P05-01

Multi-Task Deep Learning using Graph Convolutional Networks for Predicting the Unbound Fraction in Human, Mouse, and Rat Plasma

Harutoshi KATO *, Yuki DOI, Akira SASAKI

DMPK Research Laboratories, Mitsubishi Tanabe Pharma Corporation
( * E-mail: kato.harutoshi@mb.mt-pharma.co.jp )

The unbound fraction in plasma (fu), calculated from plasma protein binding (PPB), is a crucial pharmacokinetic parameter that have significant impact on pharmacological and toxicological effects of a drug. The computational prediction of fu is considered to be effective in drug discovery cycle as it can reduce the evaluation period and support drug design. In this study, we aimed to develop a prediction model for the fu in human, mouse, and rat using multi–task deep learning with graph convolution networks (GCN).
The fu data of approximately 5000 compounds each in human, mouse, and rat were used as the proprietary internal dataset for model development and validation． In addition, human fu data of approximately 2000 compounds published were added as an external dataset to expand the chemical space of the training dataset. As regression model, several deep learning models were constructed by single–task and multi–task learning using GCN, and the prediction performance of the models were evaluated using the test datasets.
As a result, we confirmed that the multi–task deep learning model for predicting fu using the internal dataset (human, mouse, rat) and external dataset (human) as training dataset achieved the best prediction performance based on the coefficient of determination (R²) on the test dataset, with values of 0.65, 0.78, and 0.89 for human, mouse, and rat, respectively.