Biological tissues are complex, heterogeneous communities of cells. While bulk RNA sequencing and spatial transcriptomics allow us to capture gene expression profiles from these tissues, they often do so by averaging signals across thousands of cells or multi-cellular "spots." To understand the underlying biology, researchers use computational deconvolution to unmix these blended signals into their constituent cell types.
Traditionally, this unmixing process relies on a closed-world assumption: it requires a pre-existing, high-quality single-cell reference atlas to serve as a ground-truth dictionary of known cell types. However, a major challenge in modern computational biology is reference dataset dependency. When the reference atlas is incomplete, biased, or unavailable, standard deconvolution models break down, introducing systematic errors that distort our understanding of tissue architecture.
The Mathematical Trap of Misattribution Bias
To understand why reference dependency degrades performance, consider how supervised deconvolution works. Mathematically, the algorithm attempts to model an observed bulk or spatial mixture profile as a linear or non-linear combination of known reference profiles. The optimization objective is simple: find the cell-type proportions that minimize the difference between the reconstructed profile and the true observed sample.
What happens when the target tissue contains a cell type, an immune infiltrate, or a disease-specific cell state that does not exist in the reference library?
Because the optimization framework must account for every transcript observed in the mixture, it cannot simply ignore the signal from an uncharacterized cell population. Instead, the model suffers from misattribution bias. As Bergenstråhle et al. (2020) demonstrate, the algorithm is forced to distribute the transcripts of the missing cell type among the most mathematically similar categories available in the reference library.
[True Tissue Mixture] ──> Contains: T-Cells, B-Cells, and an Uncharacterized Tumor State
│
▼
[Reference Library] ──> Only Contains: T-Cells and B-Cells
│
▼
[Deconvolution Error] ──> Tumor transcripts are mathematically forced into
the "closest match" (e.g., overestimating B-Cells)
This vulnerability is widespread across different modalities:
- Bulk and Spatial Transcriptomics: Yang (2021) notes that assuming a reference contains all cell types leads to incorrect assignments, masking the true cellular composition. Larsson (2021) emphasizes that any cell type omitted from the reference library is completely excluded from the resulting spatial map, yielding an incomplete representation of the tissue ecosystem.
- Tumor Purification: In clinical settings, Quon et al. (2013) show that when clinical tumor samples contain healthy tissue contaminants or immune infiltrates absent from the reference panel, the accuracy of the purification process is severely compromised.
- Epigenomics and Proteomics: The issue extends beyond RNA. Rowland et al. (2022) show that reference-based methods fail to produce stable results in bulk Hi-C data when rare populations are missing, while Wang et al. (2022) document a similar bottleneck in tissue proteomics due to the scarcity of single-cell proteomic atlases.
The Limits of Static and Proxy References
Even when a reference dataset is available, it may not match the biological reality of the target sample. This mismatch manifests in three primary ways:
1. The Proxy Approximation Layer
When a comprehensive reference atlas does not exist for a specific organism, researchers are often forced to use data from a related species as a proxy. For example, Tello‐Palencia et al. (2024) had to rely on mouse blood cell reference data to deconvolve rat blood samples due to a scarcity of rat-specific atlases. This proxy layer introduces systematic approximation errors, as cross-species genetic differences skew the proportion estimations.
2. Batch Effects and Technical Biases
Single-cell references are typically generated using different sequencing platforms, tissue preparation protocols, or patient cohorts than the target bulk or spatial data. As He et al. (2024) point out, relying on these static external references introduces severe technical batch biases that can obscure genuine biological signals.
3. Discrete Categories vs. Continuous States
Static references treat cell types as discrete, rigid categories. However, biology operates on a spectrum. Lopez et al. (2021) argue that static references fail to capture continuous phenotypic gradients, such as progressive inflammation or oxygen deprivation. If a specific cellular response or intermediate state is missing from the reference library, the model cannot effectively map that variation in the target tissue.
Shifting to Reference-Free and Weakly Supervised Inference
To overcome these constraints, the field is shifting away from strict reliance on external single-cell transcriptomes. Modern architectures use three primary strategies to achieve reference-free or weakly supervised deconvolution.
Multimodal and Spatial Regularization
Instead of relying on an external transcriptomic library, newer methods treat deconvolution as an unsupervised matrix factorization problem, using the intrinsic spatial and physical properties of the target tissue as regularizers.
For instance, Zhang et al. (2025) developed FAST, which uses regularized Non-negative Matrix Factorization (NMF) to identify cell types by integrating gene expression directly with spatial coordinates and histology image features. Similarly, He et al. (2024) introduce Starfysh, a deep generative framework that bypasses external references entirely by pairing transcriptomic data with morphology from histology images to isolate context-specific cell states unique to the sample.
Cross-Sample Information Sharing
When analyzing rare pathologies or specific developmental time points where references are completely unavailable, algorithms can leverage shared information across multiple target samples. Niyakan et al. (2023) demonstrate with MUSTANG that combining local spatial proximity with cross-sample transcriptional similarity allows for accurate cell-type identification without an external atlas.
Qualitative Knowledge over Quantitative Matrices
Another alternative is to replace full, quantitative single-cell expression matrices with qualitative biological knowledge, such as marker gene lists. The authors of Celloscope (2023) utilize a Bayesian probabilistic model driven solely by marker lists, eliminating platform-specific batch effects.
However, even these weakly supervised methods face constraints: if a major cell type is omitted from the marker list, the model will still experience proportion biases. To mitigate this, modern architectures often include "placeholder" or uncharacterized categories to absorb unexplained signals, preventing them from corrupting the proportions of known cell types.
By relaxing the requirement for static, external references, these emerging computational strategies allow researchers to discover novel cellular phenotypes and map tissue architectures exactly as they exist in situ.
Key Foundational Work
Computational purification of individual tumor gene expression profiles leads to significant improvements in prognostic prediction (2013) — contributed 1 concept elements to this direction.
Comprehensive analyses of tumor immunity: implications for cancer immunotherapy (2016) — contributed 1 concept elements to this direction.
A Self-Directed Method for Cell-Type Identification and Separation of Gene Expression Microarrays (2013) — contributed 1 concept elements to this direction.