Cytoplasmic dynamics are overlooked in single nuclei RNA-seq but can be rescued by CytoRescue, a generative AI model to recover cytoplasm enriched gene

Cytoplasmic dynamics are overlooked in single nuclei RNA-seq but can be rescued by CytoRescue, a generative AI model to recover cytoplasm enriched gene | bioRxiv /* */ /* */ <!-- <!-- /*! * yepnope1.5.4 * (c) WTFPL, GPLv2 */ (function(a,b,c){function d(a){return"[object Function]"==o.call(a)}function e(a){return"string"==typeof a}function f(){}function g(a){return!a||"loaded"==a||"complete"==a||"uninitialized"==a}function h(){var a=p.shift();q=1,a?a.t?m(function(){("c"==a.t?B.injectCss:B.injectJs)(a.s,0,a.a,a.x,a.e,1)},0):(a(),h()):q=0}function i(a,c,d,e,f,i,j){function k(b){if(!o&&g(l.readyState)&&(u.r=o=1,!q&&h(),l.onload=l.onreadystatechange=null,b)){"img"!=a&&m(function(){t.removeChild(l)},50);for(var d in y[c])y[c].hasOwnProperty(d)&&y[c][d].onload()}}var j=j||B.errorTimeout,l=b.createElement(a),o=0,r=0,u={t:d,s:c,e:f,a:i,x:j};1===y[c]&&(r=1,y[c]=[]),"object"==a?l.data=c:(l.src=c,l.type=a),l.width=l.height="0",l.onerror=l.onload=l.onreadystatechange=function(){k.call(this,r)},p.splice(e,0,u),"img"!=a&&(r||2===y[c]?(t.insertBefore(l,s?null:n),m(k,j)):y[c].push(l))}function j(a,b,c,d,f){return q=0,b=b||"j",e(a)?i("c"==b?v:u,a,b,this.i++,c,d,f):(p.splice(this.i++,0,a),1==p.length&&h()),this}function k(){var a=B;return a.loader={load:j,i:0},a}var l=b.documentElement,m=a.setTimeout,n=b.getElementsByTagName("script")[0],o={}.toString,p=[],q=0,r="MozAppearance"in l.style,s=r&&!!b.createRange().compareNode,t=s?l:n.parentNode,l=a.opera&&"[object Opera]"==o.call(a.opera),l=!!b.attachEvent&&!l,u=r?"object":l?"script":"img",v=l?"script":u,w=Array.isArray||function(a){return"[object Array]"==o.call(a)},x=[],y={},z={timeout:function(a,b){return b.length&&(a.timeout=b[0]),a}},A,B;B=function(a){function b(a){var a=a.split("!"),b=x.length,c=a.pop(),d=a.length,c={url:c,origUrl:c,prefixes:a},e,f,g;for(f=0;f<d;f++)g=a[f].split("="),(e=z[g.shift()])&&(c=e(c,g));for(f=0;f<b;f++)c=x[f](c);return c}function g(a,e,f,g,h){var i=b(a),j=i.autoCallback;i.url.split(".").pop().split("?").shift(),i.bypass||(e&&(e=d(e)?e:e[a]||e[g]||e[a.split("/").pop().split("?")[0]]),i.instead?i.instead(a,e,f,g,h):(y[i.url]?i.noexec=!0:y[i.url]=1,f.load(i.url,i.forceCSS||!i.forceJS&&"css"==i.url.split(".").pop().split("?").shift()?"c":c,i.noexec,i.attrs,i.timeout),(d(e)||d(j))&&f.load(function(){k(),e&&e(i.origUrl,h,g),j&&j(i.origUrl,h,g),y[i.url]=2})))}function h(a,b){function c(a,c){if(a){if(e(a))c||(j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}),g(a,j,b,0,h);else if(Object(a)===a)for(n in m=function(){var b=0,c;for(c in a)a.hasOwnProperty(c)&&b++;return b}(),a)a.hasOwnProperty(n)&&(!c&&!--m&&(d(j)?j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}:j[n]=function(a){return function(){var b=[].slice.call(arguments);a&&a.apply(this,b),l()}}(k[n])),g(a[n],j,b,n,h))}else!c&&l()}var h=!!a.test,i=a.load||a.both,j=a.callback||f,k=j,l=a.complete||f,m,n;c(h?a.yep:a.nope,!!i),i&&c(i)}var i,j,l=this.yepnope.loader;if(e(a))g(a,0,l,0);else if(w(a))for(i=0;i (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0];var j=d.createElement(s);var dl=l!='dataLayer'?'&l='+l:'';j.src='//www.googletagmanager.com/gtm.js?id='+i+dl;j.type='text/javascript';j.async=true;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-M677548'); Skip to main content Home About Submit ALERTS / RSS Search for this keyword Advanced Search New Results Cytoplasmic dynamics are overlooked in single nuclei RNA-seq but can be rescued by CytoRescue, a generative AI model to recover cytoplasm enriched gene View ORCID Profile Wankun Deng , View ORCID Profile Zhongming Zhao doi: https://doi.org/10.1101/2025.08.15.670239 Wankun Deng 1 Center for Precision Health, McWilliams School of Biomedical Informatics, The University of Texas Health Science Center at Houston , Houston, TX 77030, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Wankun Deng Zhongming Zhao 1 Center for Precision Health, McWilliams School of Biomedical Informatics, The University of Texas Health Science Center at Houston , Houston, TX 77030, USA 2 MD Anderson Cancer Center UTHealth Graduate School of Biomedical Sciences , Houston, TX 77030, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Zhongming Zhao For correspondence: zhongming.zhao{at}uth.tmc.edu Abstract Full Text Info/History Metrics Supplementary material Preview PDF Abstract Single-nucleus RNA sequencing (snRNA-seq) generates single cell data from nuclei. It provides valuable compatibility with frozen or difficult-to-dissociate tissues while avoiding stress responses in fresh samples. However, cytoplasmic depletion inherently limits quantification of cytoplasm-enriched genes. Here, we present CytoRescue, a novel generative AI model designed to recover attenuated cytoplasmic signals in snRNA-seq data. Our results demonstrate that CytoRescue effectively restores expression of cytoplasm-enriched genes while preserving underlying gene expression signatures. Taking advantaging of the raw-in-raw-out design, CytoRescue can be easily integrated into the existing pipelines for single-cell sequencing analysis. Notably, CytoRescue successfully recovers EGF signaling pathway components, a critical cell-cell communication pathway in lung cancer, in an independent dataset. CytoRescue addresses a fundamental limitation of snRNA-seq technology, enhancing its utility for comprehensive transcriptomic profiling while maintaining the advantages of single nucleus-based approaches. Background Single-nucleus RNA sequencing (snRNA-seq) is considered more flexible with sample preparation when comparing to single cell RNA-seq (scRNA-seq), which requires live cells from fresh tissue followed by quick dissociation in most cases 1 . Benefitting from the compatibility of frozen sample, snRNA-seq is widely used in single cell research 1 . Meanwhile, in snRNA-seq experiments, cytoplasm content will be discarded during nuclei isolation 1 ( Figure 1A ), resulting in missingness of cytoplasm dynamic 1 – 4 . However, various cytoplasm processes exhibit high dynamic in certain cells while related genes showing no or small variation in nuclei 1 – 4 . For example, microglia activation related genes are enriched in cytoplasm and snRNA-seq cannot accurately detect microglia activation 4 . Furthermore, dynamic in MHC-1 signaling pathway, which is crucial in immune recognition and response, can barely be detected in snRNA-seq in neither normal nor tumor lung samples 3 . Download figure Open in new tab Figure 1. Diagram and evaluation of CytoRescue. (A) An overview of scRNA-seq and snRNA-seq procedures. (B) Design of CytoRescue. The model is designed with both reconstruction and prediction heads. Using paired snRNA-seq and scRNA-seq profiles as prediction targets, CytoRescue is trained on a combined loss function by incorporating reconstruction loss from the reconstruction head, Kullback-Leibler divergence of the latent space, and mean squared error from the prediction head. (C-N) Mean gene expression comparison between scRNA-seq and snRNA-seq by three methods: CytoRescue-predicted, scVI-imputed, and ALRA-imputed profiles. Results from three samples are shown: RM77_1 (brain, training sample), RM101_2 (brain, validation sample), and 8-CST-0892 (normal lung, validation sample). Disease-associated microglia genes are marked as dots with black edges in panels C, D, F, G, I, J, L, M. (O) Pearson correlation coefficients (PCC) between mean gene expression of scRNA-seq profiles and raw snRNA-seq counts, CytoRescue-predicted, scVI-imputed, and ALRA-imputed profiles across all samples in this study. (P) Adjusted Rand index (ARI), normalized mutual information (NMI), and consistency of cell type annotation in lung samples. The recent advances in generative artificial intelligence (AI) models shed lights on modeling biological structures from biased, noisy and sparse data 5 – 8 . Among many algorithms, variational autoencoder (VAE) has emerged as one of most popular frameworks in single-cell study, benefitting from its capability to model the probabilistic distribution of single cell read counts 6 , 8 – 10 . To model single-cell RNA-seq data, a zero inflated negative binomial (ZINB) distribution was proposed by adding an additional parameter π as zero-inflating probability 11 . By leveraging these progresses, we propose a VAE algorithm based on ZINB distribution, namely, CytoRescue, to rescue the lost cytoplasm dynamic in snRNA-seq experiments due to technology limitation. Results and discussion We developed CytoRescue, a deep generative model that employs probabilistic representation of single-nucleus RNA sequencing data to reconstruct the gene expression signals that are lost during nuclei isolation procedures ( Figure 1B ). The specific method is described in the Methods section. Briefly, the model utilizes a variational autoencoder (VAE) framework to transform the raw, unnormalized count data by modeling gene expression patterns according to a zero-inflated binomial distribution 8 , 10 , 12 , 13 . CytoRescue outputs raw read counts as its prediction target, facilitating seamless integration into existing analytical pipelines. To leverage available paired single-nucleus RNA-seq and single-cell RNA-seq datasets, we implemented CytoRescue as a guided VAE architecture that enables the model to learn from established distributions ( Figure 1B ) 14 . We curated matched snRNA-seq and scRNA-seq datasets from published literature to construct our training, testing, and validation datasets (Table S1) 2 – 4 . RM101 was selected as the validation dataset due to the absence of any other data from the same biological sample in the training set (Table S1). Additionally, 8-CST-0892 was randomly selected as validation dataset as well 3 , 4 . To train CytoRescue, we first aligned snRNA-seq data with paired scRNA-seq data from identical samples using scmap for cell mapping (Figure S1) 15 . The model employs raw read counts from scRNA-seq as prediction targets to guide the learning of gene expression patterns in snRNA-seq data and recover signals from cytoplasmic content lost during nuclei isolation. Previous studies have demonstrated that microglia activation signals are predominantly lost in snRNA-seq data, as genes associated with microglia activation are enriched in the cytoplasm 4 . We compared pseudobulk gene expression profiles between snRNA-seq and paired scRNA-seq data in microglia from sample RM77 (training dataset) and RM101 (validation dataset) ( Figure 1C, 1D ). Overall, snRNA-seq data exhibited strong correlation with scRNA-seq data, with Pearson’s correlation coefficients (PCC) being approximately 0.85 ( Figure 1C-E ). However, disease-associated microglia (DAM) genes were significantly enriched in those genes having higher expression in scRNA-seq than those in snRNA-seq ( Figure 1C, 1D ). Notably, the majority of DAM genes exhibited greater than 2-fold expression differences, supporting their higher expression in scRNA-seq (sample RM77: 36 vs. 0; sample RM101: 41 vs. 0). This observation is consistent with a previous finding 4 . This trend of a disproportionately high number of genes with elevated expression in scRNA-seq, but not in snRNA-seq, was also observed in lung samples ( Figure 1E ). We subsequently applied CytoRescue to snRNA-seq data and compared the results with paired scRNA-seq data across all training and validation samples ( Figure 1F-H , Figure S3, S4). The rescued gene expression profiles demonstrated significantly improved alignment with scRNA-seq data, with PCC values being increased from 0.894 to 0.971 in sample RM77_2, and from 0.882 to 0.955 in RM101_2, and from 0.825 to 0.915 in 8-CST-0892 ( Figure 1F-H ). Consistent improvements were observed across all other samples. The uniformly enhanced PCC values in both training and testing datasets indicate that CytoRescue can effectively recover the lost counts in snRNA-seq experiments ( Figure 1P ). Given the availability of numerous imputation and denoising tools for single-cell/single-nucleus sequencing data, we investigated whether the lost signals in snRNA-seq could be recovered through conventional imputation or denoising approaches. We selected two widely used tools for comparison: scVI 16 (based on VAE) and ALRA 17 (based on low-rank approximation) ( Figure 1L-P ). As demonstrated, these imputation and denoising tools were not designed to restore discarded cytoplasmic mRNAs from snRNA-seq data and failed to effectively recover the lost signals ( Figure 1L-P ). Furthermore, we evaluated the consistency of cell type annotations between raw snRNA-seq counts and CytoRescue-predicted counts using adjusted Rand index (ARI), normalized mutual information (NMI), and proportion of consistency metrics 18 , 19 (see Methods). Our results demonstrate that while effectively recovering cytoplasmic dynamics, CytoRescue-predicted counts preserve cell type signatures well ( Figure 1P , Table S2). We further investigated the extent to which biological structures were captured by the model and preserved in the output. A standard approach for this assessment involves the use of Uniform Manifold Approximation and Projection (UMAP) for dimensionality reduction and cell clustering, combined with cell subpopulation annotation based on prior knowledge 20 – 22 . When compared with cell clustering using original read counts ( Figure 2A-C ), UMAP visualization of predicted read counts effectively preserved cell cluster structure ( Figure 2D-F ), particularly when it was stratified by sample type ( Figure 2A, 2D ). When clustered by original read count, microglia cells from brain tissue were intermixed with other immune cells from normal lung and tumor samples ( Figure 2B ). Following CytoRescue prediction, microglia cells were appropriately separated from other immune cell types as expected ( Figure 2E ). Additionally, CytoRescue was designed in a raw-in-raw-out manner, ( Figure 2C, 2F ), which facilitates seamless integration into existing analytical pipelines. We extended our analysis to examine the latent space representation to further verify whether CytoRescue can capture the underlying biological features in snRNA-seq data, but not just memorizing the training data 7 . As demonstrated in Figure 2G-I , the cells were well-clustered and appropriately separated regardless sample types, cell types, and patients we examined, with the most distinct separation being observed when stratified by sample type ( Figure 2G ). In summary, our evaluation demonstrated that CytoRescue could effectively capture biological structures in snRNA-seq data and recover the signals from discarded cytoplasmic content while preserving biological signatures. Download figure Open in new tab Figure 2. CytoRescue preserves biological features in original data and recovers cytoplasmic dynamics. Uniform Manifold Approximation and Projection (UMAP) of all cells in this study using raw snRNA-seq counts (A-C), CytoRescue-predicted counts (D-F), and latent representations from CytoRescue (G-I). Cells are colored by sample type (A, D, G), cell type (B, E, H), and patient ID (C, F, I) to examine different aspects of biological structure. (J) Gene Ontology Biological Process terms enriched in the genes with greater than 2-fold expression change in CytoRescue-predicted profiles compared to raw snRNA-seq counts. Dot size represents enrichment ratio; color represents significance (−log □ □ (P value)) as determined by Fisher’s exact test. (K-M) Cell-cell communication networks identified from validation sample 8-CST-0892. Interaction weights across all signaling pathways are illustrated for scRNA-seq counts (K), snRNA-seq counts (L), and CytoRescue-predicted counts (M). (O-Q) Cell-cell communication interactions through the EGF signaling pathway identified from scRNA-seq counts (O), snRNA-seq counts (P), and CytoRescue-predicted counts (Q). We compared the expression profiles before and after the CytoRescue prediction. This comparison identified the genes exhibiting greater than 2-fold increased expression in tumor and normal lung samples across three cell types with more than 100 cells present in all analyzed samples (Figure S5). For these differentially expressed genes, we performed Gene Ontology (GO) based gene set enrichment analysis to identify overrepresented biological processes 23 . As anticipated, the most significantly enriched pathways across samples were associated with translation, mitochondrial energy production, and response to external stimuli ( Figure 2J ). Notably, the primary differences in enriched pathways among distinct cell types and samples were related to external stimulation responses, which is consistent with the diverse biological contexts of different cell types and sample conditions ( Figure 2J , Figure S5). Previous studies have demonstrated that cell-cell communications (CCC), particularly those mediated by ligand-receptor (L-R) interactions, are essential for coordinating multiple cell populations in lung tissue to regulate responses to stimuli 24 , 25 . To evaluate whether CytoRescue can restore CCC signals, we analyzed the validation lung sample 8-CST-0892, which was withheld from training. Following the standard CellChat pipeline, we inferred CCC between major lung cell types using raw scRNA-seq counts, snRNA-seq counts, and CytoRescue-predicted counts ( Figure 2K-Q ) 26 . Compared to CCC inferred from snRNA-seq counts, scRNA-seq data yielded significantly higher numbers of interactions ( Figure 2K-M ), consistent with previous findings 3 . Notably, previous studies have shown that epithelial cells dominate in the EGF-based L-R interaction networks 27 . However, interactions between alveolar epithelium and other cell types were predominantly lost in snRNA-seq data ( Figure 2P ), whereas these signals were successfully identified in both scRNA-seq data and CytoRescue-predicted counts ( Figure 2O, 2Q ). Collectively, CytoRescue successfully recovered essential biological signals in a dataset never encountered during training. Conclusions We present CytoRescue, a generative AI model designed to recover lost signals in single-nucleus RNA-sequencing experiments. We demonstrate that CytoRescue effectively restores lost cytoplasmic dynamics during nuclei isolation while preserving biological structures in both latent space and output representations. Additionally, CytoRescue employs a raw-in-raw-out architecture that enables seamless integration into existing single-cell RNA-seq analytical pipelines. When applied to the samples excluded from our training, CytoRescue could successfully restore cell-cell communication signals between alveolar epithelium and other cell types. While more evaluation work based on paired scRNA-seq and snRNA-seq datasets is needed, CytoRescue is demonstrate useful for detecting and evaluation of cytoplasmic gene expression signals from snRNA-seq data. Methods Paired scRNA-seq and snRNA-seq data sets Paired scRNA-seq and snRNA-seq datasets from brain and lung tissues were utilized in this study 3 , 4 . For brain samples, raw count data from single microglia cells were obtained from Gene Expression Omnibus (GEO) datatset GSE137444 (scRNA-seq) and GSE153807 (snRNA-seq) 4 . For lung samples, paired scRNA-seq and snRNA-seq count data from normal and tumor lung tissues were obtained from Zenodo (accession: 10.5281/zenodo.1120562) 3 . Cell type annotation and mapping Cell type annotation for the collected datasets was performed following a standard reference-based pipeline implemented in Seurat (v5.0.0) 28 . For brain snRNA-seq data, cell types were annotated by anchoring to a reference subsampled to a maximum of 2,000 cells per cell type from the SEA-AD reference 29 For lung samples, cell types were annotated using a reference subsampled to a maximum of 2,000 cells per cell type from HLCA (v1.0) 30 . Following annotation, cells in snRNA-seq data were mapped to corresponding cells in paired scRNA-seq data using scmap 15 . Model structure and training In VAE architectures, Gaussian distributions are commonly employed due to their differentiability during reparameterization 22 . Here, we designed the latent representation z to be Gaussian distribution, and the zero-inflated binomial parameters were inferred from input and output layers, effectively implemented ZINB modeling of single cell/nuclei data 6 , 22 . We first define negative binomial probability mass function as: To guide the model to learn from matched snRNA-seq and scRNA-seq data, ZINB distribution is defined: x denotes expression data, μ as mean, θ as the dispersion parameter and π as zero-inflation probability. We designed the loss function as: Where ℒ ziNB denotes the ZINB loss of observed count x, which is defined by: where 10 is the indicator function equals to 1 when x>0, otherwise 0. W KL is the loss weight parameter, while ℒ KL denotes the Kullback-Leibler (KL) divergence loss measured between approximate posterior (defined by μ as mean, and σ as variance) and standard Gaussian distribution in latent space by: Meanwhile W target is the loss weight parameter of target loss ℒ KL , which is difined by: For model training, we held out 2 samples from training as validation dataset for independent test. Further, the training datasets were shuffled and randomly selected 80% as training data while 20% as testing data. The model is trained on Ubuntu server equipped with 1 AMD EPYC 7742 64-Core Processor and 8 NVIDIA A100-SXM4-40GB GPUs, only 1 GPU was used during training and prediction. Evaluation We evaluated how CytoRescue recovers cytoplasmic gene expression dynamics while preserving biological structures in training, testing, and validation datasets through: 1) assessing concordance between predicted counts and paired scRNA-seq data, 2) comparing concordance between counts processed by denoising/imputation tools and paired scRNA-seq data, 3) evaluating clustering consistency between snRNA-seq counts and CytoRescue-predicted counts, and 4) comparing dimensionality reduction maps of all cells across raw snRNA-seq, scRNA-seq, and predicted count data. The consistency between different processing procedures is measured by Pearson’s correlation of the mean expression for each gene as: where X denotes the gene expression matrix in each sample. denotes mean expression of a specific gene g,N is the total number of cells. The clustering consistence is measured by adjusted Rand index (ARI) 19 . ARI, which is normalized mutual information (NMI) 18 and proportion of consistent (Consistency) label between unprocessed and predicted profiles, is calculated as follows: Where C * denotes the cell type annotation from CytoRescue’s predicted counts and C denotes the cell type annotation from raw snRNA-seq counts. N denotes total number of data points. N i,j denotes number of data points labeled as both cell types i in C * and j in C. N i denotes number of data points labeled as cell type j in . C where Then, we calculate where N i=j denotes number of cells with the same cell type in C * and C . And N denotes total number of data points in C * . The dimension reduction of cells was done by umap-learn package (v0.5.7) on log normalized read counts for each group. Gene set enrichment analysis The most recent Gene Ontology annotation and mapping files were downloaded from the Gene Ontology Consortium (GOC) 23 . Biological Process (BP) terms were used for pathway enrichment analysis of differentially expressed genes. Fisher’s exact test was employed to identify significantly overrepresented GO terms. Cell-cell communication analysis Raw read counts from scRNA-seq, snRNA-seq, and CytoRescue-processed data were normalized to a total count of 10,000 per cell and subsequently log-transformed. CellChat (v2) was then applied to the normalized read counts following the standard pipeline with all signaling pathways 26 . The number of connections and contribution of ligand-receptor (L-R) pairs in each circle plot were also inferred. Code availability The source code of CytoRescue and analysis code involved in this paper is deposited at GitHub and can be accessed. Data availability All the data are publicly available as we described in the Methods section. Declarations Competing interests Non declared. Funding This work was partially supported by National Institutes of Health (NIH) grants (U01AG079847, R01LM012806 and R03AG077191). We thank the technical support from the Cancer Prevention and Research Institute of Texas (CPRIT) funded UTHealth Cancer Genomics Core (CPRIT RP240610). The funders had no role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript. Supplementary Figures and Tables Figure S1 . Data processing workflow in this study. The snRNA-seq data retrieved from the literature were processed using standard filtering and cell type annotation procedures. All cell types from lung samples and microglia cells from brain samples were retained for scMAP mapping to paired scRNA-seq data. A snRNA-to-scRNA paired dataset was ultimately prepared for training, testing, and analysis. Figure S2 . Training process of CytoRescue. Training was terminated at 30 epochs when both training and validation losses converged after 25 epochs. Figure S3 . Log-normalized mean gene expression comparison between scRNA-seq and snRNA-seq, CytoRescue-predicted counts, scVI-imputed counts, and ALRA-imputed counts in lung samples. Figure S4 . Log-normalized mean gene expression comparison between scRNA-seq and snRNA-seq, CytoRescue-predicted counts, scVI-imputed counts, and ALRA-imputed counts in brain samples. Figure S5 . Gene Ontology Biological Process terms enriched in the genes with greater than 2-fold change in mean expression in CytoRescue-predicted counts compared to snRNA-seq counts in samples from (A) patient #8, and (B) patient #10. Figure S6 . ERBB2 expression in (A) scRNA-seq, (B) snRNA-seq, and (C) CytoRescue-predicted counts. Expression values are log 10 -normalized. Table S1 . Statistics and summary of training and validation datasets in this study. Table S2 . Pearson’s correlation coefficients of mean gene expression between scRNA-seq and snRNA-seq by three methods: CytoRescue-predicted, scVI-imputed, and ALRA-imputed counts. Acknowledgement We thank Dr. Xiaoqian Jiang for allowing us to access to GPU resources. Funder Information Declared National Institutes of Health , U01AG079847 , R01LM012806 , R03AG077191 Cancer Prevention and Research Institute of Texas, https://ror.org/0003xa228 , RP240610 Reference 1. ↵ Slyper , M. et al. A single-cell and single-nucleus RNA-Seq toolbox for fresh and frozen human tumors . Nat Med 26 , 792 – 802 ( 2020 ). OpenUrl CrossRef PubMed 2. ↵ Alvarez , M. et al. Human liver single nucleus and single cell RNA sequencing identify a hepatocellular carcinoma-associated cell-type affecting survival . Genome Med 14 , 1 – 20 ( 2022 ). OpenUrl CrossRef PubMed 3. ↵ Renaut , S. et al. Single-cell and single-nucleus RNA-sequencing from paired normal-adenocarcinoma lung samples provide both common and discordant biological insights . PLoS Genet 20 , e1011301 ( 2024 ). OpenUrl CrossRef PubMed 4. ↵ Thrupp , N. et al. Single-Nucleus RNA-Seq Is Not Suitable for Detection of Microglial Activation Genes in Humans . Cell Rep 32 , 108189 ( 2020 ). OpenUrl CrossRef PubMed 5. ↵ Cui , H. et al. scGPT: toward building a foundation model for single-cell multi-omics using generative AI . Nature Methods 2024 21:8 21 , 1470 – 1480 ( 2024 ). OpenUrl PubMed 6. ↵ Tian , T. , Min , M. R. & Wei , Z. Model-based autoencoders for imputing discrete single-cell RNA-seq data . Methods 192 , 112 – 119 ( 2021 ). OpenUrl PubMed 7. ↵ Gomari , D. P. et al. Variational autoencoders learn transferrable representations of metabolomics data . Communications Biology 2022 5:1 5 , 1 – 9 ( 2022 ). OpenUrl PubMed 8. ↵ Gayoso , A. et al. A Python library for probabilistic analysis of single-cell omics data . Nat Biotechnol 40 , 163 – 166 ( 2022 ). OpenUrl CrossRef PubMed 9. Kingma , D. P. & Welling , M. Auto-Encoding Variational Bayes . 2nd International Conference on Learning Representations, ICLR 2014 - Conference Track Proceedings ( 2013 ) doi: 10.61603/ceas.v2i1.33 . OpenUrl CrossRef 10. ↵ Grün , D. , Kester , L. & Van Oudenaarden , A. Validation of noise models for single-cell transcriptomics . Nature Methods 2014 11:6 11 , 637 – 640 ( 2014 ). OpenUrl PubMed 11. ↵ Choi , K. , Chen , Y. , Skelly , D. A. & Churchill , G. A. Bayesian model selection reveals biological origins of zero inflation in single-cell transcriptomics . Genome Biol 21 , 1 – 16 ( 2020 ). OpenUrl CrossRef PubMed 12. ↵ Risso , D. , Perraudeau , F. , Gribkova , S. , Dudoit , S. & Vert , J. P. A general and flexible method for signal extraction from single-cell RNA-seq data . Nature Communications 2018 9:1 9 , 1 – 17 ( 2018 ). OpenUrl PubMed 13. ↵ Love , M. I. , Huber , W. & Anders , S. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2 . Genome Biol 15 , 1 – 21 ( 2014 ). OpenUrl CrossRef PubMed 14. ↵ Carbajal , G. , Richter , J. & Gerkmann , T. Guided Variational Autoencoder for Speech Enhancement With a Supervised Classifier . ICASSP, IEEE International Conference on Acoustics, Speech and Signal Processing - Proceedings 2021-June , 681 – 685 ( 2021 ). 15. ↵ Kiselev , V. Y. , Yiu , A. & Hemberg , M. Scmap: Projection of single-cell RNA-seq data across data sets . Nat Methods 15 , 359 – 362 ( 2018 ). OpenUrl CrossRef PubMed 16. ↵ Gayoso , A. et al. A Python library for probabilistic analysis of single-cell omics data . Nat Biotechnol 40 , 163 – 166 ( 2022 ). OpenUrl CrossRef PubMed 17. ↵ Linderman , G. C. et al. Zero-preserving imputation of single-cell RNA-seq data . Nat Commun 13 , 1 – 11 ( 2022 ). OpenUrl CrossRef PubMed 18. ↵ Xuan Vinh , N. , Epps , J. & Bailey , J. Information Theoretic Measures for Clusterings Comparison: Variants, Properties, Normalization and Correction for Chance . Journal of Machine Learning Research 11 , 2837 – 2854 ( 2010 ). OpenUrl 19. ↵ Halkidi , M. , Batistakis , Y. & Vazirgiannis , M. Cluster validity methods: Part I . SIGMOD Rec 31 , 40 – 45 ( 2002 ). OpenUrl 20. ↵ Becht , E. et al. Dimensionality reduction for visualizing single-cell data using UMAP . Nat Biotechnol 37 , 38 – 47 ( 2019 ). OpenUrl CrossRef 21. Uniform manifold approximation and projection . Nature Reviews Methods Primers 2024 4:1 4 , 1 – 1 ( 2024 ). OpenUrl CrossRef 22. ↵ Lopez , R. , Regier , J. , Cole , M. B. , Jordan , M. I. & Yosef , N. Deep generative modeling for single-cell transcriptomics . Nat Methods 15 , 1053 – 1058 ( 2018 ). OpenUrl CrossRef PubMed 23. ↵ Consortium , T. G. O. et al. The Gene Ontology knowledgebase in 2023 . Genetics 224 , ( 2023 ). 24. ↵ Su , J. et al. Cell–cell communication: new insights and clinical implications . Signal Transduction and Targeted Therapy 2024 9:1 9 , 1 – 52 ( 2024 ). OpenUrl 25. ↵ Zhou , H. , Fan , E. K. & Fan , J. Cell-cell interaction mechanisms in acute lung injury . Shock 55 , 167 – 176 ( 2021 ). OpenUrl PubMed 26. ↵ Jin , S. et al. Inference and analysis of cell-cell communication using CellChat . Nat Commun 12 , 1 – 20 ( 2021 ). OpenUrl CrossRef PubMed 27. ↵ Raredon , M. S. B. et al. Single-cell connectomic analysis of adult mammalian lungs . Sci Adv 5 , ( 2019 ). 28. ↵ Hao , Y. et al. Dictionary learning for integrative, multimodal and scalable single-cell analysis . Nat Biotechnol 42 , 293 – 304 ( 2024 ). OpenUrl CrossRef PubMed 29. ↵ Gabitto , M. I. et al. Integrated multimodal cell atlas of Alzheimer’s disease . Nature Neuroscience 2024 27:12 27 , 2366 – 2383 ( 2024 ). OpenUrl CrossRef PubMed 30. ↵ Sikkema , L. et al. An integrated cell atlas of the lung in health and disease . Nature Medicine 2023 29:6 29 , 1563 – 1577 ( 2023 ). OpenUrl CrossRef PubMed View the discussion thread. Back to top Previous Next Posted August 21, 2025. Download PDF Supplementary Material Email Thank you for your interest in spreading the word about bioRxiv. NOTE: Your email address is requested solely to identify you as the sender of this article. Your Email * Your Name * Send To * Enter multiple addresses on separate lines or separate them with commas. You are going to email the following Cytoplasmic dynamics are overlooked in single nuclei RNA-seq but can be rescued by CytoRescue, a generative AI model to recover cytoplasm enriched gene Message Subject (Your Name) has forwarded a page to you from bioRxiv Message Body (Your Name) thought you would like to see this page from the bioRxiv website. Your Personal Message CAPTCHA This question is for testing whether or not you are a human visitor and to prevent automated spam submissions. Share Cytoplasmic dynamics are overlooked in single nuclei RNA-seq but can be rescued by CytoRescue, a generative AI model to recover cytoplasm enriched gene Wankun Deng , Zhongming Zhao bioRxiv 2025.08.15.670239; doi: https://doi.org/10.1101/2025.08.15.670239 Share This Article: Copy Citation Tools Cytoplasmic dynamics are overlooked in single nuclei RNA-seq but can be rescued by CytoRescue, a generative AI model to recover cytoplasm enriched gene Wankun Deng , Zhongming Zhao bioRxiv 2025.08.15.670239; doi: https://doi.org/10.1101/2025.08.15.670239 Citation Manager Formats BibTeX Bookends EasyBib EndNote (tagged) EndNote 8 (xml) Medlars Mendeley Papers RefWorks Tagged Ref Manager RIS Zotero Tweet Widget Facebook Like Google Plus One Subject Area Bioinformatics Subject Areas All Articles Animal Behavior and Cognition (7642) Biochemistry (17708) Bioengineering (13904) Bioinformatics (41992) Biophysics (21466) Cancer Biology (18618) Cell Biology (25531) Clinical Trials (138) Developmental Biology (13387) Ecology (19924) Epidemiology (2067) Evolutionary Biology (24337) Genetics (15615) Genomics (22521) Immunology (17749) Microbiology (40424) Molecular Biology (17194) Neuroscience (88673) Paleontology (667) Pathology (2839) Pharmacology and Toxicology (4827) Physiology (7650) Plant Biology (15160) Scientific Communication and Education (2046) Synthetic Biology (4302) Systems Biology (9826) Zoology (2271)