SmartImpute is a targeted imputation framework for single-cell transcriptome data

Summary

In this study, we present SmartImpute, a targeted imputation framework for single-cell RNA sequencing (scRNA-seq) data. SmartImpute focuses on predefined marker genes to improve biological relevance and computational efficiency. It employs a modified generative adversarial imputation network (GAIN) with a multi-task discriminator that imputes missing values while preserving true biological zeros. We applied SmartImpute to scRNA-seq datasets from head and neck squamous cell carcinoma, human bone marrow, and lung cancer. Across these datasets, SmartImpute improved clustering, cell type annotation, and trajectory inference and successfully scaled to datasets with over one million cells. In addition, SmartImpute was applied to spatial transcriptomics data, where it improved spatial gene expression patterns and clustering. These results demonstrate that SmartImpute facilitates deeper insights into cellular heterogeneity and disease progression.

Graphical abstract

Highlights

• •Target imputation framework for scRNA-seq data
• •Multi-task GAIN architecture performs efficient imputation and keeps biological zeros
• •GPT-based function for customizable marker gene panel selection
• •SmartImpute improves downstream analyses and can scale to large datasets

Motivation

Single-cell RNA sequencing (scRNA-seq) has revolutionized our understanding of cellular heterogeneity and tissue complexity. However, high dropout rates in scRNA-seq data complicate downstream analyses, such as cell type annotation and trajectory inference. Existing imputation methods are computationally expensive and may produce artificial noise by filling in values that should remain as true biological zeros. Our aim was to develop an efficient, biologically relevant imputation strategy that focuses on informative marker genes, preserves meaningful zeros, and can scale to large scRNA-seq and spatial omics datasets.

Yao et al. propose SmartImpute, a target imputation framework for scRNA-seq data. They demonstrate that the integration of marker-gene-focused imputation with multi-task GAIN enables efficient and biologically relevant imputation in a range of single-cell modalities, including multiple cancer types and spatial transcriptomics.

Introduction

Over the past decade, the rapid advancement of single-cell RNA sequencing (scRNA-seq) has revolutionized our understanding of cellular heterogeneity and tissue transcriptomic complexity. By delivering high-resolution gene expression data at the individual cell level, scRNA-seq has enabled the in-depth study of cellular subtypes, lineage relationships, and dynamic biological processes with unparalleled detail. However, a significant challenge in scRNA-seq data is the high frequent occurrence of dropout events, where gene expression measurements are recorded as zero for several reasons: (1) low amounts of mRNA in individual cells, (2) technical or sequencing artifacts, and (3) inherent cell type differences, with some cells exhibiting low expression levels for certain genes. These dropouts complicate downstream analyses, such as cell type identification,^1,2 differential expression analysis, and trajectory inference,^3,4,5 therefore necessitating imputation methods to infer the missing values before proceeding with downstream scRNA-seq analysis.

Existing imputation methods, such as SAVER,⁶ SAVER-X,⁷ DCA,⁸ and DeepImpute,⁹ have made significant progress in addressing the dropout problem. SAVER uses a Bayesian framework, assuming a Poisson-gamma mixture model to estimate gene expression levels, while DCA employs a zero-inflated negative binomial noise model within an autoencoder framework.^10,11 DeepImpute, on the other hand, utilizes a deep neural network approach to impute target genes. SAVER-X combines a deep autoencoder with a Bayesian model and extracts transferable gene-gene relationships across different datasets and even different species to denoise new target datasets. Despite their advancements, these methods face several limitations that restrict their applicability in daily RNA-seq or large-scale studies. They often impose a high computational cost, risk overfitting and producing biased results, and require users to manage larger imputed matrices compared to sparse ones. Additionally, in many real data analyses, imputation may offer limited improvement in downstream analyses, such as cell typing.^12,13,14 These issues underscore an urgent need for more efficient and biologically relevant imputation methods to address the challenges posed by dropout events in scRNA-seq data.

Building on these earlier approaches, a number of recent imputation tools have been developed that combine probabilistic models, graph-based structures, and adversarial learning to improve both accuracy and scalability. For example, VAE-based tools like scVI¹⁵ model gene expression noise using Bayesian frameworks and low-dimensional representations to account for dropout. Graph-based methods such as GraphSCI¹⁶ and scGNN¹⁷ use cell-cell relationships to smooth expression profiles across similar cells. Ensemble methods like SCRABBLE¹⁸ integrate outputs from multiple models to better balance sensitivity and specificity across different dropout levels. Adversarial tools like scIGANs¹⁹ and scSemiGAN ²⁰ have shown strong performance in recovering missing data by training with a generative adversarial approach. While these methods help reduce bias and computational load, they often face challenges when scaled to large datasets or when trying to preserve biologically meaningful patterns needed for downstream analysis.

In this study, we present SmartImpute, a computational framework specifically designed for targeted scRNA-seq data imputation. SmartImpute is available for public access on GitHub (https://github.com/wanglab1/SmartImpute). SmartImpute focuses on a predefined set of marker genes most informative for cellular behavior. This approach is based on observations in cancer genomics data analysis, where, despite the heterogeneity in both cancer and immune cells, most cell types and subtypes can be identified using fewer than 1,000 genes.^21,22,23 This targeted strategy not only improves the biological relevance of the imputation process but also enhances computational efficiency by limiting the imputation to the most informative subset of genes. Moreover, SmartImpute leverages a modified generative adversarial network (GAN) architecture,²⁴ known as a generative adversarial imputation network (GAIN),²⁵ to accurately distinguish between true biological zeros and missing values. By integrating a multi-task discriminator,^26,27 SmartImpute prevents overfitting and maintains the integrity of biologically relevant zeros, thereby ensuring high imputation accuracy.

We demonstrate the performance and efficiency of SmartImpute using scRNA-seq data from head and neck squamous cell carcinoma (HNSCC) and human bone marrow. Our results show that SmartImpute significantly enhances the visualization of cell type clusters, improves cell type annotation, and preserves biologically meaningful gene expression patterns. Benchmarking against other imputation methods highlights SmartImpute’s superior performance in terms of both accuracy and computational efficiency. Overall, SmartImpute provides a lightweight and more realistic solution for addressing dropout events in scRNA-seq data. By focusing on targeted gene panels and employing an advanced GAN architecture, SmartImpute enhances the biological interpretability of single-cell data, facilitating deeper insights into cellular heterogeneity and disease progression.

Results

Targeted marker gene panel approach

SmartImpute (https://github.com/wanglab1/SmartImpute) introduces a targeted approach to the imputation of scRNA-seq data by selectively focusing on a predefined set of marker genes. This approach offers two key benefits. First, it improves the biological relevance of the imputation process by concentrating on genes important for understanding cellular behavior. Second, it makes the computation more efficient by only imputing on the most informative gene subset of the data. Starting with a group of 580 well-established marker genes (BD Rhapsody Immune Response Targeted Panel [Human]),²⁸ researchers have the flexibility to adjust a customized panel to fit the needs of their specific projects. This customization enables a tailored imputation that not only aligns with specific research objectives but also embeds biological insights directly into the imputation process, setting it apart from methods such as DeepImpute that may overlook biological significance. To aid in the selection of an effective gene panel with specific information about data, we offer an R package (https://github.com/wanglab1/tpGPT) that utilizes a generative pre-trained transformer (GPT) model²⁹ integrated with the SmartImpute framework (STAR Methods).

Modified GAN architecture for imputation

SmartImpute employs a modified version of GANs (STAR Methods) to address the imputation challenge in scRNA-seq. The process begins with data preparation by identifying the target gene panel in the gene expression data. The training process is refined by incorporating a proportion of non-target genes, without losing information from the dataset and ensuring robust model generalizability. SmartImpute uses a multi-task discriminator to improve the imputation accuracy and the proper identification of biological zeros (Figure 1). This modified generative model allows SmartImpute to impute values without making any assumptions about the underlying data distribution. Consequently, the imputed values produced by SmartImpute exhibit greater accuracy compared to other methods, such as SAVER, which relies on a Poisson-gamma mixture model, and DCA, which utilizes an autoencoder model.

Figure 1.

Overview of SmartImpute

SmartImpute uses a targeted approach for imputing missing values in scRNA-seq data. It adapts the GAIN architecture, adding a multi-task discriminator to distinguish between biological zeros and missing values.

Performance of SmartImpute in HNSCC normalized data

To evaluate SmartImpute’s performance in normalized data, we conducted a comparison of UMAP clustering³⁰ using HNSCC data³¹ with and without imputation. The annotation provided by a previous study was used as a cell type true label (STAR Methods). The visual comparison demonstrates SmartImpute’s ability to distinguish between different cell types, which is not feasible without imputation. Particularly, cell types such as CD4 conventional T (CD 4 Tconv), CD8 exhausted T, and CD8 conventional T (CD8 Tconv) cells are indistinguishable without imputation. Moreover, SmartImpute is able to separate closely related cell types like myocytes, fibroblasts, and myofibroblasts while preserving their biological distinctiveness (Figure 2A).

Figure 2.

Performance of SmartImpute in HNSCC normalized data

(A) UMAP clustering without imputation (left) and with SmartImpute (right) with a true label.

(B), Gene expression heatmap for cell type-associated marker genes without imputation (left) and with SmartImpute (right).

(C) SingleR cell type prediction accuracy in fibroblasts.

(D) Comparison with different imputation methods: Jaccard index (left) and running time (right).

Next, we demonstrate the accuracy of SmartImpute’s imputation by comparing heatmaps of gene expression with and without imputation. Ideally, marker genes should be explicitly highly expressed by corresponding cell types they present (the diagonal blocks), with zero or low expression in other cell types (the off-diagonal blocks).³² In the heatmap without imputation, due to the dropouts, the diagonal blocks shows a sparse expression pattern. In contrast, after applying SmartImpute, the heatmap shows significant improvement, with the imputation process filling in the missing values in diagonal blocks and producing more complete and continuous expression patterns across all cell types. Specifically, the marker gene blocks are distinctly filled, particularly for T cells (CD4 Tconv, CD8 exhausted, and CD8 Tconv), showcasing well-defined clusters of gene expression. Genes important for specific cell types, such as CCR7/CXCR4 for Tconv cells and ICOS for regulatory T cells, show more consistent and biologically relevant expression patterns after imputation. Moreover, the off-diagonal blocks remain empty after imputation, indicating that SmartImpute prevents overfitting without generating false positives (Figure 2B).

To assess how SmartImpute influences downstream analyses such as cell type prediction, we employed SingleR³³ with a BLUEPRINT cell type reference,³⁴ which contains most cell types present in this dataset. We compared the prediction with and without imputation against the true labels. Imputation consistently improved prediction accuracy, with a notable enhancement observed in fibroblasts. Without imputation, SingleR correctly identifies only 57.3% of true fibroblasts, whereas SmartImpute increases this accuracy to 86.5% (Figure 2C). Additional results highlighting improvements in cell typing across other cell types are provided in Figure S1A. Similarly, enrichment scores of cell type-specific marker sets calculated by AUCell³⁵ further demonstrate SmartImpute’s effectiveness in enhancing cell typing accuracy, particularly for malignant cells (Figure S1B).

To highlight the advantage of SmartImpute’s targeted imputation approach, we benchmarked it against other imputation methods in terms of both accuracy and efficiency. In the accuracy comparison, SmartImpute achieves the highest Jaccard index, scoring 0.67 and outperforming DCA (0.637), DeepImp (0.649), and SAVER (0.623) (Figure 2D). UMAP plots generated from data imputed by these methods are shown in Figure S1C. To illustrate SmartImpute’s computation efficiency, we compare running times as the number of genes increases. SAVER is unable to process datasets with more than 5,000 genes, while SmartImpute consistently maintains the shortest running time compared to the other methods (Figure 2D).

We next performed a series of simulations to evaluate SmartImpute’s ability to recover true expression values and preserve cell type structure under varying levels of dropout. Using scDesign2,³⁶ we generated datasets of 10,000 cells by 10,000 genes and introduced “technical” zeros at rates of 20%, 40%, 60%, and 80% by randomly masking non-zero entries. Across all dropout levels, SmartImpute consistently improved clustering accuracy compared to raw data (Figure S2A). To examine the multi-task discriminator’s behavior, we tracked the dynamic losses of D₁ and D₂ at a moderate 50% dropout (Figure S2B). Finally, we conducted paired Wilcoxon signed-rank tests on Jaccard indices across simulations. SmartImpute outperformed all others with p < 1 × 10⁻¹⁰ in every pairwise comparison (Figure S2C).

Performance of SmartImpute in HNSCC UMI count data

Next, we evaluate SmartImpute’s performance under a more complicated scenario using unique molecular identifier (UMI) count data. Unlike normalized data, the raw count is not adjusted for biases such as sequencing depth, batch effects, and other technical artifacts, making imputing raw count data more challenging.^37,38 To assess SmartImpute’s performance on non-normalized data, we use raw UMI single-cell count data from immune cells extracted from 26 HNSCC tumors³⁹ (STAR Methods).

Consistent with the results obtained from normalized data, we demonstrate SmartImpute’s efficacy in resolving different cell types into coherent clusters by UMAP clustering (Figure 3A). After imputation, we observe a better separation between myeloid lineage (monocytes, macrophages, and dendritic cells) with B cells and plasma cells. Within the myeloid lineage, the CD14⁺ and CD16⁺ monocytes are distinctly separated from each other, and two clear trajectories are observed: one from monocytes to dendritic cells and one from monocytes to macrophages. Furthermore, to evaluate SmartImpute’s impact on prediction of T cell subtypes, we applied Azimuth,⁴⁰ a reference-based mapping and annotation method, using publicly available peripheral blood mononuclear cell (PBMC) data as in Hao et al.⁴⁰ The CD8 T cell subtypes—CD8 TExh (exhausted), CD8 TpreExh (pre-exhausted), and CD8 TEM (effector memory), play distinct roles in immune response modulation but have closely related phenotypic and functional characteristics, making them challenging to differentiate in scRNAseq data analysis. The pre-exhausted CD8T cells, characterized by co-expression of the native marker TCF7 and exhaustion marker TOX,^41,42 are particularly prone to misclassification. In the PBMC reference, the closest annotation for these three CD8T subtypes is CD8 TEM (PAN) cells due to their shared effector functions and memory characteristics. Without imputation, a significant number of CD8 TExh cells are misclassified as CD8 naive, and roughly one-fourth of CD8 TpreExh and CD8 TEM cells are incorrectly labeled as CD4, CD8 TCM, or other T cells (Figure 3C). After imputation, the prediction accuracy of Azimuth is significantly improved. Nearly all CD8 TExh cells are correctly identified, with misclassifications of CD8 TpreExh and CD8 TEM to other cell types reduced to one-third of the original rate. To quantitively measure this improvement, we calculate the F1 score for cell typing accuracy before and after imputation. SmartImpute improves the F1 score from 0.61 to 0.73 for CD8 TExh, from 0.63 to 0.71 for CD8 TpreExh, and from 0.62 to 0.75 for CD8 TEM. Notably, other cell types also show a general improvement, with the F1 score increasing from 0.71 to 0.75 after imputation (Figure 3D).

Figure 3.

Performance of SmartImpute in HNSCC UMI count data

(A) UMAP clustering without imputation (left) and with SmartImpute (right) with a true label.

(B) Cell type prediction for CD8 T cells without imputation (left) and with SmartImpute (right).

(C) Cell type prediction accuracy measured by F1 score.

SmartImpute improves classification of T cell subtypes

Furthermore, we apply SmartImpute to a more challenging problem: distinguishing closely related T cell subtypes. These subtypes are biologically similar, making it difficult to distinguish them in downstream analyses.^43,44,45 We demonstrate SmartImpute’s effectiveness in separating these T cell subtypes using scRNAseq data of purified T cells from HNSCC tumors and site-matched inflamed non-malignant tissues.⁴⁶ The clarity of cluster distinction among various T cell populations is significantly enhanced with SmartImpute. This improvement is evident in a better separation between CD4T and CD8T cells as well as in the clear trajectories observed from naive to regulatory cells in CD4T lineage and from effector to exhausted cells in CD8T lineage (Figure S3A). Quantitative improvements in T cell subtype clustering are also evident, with the Jaccard index increasing from 0.48 to 0.63 and the adjusted Rand index (ARI) increasing from 0.46 to 0.65 after SmartImpute (Figure S3B).

Mapping spatial architecture with SmartImpute

We applied SmartImpute to a 10× Visium spatial transcriptomics dataset of the human dorsolateral prefrontal cortex (DLPFC),⁴⁷ using the top 500 spatially variable genes as a target panel. After imputation, the separation between cortical layers became much clearer (Figure 4A). Spatial autocorrelation also improved significantly, with the average Moran’s I increasing from 0.01 to 0.03 (Wilcoxon p = 1.1 × 10⁻¹⁰). This stronger spatial structure led to better detection of layer-specific gene expression, particularly between layers 1 and 3. The three genes showing the largest gains in Moran’s I, LRIG1, CALD1, and CDK14 (Figure S3C), are all known to mark distinct cortical features. LRIG1 labels oligodendrocyte-lineage cells in white matter and layer 6,⁴⁸ CALD1 marks vascular cells in layer 4,⁴⁹ and CDK14 is expressed in proliferative cells in layers 2 and 3.⁵⁰ By restoring the spatial patterns of these informative genes, SmartImpute enhances both global clustering and fine-scale layer organization, which are otherwise difficult to detect.

Figure 4.

Performance of SmartImpute in Human DLPFC Visium data and human bone marrow data

(A) Spatial clustering of DLPFC cortical layers for sample 151,673 with ground truth, no imputation, and SmartImpute.

(B) UMAP clustering of human bone marrow data and cellular trajectories defined with SmartImpute.

Unraveling cellular trajectories with SmartImpute

The targeted imputation approach in SmartImpute is further demonstrated using bone marrow single-cell data generated with a custom gene panel with 462 genes.⁵¹ The original study utilized both mRNA and protein surface markers to aid cell clustering and trajectory analysis. In contrast, our goal is to reach comparable levels of performance in downstream analysis using only mRNA data imputed by SmartImpute. After imputation, the UMAP plot clearly shows how hematopoietic stem cells (HSCs) and multipotent progenitors differentiate into various cell lineages. This includes the trajectories from HSCs to blood cells, B cells, and natural killer cell progenitors. Additionally, the progression of pro-B cells into more mature B cell stages is clearly structured on UMAP. Furthermore, the imputed data reveal the potential differentiation pathways of monocytes into various types of dendritic cells, presenting a comprehensive picture of immune cell development. (Figure 4B). In addition, we evaluated how SmartImpute recovers gene dynamics along inferred trajectories. Pseudotime was estimated using Monocle3.⁵² For each gene, we calculated Spearman and Pearson correlations between gene expression and pseudotime before and after imputation. Genes were then ranked by the absolute improvement in correlation. Many of the top-ranked genes—those with the largest gains in pseudotime association, such as MDK, IL2RA, and DLGAP5—showed no clear pattern in the original data but became strongly linked to pseudotime after imputation (Figure S4).

Scalability and performance on million-cell datasets

To demonstrate SmartImpute’s scalability, we applied it to the Lung Cancer Atlas from CELLxGENE (https://cellxgene.cziscience.com/), comprising 1,283,972 single cells. Using the same 580-gene immune response panel, we generated UMAP embeddings on both the original data and the SmartImpute output. Imputation was completed in 2.6 h on a single GPU (NVIDIA A30, peak memory 75 GB). The ARI increased from 0.65 to 0.79 after imputation (Figure S3D).

Collectively, our evaluation of SmartImpute’s performance reveals its capability to impute scRNA-seq data across various datasets. By customizing the gene panel selection process and integrating GAIN architecture, SmartImpute consistently delivers highly accurate results that have significant implications for the interpretation and analysis of single-cell data.

Discussion

The targeted imputation approach employed by the proposed SmartImpute offers significant advantages in enhancing both the biological relevance and computational efficiency of scRNA-seq data analysis. By focusing on a predefined set of marker genes, SmartImpute aligns the imputation process with specific biological contexts, ensuring that the imputed values are meaningful for understanding cellular behavior. This strategy is particularly beneficial for studies where certain genes are known to be critical for distinguishing between cell types, subtypes, or states. For example, in the single-cell sequencing datasets evaluated in this study, SmartImpute exhibited superior or comparable performance in improving cell type annotation and clustering accuracy relative to other imputation methods. The use of a targeted gene panel not only improves the interpretability of the results but also significantly reduces the computational burden, allowing for faster processing and scalability to larger datasets. Another benefit of this approach is that, since only a subset of genes is imputed, the risk of overfitting is minimized, and data integrity is better preserved compared to full-panel imputation.

Preserving biological zeros in scRNA-seq data is crucial for accurate biological interpretation,¹⁷ as it maintains the integrity of data where gene non-expression is as significant as expression. SmartImpute’s multi-task discriminator design in GAIN architecture effectively preserves these zeros, avoiding the complete imputation for every dropout typical of other methods, which can lead to incorrect conclusions. For example, in studies like the differentiation of cells, where specific gene absence is indicative of biological states, SmartImpute’s capability to maintain true zeros allows for precise tracking of gene expression changes without false imputation. This precision is essential for robust biological conclusions and enhances the understanding of complex cellular functions, making SmartImpute a valuable tool in genomic research and precision medicine.

The integration of the GPT model into SmartImpute marks a significant enhancement in refining gene panel selection for targeted imputation approaches. The GPT model has already been used in cell type annotation,⁵³ demonstrating its utility by providing robust and accurate estimation in scRNA-seq analysis. The potential integration of the GPT model with SmartImpute can extend far beyond cell type annotation and gene panel selection, such as harmonizing terminologies and reconciling discrepancies in nomenclature across different datasets and studies. This is particularly important in genomics, where researchers may use different terms or abbreviations to describe the same cell types, or when identifying new subtypes where standardized naming conventions have not yet been established. By recognizing and reconciling various naming conventions for cell types and genomic markers, the GPT model can unify disparate terms into a consistent nomenclature. This standardization allows researchers to more easily integrate and compare data from diverse studies, as all names are converted to a uniform format suitable for coding and analysis. This capability not only streamlines workflows but also enhances the reproducibility and comparability of genomics research.

Despite its promising strengths, the targeted approach also has limitations that must be considered. One potential drawback is the reliance on prior biological knowledge to select the marker genes, which may introduce biases if the selected panel does not capture all relevant genes for a given study. This could result in the omission of important biological signals and potentially impact downstream analyses. Moreover, the predefined gene panels may need to be frequently updated to reflect new discoveries in the field, adding an additional layer of complexity to the imputation process. To encourage reproducibility, we have provided a GPT model for automatic gene panel selection, which partially addresses the issue by enabling the customization of gene panels. However, it is essential to rigorously validate the selected genes to ensure their relevance and comprehensiveness for specific research contexts and experimental designs. Additionally, it is important to conduct further research to identify more informative biomarkers or panels for different cell types or cell states, thereby enhancing the overall effectiveness and applicability of targeted imputation in various scRNA-seq studies.

The same targeted imputation method can also be extended to spatial omics data, such as NanoString CosMx⁵⁴ and 10× Xenium,⁵⁵ which often contain many missing values. It is important to note that many spatial omics panels are targeted by nature, and with larger panels, the proportion of missing values tends to increase. By applying SmartImpute’s targeted approach to spatial transcriptomics, researchers can achieve similar improvements in data quality and interpretability, enabling deeper insights into the molecular mechanisms associated with the spatial organization of tissues.

Emerging single-cell foundation models, such as scFoundation⁵⁶ and scGPT,⁵⁷ offer promising advantages for imputation in single-cell transcriptomics. These methods were trained on large datasets to capture broad transcriptional patterns and to generate biologically meaningful embeddings, making them well-suited for transferring knowledge across datasets. However, foundation models are often resource-intensive to train and fine-tune, and their ability to generalize may be affected by batch effects or differences in experimental protocols. Additionally, because these models perform imputation implicitly through learned representations rather than direct value prediction, they can lack transparency, making it difficult to interpret how individual expression values are inferred or to assess the quality of the imputation.

Limitations of the study

While SmartImpute demonstrates strong performance across different datasets, several limitations remain. First, its performance depends on the proper selection of marker genes, which are expected to capture key biological variation relevant to the analysis task. While the integrated GPT-based tool facilitates marker selection, its reliability decreases when generating large panels (more than 400 genes) and is limited by the scope and accuracy of the GPT model’s training data. This may lead to inclusion of suboptimal or irrelevant genes. Incorporating cell type specificity and coverage analysis may help refine and validate the selected gene panels for improved robustness. Second, although SmartImpute can be applied to spatial transcriptomics data, it does not incorporate spatial coordinates or tissue morphology. This limits its ability to fully capture spatial organization and cell-cell interactions in spatial transcriptomics data. As a result, while SmartImpute improves clustering and annotation in spatial data, it fails to recover the two narrow cortical layers (L2 and L4) present in the ground truth (Figures 4A and 4B). Future work is needed to extend the model by integrating spatial architectures to better capture tissue organization.

Resource availability

Lead contact

Requests for further information and resources should be directed to the lead contact, Xuefeng Wang (xuefeng.wang@moffitt.org).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •All datasets used in this study have been summarized in the key resources table.
• •The code for SmartImpute is public available on GitHub (https://github.com/wanglab1/SmartImpute) and Zenodo (https://doi.org/10.5281/zenodo.15758184). The code for tpGPT is public available on GitHub (https://github.com/wanglab1/tpGPT) and Zenodo (https://doi.org/10.5281/zenodo.15758205).
• •Any additional information needed to re-analyze the results reported in this paper is available from the lead contact upon request.

Acknowledgments

This work has been supported in part by a National Institutes of Health grant (R01DE030493 to X.W.), a National Cancer Institute grant (CA274489 to X.Y.), a T32 training grant from the National Institutes of Health (T32CA23339 to X.W.), and Biostatistics and Bioinformatics Shared Resources at the H. Lee Moffitt Cancer Center and Research Institute, an NCI-designed Comprehensive Cancer Center (P30-CA076292).

Author contributions

Conceptualization, X.W., X.Y., and S.Y.; data curation, X.Y., T.L., and S.Y.; method development, S.Y. and X.W.; intellectual contributions to study design and review of results, J.T.D. and T.I.S.; writing – original draft, S.Y., X.W., and X.Y.; funding acquisition, X.W. and X.Y. All authors reviewed and edited the manuscript.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Deposited data
HNSCC tumor normalized data	GEO: GSE103322, Yu et al.32	https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE103322
HNSCC tumor-infiltrated immune cells count data	GEO: GSE139324, Yu et al.32	https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE133324
HNSCC tumor-infiltrated T cells count data	GEO: GSE163633	https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE163633
Human DLPFC 10x Visium spatial transcriptomics datasets	Mao et al.48	https://research.libd.org/spatialLIBD/
Human bone marrow data	Triana et al.51	https://figshare.com/articles/dataset/Expression_of_97_surface_markers_and_RNA_transcriptome_wide_in_13165_cells_from_a_healthy_young_bone_marrow_donor/13397987
Lung Cancer Atlas (LuCA)	CELLxGENE	https://cellxgene.cziscience.com
Software and algorithms
SmartImpute	This paper	https://github.com/wanglab1/SmartImpute and https://doi.org/10.5281/zenodo.15758184
tpGPT	This paper	https://github.com/wanglab1/tpGPT and https://doi.org/10.5281/zenodo.15758205
SingleR	Aran et al.33	https://github.com/dviraran/SingleR
AUCell	Aibar et al.34	https://github.com/aertslab/AUCell
Seurat	Satija lab	https://satijalab.org/seurat/
scDesign2	Sun et al.36	https://github.com/JSB-UCLA/scDesign2
Monocle3	Cao et al.52	https://github.com/cole-trapnell-lab/monocle3
SAVER	Huang et al.6	https://github.com/mohuangx/SAVER
DCA	Eraslan et al.8	https://github.com/theislab/dca
DeepImpute	Arisdakessian et al.9	https://github.com/lanagarmire/DeepImpute

Method details

SmartImpute overview

SmartImpute employs a targeted imputation approach for single-cell RNA sequencing data, using a select panel of marker genes essential for identifying distinct cell types and states. This approach leverages prior biological knowledge to enhance imputation accuracy and computational efficiency. SmartImpute adapts a modified Generative Adversarial Network (GAN) architecture, specifically tailored for imputation tasks, known as GAIN. The original GAIN framework is limited in scRNA-seq applications where dropouts are represented as zero values, lacking the capability to differentiate between biological zeros and missing data. To address this, SmartImpute incorporates an additional discriminator aiming for distinguishing between the biological zeros and missing values in scRNA-seq data. In the following, we will discuss (1) the implementation of target imputation approach, (2) the multi-task discriminator for imputation in scRNA-seq data (3) target panel genes selection, (4) training details and (5) evaluation metrics.

Targeted imputation with attention mechanism

SmartImpute improves accuracy and computational efficiency by incorporating biological information through its targeted imputation approach. This approach is implemented in the generator (G) by incorporating an attention mechanism. For clarity, the following notation is introduced: the expression matrix of single cell data is denoted by a $P$ by $N$ matrix that $X=\left({\mathit{x}}_1,\dots ,{\mathit{x}}_N\right)$, ${\mathit{x}}_i={\left(x_{i1},..,x_{iP}\right)}^{\prime }$. Without losing generality, the first $q$ genes are designated as the target genes. In the targeted imputation setting, only dropouts in ${\mathit{x}}_i$ where $i\le q$ will be imputed.

An attention parameter $\gamma$ is incorporated to refine the generating process. The attention parameter $\gamma$ takes range between 0 and 1, allowing for the integration of additional genes from outside the target panel to aid in the generation of missing values. During model training, each batch is composed of target genes ${\mathit{x}}_1$ to ${\mathit{x}}_q$, and a supplementary set of $\left[\gamma ·\left(p-q\right)\right]$ randomly chosen genes with [$·$] denoted as the ceiling function. In SmartImpute, the generator G consists of K subnets, denoted as $G_k,k=1,\dots ,K\text{.}$ During the training of the k-th batch, the subnet $G_k$ functions by taking the expression matrix $X$, random noise Z, and attention parameter $\gamma$ as inputs:

$$X^{G_k}=G_k\left(X,\left(1-M\right)⨀Z,\gamma \right)\text{.}$$

Here $M$ is the mask matrix which indicating the dropout events with: $m_{ij}=1$ for $x_{ij}\ne 0$ and $m_{ij}=0$ otherwise. The term $\left(1-M\right)⨀Z$ ensures that the generator only incorporates random noise corresponding to the positions of dropouts during the training process. Specifically, scRNA-seq data used in $G_k$ includes not $P$ genes, but $q+\left[\gamma ·\left(p-q\right)\right]$ genes. The final output of generator G is the average of all subnets output $X^{G_k}$: $X^G=\frac{1}{K}\sum _{k=1}^K{X}^{G_k}$.

Biological zero recognition with multi-task discriminator

Although the target imputation approach already considers the issue of overfitting, SmartImpute further prevents it by adding a multi-task discriminator. Some Imputation methods such as DCA and DeepImpute consider every dropout as a missing value and impute expression values for all genes in every cell. While SAVER applies a lasso penalization to preserve biological zeros but identifies relatively few of them. To address this challenge, we introduce a biological zero recognition task with SmartImpute using a multi-task discriminator.

The multi-task discriminator functions to address two specific tasks. The first discriminator $D_1$ serves a similar role to the original discriminator in the GAIN framework, degerming whether elements of $X^G$ are real or imputed. To accomplish the task, $D_1$ takes the mask matrix M of original data as one of its inputs. The primary objective of $D_1$ is to accurately reproduce the mask matrix M from the imputed dataset $X^G$. While $D_1$ can distinguish between real and imputed data, it will impute expression values of target genes of non-related cell types. To address this issue and prevent the overfitting of imputations on biological zeros, an additional discriminator $D_2$ is incorporated in the framework.

The second discriminator $D_2$, addresses a unique challenge in scRNA-seq data by determining whether the observed dropouts represent true zeros, indicating no gene expression, or missing values due to technical limitations. Different from $D_1$, $D_2$ employs a specifically designed missing assignment matrix $M^{D_2}$ to allocate missing zeros for the imputed data with $X^{G^{\prime }}=X^G⨀M^{D_2}$. To enhance the identification of biological zeros, $D_2$ incorporates a threshold value $\tau$ which is used to classify gene expression values. To distinguish biological zeros from missing values, we model the residual imputation noise at true biological zeros as following a Uniform [0,1] distribution on the raw UMI scale, with $\sigma =\sqrt{\frac{1}{12}}$ . This translates to a threshold of $\tau =\sigma \times \ln 2\approx 0.20$ on the log-transformed UMI scale. We set $\tau$ = 0.2 as the default threshold, though users can choose a more conservative threshold. Expression values below $\tau$ are considered biological zeros: $x_{ij}^{D_2}=0$ if $x_{ij}^G\le \tau$. Otherwise, it keeps its value from $X^{G^{\prime }}$: $x_{ij}^{D_2}=x_{ij}^{G^{\prime }}$. To prevent $D_2$ from disrupting $D_1$’s training, we implement a simple warm-up period: for the first T₁ epochs, we train only $D_1$ against the generator. We then activate $D_2$, which allows the generator to first learn the overall data distribution before $D_2$ begins identifying biological zeros. This approach ensures that $X^{D_2}$ contains the true biological zeros information even when there are no zeros in the imputed data $X^G$. The mask matrix for biological zero recognition $M^{X^{G^{\prime }}}$ is further updated with $m_{ij}^{X^{G^{\prime }}}=1$ if $x_{ij}^G\le \tau$, and $m_{ij}^{X^{G^{\prime }}}=m_{ij}^{D_2}$, otherwise.

Double hint mechanism

Since two discriminators $D_1$ and $D_2$ makes discriminations on different input data, a double hint mechanism is used in SmartImpute framework. Following the hint mechanism in GAIN, each discriminator in SmartImpute will be assigned with distinct hint matrices. By providing partial information about the missing entries, the hint mechanism guides the discriminator toward a more well-defined and unique optimal solution. This refinement in the optimization landscape enhances the robustness by the convergence of training process and subsequently improves the quality of the imputations.

For $D_1$, the hint matrix $H_1$ is derived from the mask matrix of the original data. $H_1$ provides part of information by indicating which values in the data matrix are observed versus those that are missing:

$$H_1=B_1⨀M+0.5\left(1-B_1\right)\text{.}$$

Where $B_1$ is a random variable taking the value from {0,1}. The hint matrix is used as one of the inputs for discriminator, and $D_1$ is constructed as a function of $X^G,H_1$ and $M$: $Y^{D_1}=D_1\left(X^G,H_1,M\right)$.

For the discriminator $D_2$, the corresponding hint matrix $H_2$ is defined by $M^{X^{G^{\prime }}}$,

$$H_2=B_2⨀M^{X^{G^{\prime }}}+0.5\left(1-B_2\right)\text{.}$$

$B_2$ performs the same function as $B_1$ for a random assignment of where to give hint and $D_2$ is the function of $X^{G^{\prime }}$, $H_2$ and $M^{X^{G^{\prime }}}$: $Y^{D_2}=D_2\left(X^{G^{\prime }},H_2,M^{X^{G^{\prime }}}\right)$.

Optimizing imputation with adversarial training by WGAN

SmartImpute takes the advantage of GAN model in improving the imputation accuracy with an adversarial training process. SmartImpute’s training procedure involves an adversarial training strategy, where the generator competes with the discriminators to improve imputation accuracy. The imputation model of SmartImpute is iteratively trained by fixing one of the generator or discriminators at a time with back propagation. For the training of generator, two loss function is used.

• (1)Adversarial Loss (Wasserstein Distance): The first loss function of generator reflects the adversarial process of GAN architecture. In the adversarial process, the generator is designed to fool the discriminator $D_1$ by minimizing this loss, while $D_1$ aims to accurately differentiate between the data from generator and the original data by maximizing the same function. This adversarial loss is measured using the Wasserstein distance,⁵⁸ known as WGAN.⁵⁹ WGAN’s advantages include more stable training, meaningful and interpretable loss metrics, reduced mode collapse, robustness to hyperparameters, and ultimately higher-quality generated samples. The loss function with Wasserstein distance is defined as:

$$L_1=E_{x\sim P_G}\left[D_1\left(X^G,H_1,M\right)\right]-E_{x\sim P_r}\left[D_1\left(X,H_1,M\right)\right]$$

In this context, $P_r$ and $P_G$ denote the real and generated data distributions, respectively.

• (2)Reconstruction Loss: The second part of the generator’s loss function focuses on the accuracy of the imputed data. The generated values for the observed elements should be close to the original values. This objective is defined using the mean squared error (MSE) for normalized (continuous) data:

$$L_2=\frac{1}{N_{obs}}\sum {\left(x_{obs}-x_{obs}^G\right)}^2+{\lambda }_G\left|x_{obs}^G\right|$$

The reconstruction loss is calculated only with the observed entries. Here, $N_{obs}$ is the number of observed entries, $x_{obs}$ are the original observed values, and $x_{obs}^G$ are the generated values. We add an extra regularization term ${\lambda }_G\left|x_{obs}^G\right|$ to threshold the biological zeros.

When dealing with the UMI count data, SmartImpute adapts the reconstruction loss by a zero-inflated negative binomial (ZINB) likelihood:

$$L_2=-\sum _{i=1}^{N_{obs}}\log \left[{\pi }_i+\left(1-{\pi }_i\right)·\text{NB}\left(x_i^G;r,p_i\right)\right]+{\lambda }_G\left|x_{obs}^G\right|\text{.}$$

Here ${\pi }_i$ is the probability of an excess zero for the i-th gene, and $\text{NB}\left(x_i^G;r,p_i\right)$ is the probability mass function of the negative binomial distribution with success probability $p_i$ and number of failures r before the i-th success.

The loss function for generator is then the weighted sum of $L_1$ and $L_2$:

$$L_G=L_1+\alpha L_2\text{.}$$

For discriminator $D_1$, the loss function is the negative of $L_1$,

$$L_{D_1}=-L_1=E_{x\sim P_r}\left[D_1\left(X,H_1,M\right)\right]-E_{x\sim P_G}\left[D_1\left(X^G,H_1,M\right)\right]\text{.}$$

And for discriminator $D_2$, we train $D_2$ to maximize the probability of $D_2$ predicting $M^{G^{\prime }}$,

$$L_{D_2}=E_{M^{G^{\prime }}}\left[{M^{G^{\prime }}}^T\log D_2\left(X^{G^{\prime }},H_2,M^{X^{G^{\prime }}}\right)+{\left(1-M^{X^{G^{\prime }}}\right)}^T\log \left\{1-D_2\left(X^{G^{\prime }},H_2,M^{X^{G^{\prime }}}\right)\right\}\right]\text{,}$$

where log is element-wise logarithm, and the expectation is calculated based on the random variable $M^{G^{\prime }}$.

Hyperparameter tuning

We conducted a hyperparameter search for different datasets using SmartImpute. To maintain robustness and computational efficiency, we fixed as many hyperparameters as possible. The main hyperparameters to search in SmartImpute includes the attention parameter $\gamma$, biological zero regularization parameter ${\lambda }_G$, hint rate for $H_1$: $h_1$, hint rate for $H_2$: $h_2$, learning rate ${\epsilon }_E$, and weighting parameter $\alpha$.

For the normalized HNSCC data, we first performed a 5-fold cross-validation to train the multi-task targeted GAIN. Two critical parameters were decided with ${\epsilon }_E=0.05$ and ${\lambda }_G=0.5$ . In this dataset, we set the number of hidden layers to 2 with nodes number 64. For SmartImpute, we found that three hidden layers for G and two for both $D_1$ and $D_2$ gave the relatively robust results with the number of nodes: 128, 128, 128, 32, and 32. Results were produced with learning late 0.001, $h_1$ = $h_2=$ 0.9, attention parameter $\gamma$ 0.005, number of subnets 4, number of sample batch 64, penalty ${\lambda }_G$ = 1, and weighting parameter $\alpha =90\text{.}$

For the HNSCC count data, we train the generator with ZINB likelihood function for $L_2$. The hyper-parameters values were given with with learning late 0.001, $h_1$ = $h_2=$ 0.8, attention parameter $\gamma$ 0.002, number of subnets 4, number of sample batch 64, penalty ${\lambda }_G$ = 1, and weighting parameter $\alpha =75\text{.}$

GPT model for target gene panel selection

To improve the targeted imputation approach of SmartImpute, the selection of an informative gene panel is performed using a GPT model with the R software package tpGPT. A well-established default base gene panel of 580 marker genes (BD Rhapsody Immune Response Targeted Panel (Human)) is provided as a starting point for constructing a data-specific target panel. In tpGPT, researchers can specify parameters such as cancer type, tissue type, and cell type to enrich the default panel with additional genes relevant to their particular research focus. The gene list is generated using structured prompts based on the following template:

‘’Generate a list to Number marker genes in Cancer_type associating with Tissue_type. \n Only provide the gene list. \n Ensure each gene is listed in a new line without any prefixes such as numbers.” In this template, “Number”, “Cancer_type” and “Tissue_type” are placeholders that will be replaced with the corresponding parameters. (For example, 250 for “Number”, “head and neck cancer” for “Cancer_type” and “epithelial tissue” for Tissue type.)

For generating a cell type-specific gene panel, the following prompt template is used: “Generate a list of marker genes for following cell types: Cell_type. \n Each cell type should contain at least N_celltype genes. \n Only provide the gene list. \n Do not provide list under input cell types. \n Ensure each gene is listed in a new line without any prefixes such as numbers.” In this template, “Cell_type” and “N_celltype” can be replaced with the corresponding parameters.

To evaluate the reliability of tpGPT, we compared gene lists generated by different OpenAI accounts across multiple GPT model versions (GPT-4 Turbo, GPT-4.1, o4-mini, and o3-mini) for 11 cancer types: Melanoma, Lung Cancer (NSCLC vs. SCLC), Head and Neck Cancer; Breast Cancer, Glioblastoma, Liver Cancer, Leukemias, Bladder Cancer, Colorectal Cancer, Prostate Cancer, and Pancreatic Cancer. On average, 86.91% of the top 100 genes were retained across different accounts and time points (Table S1). The completed lists are provided in Tables S2 and S3. The default GPT model used in tpGPT is gpt-o4-mini which is chosen for its speed and robustness. However, users can specify other models to generate the gene panel, such as gpt-4-turbo and gpt-3.5-turbo or the new updated GPT models, within the tpGPT function.

Single-cell RNAseq data downloading and processing

HNSCC tumor normalized data

Normalized gene count matrix of 6,000 single cells from neck squamous cell carcinoma tumors³¹ were downloaded from GEO: GSE103322 and analyzed as described in.³² As described in Methods in Puram et al., the normalized gene expressions were quantified as y = log2 (TPM + 1), where TPM refers to transcripts per million. The annotation provided by³² were used as true label of cells to evaluate the performance of SmartImpute.

HSNCC tumor-infiltrated immune cells count data

Raw UMI count data of CD45⁺ cells in 26 neck squamous cell carcinoma tumors³⁹ were downloaded from GEO: GSE139324 and processed with Seurat v4.0 for further analysis. Barcodes with fewer than 200 genes expressed or more than 10% UMIs originated from mitochondrial genes were filtered out; genes expressed in fewer than 3 barcodes were also excluded. For each tumor, standard library size and log-normalization was performed on raw UMI counts using NormalizeData(), and top 4000 most variable genes were identified by the “vst” method in FindVariableFeatures(). Data from individual tumors were further integrated to remove batch effects using FindIntegrationAnchors() and IntegrateData() functions in Seurat with 8,000 “anchors” and top 40 principal components. From the integrated data, scaled z-scores for each gene were calculated using ScaleData() function in Seurat by regressing against the percentage of UMIs originated from mitochondrial genes, S and G2/M phases scores, and total reads count. A shared nearest neighbor (SNN) graph was constructed based on the first 40 principal components computed from the scaled integrated data. Louvain clustering was performed using the FindClusters() function at resolution 1 for major cell type scRNA-seq data. Differential expression analysis was performed using FindAllMarkers() function with logfc.threshold = 0.25, min.pct = 0, and test.use = ”wilcox”. Clusters were annotated by comparing differential genes with canonical markers for major populations: B cells (CD79A, CD19), Plasma cells (IGHA1, IGHG1), T cells (CD3E, CD3D, CD2), Natural Killer cells (KLRC1, FCGR3A, NCAM1), Macrophages (CD68, CD163, MRC1, C1QC, C1QB), Monocytes (CD14, FCGR3A, FCN1), CD14⁺ Monocytes (FCN1, VCAN, LYZ), CD16⁺ Monocytes (FCGR3A, FCGR3B), cDC1 (XCR1, CLEC9A, FLT3), cDC2 (FCER1A, CD1C, HLA-DQA1), mregDC (LAMP3, CCR7, FSCN1), pDC (LILRA4, IL3RA, SERPINF1), Mast cells (KIT, TPSAB1, CPA3). T cells were extracted for further clustering analysis. The T cells clusters were subdivided into four subtypes: CD4⁺ naive T (CD4, TCF7, SELL, IL7R), Treg (CD4, FOXP3, TNFRSF18, IL2RA), CD8⁺ effector-memory (EM) T (CD8A, GZMK, EOMES, KLRG1), CD8⁺ pre-exhausted (preExh) T (CD8A, TCF7, TOX, HAVCR2), CD8⁺ exhausted (Exh) T (CD8A, TOX, LAG3, HAVCR2), proliferative T (STMN1, MKI67, CENPF, CCNA2). This annotation of T cell subtypes and other major cell types were used as true label.

HSNCC tumor-infiltrated T cells count data

Raw UMI count of tumor-infiltrated T cells from human oral squamous cell carcinomas (OSCC) and inflamed non-malignant oral tissues⁴⁶ were downloaded from GEO: GSE163633. The raw data were normalized, scaled, batch-corrected, integrated, and clustered as described above. Clusters were annotated into different T cell subtypes based on canonical markers and used as true label of cells in downstream analysis. The markers used were as follows: Th2: CD4, TNFSF11, GATA3, IL4; Treg: CD4, FOXP3, TNFRSF18, IL2RA; CD4⁺ naive T (CD4, TCF7, SELL, IL7R); CD8⁺ naive T (CD8A, TCF7, SELL, IL7R); CD8⁺ effector T (CD8A, GZMB, GZMA, PRF1, CCL3), CD8⁺ exhausted T (CD8A, TOX, LAG3, HAVCR2), proliferative T (STMN1, MKI67, CENPF, CCNA2).

Bone marrow data

We downloaded raw UMI counts and cell annotation of single-cell RNAseq data generated from 13,000 bone marrow cells from a health donor⁵¹ (https://figshare.com/articles/dataset/Expression_of_97_surface_markers_and_RNA_transcriptome_wide_in_13165_cells_from_a_healthy_young_bone_marrow_donor/13397987). Raw UMI counts were normalized and filtered as described above with Seurat. Cell annotation provided by the original study was used as true label of cells to evaluate the performance of SmartImpute.

Quantification and statistical analysis

Clustering performance evaluation

To evaluate the clustering performance, we calculated two metrics with known cell type labels.

The first metric is Jaccard index (JI). This metric is used to evaluate the similarity between the ground truth set and the set obtained by clustering. It is defined as the size of the intersection divided by the size of the union of the sample sets. Let TP, FN, and FP denote True Positives, False Negatives and False Positives.The Jaccard Index is defined as follow: $JI=\frac{TP}{TP+FN+FP}\text{.}$

The second metric is Adjusted Rand Index (ARI). Unlike the Jaccard Index, the ARI adjusts for the chance grouping of elements, providing a more rigorous assessment of clustering. It considers all pairs of samples and counts pairs that are assigned in the same or different clusters in the predicted and true clusters. The ARI is the corrected-for-chance version of the Rand index. Let $U=\left\{u_1,\dots ,u_r\right\}$ and $\mathrm{V}=\left\{v_1,\dots ,v_s\right\}$ be two partitions on dataset X. And ARI is defined as,

$$ARI=\frac{\sum \left(\begin{array}{c}{n}_{ij}\\ 2\end{array}\right)-\left[{\sum }_i\left(\begin{array}{c}{a}_i\\ 2\end{array}\right){\sum }_j\left(\begin{array}{c}{b}_j\\ 2\end{array}\right)\right]/\left(\begin{array}{c}n\\ 2\end{array}\right)}{\frac{1}{2}\left[{\sum }_i\left(\begin{array}{c}{a}_i\\ 2\end{array}\right)+{\sum }_j\left(\begin{array}{c}{b}_j\\ 2\end{array}\right)\right]-\left[{\sum }_i\left(\begin{array}{c}{a}_i\\ 2\end{array}\right){\sum }_j\left(\begin{array}{c}{b}_j\\ 2\end{array}\right)\right]/\left(\begin{array}{c}n\\ 2\end{array}\right)}\text{,}$$

where $n_{ij}=\left|u_i\cap v_j\right|$, $a_i=\sum _{k=1}^s{n}_{ik}$ and, $b_j=\sum _{k=1}^r{n}_{kj}$.

Data processing and Seurat object: To proceed the downstream analysis such as UMAP clustering, we transfer the original scRNA-data and imputed data into a Seurat object with meta information.⁶⁰ The UMI count data is normalized with Normalization function in Seurat package.

UMAP visualization with Seurat: To visualize the high-dimensional scRNA-seq data post-imputation, we utilized the Uniform Manifold Approximation and Projection (UMAP) technique implemented in the Seurat package. UMAP is a non-linear dimensionality reduction method that helps to uncover the intrinsic structure and relationships within complex datasets. Each dataset is first scaled to normalize the gene expression measurements, ensuring that variable genes are appropriately weighted. The scaled data is then subjected to principal component analysis (PCA) to reduce the dimensions of the dataset to the principal components that capture the most variation. In our study, we use the first 50 PCAs for analysis.

Cell type prediction evaluation

To evaluate how SmartInput improves the cell type prediction, we treated the provided annotation as true cell type label (as described above) and employed several automated cell typing algorithms for prediction.

HNSCC tumor normalized data: Cell types within the normalized HNSCC data were predicted using SingleR,³³ which leverages reference data to assign cell type labels to individual cells in the experimental dataset. We search with the default Human references used in SingleR package. The results of the cell type annotation are presented with the reference of BLUEPRINT which contains the most cell types in the normalized HNSCC dataset. We also utilized the AUCell package³⁵ to verify cell types within this dataset. AUCell calculates the Area Under the Curve (AUC) to assess the activity of predefined gene sets within individual cells. We constructed marker gene sets for each cell type by selecting the top five highly expressed genes characteristic of the corresponding cell type. These marker gene sets were then used to build gene rankings for each cell based on their expression profiles. By calculating the AUC scores for these gene sets, we were able to determine the relative activity of the marker genes in each cell. Cells exhibiting high AUC scores for a particular gene set were identified as belonging to the corresponding cell type.

HSNCC tumor-infiltrated immune cells count data: We employed Azimuth⁴⁰ cell type prediction application built within Seurat package (https://satijalab.org/azimuth/). The publicly available peripheral blood mononuclear cell (PBMC) reference data were selected and incorporated as the reference data in “RunAzimuth” function. The choice of the PBMC dataset was guided by its alignment with the expected cell types inherent in the HNSCC count data.

The annotation accuracy is calculated using the F1 score which is defined as: $F1=2\frac{TP}{2TP+FP+FN}\text{.}$

TP, FN, and FP denote True Positives, False Negatives, and False Positives, respectively, as defined in the calculation of the Jaccard Index.

Performance benchmarking against existing methods

We benchmark the performance of SmartImpute with other imputation frameworks as follows:

SAVER (v1.1.3): Single-cell analysis via expression recovery (SAVER) is an expression recovery method using a Bayesian Framework. SAVER assumes a Poisson-Gamma mixture model to estimate the posterior distribution of the gene expression levels. SAVER employs a LASSO penalty which has the similar functions to the penalization we used in the loss function $L_2$. We used the SAVER R package with default parameters, as detailed in the SAVER documentation (https://github.com/mohuangx/SAVER).

DCA: The Deep Count Autoencoder (DCA) framework uses a zero inflated negative binomial noise model to impute the scRNA-seq data. By incorporating the ZNIB distribution model with an autoencoder framework, DCA can capture the sparsity and overdispersion of the count data. We followed the default setting with learning rate 0.001, batch size 32 and a latent representation space of 32 dimensions. All other hidden layers are 64 dimensions. (https://github.com/theislab/dca).

DeepImpute: DeepImpute employs a data driven targeted imputation approach with deep neural network. The first step is using the variance over mean ration with default value 0.5 to select the target genes. DeepImpute splits the genes into random subsets with each containing S genes. We follow the default setting of S = 512, aligning with the number of target genes used in SmartImpute. We set the learning rate to 0.0001 and a batch size 64 as suggested in default settings. DeepImpute contains a dropout layer and we set the dropout rate as 20% as suggested by the author. (https://github.com/lanagarmire/DeepImpute).

The computation time for each method was evaluated using a subset of the normalized HNSCC data on a 3.3 GHz 12-Core Xeon Processor machine. We calculated the average running time five scenarios involving different sizes of gene subsets. In each subset, we included the targeted panel genes of SmartImpute (size = 580) and randomly added additional genes, resulting in total gene numbers for each scenario of 1,000, 5,000, 10,000, 15,000, and 18,241. All cells (total = 5,357) were used in each scenario. Each method was run five times to obtain the average computation time. All imputation results were converted into Seurat objects for subsequent UMAP clustering.

Published: August 4, 2025