The abstract is read at the publisher; the summary is JClub's.
scUNVEIL learns scRNA-seq cell embeddings from next-UMI prediction, achieving best zero-shot fine-grained cell-type annotation on ~92 million human cells.
- Why it matters: Large-scale pretraining can improve scRNA-seq representations, but measurement effects from limited UMI depth limit what models can infer from sparse counts.
- What they did: scUNVEIL is a self-supervised model that predicts the next-UMI from count vectors via a residual multilayer perceptron, using a single low-dimensional hidden state as the embedding.
- The result: In linear evaluation, its zero-shot representations lead among tested methods, with over 1000-fold faster inference than the strongest transformer baseline and matching predictive performance using 88% fewer input UMIs per cell.
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.