Korean AI-driven drug discovery speeds up

with data sharing and convergence

The Ministry of Science and ICT·MOHW will allocate KRW 34.8 billion in a federate learning-based project to accelerate new drug development

A multi-institution collaborative body for federate learning-based drug discovery is expected to be formed within the first half of the year

'High expandability of platform,' 'Securing ownership of data utilizing technology,' and 'Reducing new drug development cost' expected

Ministry of Health and Welfare (MOHW) and the Ministry of Science and ICT will allocate KRW 34.8 billion in AI-driven new drug discovery.

AI-driven new drug discovery through data collaboration among the biopharmaceutical industry, medical institutions, research institutions, universities, and public institutions has accelerated and gained significant attention.

 

Previously, individual biopharmaceutical companies in South Korea conducted AI-driven new drug discovery through their in-house protocol, and the process was kept confidential.

 

Starting with the recently organized business group called K-MELLODDY (Machine Learning Orchestration for Drug Discovery), a new system is expected to be implemented within the first half of the year to build upon a system to utilize sources and information related to various clinical and substance owned by 20 institutions.

 

An individual company’s AI-driven new drug discovery could predict a new drug candidate’s ADMET (Absorption, Distribution, Metabolism, Excretion, Toxicity) value via in-vitro studies.

 

However, these test results were insufficient to ensure the drug would pass in-vivo (pre-clinical) or clinical trials.

 

Furthermore, ADMET prediction for clinical trial outcomes requires learning from clinical trial (in-human) data, but not enough clinical or in vivo data makes implementation extremely difficult.

 

Furthermore, beyond predicting ADMET for a single drug molecule, solutions were needed that could expand to include interactions with specific targets, drug-drug interactions, responses based on user types, and various toxicity predictions.

 

Therefore, optimized models based on federated learning have been considered for AI-driven drug development.

 

K-MELLODDY aims to utilize new drug development data without leaking internal data to the outside world.

 

In other words, federated learning and blockchain, which securely maintain companies' intellectual property data while leveraging AI, are evaluated as the core technologies of this project.

 

The project’s characteristics and advantage are differentiated by 'Obtaining reliability of AI-driven new drug discovery,' 'High expandability of platform,' 'Securing ownership of data utilizing technology,' and 'Reducing new drug development cost.' Upon completion of the platform, utilizing AI predictions for the candidate product's metabolism and toxicity experiment results could reduce the number of experiments by half, leading to more than 50% cost savings.

 

“K-MELLODDY project aims to build a federated learning-based AI drug development platform by utilizing the power of data held by Korea's academia, industry, and research institutes related to drug development.

 

It aims to derive successful AI case studies in the drug discovery stage,” Hong Sungeun, Senior Researcher at the Convergence AI Institute for Drug Discovery (CAIID), stated.

 

He added, “A budget of KRW 34.8 billion will be allocated over the next five years to complete the development of a Korean federated learning-based AI new drug discovery model.” Meanwhile, K-MELLODDY is described as a Korean AI new drug platform project that has been improved to suit the needs of the Korean situation while benchmarking the EU-MELLODDY project.

 

EU-MELLODDY is an initiative for data collaboration among 10 pharmaceutical companies, first implemented in the EU.

 

The initial business goal was to prove that data-based cooperation is possible between competing pharmaceutical companies and spread technological safety.

 

Based on the results of EU-MELLODDY, which was conducted in 2019 for three years with KRW 25.6 billion put into, it was observed that the machine learning-based cooperative ADME/Tox prediction model was learned without leaking pharmaceutical companies’ confidential information.

 

In addition, it outperformed the single model formulated by individual institutions without data leakage.