Nephrobase Cell+

Single-cell foundation models such as Geneformer, scGPT, and UCE learn transferable representations of cellular state from large pan-tissue corpora. While powerful, their breadth can dilute the resolution needed to distinguish the closely related cell types and states within a single organ. Nephrobase Cell+, developed by the Susztak laboratory at the University of Pennsylvania and released as a preprint in October 2025, takes the complementary approach of building a foundation model specialized for the kidney.

The model is trained across four mammalian species and several assay modalities, giving it a multimodal, cross-species view of kidney biology. Its central claim is that a domain-focused corpus, combined with an architecture designed to absorb technical variation, yields representations that transfer to new datasets and even new species without task-specific retraining, addressing the batch-effect and annotation-consistency problems that dominate kidney single-cell analysis.

#Key Features

• Kidney specialization: Trained specifically on kidney data rather than a pan-tissue corpus, targeting fine-grained resolution of renal cell types and states.
• Multimodal training: Integrates scRNA-seq, snRNA-seq, snATAC-seq, and spatial transcriptomics within a single model.
• Cross-species coverage: Spans human, mouse, rat, and pig, enabling zero-shot alignment of cell types across organisms.
• Mixture-of-experts architecture: A transformer encoder-decoder with gene-token cross-attention and a mixture-of-experts module scales capacity while separating technical from biological variation.
• Batch-robust embeddings: Removes donor and assay batch effects while preserving cell-type structure, reporting the highest cluster concordance and batch-mixing scores among the compared methods.

#Technical Details

Nephrobase Cell+ is a transformer-based encoder-decoder with gene-token cross-attention and a mixture-of-experts module, offered in 1-billion-parameter and 500-million-parameter variants. It was pretrained on approximately 100 billion tokens drawn from about 39.5 million single-cell and single-nucleus profiles across 4,319 samples, covering human, mouse, rat, and pig and four assay modalities. In benchmarking, it outperforms Geneformer, scGPT, UCE, PCA, and autoencoder baselines on cluster concordance and batch mixing, and achieves over 90% zero-shot annotation accuracy for major kidney lineages in both human and mouse.

#Applications

The model supports kidney research workflows including annotation of newly generated single-cell and single-nucleus datasets without retraining, integration of data across donors, assays, and species, and construction of harmonized kidney cell atlases. Its cross-species alignment is useful for translating findings between animal models and human tissue, with relevance to studies of chronic and diabetic kidney disease where consistent cell-type definitions are critical.

#Impact

Nephrobase Cell+ exemplifies the shift toward organ-specialized single-cell foundation models that trade generality for depth in a specific biological domain. Its practical reach is currently constrained: it is a preprint awaiting peer review, its benchmarks are computational, and it is released under a non-permissive license (CC No reuse) with no public code or weights, which limits independent evaluation and downstream adoption.