Zero shot molecular generation via similarity kernels.

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Generative models, particularly diffusion models, are advancing novel chemical discovery.
• Their success relies on the score function's link to physical forces, enabling equivariant neural networks.
• The precise behavior of the learned score function in molecular generation remains underexplored.

Purpose of the Study:

• To analyze the score function's behavior during energy-based diffusion model training for molecular generation.
• To develop a novel, zero-shot molecular generation method based on these insights.
• To enable shape control and fragment-biased generation for enhanced molecular design.

Main Methods:

• Trained an energy-based diffusion model to analyze the score function during molecular generation.
• Developed Similarity-based Molecular Generation (SiMGen), a zero-shot method combining similarity kernels and many-body descriptors.
• Integrated point cloud priors for shape control and utilized the method to guide existing trained models.

Main Results:

• The score function transitions from a restorative potential to a quantum-mechanical force during generation.
• Identified unique intermediate score properties facilitating the construction of large molecules.
• SiMGen successfully generated molecules without retraining, offering shape control and fragment-bias.

Conclusions:

• Understanding the score function's dynamics is crucial for advancing generative molecular modeling.
• SiMGen provides a versatile, training-free approach for novel molecular generation and property control.
• The developed method and associated web tool (ZnDraw) offer practical applications in chemical design.