Particularly in oncology, combination drug treatments can offer higher patient survival rates and a lower risk of drug resistance than monotherapy. Either by simultaneously inhibiting key proteins of the same pathway or by inhibiting molecules of separate cancer pathways in parallel, clinicians can maximise the efficacy of chemotherapy.
The scope of potential combination therapeutic options is enormous but lacks thorough and accurate inquiry. Therefore, synergy predictions are essential for compound prioritisation to reduce the search space and cost of large-scale combination drug screenings.
Previous approaches to predicting drug synergies have included the use of chemical structure and genomic information, with an increasing number of teams applying deep learning to predict synergies of studied drugs and cancer cell lines. Notably, a 2019 DREAM challenge study, using supervised machine learning to predict drug synergies in a pharmacogenomic screen, performed well but was limited practically by the availability of combinatorial molecule screening data.
An international team led by Mi Yang of Stanford University published their methodology last week in Nature for prioritising drug combinations for high-throughput screening and stratifying cancer cell line response. Central to their approach is the notion that the chances of drug synergy increases when targeting proteins with either a strong functional similarity (when targeting proteins of a singular pathway) or strong dissimilarity (to target proteins belonging to unrelated pathways). This similarity metric was used to construct a compound prioritisation methodology, using basal gene expression data, drug response (IC50) of monotherapy data and information on drug targets.
The first workflow established by the group to generate synergy predictions used multitask machine learning on basal gene expression levels and response data to the monotherapy treatment. The approach was validated in 29 combinations in 33 cancer cell lines.
The second workflow for synergy stratification, the authors say, has the potential to maximise drug efficacy by matching synergy-inducing drugs to tumour types, and eventually, patients, based on transcriptomic profiling.
In concluding their study, the authors summarised the implications of this work: “exploring the interactions between drug targets and signalling pathways in a tissue-specific manner can provide a novel in-depth view of cellular mechanisms and drug modes of action, which can ultimately rationalize drug combination strategies in cancer. Target functional similarity could be used as a metric for compound prioritization. Synergy by similarity hypothesis could be a rationale for first-line treatment, while synergy by opposite effect could potentially fit patients having acquired resistance.”