<ul><li>Medicinal chemists often optimize drugs considering 3D structures and designing structurally distinct molecules while retaining key features like shapes, pharmacophores, or chemical properties.</li><li>A new flexible zero-shot molecule manipulation method has been proposed by navigating in a shared latent space of 3D molecules using MolFLAE, a Variational AutoEncoder for 3D molecules.</li><li>MolFLAE encodes 3D molecules into a fixed-dimensional, SE(3)-equivariant latent space, allowing for tasks such as atom editing, structure reconstruction, and coordinated latent interpolation.</li><li>The approach showcases competitive performance on unconditional 3D molecule generation benchmarks and demonstrates efficacy in drug optimization tasks, such as generating molecules with improved hydrophilicity while maintaining key interactions.</li></ul>

Manipulating 3D Molecules in a Fixed-Dimensional SE(3)-Equivariant Latent Space

Discover more