O01_05
Development of a Molecular Generative model via the Decoupled Setting on Multi-objective Bayesian Optimization
Takamasa SUZUKI *1, Nobuaki YASUO2, Masakazu SEKIJIMA1
1School of Computing, Tokyo Institute of Technology
2Tokyo Tech Academy for Convergence of Materials and Informatics (TAC-MI), Tokyo Institute of Technology
( * E-mail: suzuki.t.dq@m.titech.ac.jp )
The cost of developing a new drug has risen dramatically year after year. To address that problem, studies on molecular generative models that can generate compounds with novel structures have been raised in recent years. Thousands of computer-aided drug discovery (CADD) methods have been practiced in the pharmaceutical industry. Existing generative models achieve multi-objective optimization with the weighted sum or product. In other methods, generative models consider the Pareto optimality to avoid the problem of linear aggregation. However, in calculating the Pareto frontier, all candidates are required to have all objective values.
In this study, we have developed a new deep learning-based multi-objective de novo molecular generative model, which could simultaneously optimize molecules with the "decoupled setting" in entropy-based multi-objective Bayesian optimization (MBO) frameworks. MBO has several approaches, such as entropy-based, hypervolume-based, and scaler-based methods. Entropy-based methods maximize the entropy of the distribution of candidates as an acquisition function and enable to evaluation of objective function separately. The "decoupled setting" realized avoiding costs of repeatedly calculating high computationally complicated objective functions. With the setting, it is not necessary for all candidate points to be calculated in all objective functions. The setting decreases computational costs and makes molecule searches efficient. The proposed method is one of the multi-objective models and optimizes multiple objective functions, binding affinity, and drug-likeness. This method guides the discovery of high-efficiency drugs.