Molecular design beyond training data with novel extended objective functionals of generative AI models driven by...

Area of Science:

• Computational chemistry
• Quantum computing applications
• Drug discovery and development

Background:

• Deep generative models are emerging for small molecule design, but often generate compounds with low drug-likeness.
• Improving the quality and drug-likeness of generated molecules remains a key challenge in computational drug discovery.

Purpose of the Study:

• To develop a novel quantum annealing-based generative model approach to optimize deep generative models for small molecule design.
• To enhance the frequency of drug-like compounds generated by molecular generative models.

Main Methods:

• Integration of a D-Wave annealing quantum computer with deep generative models.
• Development of a neural hash function (NHF) for simultaneous regularization and binarization, enabling classical-quantum signal transformation.
• Utilizing quantum annealing as a stochastic generator within the generative model framework.

Main Results:

• Quantum-annealing generative models produced compounds with higher validity and drug-likeness compared to fully-classical models.
• Generated molecules exceeded the training data in drug-likeness features without explicit constraints.
• The NHF effectively facilitated error evaluation by transforming continuous and discrete signals.

Conclusions:

• Quantum annealing offers a promising approach for optimizing deep generative models in drug design.
• This method can improve feature space sampling and characteristic feature extraction for accelerated drug discovery.
• The developed quantum-classical neural network architecture shows potential for advancing molecular generative capabilities.