SigPrimedNet: A Signaling-Informed Neural Network for scRNA-seq Annotation of Known and Unknown Cell Types

SIMPLE SUMMARY: Single-cell data has enabled the study of cell dynamics at an unprecedented resolution. Cell type and functional annotation are crucial to address during any analysis involving transcriptomic data at the cell level since both annotations provide the basis to understand the complex biological processes behind the communication machinery. We propose SigPrimedNet, a data-driven solution to identify cells while learning a functional summarization of signaling measurements by incorporating the knowledge stored in pathway databases. To do so, we decompose each signaling pathway into canonical effector circuits, which act as a minimal functional unit. These circuits inform the design of a cell-type classification neural network model, which allows us to extract meaningful features that act as a proxy of the signaling activity of any given cell. Furthermore, we train an unsupervised anomaly detection algorithm on the inferred activities, which enables the model to identify unknown cells when working with previously unseen cells. To illustrate the performance of the proposed model we conduct a series of experiments over publicly available data with promising results across every task: cell-type annotation, unknown cell-type identification, and clustering. Finally, we showcase the biological richness of the signaling activity learned by the model. ABSTRACT: Single-cell RNA sequencing is increasing our understanding of the behavior of complex tissues or organs, by providing unprecedented details on the complex cell type landscape at the level of individual cells. Cell type definition and functional annotation are key steps to understanding the molecular processes behind the underlying cellular communication machinery. However, the exponential growth of scRNA-seq data has made the task of manually annotating cells unfeasible, due not only to an unparalleled resolution of the technology but to an ever-increasing heterogeneity of the data. Many supervised and unsupervised methods have been proposed to automatically annotate cells. Supervised approaches for cell-type annotation outperform unsupervised methods except when new (unknown) cell types are present. Here, we introduce SigPrimedNet an artificial neural network approach that leverages (i) efficient training by means of a sparsity-inducing signaling circuits-informed layer, (ii) feature representation learning through supervised training, and (iii) unknown cell-type identification by fitting an anomaly detection method on the learned representation. We show that SigPrimedNet can efficiently annotate known cell types while keeping a low false-positive rate for unseen cells across a set of publicly available datasets. In addition, the learned representation acts as a proxy for signaling circuit activity measurements, which provide useful estimations of the cell functionalities.