Deep generative models, such as variational autoencoders (VAEs) or deep Boltzmann machines (DBMs), can generate an arbitrary number of synthetic observations after being trained on an initial set of samples. This has mainly been investigated for imaging data but could also be useful for single-cell transcriptomics (scRNA-seq). A small pilot study could be used for planning a full-scale experiment by investigating planned analysis strategies on synthetic data with different sample sizes. It is unclear whether synthetic observations generated based on a small scRNA-seq dataset reflect the properties relevant for subsequent data analysis steps. We specifically investigated two deep generative modeling approaches, VAEs and DBMs. First, we considered single-cell variational inference (scVI) in two variants, generating samples from the posterior distribution, the standard approach, or the prior distribution. Second, we propose single-cell deep Boltzmann machines (scDBMs). When considering the similarity of clustering results on synthetic data to ground-truth clustering, we find that the $$scVI_{posterior}$$ s c V I posterior variant resulted in high variability, most likely due to amplifying artifacts of small datasets. All approaches showed mixed results for cell types with different abundance by overrepresenting highly abundant cell types and missing less abundant cell types. With increasing pilot dataset sizes, the proportions of the cells in each cluster became more similar to that of ground-truth data. We also showed that all approaches learn the univariate distribution of most genes, but problems occurred with bimodality. Across all analyses, in comparing 10$$\times$$ × Genomics and Smart-seq2 technologies, we could show that for 10$$\times$$ × datasets, which have higher sparsity, it is more challenging to make inference from small to larger datasets. Overall, the results show that generative deep learning approaches might be valuable for supporting the design of scRNA-seq experiments.

Metastatic disease is a major cause of death for patients with melanoma. Melanoma cells can become metastatic not only due to cell-intrinsic plasticity but also due to cancer-induced protumorigenic remodeling of the immune microenvironment. Here, we report that innate immune surveillance by natural killer (NK) cells is bypassed by human melanoma cells expressing the stem cell marker NGFR. Using in vitro and in vivo cytotoxic assays, we show that NGFR protects melanoma cells from NK cell–mediated killing and, furthermore, boosts metastasis formation in a mouse model with adoptively transferred human NK cells. Mechanistically, NGFR leads to down-regulation of NK cell activating ligands and simultaneous up-regulation of the fatty acid stearoyl–coenzyme A desaturase (SCD) in melanoma cells. Notably, pharmacological and small interfering RNA–mediated inhibition of SCD reverted NGFR-induced NK cell evasion in vitro and in vivo. Hence, NGFR orchestrates immune control antagonizing pathways to protect melanoma cells from NK cell clearance, which ultimately favors metastatic disease.

Motivation: When designing experiments, it is advised to start with a small pilot study for determining the sample size of full-scale investigations. Deep learning techniques for single-cell RNA-sequencing data that can uncover low-dimensional representations of expression patterns within cells could be useful also with pilot data. Here, we examine the ability of these methods to learn the structure of data from a small pilot study and generate synthetic expression datasets useful for planning full-scale experiments. Results: We investigate two deep generative modeling approaches. First, we consider single-cell variational inference (scVI) in two variants, either generating samples from the posterior distribution, which is the standard approach, or from the prior distribution. Second, we propose single-cell deep Boltzmann machines (scDBM), which might be particularly suitable for small datasets. When considering the similarity of clustering results on synthetic data to ground-truth clustering, we find that scV Iposterior exhibits high variability. Expression patterns from scV Iprior and scDBM perform better. All approaches show mixed results for cell types with different abundance by sometimes overrepresenting highly abundant cell types and missing less abundant cell types. Taking such tradeoffs in performance into account, we conclude that for making inference from a small pilot study to a larger experiment, it is advantageous to use scV Iprior or scDBM, as scV Iposterior produces signals that are not justified by the original data. The proposed scDBM seems to have an advantage for small pilot datasets. Overall, the results show that generative deep learning approaches might be valuable for supporting the design of scRNA-seq experiments.

Despite conceptual research on hippocampus development and the application of single-cell-resolved technologies, the nature and maturation of its diverse progenitor populations are unexplored. The chromatin modifier DOT1L balances progenitor proliferation and differentiation, and conditional loss-of-function mice featured impaired hippocampus development. We applied single-cell RNA sequencing on DOT1L-mutant mice and explored cell trajectories in the E16.5 hippocampus. We resolved in our data five distinct neural stem cell populations with the developmental repertoire to specifically generate the cornu ammonis (CA) 1 field and the dentate gyrus (DG). Within the two developing CA1- and CA3-fields, we identified two distinct maturation states and we thus propose CA1- and CA3-differentiation along the radial axis. In the developing hippocampus, DOT1L is primarily involved in the proper development of CA3 and the DG, and it serves as a state-preserving epigenetic factor that orchestrates the expression of several important transcription factors that impact neuronal differentiation and maturation.