Scvi-hub: an actionable repository for model-driven single-cell analysis (www.nature.com)

scvi-hub is a new platform for sharing and reusing pretrained probabilistic single‑cell omics models within the scvi-tools/scverse ecosystem. It packages state-of-the-art parametric models (e.g., scVI/scANVI conditional VAEs) on the Hugging Face Model Hub and cloud buckets, and exposes them through a streamlined Python/R API. Users can immediately run common tasks—visualization, batch-corrected embedding, imputation, automated annotation, differential expression and spatial deconvolution—on new query data without downloading large reference count matrices. The repository is already seeded with 90+ pretrained models from Tabula Sapiens and full access to the CZI CELLxGENE Census, and includes tooling for contributors (model upload/versioning) and consumers (model discovery and integration with Seurat/scvi-tools workflows). Technically, scvi-hub’s key innovation is “data minification”: storing low‑dimensional posterior parameters (latent means/variances) rather than full count matrices, and using the generative model to reconstruct normalized counts on demand. This cuts storage and compute needs, letting researchers run atlas-scale analyses on conventional hardware. scvi-hub also implements standardized model evaluation via posterior predictive checks (gene/cell coefficient of variation, differential-expression metrics) in the scvi.criticism module so users can assess fidelity before reuse. By standardizing model deposition, evaluation and access, scvi-hub enables scalable transfer learning across single‑cell atlases and broadens reproducible, atlas-level analyses for the community.