APBN New Site

APBN Developing Site

Using AI to Navigate Through the “Galaxy” of Potential Drug Candidates

This novel framework provides more accurate and robust drug-target interaction predictions than conventional methods, potentially bringing down the cost and time spent in drug development.

In drug development, being able to recognise interactions between chemical compounds and protein targets plays an important role. In 2017, chemists estimate that there are as many compounds with drug-like properties as there are many atoms in the Solar System. This number makes up to be more than 1060. Predicting which of these drug candidates can interact with which target is crucial in virtual screening, drug repurposing, and identifying potential drug effects.

Drug development typically takes about 14 years and can cost up to $1.5 billion. It is apparent that traditional biological experiments for drug-target interaction (DTI) detection are insufficient to navigate through this “galaxy” of potential drug-like compounds.

Professor Hou Tingjun is an expert in computer-aided drug design at the Zhejiang University College of Pharmaceutical Sciences. For decades, he has been committed to developing drugs with computer technology. “The biggest challenge lies in the interactions between unknown targets and drug molecules. How can we discover them more efficiently? This involves a new breakthrough in method,” said Prof. Hou.

With artificial intelligence (AI) presenting new possibilities in recent years, scientists now look to leverage AI in drug discovery. “With artificial intelligence, we may be able to reach the more upstream stage in drug discovery, thus improving the efficiency and success rate of the drug development.”

In addition to AI, multi-omics data like genomics, proteomics, and pharmacology, have also seen considerable growth. In each field, there has been an influx of information about diseases, drugs, side effects, etc. that has been stored in specialised databanks. However, their value for drug discovery remains obscure.

Professor He Shibo, who is a scholar specialising in big data and network sciences at the Zhejiang University College of Control Science and Engineering said, “This domain is particularly suited for inter-disciplinary research. This considerable body of biological information can be abstracted into a multi-layered and heterogeneous network system,” said Prof. He.

In a study published in Nature Communications, a team led by Hou Tingjun and He Shibo from Zhejiang University and Cao Dongsheng from Central South University combined knowledge graph and recommendation system techniques to develop a unified framework called knowledge graph embedding and neural factorisation machine (KGE_NFM) for DTI prediction.

This framework first learns a low-dimensional representation for various entities in the knowledge graph and then incorporates the multimodal information through neural factorisation machine. Their novel framework is then evaluated under three real-world scenarios: the warm start, the cold start for drugs, and the cold start for proteins. Based on four benchmark datasets, the AI algorithms were on par or slightly inferior to conventional methods in the first two scenarios. However, in the third scenario, the KGE_NFM achieved accurate and robust predictions, outdistancing its rivals by 30 per cent.

“This demonstrates the remarkable ability and superiority of AI in predicting the unknown protein targets. Discovering the ‘unknown drug-target interactions’ from the ‘unknown protein targets’ is an undeniably important undertaking in the future of drug discovery,” observed Prof. Hou. “The use of KGE can not only expand the dimension of information but also promote the interpretability and credibility of algorithmic systems.” [APBN]

Source: Ye et al. (2021). A unified drug–target interaction prediction framework based on knowledge graph and recommendation system. Nature communications, 12(1), 1-12.