Janus-QUBO: A Duality-Aware Framework for Navigating Chemical Space with a Tunable Quantum-Inspired Landscape

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Quantum computing applications

Background:

• Discovering novel molecules is a key challenge in chemistry.
• Deep generative models struggle with discrete molecular structures.
• Existing models often lack transparency and global exploration capabilities.

Purpose of the Study:

• Introduce Janus, a novel framework for transparent and interpretable molecular design.
• Address the limitations of continuous latent spaces in deep generative models.
• Enable efficient global discovery and rational optimization of molecules.

Main Methods:

• Utilize a Transformer-based autoencoder with a binary bottleneck for discrete latent space representation.
• Reformulate molecular generation and optimization as a Quadratic Unconstrained Binary Optimization (QUBO) problem.
• Leverage classical and quantum annealers for combinatorial optimization.

Main Results:

• Achieve superior multiobjective performance in molecular design compared to state-of-the-art methods.
• Preserve scaffold integrity during optimization, avoiding structural fragmentation.
• Demonstrate the efficacy of quantum-assisted molecular design for drug-like properties.

Conclusions:

• Janus provides a robust framework for rational, quantum-assisted molecular design.
• Unifies combinatorial exploration with deep model interpretability.
• Enables efficient and transparent discovery of novel chemical scaffolds.