Single-cell sequencing provides deep cell-specific information, but the method suffers from technical constraints, most notably a limited number of expressed genes per cell, which leads to suboptimal clustering and cell type identification. DISCERN, a novel deep generative network, can be used to reconstruct missing single-cell gene expression using a reference dataset. It combines techniques already used for deep learning-based analysis of expression data with techniques originally developed for image analysis tasks. DISCERN greatly improves cell clustering, cell type, and activity detection, and can give insights into the cellular regulation of disease. It is able to correct batch effects, align expression values, and enable better detection of cell types verified by CITE-seq information. A unique feature of DISCERN is using an hq reference to infer biologically meaningful gene expression. Its usability and robustness should enable even non-expert users to perform gene expression reconstruction.