CytoRescue

Single-nucleus RNA sequencing is often the only viable option: frozen archival tissue, fibrotic or fatty samples, and cell types that refuse to dissociate all rule out whole-cell protocols, and working from nuclei avoids the dissociation-induced stress response that contaminates fresh preparations. The price is structural. Transcripts that have left the nucleus are simply not captured, so cytoplasm-enriched genes are systematically undercounted — a bias that falls hardest on activation programs and on the secreted ligands that drive cell-cell communication.

CytoRescue, from the Center for Precision Health at UTHealth Houston, treats this as a supervised generative problem rather than a denoising one. Where general imputation tools smooth over dropout within a dataset, CytoRescue is trained on paired snRNA-seq and scRNA-seq profiles of the same biological samples, using the whole-cell counts as an explicit prediction target. The model therefore learns what the missing cytoplasmic fraction looks like, instead of redistributing signal that is already present.

Its design choice with the most practical consequence is raw-in-raw-out: CytoRescue consumes unnormalized counts and emits unnormalized counts, so it slots in ahead of any existing analysis pipeline without changing downstream normalization, clustering, or differential expression conventions.

#Key Features

• Guided VAE with paired supervision: Matched scRNA-seq counts serve as the prediction target during training, so the model learns the specific cytoplasmic deficit of nuclear preparations rather than a generic smoothing operator.
• Raw-in-raw-out interface: Inputs and outputs are unnormalized read counts, allowing drop-in insertion into standard single-cell workflows.
• ZINB likelihood over counts: Zero-inflated negative binomial modeling handles the overdispersion and excess zeros characteristic of single-cell count data.
• Cell identity preserved: Adjusted Rand index, normalized mutual information, and consistency metrics confirm that recovered profiles retain the original cell type structure while the cytoplasmic signal is restored.
• Fixed checkpoint, zero-shot transfer: A single trained model was applied without retraining to a lung sample held out of training entirely, where it recovered signaling that snRNA-seq alone had lost.

#Technical Details

CytoRescue is a variational autoencoder with a Gaussian latent space and zero-inflated negative binomial parameters inferred at the input and output layers, trained with a composite objective combining the ZINB reconstruction loss, the KL divergence of the latent space, and a mean squared error term from a prediction head supervised by the paired scRNA-seq counts. Training data comprised paired snRNA-seq and scRNA-seq datasets from human brain and lung tissue: paired microglia counts from GEO series GSE137444 (scRNA-seq) and GSE153807 (snRNA-seq), and paired normal and tumor lung data from a public Zenodo deposit. Cells were annotated against the SEA-AD reference for brain and Human Lung Cell Atlas v1.0 for lung, then matched between assays using scmap. Two samples were withheld entirely as validation; the remainder was split 80/20 into training and testing. Training ran on a single NVIDIA A100-SXM4-40GB GPU. Concordance with paired whole-cell data, measured as Pearson correlation of per-gene mean expression, rose from 0.894 to 0.971 in brain sample RM77_2, from 0.882 to 0.955 in RM101_2, and from 0.825 to 0.915 in the withheld lung sample 8-CST-0892. The general-purpose imputation and denoising tools scVI and ALRA, which were not built for this problem, did not recover the lost signal. Genes gaining more than twofold expression after recovery were enriched for translation, mitochondrial energy production, and response to external stimulus.

#Applications

The model targets any study where snRNA-seq is the only practical assay but cytoplasmic biology matters: neurodegeneration work relying on disease-associated microglia signatures, tumor microenvironment studies in frozen resections, and any analysis of ligand-receptor signaling from nuclear data. On the withheld lung sample, CellChat inference over CytoRescue outputs restored EGF-mediated interactions between alveolar epithelium and other cell types that were largely absent from the raw snRNA-seq counts but present in the paired whole-cell data. Because the interface is raw counts, adoption requires no change to an established pipeline.

#Impact

CytoRescue reframes a well-known snRNA-seq artifact as something learnable from paired data rather than an irreducible limit of the assay, and its recovery of cell-cell communication on a sample never seen in training is the most consequential result: signaling analyses are precisely where nuclear-only data has been least trustworthy. The scope is currently narrow. Training and validation cover brain and lung tissue only, so generalization to other organs is not yet demonstrated, and the approach depends on the continued availability of paired snRNA-seq/scRNA-seq datasets, which remain uncommon. The work is a preprint that has not been peer reviewed and is distributed under a CC BY-NC-ND license; source and analysis code are available under the MIT license.