Abstract
Single-cell transcriptomic technology has now matured, allowing quantification of mRNA transcripts corresponding to tens of thousands of genes within a cell. However, still only a small fraction of these mRNA is captured and measured by today’s single-cell assays. There are likely hundreds of thousands of mRNA copies present within a typical human cell, yet these assays omit a majority of the transcripts that are actually present. This introduces technical noise, especially non-biological variability and excessive sparsity, which frustrates downstream analysis and potentially skews biological conclusions. To overcome these challenges, we here develop HyperCell, a probabilistic deep learning approach that explicitly models this undersampling to produce estimates of each cell’s original gene transcript abundances across the whole transcriptome. We demonstrate that our framework offers benefits in various mRNA modeling settings, by i) correctly differentiating between spurious sampling-induced and real biological zeros, outperforming existing approaches, ii) estimating the total mRNA content of cells across states to reduce contamination due to background transcripts, iii) reducing contamination due to background transcripts, and iv) helping to counteract biases that may appear during typical differential gene expression analyses using widespread normalization approaches. Our approach to correcting for the technical noise introduced by the single-cell experimental process brings us closer to studying biology, starting from the true transcriptome of cells.
Introduction
Single-cell transcriptomics (scRNA-seq) has revolutionized biological research, yet it suffers from a significant "missing data" problem. Current experimental assays typically capture only 10% to 30% of the mRNA molecules present in a cell. This massive undersampling leads to "dropout"—where expressed genes appear as zeros—and technical noise that makes identical cells appear distinct simply due to stochastic sampling.
Traditional tools often address this by analyzing relative abundances or assuming infinite transcript pools (e.g., Negative Binomial distributions). However, these methods frequently fail to distinguish between biological zeros and technical artifacts. Furthermore, standard normalization techniques that force all cells to a uniform "library size" can inadvertently erase critical biological signals, such as hypertranscription, where a cell increases its total RNA production.
The HyperCell Framework
HyperCell is a deep generative model based on the Information Maximizing Variational Autoencoder (InfoVAE) framework. Its primary innovation is the use of the multivariate hypergeometric distribution to model the measurement process. Unlike Poisson or Multinomial approximations, the hypergeometric distribution accurately models sampling without replacement from a finite population—a precise mathematical representation of mRNA capture in droplets.
To enable optimization within neural networks, the authors implemented a continuous relaxation of the hypergeometric distribution using Gamma functions. Additionally, the model employs Lagrangian constrained optimization to ensure that estimated "true" counts are never lower than the observed data. HyperCell also incorporates a background modeling component to identify and remove "ambient RNA" contamination from the surrounding cellular "soup."
Validation and Performance
The model was validated across a massive scale, encompassing over 6 million cells and nuclei, including the Human Brain Cell Atlas and large-scale immune datasets. In synthetic spike-in experiments where the ground truth was approximately 50,000 molecules, HyperCell recovered a median of 47,312 molecules, a significant improvement over the raw observed count of roughly 13,942.
Benchmarking against state-of-the-art methods like scVI, MAGIC, and ALRA showed that HyperCell superiorly preserves sex-specific gene patterns and removes background contamination (achieving an F1 score of 0.97 for monocyte marker localization). Furthermore, HyperCell improved the correlation of single-cell differential expression (DE) results with "gold standard" bulk RNA-seq by over 50% in several immune cell types.
Biological Insights: The Transcriptome Size Census
By estimating absolute counts, HyperCell allows researchers to treat "transcriptome size" (total mRNA content) as a distinct biological feature. The study revealed that this size varies significantly across cell types; for instance, oligodendrocytes in the brain exhibit much larger transcriptomes than ependymal cells due to metabolic hyperactivity.
In a disease context, HyperCell identified specific dopamine neuron subtypes in Parkinson’s Disease patients that show a significant increase in total transcriptome size. This nuance is typically masked by standard relative abundance normalization, which assumes constant total RNA across cells. By recovering these absolute values, HyperCell reveals biological shifts that were previously invisible.
Limitations and Future Directions
As a deep learning model, HyperCell requires significant GPU resources to scale to multi-million cell datasets. It also relies on Unique Molecular Identifier (UMI) counts, making it less applicable to older non-UMI datasets. While it effectively handles ambient background RNA, integrating data across disparate laboratories still requires careful attention to batch effects. Future work may extend absolute transcript estimation to study cell cycle states, cancer cell hypertranscription, and high-resolution drug discovery.
Summary updated