Single-cell genomics technologies enable multimodal profiling of millions of cells across temporal and spatial dimensions. Experimental limitations prevent the measurement of all-encompassing cellular states in their native temporal dynamics or spatial tissue niche. Optimal transport theory has emerged as a powerful tool to overcome such constraints, enabling the recovery of the original cellular context. However, most algorithmic implementations currently available have not kept up the pace with increasing dataset complexity, so that current methods are unable to incorporate multimodal information or scale to single-cell atlases. Here, we introduce multi-omics single-cell optimal transport (moscot), a general and scalable framework for optimal transport applications in single-cell genomics, supporting multimodality across all applications. We demonstrate moscot's ability to efficiently reconstruct developmental trajectories of 1.7 million cells of mouse embryos across 20 time points and identify driver genes for first heart field formation. The moscot formulation can be used to transport cells across spatial dimensions as well: To demonstrate this, we enrich spatial transcriptomics datasets by mapping multimodal information from single-cell profiles in a mouse liver sample, and align multiple coronal sections of the mouse brain. We then present moscot.spatiotemporal, a new approach that leverages gene expression across spatial and temporal dimensions to uncover the spatiotemporal dynamics of mouse embryogenesis. Finally, we disentangle lineage relationships in a novel murine, time-resolved pancreas development dataset using paired measurements of gene expression and chromatin accessibility, finding evidence for a shared ancestry between delta and epsilon cells. Moscot is available as an easy-to-use, open-source python package with extensive documentation at https://moscot-tools.org.

Motivation Protein design has become increasingly important for medical and biotechnological applications. Because of the complex mechanisms underlying protein formation, the creation of a novel protein requires tedious and time-consuming computational or experimental protocols. At the same time, machine learning has enabled the solving of complex problems by leveraging large amounts of available data, more recently with great improvements on the domain of generative modeling. Yet, generative models have mainly been applied to specific sub-problems of protein design. Results Here we approach the problem of general purpose protein design conditioned on functional labels of the hierarchical Gene Ontology. Since a canonical way to evaluate generative models in this domain is missing, we devise an evaluation scheme of several biologically and statistically inspired metrics. We then develop the conditional generative adversarial network ProteoGAN and show that it outperforms several classic and more recent deep learning baselines for protein sequence generation. We further give insights into the model by analysing hyperparameters and ablation baselines. Lastly, we hypothesize that a functionally conditional model could generate proteins with novel functions by combining labels and provide first steps into this direction of research. Availability Code and data is available at https://github.com/timkucera/proteogan Supplemental information Supplemental data are available at Bioinformatics online.

The emergence of single-cell co-assays enables us to learn to translate between single-cell modalities, potentially offering valuable insights from datasets where only one modality is available. However, the sparsity of single-cell measurements and the limited number of cells measured in typical co-assay datasets impedes the power of cross-modality translation. Here, we propose Polarbear, a semi-supervised translation framework to predict cross-modality profiles that is trained using a combination of co-assay data and traditional “single-assay” data. Polarbear uses single-assay and co-assay data to train an autoencoder for each modality and then uses just the co-assay data to train a translator between the embedded representations learned by the autoencoders. With this approach, Polarbear is able to translate between modalities with improved accuracy relative to state-of-the-art translation techniques. As an added benefit of the training procedure, we show that Polarbear also produces a matching of cells across modalities.

No abstract