A data-driven generative strategy to avoid reward hacking in multi-objective molecular design.

Area of Science:

• Computational chemistry
• Data-driven molecular design
• Generative artificial intelligence

Background:

• Data-driven generative models are revolutionizing molecular design in drug discovery and materials science.
• Reward hacking, caused by prediction model extrapolation failure, poses a significant challenge.
• Existing applicability domain (AD) methods struggle with multi-objective optimization due to complex reliability assessments.

Purpose of the Study:

• To propose a reliable molecular design framework for generative models.
• To prevent reward hacking during multi-objective optimization.
• To enable automatic adjustment of reliability levels based on user-defined property prioritization.

Main Methods:

• Development of a novel framework integrating generative models with reliability estimation.
• Implementation of a strategy to manage overlapping applicability domains in multi-objective optimization.
• Demonstration using anticancer drug candidate design as a case study.

Main Results:

• Successful design of molecules with high predicted properties and reliability.
• Identification of potential drug candidates, including an approved drug, through multi-objective optimization.
• Validation of the framework's ability to prevent reward hacking in complex molecular design tasks.

Conclusions:

• The proposed framework effectively addresses reward hacking in generative molecular design.
• Reliable multi-objective optimization is achievable even with complex property predictions.
• The framework offers adaptable reliability adjustments for diverse applications, exemplified by drug discovery.