Abstract
An inherent challenge in interpreting single-cell transcriptomic data is the high frequency of zero values. This phenomenon has been attributed to both biological and technical sources, although the extent of the contribution of each remains unclear. Here, we show that the underlying gene presence/absence sparsity patterns are by themselves highly informative. We develop an algorithm, called DECODE, to assess the extent of joint presence/absence of genes across different cells, and to infer a gene dependency network. We show that this network captures biologically-meaningful pathways, cell-type specific modules, and connectivity patterns characteristic of complex networks. We develop a model that uses this network to discriminate biological vs. technical zeros, by exploiting each gene’s local network neighborhood. For inferred non-biological zeros, we build a predictive model that imputes the missing value of each gene based on activity patterns of its most informative neighbors. We show that our framework accurately infers gene-gene functional dependencies, pinpoints technical zeros, and predicts biologically-meaningful missing values in three diverse datasets.
Introduction
Single-cell RNA-sequencing (scRNA-seq) has revolutionized our ability to study cellular heterogeneity at high resolution. However, the technology is plagued by "zero-inflation," where a significant portion of the data matrix consists of zeros. These zeros are a mixture of biological zeros (true lack of expression) and technical zeros (often called "dropouts," where a gene is expressed but not detected).
Distinguishing between these two is a fundamental challenge. Traditional imputation methods often rely on global cell-to-cell similarity, which can lead to "smearing"—a phenomenon where distinct cell states are blurred together into a statistical average. DECODE (DEpendency-based COmputational DE-noising) offers a different approach: it treats the sparsity pattern itself as a signal, using the statistical patterns of gene presence and absence to reconstruct biological networks and accurately impute missing data.
The DECODE Framework
DECODE operates through a three-stage pipeline designed to leverage gene-gene dependencies rather than global cell similarities.
1. The Gene Co-dependency Index (CDI)
Standard correlation metrics often perform poorly on sparse single-cell data. To address this, DECODE uses an information-theoretic measure called the Gene Co-dependency Index (CDI). CDI quantifies the likelihood that if Gene A is detected, Gene B will also be detected. This allows for the construction of a weighted, undirected gene dependency network that captures functional relationships even in highly sparse environments.
2. Discriminating Zeros via "Guilt-by-Association"
Once the CDI network is established, DECODE applies a "guilt-by-association" principle to identify dropouts. For any given cell where a gene is recorded as zero, DECODE examines that gene's functional neighbors in the network. If the neighbors are active but the gene is not, the zero is statistically flagged as a technical dropout using a minimum hypergeometric (mHG) p-value. This ensures that biological zeros—where both a gene and its functional module are silent—remain untouched.
3. Targeted Imputation
For zeros identified as technical, DECODE employs $ε$-Support Vector Regression ($ε$-SVR). The model is trained on the expression patterns of a gene's most informative neighbors to predict the likely missing value. By focusing on local neighborhoods rather than global cell clusters, DECODE avoids the over-smoothing common in other tools.
Validation and Key Results
The authors validated DECODE across three diverse datasets: Mouse Brain (microfluidic), Human PBMC (10x Genomics), and Human Pancreas (inDrop).
- Superior Functional Inference: CDI achieved an Area Under the Curve (AUC) of 0.91 for the Mouse Brain and 0.89 for the Pancreas datasets when recovering known gene-gene functional dependencies, significantly outperforming standard correlation and mutual information.
- Accuracy in Dropout Detection: Compared to the popular tool
scImpute, DECODE showed a 20.4% increase in sensitivity and a 29.4% increase in specificity in identifying true expression versus dropouts. - Preservation of Data Structure: In cross-validation, DECODE was uniquely successful (compared to
MAGICandscImpute) in faithfully recovering both gene-gene and cell-cell relationships without introducing artificial signals. - Clinical Application: Analyzing Type 2 Diabetes (T2D) data, DECODE identified three distinct gene modules related to insulin signaling. These modules were corroborated by physical protein-protein interaction networks, demonstrating the framework's ability to extract biologically relevant insights from noisy clinical data.
Discussion
The core insight of DECODE is that sparsity is not merely a technical hurdle but a source of information. The presence/absence patterns of genes are sufficient to identify canonical marker genes—which emerge as "hubs" in the CDI network—and to reconstruct the functional architecture of the cell.
While DECODE is highly effective, its performance depends on the initial quality of the gene-gene network. In extremely sparse datasets where co-dependency cannot be established, the framework may face limitations. Future developments may include integrating multi-omic data, such as chromatin accessibility, to further refine the dependency networks and exploring deep learning architectures to capture non-linear gene interactions during the imputation phase.
Summary updated