Combination therapies are fast becoming the cornerstone of a next wave cancer treatments, but why are they so important? We know that precisely targeting drivers of cancer can improve patient responses – a good example is the use of EGFR inhibitors in EGFR mutated tumours – however, cancer is a complex and evolving disease which can limit the effectiveness of monotherapy. Using combinations of drugs to inhibit different cancer pathways has the potential to target distinct tumour cell populations and shut down bypass mechanisms that enable cancer cells to escape death. The concept sounds simple but the vast number of possible combinations and complexity of the cancer genome means that identifying the most promising combinations is no small task.
Today we can readily screen large numbers of drug-drug combinations across cell panels, however we need to find more efficient ways to make sense of the resulting data, to predict optimal combinations and understand which patients are most likely to benefit. This is where data science takes centre stage.
Pooling the expertise of 160 research teams, the AstraZeneca-Sanger Combination Drug DREAM Challenge has advanced computational approaches for predicting cancer drug combinations and identifying biomarkers of synergy.
As part of the AstraZeneca-Sanger Drug Combination Prediction DREAM Challenge, we shared drug target information for over 11,000 experimentally tested drug combinations across 85 molecularly characterised cancer cell lines. One of the largest and most diverse data-sets shared by any company or organisation to date, this enabled participants to train and test computational models to predict drug synergy. Results of the Challenge are reported in a recent publication in Nature Communications, and the data is now available for further research through Synapse.
The power of crowd sourcing was in full effect for this project – we saw >100 diverse methods we might never have tried in-house. There is so much to be learned by understanding which methods perform better in different cell and drug contexts. The most successful models paired machine learning with innovative approaches to filter molecular features associated with cancer or the drug target, and grouped drugs by target or signalling pathway to enable algorithms to share learning between them. Among the winning models, combination synergy between drugs was well predicted for over 60% of combinations. A particularly interesting observation was the greater universal predictability when two drugs target parallel nodes downstream of a common regulator.
The power of crowd sourcing was in full effect for this project – we saw >100 diverse methods we might never have tried in-house. There is so much to be learned by understanding which methods perform better in different cell and drug contexts
The Challenge has also provided insight into putative biomarkers of response with potential translatability to the clinic. Synergy was consistently seen between AKT inhibitors and pan-PI3K inhibitors in cells with activating mutations in PIK3CA or PTEN deletions, and with EGFR inhibitors in cells habouring activating EGFR or ERBB2 mutations. Of course, there are many aspects of cancer biology not captured by cell models – for example immune mediated mechanisms of resistance, but by making these data and algorithms publicly available we hope this provides a platform to accelerate further research in the field.
Veeva ID: Z4-17750
Date of next review 19/6/2021