Single-cell data are crucial for biomedical discovery, facilitated by low-dimensional latent space encoding of single-cell RNA-seq profile. Obtained latent codes can then be plotted into 2-D space via t-SNE or UMAP allowing practitioners to infer new knowledge. Alternatively, downstream deep learning applications can also be trained from the latent codes, one of the common being the cell type classifier. The usefulness of the 2-D plot or downstream application depends critically on the structure of the latent space and whether it encodes biological information or noise. The proposed approach aims to improve the latent space via injecting a bit of label information, thus denoting our approach as a semi-supervised one. We include a novel dual-VAE structure, where information flows from the controller VAE to the main model. Our results demonstrate that the incorporation of SemafoVAE improves the performance of the existing scANVI model, therefore offering a refined model structure with disentangled latent representations for robust biological insights.