Enlightening cancer drug discovery with challenge-winning AI solutions

A team of researchers (NetPhar) from the University of Helsinki has recently won two international competitions (DREAM challenge) for providing the best AI algorithms to predict protein targets and cellular sensitivities of anti-cancer drugs.

What are DREAM challenges?

DREAM is an international community effort solving key questions in system biology and translational medicine powered by Sage Bionetworks. Recently, together with NIH-CTD2, one of the world’s largest consortiums on cancer target discovery, DREAM has launched two competitions exploring the best algorithms for drug target and drug sensitivity prediction.

“We provided participants with gene expression changes after drug treatment, which are considered to be more informative data for drug target and drug sensitivity prediction. Together with the wisdom of the crowd, we hope to get a better understanding of the mechanisms of action of oncology drugs,” says the main organiser Dr. Eugene Douglass from University of Columbia.

Each challenge has attracted more than 100 researchers all over the world (Figure 1), including those from renowned academic institutes as well as from AI start-up companies. The algorithms developed by Team NetPhar performed the best in the drug sensitivity prediction as well as one sub-challenge of the drug target prediction.

“Our team has the privilege to gather talented researchers with both domain knowledge and strong analytic mindsets, with backgrounds varying from computational biomedicine and pharmacology to statistics and data science,” emphasized by Wenyu Wang, PhD student and leading member of the winning team.

Maailmankartta johon merkitty DREAM-haastekilpailun osallistujien kotimaat

Figure1. Geographic distribution of the teams in the DREAM Challenge (Reference: synapse.org, https://www.synapse.org/#!Map:3406893

How does the DREAM challenge help cancer drug discovery?

It is reported that 97% cancer drugs failed due to low efficacy in clinical trials, reflecting our poor understanding of drugs’ mechanisms of action. Researchers have recently found that drugs often have misidentified targets, impeding the success of clinical trials in cancer drug discovery. 

“These two challenges have offered a valuable platform to study drug mechanisms in tumor cells. Despite its power, the drug perturbed expression data should be used with care. This type of data is intrinsically noisy and thus we chose carefully our methods to avoid bias irrelevant to the research question,” commented by Dr. Alberto Pessia, Statistician of the group.

“The DREAM challenges are quite connected to our several ongoing cancer projects. We utilized data portals (DrugComb, DTC) and tools (TidyComb, SynergyFinder) developed by our group when preparing data for training the AI models. These highly-demanding international competitions helped validate the robustness of these established data infrastructure for AI development. On the other hand, as a member of iCAN Digital Precision Cancer Medicine Flagship, we will further extend these algorithms to the study of ovarian cancer and leukemia. We hope that these applications can provide us more insights on cancer drug discovery and precision medicine,” concluded Assistant professor Jing Tang, principal investigator of the research group.

Funding sources: European Research Council (ERC) starting grant DrugComb (Informatics approaches for the rational selection of personalized cancer drug combinations) [No. 716063]; Academy of Finland Research Fellow grant [No. 317680]; FIMM-EMBL In­ter­na­tional PhD programme and Doctoral program of Biomedicine. The study was carried out under the iCAN Digital Precision Cancer Medicine flagship platform funded by the Academy of Finland.