Imputation

These methods aggregate expression values across similar cells to correct zeros, operating on the principle that technical dropouts are stochastic while biological zeros are systematic.

• MAGIC (Markov Affinity-based Graph Imputation): Diffuses gene expression values across a cell-cell affinity graph constructed from diffusion maps, effectively sharing information between phenotypically similar cells.
• kNN-Smoothing: Averages expression from a cell's k-nearest neighbors in PCA or latent space, providing a computationally lightweight correction.
• Key limitation: May over-smooth genuine biological heterogeneity, blurring the boundaries between closely related but distinct cell states or obscuring rare transient populations.

These approaches fit explicit probabilistic models to the zero-inflated negative binomial (ZINB) or hurdle distributions characteristic of scRNA-seq data, distinguishing technical from biological zeros through likelihood estimation.

• scImpute: Identifies dropout candidates by fitting a Gamma-Normal mixture model to each gene, then imputes values only for likely dropout events using non-negative least squares regression on similar cells.
• SAVER (Single-cell Analysis Via Expression Recovery): Borrows information across genes using an empirical Bayes approach with a Poisson-Gamma mixture, recovering true expression by estimating posterior means.
• Key advantage: Preserves biological zeros by explicitly modeling the dropout probability rather than blindly smoothing all low values.

Neural network architectures learn compressed latent representations of the full transcriptomic profile, reconstructing gene expression from a lower-dimensional manifold where dropouts are implicitly corrected.

• Autoencoders (e.g., DCA, scVI): Encode the count matrix into a bottleneck layer and decode back to the original gene space, with the reconstruction naturally filling in missing values based on learned gene-gene covariance structures.
• Deep Count Autoencoder (DCA): Specifically parameterizes the decoder with a ZINB loss function, directly modeling the count distribution, dropout probability, and dispersion for each gene.
• Key advantage: Captures non-linear gene-gene relationships that linear methods miss, enabling imputation of complex co-expression patterns without explicit neighborhood definition.

These methods leverage paired multi-omic measurements from the same cell to impute missing modalities, exploiting the fact that chromatin accessibility or protein abundance can predict transcript levels.

• Seurat v3/v4 Transfer: Projects cells into a shared latent space defined by canonical correlation analysis (CCA) anchors, enabling imputation of RNA from ATAC or protein data in multimodal assays like CITE-seq.
• TotalVI (scVI-tools): A variational autoencoder that jointly models RNA and protein counts, imputing missing proteins for cells where only the transcriptome was measured by learning a joint probabilistic representation.
• Key limitation: Requires paired training data where both modalities are measured simultaneously, limiting applicability to experimental designs that include multimodal profiling.

Rather than borrowing from similar cells, these methods impute dropouts by leveraging co-expression patterns across genes within the same cell, exploiting the fact that functionally related genes exhibit correlated expression.

• ALRA (Adaptively-thresholded Low-Rank Approximation): Performs a truncated SVD on the normalized count matrix, then adaptively thresholds the low-rank reconstruction to restore zeros that represent likely dropouts while preserving true biological zeros.
• ENHANCE: Uses a denoising autoencoder trained on bulk RNA-seq data to learn gene regulatory relationships, then applies these patterns to correct single-cell dropouts.
• Key advantage: Avoids the over-smoothing pitfall of cell-based methods, preserving sharp boundaries between cell types while recovering within-cell co-expression structure.

Evaluating imputation accuracy requires careful experimental design, as no ground truth exists for endogenous transcriptomes. Common validation strategies include:

• Spike-in controls: Synthetic RNA molecules added at known concentrations (e.g., ERCC spike-ins) provide absolute truth for a limited set of transcripts, enabling quantification of false positive rate and over-imputation bias.
• Downsampling experiments: Artificially subsampling reads from deeply sequenced cells creates synthetic dropouts with known true values, though this only tests recovery of high-expression genes.
• Biological validation: Testing whether imputed values recover known gene regulatory relationships, pathway memberships, or protein-level measurements from orthogonal assays like CITE-seq.
• Key pitfall: Over-imputation can inflate gene-gene correlations and create spurious co-expression patterns, necessitating evaluation of downstream differential expression and trajectory inference results.

The biological and technical phenomenon that imputation seeks to correct. A dropout occurs when a transcript is present in the cell but fails to be reverse-transcribed or amplified during library preparation, resulting in a false zero in the count matrix. This is distinct from true biological zeros where a gene is genuinely not expressed. Dropout rates are inversely correlated with gene expression level and are the primary source of zero-inflation in scRNA-seq data. Distinguishing dropouts from true zeros is the central statistical challenge that imputation models must solve.

The raw numerical substrate upon which imputation operates. A count matrix is a genes-by-cells matrix where each entry records the number of Unique Molecular Identifiers (UMIs) detected. In typical scRNA-seq experiments, over 90% of entries are zero. This extreme sparsity violates assumptions of normality in downstream statistical tests and distorts distance metrics used in clustering. Imputation algorithms smooth this matrix by borrowing information across the k-nearest neighbor graph, effectively filling in the structural zeros while attempting to preserve the stochasticity of genuine biological absence.

The foundational algorithmic paradigm underlying most imputation methods. The core assumption is that cells with similar global transcriptomic profiles share similar expression values for individual genes. The algorithm proceeds by:

• Constructing a nearest-neighbor graph in a reduced-dimensional space (typically PCA)
• For each cell, identifying its k most similar neighbors
• Aggregating expression values from these neighbors (via averaging, weighted pooling, or matrix factorization) to correct the target cell's dropouts

Methods like MAGIC and sctransform implement variations of this principle, differing in their distance metrics and aggregation functions.

A critical preprocessing step that must be performed before imputation in multi-sample studies. Batch effects—systematic technical variation between experimental runs—can cause cells from different batches to appear artificially dissimilar. If imputation is applied across uncorrected batches, the algorithm will smooth within batch-specific artifacts rather than true biological neighbors. Tools like Harmony, Scanorama, or Seurat's CCA-based integration must first align datasets in a shared latent space. Only then can the k-nearest neighbor graph used for imputation span genuinely biologically similar cells across conditions.

The most consequential downstream analysis affected by imputation choices. Imputing gene expression values artificially reduces within-group variance, which can inflate statistical significance in differential expression tests and produce false positives. Best practices include:

• Using imputed values only for visualization and clustering, not for hypothesis testing
• Applying pseudobulk approaches that aggregate cells to sample-level counts before testing
• Employing methods like DESeq2 or edgeR on raw counts, which natively model the negative binomial distribution of sparse data

Regulatory submissions to the FDA require explicit documentation of whether and how imputation was applied to data used for biomarker claims.

An emerging class of methods that use variational autoencoders (VAEs) or diffusion models to learn the underlying data distribution and probabilistically reconstruct missing values. Unlike kNN-based smoothing, these models:

• Capture non-linear gene-gene correlations beyond local neighborhood similarity
• Model the uncertainty of imputed values, providing confidence intervals
• Can be pre-trained as foundation models (e.g., scGPT, scFoundation) on massive atlases and fine-tuned for imputation on specific datasets

These methods are computationally intensive but offer state-of-the-art performance for recovering subtle biological signals in rare cell populations.