinatorial screening in precision medicine projects with decreased experimental costs, translational time and number of patient-derived cells required.
Combination therapies have become a standard treatment of several complex diseases. High-throughput screening (HTS) makes it possible to profile phenotypic effects of thousands of drug combinations in patient-derived cells and other pre-clinical model systems. However, due to the massive number of potential drug and dose combinations, large-scale multi-dose combinatorial screening requires extensive resources and instrumentation, beyond the capability of most academic laboratories. Testing of hundreds of combinations is also impossible in limited cell numbers from patient samples.
To make HTS combinatorial screening feasible in translational projects, FIMM researchers implemented a machine learning-based model for systematic prediction of drug combination effects with a minimal set of experimentation. Using a compendium of 23,595 drug combination matrices tested in various cancer cell lines, as well as in malaria and Ebola infection models, they demonstrated how cost-effective experimental designs with machine learning capture almost the same degree of information for synergy and antagonism detection as the fully-measured dose-response matrices.
“We hope the method will become useful in various studies aiming at identification of anti-cancer, bacterial, fungal or antiviral drug combination synergies”, says doctoral student Aleksandr Ianevski, the lead author of the study who implemented the tool.
The only input needed for the DECREASE model is a sub-matrix of the drug combination dose-response measurements, without the need to have any structural or target information of compounds, nor molecular profiles of the cell samples that are often unavailable or take longer time to profile. This enables users to run the DECREASE model for each patient sample and drug combination separately, trained on the fly based on user-provided incomplete dose-response matrix, therefore enabling real-time personalized medicine applications.
“DECREASE reduces experimental costs by 62% when using an 8x8 dose-matrix assay. This a nice example of clever utilization of machine learning to decrease the cost of combinatorial HTS”, says postdoctoral researcher Anil K Giri, another lead author of the study.
“Instead of using any fixed drug concentration levels, such as IC50, we showed that measuring the dose-response matrix diagonal provides most accurate and robust option for synergy screening”, says postdoctoral fellow Prson Gautam, an expert in combinatorial drug screening.
The method development was carried out with the FIMM High Throughput Biomedicine (HTB) unit to support future combinatorial screens. To promote its wide application, the method is implemented as an interactive web-tool with minimal user requirements. The source codes are also freely available to support its further extensions in other applications. The method makes use of a novel composite nonnegative matrix factorization (cNMF)-based machine learning model to predict the missing elements of the multi-dose response matrix.
Original publication: Aleksandr Ianevski, Anil K Giri, Prson Gautam, Alexander Kononov, Swapnil Potdar, Jani Saarela, Krister Wennerberg and Tero Aittokallio. Prediction of drug combination dose-response landscapes with a minimal set of experiments. Nature Machine Intelligence. https://www.nature.com/articles/s42256-019-0122-4