Even for the same cancer patient, the efficacy of anticancer drugs can vary. This is because the responsiveness of anticancer drugs varies depending on individual genetic mutations. Artificial intelligence (AI) technology that provides advance information on such patient-specific results is being developed and is gaining attention.

Ulsan National Institute of Science and Technology (UNIST) announced on the 19th that Professor Lee Se-min's team in the Department of Biomedical Engineering, together with Professor Jeong Won-gi's team from Korea University and Professor Seo Ji-won's team from Hanyang University, have developed a 'machine learning model for predicting patient-tailored anticancer drug responsiveness based on multi-omics1) data.'

The research team reported that they achieved much better performance than existing anticancer drug response prediction models by utilizing large-scale anticancer drug response data and multi-omics data. This was made possible through network embedding technology and the latest deep learning model.

Cancer is a typical genome-related disease, or 'genome disease'. Disease occurs when mutations continue to accumulate in the genome, the 'blueprint of life' that each person possesses. In cancer tissue, the gene expression pattern is also different from that of normal tissue. These genetic mutations and gene expression profiles show significant differences even among patients with the same cancer, and are known to have a significant correlation with patient-specific anticancer drug responsiveness.

Accordingly, there have been many recent attempts to develop 'patient-tailored anticancer drug response prediction models' based on multi-omics data that encompass cancer patient-specific genetic mutations and gene expression patterns. However, the biological data for learning these models has many types and factors, but the number of samples is insufficient, which limits the accuracy of machine learning models.

To overcome this, the research team applied ‘network embedding technology’ to effectively reflect the correlation between multidimensional data. Extracting embedding that reflects the correlation of the network set formed by nodes and edges is the core of this study.

This is because by using embedding vectors, we can know the representative value of each node, so we can effectively handle high-dimensional data. The research team trained the embedding vector using a deep neural network, an AI technique,5) to derive the efficacy of anticancer drugs tailored to patients.

“We applied various artificial intelligence techniques to supplement the biased reactivity data toward resistance,” said first author Dr. Kang-Geun Lee of Korea University. “The new model’s anticancer drug reactivity prediction performance was shown to be approximately 93% accurate, which is significantly improved from the existing model.”

“We achieved excellent performance through deep neural networks and network embedding technology that effectively extracts interactions between elements existing in high-dimensional multi-omics data,” said co-author Cho Dong-bin, a researcher at Hanyang University.

Dr. Jin-Ho Jang of UNIST, a co-author, also expressed his expectations, saying, “This technology will accelerate personalized treatment by suggesting suitable drug candidates for cancer patients.”

Meanwhile, the results of this study were published in 'Briefings in Bioinformatics', the top academic journal in the field of bioinformatics, and were supported by the 'Next Generation Information Computing Technology Development Project' and the 'University Key Research Institute Support Project' of the National Research Foundation of Korea.