GAN-Based Training of Semi-Interpretable Generators for Biological Data Interpolation and Augmentation

Single-cell measurements incorporate invaluable information regarding the state of each cell and its underlying regulatory mechanisms. The popularity use single-cell are constantly growing. Despite typically large number collected data, under-representation important (sub-)populations negatively affects down-stream analysis robustness. Therefore, enrichment biological datasets with samples that belong to a rare or manifold is overall advantageous. In this work, we train families generative models via minimization Rényi divergence resulting in an adversarial training framework. Apart from standard neural network-based models, propose semi-interpretable models. proposed further tailored generate realistic gene expression measurements, whose characteristics include zero-inflation sparsity, without need any data pre-processing. Explicit factors such as measurement time, cluster taken into account by our conditional variables. We compare them against state-of-the-art on range synthetic real demonstrate their ability accurately perform interpolation augmentation.