MolBit: De novo Drug Design via Binary Representations of SMILES for avoiding the Posterior Collapse Problem
Many drug design studies have proposed combinations of VAEs and RNNs to generate SMILES strings.
Although those RNN-VAE models have good validity performance, they suffer from the posterior collapse problem, in which every latent vector has an identical molecular property distribution.
We proposed a Gumbel-Softmax-based generative model called MolBit to avoid the posterior collapse problem.
https://doi.org/10.1109/BIBM52615.2021.9669668