De Novo Molecular Design via Shape-Constrained Diffusion Models.

Area of Science:

• Computational chemistry
• Drug discovery
• Artificial intelligence in chemistry

Background:

• De novo molecular design faces challenges in balancing 3D shape similarity with structural novelty.
• Existing generative models struggle to control molecular shape during the design process.

Purpose of the Study:

• To introduce Diff-Shape, a novel diffusion-based framework for controllable, shape-guided de novo molecular generation.
• To enable the design of molecules with high 3D shape fidelity and structural novelty.

Main Methods:

• Diff-Shape couples a pretrained 3D generator with a Graph ControlNet, using 3D reference shapes as input.
• The framework incorporates substructure inpainting for tasks like scaffold hopping, decoration, and linker generation.

Main Results:

• Diff-Shape significantly improves 3D shape fidelity compared to state-of-the-art methods while maintaining low 2D graph similarity.
• Performance remains robust across various novelty thresholds, noise levels, and reference shapes.
• Designed molecules for KRAS G12D and EGFR mutations were synthesized and showed nanomolar potency.

Conclusions:

• Diff-Shape offers a powerful and generalizable solution for shape-constrained molecular generation.
• The framework demonstrates translational relevance through the successful design and synthesis of potent drug candidates.