Nonnegative spatial factorization applied to spatial genomics.

Area of Science:

• Computational biology
• Bioinformatics
• Statistical modeling

Background:

• Nonnegative matrix factorization (NMF) offers interpretable parts-based representations for high-dimensional count data.
• Standard NMF struggles to integrate spatial relationships inherent in datasets like spatial transcriptomics.
• Existing real-valued methods may not fully capture the unique characteristics of count data or spatial dependencies.

Purpose of the Study:

• Introduce nonnegative spatial factorization (NSF), a novel probabilistic dimension reduction technique.
• Develop a spatially-aware model that naturally incorporates sparsity and scales to large datasets.
• Address the limitations of traditional NMF in handling structured observations.

Main Methods:

• NSF utilizes transformed Gaussian processes to create a spatially-aware probabilistic framework.
• The model is designed to encourage sparsity and handle tens of thousands of observations.
• A hybrid extension is proposed to combine spatial and nonspatial components for comprehensive analysis.

Main Results:

• NSF demonstrates superior accuracy in recovering ground truth factors compared to real-valued methods like MEFISTO in simulations.
• The model achieves lower out-of-sample prediction error on mouse brain and liver spatial transcriptomics datasets than probabilistic NMF.
• The hybrid NSF effectively quantifies the spatial importance of both observations and gene expression features.

Conclusions:

• NSF provides a powerful, spatially-explicit approach for analyzing high-dimensional count data, particularly in spatial transcriptomics.
• The method enhances interpretability and predictive performance by leveraging observed spatial structures.
• NSF offers a flexible framework, with a hybrid extension, to accommodate varying degrees of spatial correlation in biological data.