<p>In drug discovery, optimizing molecules to enhance target properties is a crucial step. Recent advances in generative machine learning have facilitated this by exploring molecular structures within a latent representation space. Those studies typically employ encoder-decoder architectures to convert molecules into latent vectors and then reconstruct them. This process often requires complex subsequent editing to produce valid molecules. In this research article, we propose a framework that eliminates this final editing step by directly mapping input molecules to a latent subspace preferred by a binary classifier aimed at the target property. Based on a variational autoencoder (VAE), we integrate an auxiliary classifier in the latent space to steer the training process toward generating molecular fragments that better align with binary labels. We test our model on three benchmark drug optimization tasks, achieving positive outcomes for various molecular properties, including improvement of the target property. This study presents a molecule optimization method particularly effective in settings with limited data and offers insights into molecule substructures that could enhance drug lead properties.</p>