Researchers at Aalto University, the University of Helsinki, and the University of Turku in Finland report they have developed a machine learning model that can predict how combinations of different cancer drugs kill various types of cancer cells. The new AI model was trained with a large set of data obtained from previous studies, which had investigated the association between drugs and cancer cells.
Their study, “Leveraging multi-way interactions for systematic prediction of pre-clinical drug combination effects,” was published in Nature Communications.
Timely detection and treatment of precancerous lesions or early-stage cancer is critical for preventing morbidity and mortality. Several ML applications may improve screening and diagnosis. The researchers presented comboFM, a machine learning framework, for predicting the responses of drug combinations in preclinical studies, such as those based on cell lines or patient-derived cells.
“The model learned by the machine is actually a polynomial function familiar from school mathematics, but a very complex one,” explained Juho Rousu, Ph.D., professor, computer science, Aalto University.
“comboFM models the cell context-specific drug interactions through higher-order tensors, and efficiently learns latent factors of the tensor using powerful factorization machines. The approach enables comboFM to leverage information from previous experiments performed on similar drugs and cells when predicting responses of new combinations in so far untested cells; thereby, it achieves highly accurate predictions despite sparsely populated data tensors,” the researchers wrote.
The researchers demonstrated high predictive performance of comboFM in various scenarios using data from cancer cell line pharmacogenomic screens. Their results demonstrate that comboFM provides an effective means for systematic pre-screening of drug combinations to support precision oncology applications.
To evaluate the comboFM model, the researchers used the anticancer drug combination response data from the Nation Cancer Institute-ALMANAC study. They considered a subset of the data consisting of 50 unique FDA-approved drugs in 617 distinct combinations screened in various concentration pairs across all the 60 cell lines originating from nine tissue types. In the data subset, a total of 333,180 drug combination response measurements and 222,120 monotherapy response measurements of single drugs are available in the form of percentage growth of the cell lines.
“The model gives very accurate results. For example, the values of the so-called correlation coefficient were more than 0.9 in our experiments, which points to excellent reliability,” said Rousu. In experimental measurements, a correlation coefficient of 0.8–0.9 is considered reliable.
“This will help cancer researchers to prioritize which drug combinations to choose from thousands of options for further research,” noted Tero Aittokallio, a researcher from the Institute for Molecular Medicine Finland (FIMM) at the University of Helsinki.
The researchers also reported how their machine learning approach could be used for non-cancerous diseases. The model would have to be re-taught with data related to that disease, but can aid in providing insight on how a combination of antibiotics affect infections or how effectively a combination of drugs can kill cells that have been infected.
“Given the high cost of the experimental screening of drug combinations, comboFM has the potential to provide time- and cost-effective means toward prioritizing the most promising drug combinations for further preclinical or clinical studies. The accurate and robust drug combination response predictions provide a promising approach to streamline the development and expansion of combination therapeutics in personalized cancer treatment,” concluded the researchers.