ScNucAdapt enables scRNA-seq to snRNA-seq cross-annotation

The arXiv preprint (arXiv:2511.08996) by Xiran Chen et al. introduces ScNucAdapt, a method designed for cross-domain cell type annotation between scRNA-seq and snRNA-seq. Per the preprint, ScNucAdapt applies partial domain adaptation to reconcile distributional differences and varying cell compositions across these data modalities. The authors report evaluation on both paired and unpaired scRNA-seq/snRNA-seq datasets and claim that ScNucAdapt outperforms existing annotation approaches on these benchmarks.

Per the preprint, ScNucAdapt frames cross-annotation as a domain-adaptation problem where only a subset of classes is shared across source and target, and the method implements mechanisms to handle unshared cell types and composition mismatch. The paper provides experimental protocols and comparative benchmarks against prior methods; specific algorithmic components and evaluation metrics are described in the PDF and accompanying code/data links listed on the arXiv entry.

Partial domain adaptation is an established technique in ML for cases where the target label space is a subset of the source, or where spurious source-only classes exist. For practitioners, adapting these techniques to single-cell genomics addresses two common issues: modality-driven expression shifts between cell and nucleus measurements, and altered cell-type prevalence due to dissociation biases. Methods that explicitly model partial overlap can reduce label-mismatch errors compared with straightforward batch-correction or naive transfer learning approaches.

Cross-annotation between scRNA-seq and snRNA-seq matters for labs that combine fresh dissociated-cell data with frozen or hard-to-dissociate tissue samples. Improvements in label transfer accuracy speed downstream analysis and integration workflows used by computational biologists. While this work is presented as a preprint, it contributes a practical adaptation of domain-adaptation ideas to a recurring problem in single-cell analysis and supplies code/data resources that can be incorporated into pipelines.

For practitioners: watch for independent benchmarks reproducing the reported improvements, the availability and usability of the authors' code and pretrained models, and extensions of the approach to multi-modal integration (for example, combining ATAC, protein, or spatial data). Also observe how the community evaluates robustness to rare cell types and to technical confounders not covered in the paper.