Date: 3rd December 2020
Combinatorial therapies are standard practice for treating cancer and can involve a mixture of surgery, radiotherapy and medication. In addition, combinatorial drug therapies often improve the effectiveness of treatment and are essential for killing cancer cells that have metastasised. However, a limitation to drug combinatorial therapy has been the slow and expensive process of screening drug combinations, which has led to the full benefits of this approach not being realised. Now, researchers have developed a machine learning model that accurately predicts how combinations of different cancer drugs kill various types of cancer cells.
Treating cancer with a combination of drugs has several benefits such as overcoming monotherapy resistance and improving therapeutic efficacy due to multi-targeting effects. Furthermore, used at the correct dosage – which is usually lower that treating with an individual drug – harmful side-effects such as toxicity can be minimised.
However, finding ideal drug combinations is difficult, and whilst recent advances in high-throughput screening methods has made this feasible, the sheer number of combinations layered onto variations in dosage of each individual drug, means that systematic evaluation of these combinations quickly becomes impractical.
Now, researchers from Aalto University, University of Helsinki and the University of Turku, Finland, led by Juho Rousu and Tero Aittokallio, have developed comboFM, a machine learning framework for predicting the responses of drug combinations in pre-clinical studies. They demonstrate the high predictive performance of comboFM in various prediction scenarios using data from cancer cell line pharmacogenomic screens and experimentally validate predicted novel drug synergies.
comboFM, is a novel machine learning framework for the systematic modelling of drug-dose combination effects in a cell context-specific manner. It was trained on large datasets from previous cancer cell line pharmacogenomic screens, which had investigated the association between cancer cells and drugs. Unlike existing machine learning models, comboFM can explore and predict the detailed landscape of drug combination responses across various doses and in particular cancer cell types.
To start, the team evaluated comboFM using the anticancer drug combination response data from a previous study. They included a subset of 50 unique FDA-approved drugs in 617 distinct combinations, screened in various concentration pairs across 60 cell lines, originating from 9 tissue types. Three prediction scenarios were considered: filling in missing entries in partially tested dose–response matrices, predicting a complete dose–response matrix in a new cell line, and making predictions for a completely new drug combination not tested so far in any cell line.
In all of the three prediction scenarios, comboFM showed the highest average prediction accuracy in each of the tissue types, and also the smallest variance across the tissue types when compared with a widely-used reference machine learning model (random forest). Furthermore, comboFM was shown to provide high accuracies across various types of combination therapies such as chemotherapies, targeted therapies, and other therapies.
As the interest in drug combination experiments often lies in discovering the most synergistic drug combinations, the team also quantified drug combination synergies, re-training comboFM with a larger dataset.
This resulted in a total of 10,320 predicted complete dose–response matrices. A subset of 16 of the top synergistic drug combinations specific for 4 cell lines were then experimentally validated. The team chose those with highly synergistic effects only in a subset of all the cell lines and tissue types, which would limit toxicity and generate a more targeted approach. All the drug combinations predicted by comboFM were experimentally validated as synergistic, and one in particular between anaplastic lymphoma kinase (ALK) inhibitor crizotinib and proteasome inhibitor bortezomib in a lymphoma cell line exhibited very high levels of synergy. The model also identified another unique drug combination for this cell line, supporting the further investigation of these drug combinations for treating lymphoma.
Conclusions and future applications
The team here have presented a novel machine learning framework, comboFM, for large-scale systematic prediction of drug combination effects in human cancer cell lines. It is hoped that the power of artificial intelligence (AI) can accelerate the experimental work by providing guidance toward identifying the most promising drug combinations for further pre-clinical or clinical studies.
The model identified many previously untested synergistic drug combinations, of which many in hindsight had biologically plausible combination effects – highlighting the potential power of comboFM. By combining these types of drugs, more efficient and less toxic cancer drug regimes could overcome monotherapy resistance and transform patient outcomes.
This latest use of AI is just one of the ways scientists are exploiting AI to deliver solutions for cancer. Just a few days ago we reported scientists have combined AI with cell engineering, to design ‘living’ medicines, or smart cell therapies as they are being dubbed, that exploit discriminatory combinatorial antigen targets that can precisely target tumours. We also announce yesterday, that DeepMind’s AI model can resolve protein structures, which represents an extraordinary advancement on how proteins fold and could revolutionise medical research including cancer treatments.
Whilst, comboFM is set to accelerate drug screening and repurposing of drugs and drug combination for cancer, its applications are far more wide-reaching. The same machine learning approach can be used for other diseases, with the model re-taught with the relevant datasets – perhaps tackling antibiotic resistance, or revealing dug combinations to treat SARS-Cov-2 for example. How this AI-driven solution evolves and supports precision healthcare applications will be an interesting and important journey.
For more information please see the press release from Aalto University
Julkunen, H., A. Cichonska, P. Gautam, S. Szedmak, J. Douat, T. Pahikkala, T. Aittokallio and J. Rousu (2020). “Leveraging multi-way interactions for systematic prediction of pre-clinical drug combination effects.” Nature Communications 11(1): 6136.