De Novo Molecular Design via Shape-Constrained Diffusion Models

Generative models that preserve 3D shape similarity while enabling structural novelty remain a central challenge in de novo molecular design. We present Diff-Shape, a diffusion-based framework that couples a pretrained unconditional 3D generator with a Graph ControlNet constraint that ingests a 3D reference shape. By integrating substructure inpainting, Diff-Shape enables controllable, shape-guided design across tasks such as scaffold hopping, scaffold decoration, and linker generation. On comprehensive benchmarks, Diff-Shape achieves substantially higher rates of 3D shape fidelity than state-of-the-art methods, while maintaining low 2D graph similarity. These improvements are robust across novelty thresholds, noise levels, and diverse references. Finally, case studies on Kirsten rat sarcoma virus (KRAS) G12D and epidermal growth factor receptor L858R/T790M/C797S demonstrate that several designed candidates were synthesized and exhibit nanomolar biochemical potency, underscoring the method's translational relevance. Together, these results position Diff-Shape as a powerful and generalizable approach for shape-constrained molecular generation.