Professor Jaewoo Kang’s Group awarded best performance prize 

at 2019 AI-Based Drug Development DREAM Challenge.


- A splendid achievement of winning first place in DREAM challenges in three consecutive years was garnered; an AI-based drug activity prediction model independently developed by the group was used.

- The group was co-champions with the Chung Hua University-University of Illinois consortium and the University of North Carolina.


▲(From the left) Prof. Julio Saez-Rodriguez (RWTH-Aachen University), Dr. Minji Jeon (Korea University), 

Sunkyu Kim (PhD student, Korea University), Dr. Robert Allaway (Sage Bionetworks), 

Sungjoon Park (PhD student, Korea University), Prof. Jaewoo Kang (Korea University) 

and Dr. Justin Guinney (Sage Bionetworks) in a commemorative photo taken after the awards ceremony.


The DREAM Challenge is an international data science competition in the field of biomedicine held by U.S. IBM and Sage Bionetworks. Starting from the first competition in 2007, it has been held more than 50 times.


Prof. Jaewoo Kang’s group won second place in the AstraZeneca-Sanger Drug Combination Prediction Challenge held in 2016, the first year they participated in a DREAM challenge. In 2017, his group won, the first Korean team to do so, first place in the cancer proteogenomic prediction challenge held by the U.S. National Cancer Institute. The next year, 2018, his group won first place in the Multi-targeting Drug DREAM Challenge held by the Icahn School of Medicine at Mount Sinai, U.S.. Winning first place in the IDG DREAM Challenge this year, Kang’s group achieved the splendid accomplishment of becoming a DREAM champion in three consecutive years.


The DREAM Challenge this year was held by the Finland Institute for Molecular Medicine with data provided by the Illuminating the Druggable Genome (IDG) consortium. The assignment given to the participating teams was to predict the 394 drug activity values between 25 new drug candidates and 207 kinases. A total of 54 teams from all over the world participated in the Challenge, including the Chung Hua University-University of Illinois consortium and the University of North Carolina, which became co-champions with Kang’s group, as well as the U.S. National Institutes of Health (NIH) and the European Molecular Biology Laboratory (EMBL), the representative government-funded research institutes of the U.S. and Europe.


Kang’s group approached the assignment by using a graph neural network, which is a deep-learning model for recognizing the chemical structures of drugs as graphs and mapping them in a mathematical vector space. The graph neural network learns the sub-structures and chemical properties of drugs that are important to protein binding and predicts the drugs’ activity with target proteins based on the learned information.


Pharmaceutical companies have generally employed the high-throughput screening (HTS) method to measure experimentally the activity of thousands of compounds with target proteins. However, that method requires a great amount of time and cost, and its success ratio is low. While it is known that more than 1.2 billion compounds can be synthesized as new drugs, the probability that an optimal lead compound is included in thousands of randomly selected compounds is extremely low. When an AI-based prediction model is used to rapidly select a small number of candidates out of the 1.2 billion compounds to perform experimental verification with the selected candidates only, the time and cost can be greatly reduced and the probability of success can be significantly increased.


Kang said, “In addition to the models that won the first prizes in previous challenges, the Best Performance Prize that we won in this year showed that the key element technologies that are essential to the drug discovery process have already been realized successfully in AI models, which were certified in the international competitions.” He explained his future plan: “We will combine these models in an organized manner to prepare an AI-based drug discovery platform, which will be applied to actual new drug development work performed by cooperating research institutes and pharmaceutical companies.”


The result of the Challenge was announced at the 2019 Research in Computational Molecular Biology/International Society for Computational Biology Regulatory and Systems Genomics Conference (RECOMB/ISCB RSG) with DREAM held at the Memorial Sloan Kettering Cancer Center, New York City, in early November. Kang’s group was invited to the conference as the Best Performance Prize winner and introduced their model in a talk entitled “In-silico Molecular Binding Affinity Prediction with Multi-Task Graph Neural Networks.”