| Summary: | The extraordinary regenerative capacity of planarians is derived from their large population of pluripotent stem cells called neoblasts. Neoblasts are a heterogenous population of stem cells that begin unspecified, but are rapidly specified within a single cell cycle and give rise to every cell in the animal. In this work, we characterize the different types of specified neoblasts in the planarian Schmidtea mediterranea by using single-cell RNA sequencing. We also profile their post-mitotic descendants that serve as migratory progenitors. Through these analyses, we uncover many novel neoblast and post-mitotic progenitor types, and the genes that define them. Because neoblasts subtypes are often specified by their unique expression of Fate Specific Transcription Factors (FSTFs), we computationally identify planarian transcription factor genes and characterize neoblasts and post-mitotic progenitors by their expression of these putative transcription factors. We find that many cell types can be identified through their unique expression of FSTF modules, which can be used to connect cells with the same fate across neoblast, post-mitotic progenitor, and differentiated cell stages. Through gene inhibition studies, we discover that many of the FSTFs discovered for these novel precursor types have a functional role in their specification. We note that transcriptional signatures of some cell-type fates become apparent at different stages in the lifetime of the progenitor and propose a model where cellular diversity arises at different times for different tissue classes.
To better uncover the genes signatures that define different cell-types, we develop methods that use neural networks to learn patterns in gene expression in planarian post-mitotic progenitors. We find that neural network classifiers are powerful predictors of cell-type based on gene expression and the network’s learned weights can be examined to uncover gene signatures that define different progenitor types. We find that autoencoders, a class of neural networks made for the efficient representation of data, can be used to learn gene signatures in cells in an unsupervised manner. We find that cells close together in UMAP space, but belonging to different cell clusters through traditional clustering, are often difficult for the neural networks to distinguish. From this, we hypothesize that cells in different clusters may be transcriptionally similar and cells existing across continuous UMAP space may exist across continuous gene-expression space, and so assigning cells to discrete clusters may not always be appropriate. We propose autoencoders, which can encode gene-expression signatures through a continuous, latent-space encoding, may be more appropriate and can be used to uncover novel gene and cell relationships.
|
|---|