pipette
ESEspañol

scUNVEIL: Learning scRNA-seq Cell Embeddings Through Measurement Enrichment

T. Honzik, L. Kuhajda, L. Houdova, D. Georgiev

Preprint

In the authors' words

Large-scale pretraining can improve scRNA-seq representations through measurement enrichment: estimating what the same measurement would reveal at greater UMI depth under the same measurement process. We introduce scUNVEIL, a self-supervised model trained through next-UMI prediction. We show that the Bayes-optimal next-UMI predictor is the posterior mean of the transcriptome composition that arbitrarily deep sequencing would reveal. Because UMI identities are exchangeable conditional on transcriptome composition, scUNVEIL operates directly on count vectors using a residual multilayer perceptron, without attention over gene tokens. The measurement-enriched transcriptome composition is predicted from a single low-dimensional hidden state, which therefore serves naturally as the cell embedding. The same model provides both reusable cell representations and measurement-enriched gene-expression estimates. scUNVEIL was pretrained without labels or auxiliary metadata on approximately 92 million human cells spanning diverse tissues. Under linear evaluation on fine-grained cell-type annotation, its zero-shot representations achieved the best performance among the evaluated methods, while providing over 1000-fold faster inference than the strongest evaluated transformer-based foundation model and matching that model's predictive performance using 88% fewer input UMIs per cell. Further evaluations show that these representations improve prediction of tumor reactivity, mitochondrial membrane potential, cell morphology, and lineage-associated treatment outcomes.

Main resultThe abstract does not state a limitation.

Appeared: Thursday, November 19. bioRxiv. Preprint, not yet peer-reviewed.

DOI: 10.64898/2026.11.17.767296