Fig. 1. Overview of generative modeling pipeline. First, a docked protein–ligand complex is converted to an atomic density representation through atom typing and gridding operations. Density grids are then provided as input to a conditional variational autoencoder (CVAE). The CVAE input branch encodes the full complex density, while its conditional branch encodes only the receptor density. The complex density is mapped to a probabilistic latent space, which is then sampled as a latent vector z ∼ N(μ, σ). This is combined with the conditional vector c output by the conditional encoder, and together they are provided to the decoder. The decoder generates an output ligand density that is converted into the final 3D molecular structure through atom fitting and bond adding.