Abstract
Tissue heterogeneity is a major confounding factor in studying individual populations that cannot be resolved directly by global profiling. Experimental solutions to mitigate tissue heterogeneity are expensive, time consuming, inapplicable to existing data, and may alter the original gene expression patterns. Here we ask whether it is possible to deconvolute two-source mixed expressions (estimating both proportions and cell-specific profiles) from two or more heterogeneous samples without requiring any prior knowledge. Supported by a well-grounded mathematical framework, we argue that both constituent proportions and cell-specific expressions can be estimated in a completely unsupervised mode when cell-specific marker genes exist, which do not have to be known a priori, for each of constituent cell types. We demonstrate the performance of unsupervised deconvolution on both simulation and real gene expression data, together with perspective discussions.
Introduction
In biological research, "tissue heterogeneity" presents a significant challenge. Global gene expression profiling of a tissue sample—such as a tumor biopsy—often results in a composite signal from a mixture of cell types, including cancer cells, immune cells, and fibroblasts. To accurately interpret this data, researchers must "unmix" or deconvolute the signal into its constituent parts.
Traditional deconvolution methods are typically supervised or semi-supervised, requiring prior knowledge of either cell proportions or specific "marker gene" signatures. However, these signatures are often unknown or highly variable in clinical settings. This paper presents a "blind" deconvolution roadmap: a method to unmix two cell types using only the mixed data itself, without any prior biological "cheat sheets."
Methods: The Geometry of Unmixing
The proposed unsupervised approach is built upon a linear latent variable model. The core innovation lies in the geometric interpretation of gene expression scatter plots. When the expression levels of two mixed samples are plotted against each other, the data points are naturally compressed into a cone-shaped sector.
The "edges" or radii of this cone are defined by marker genes—genes uniquely expressed in one cell type but not the other. Crucially, these markers do not need to be known beforehand. By identifying genes that lie on the outermost boundaries of the scatter plot, the algorithm can mathematically estimate the mixing matrix (the proportions of each cell type). Once the proportions are known, matrix inversion is used to recover the pure expression profiles for every gene in the dataset.
Key Results
The authors validated this roadmap using numerical mixtures, biological mixtures of MCF7 (breast cancer) and HS27 (fibroblasts) cell lines, and public benchmark datasets.
- Proportion Estimation: In both simulated and real biological mixtures, the method estimated cell proportions with a correlation coefficient (r) of 0.99 compared to ground truth.
- Profile Recovery: The recovered "pure" expression profiles for the constituent cell types showed an average correlation of 0.99 with the actual measured profiles of the pure cell lines.
- Benchmark Performance: On the AFFY brain-heart benchmark, this unsupervised method achieved a correlation of 0.99, slightly outperforming existing semi-supervised methods (r=0.98) that required more prior information.
Discussion and Future Directions
The primary insight of this work is that marker genes act as geometric anchors; their mere existence in the genome allows for the discovery of underlying mixing proportions. While currently optimized for two-source mixtures (e.g., tumor vs. stroma), this framework provides a mathematically grounded foundation for studying intratumor heterogeneity.
Future work will focus on expanding this geometric proof to three or more sources, which involves higher-dimensional analysis. Additionally, while the method requires the existence of unique marker genes, it remains a powerful tool for most biological contexts where cell types exhibit distinct expression signatures.
Summary updated