Keywords: Single-cell transcriptomics, transfer learning, probabilistic modelling, immunology
TL;DR: scvi-hub is a platform of pretrained scvi-tools models. We highlight case studies on lung disease biology. We introduce minified data to reduce hardware requirements, and scvi.criticism, a framework for posterior-predictive checks in scvi-tools.
Abstract: The accumulation of single-cell omics datasets in the public domain has opened
new opportunities to reuse and leverage the vast amount of information they con-
tain. Such uses, however, are complicated by the need for complex and resource-
consuming procedures for data transfer, normalization, and integration that must
be addressed prior to any analysis. Here we present scvi-hub: a platform for evalu-
ating, sharing, and accessing probabilistic models that were trained on single-cell
omics datasets. We demonstrate that these pre-trained models allow immediate
access to a slew of fundamental tasks like visualization, imputation, annotation,
outlier detection, and deconvolution of new (query) datasets with a much lower
requirement for compute resources. We also show that pretrained models can help
drive new discoveries with the existing (reference) datasets through rapid, model-
based analyses. Scvi-hub is built within scvi-tools and integrated into scverse.
Scvi-hub is publicly available to enable efficient sharing of single-cell omic stud-
ies, and also to put advanced capabilities for transfer learning at the fingertips of
a broad community of users. We provide an extended journal version on bioRxiv
Submission Number: 29
Loading