CoarsenConf: Equivariant\nCoarsening with Aggregated\nAttention for Molecular Conformer Generation

Molecular conformer\ngeneration (MCG) is an important task in cheminformatics\nand drug discovery. The ability to efficiently generate low-energy\n3D structures can avoid expensive quantum mechanical simulations,\nleading to accelerated virtual screenings and enhanced structural\nexploration. Several generative models have been developed for MCG,\nbut many struggle to consistently produce high-quality conformers\nfor meaningful downstream applications. To address these issues, we\nintroduce CoarsenConf, which coarse-grains molecular graphs based\non torsional angles and integrates them into an SE(3)-equivariant\nhierarchical variational autoencoder. Through equivariant coarse-graining,\nwe aggregate the fine-grained atomic coordinates of subgraphs connected\nvia rotatable bonds, creating a variable-length coarse-grained latent\nrepresentation. Our model uses a novel aggregated attention mechanism\nto restore fine-grained coordinates from the coarse-grained latent\nrepresentation, enabling efficient generation of accurate conformers.\nFurthermore, we evaluate the chemical and biochemical quality of our\ngenerated conformers on multiple downstream applications, including\nproperty prediction and large-scale oracle-based protein docking.\nOverall, CoarsenConf generates more accurate conformer ensembles compared\nto prior generative models.