Supervised discovery of interpretable gene programs from single-cell data

Abstract

Factor analysis can drive biological discovery by decomposing single-cell gene expression data into a minimal set of gene programs that correspond to processes executed by cells in a sample. However, matrix factorization methods are prone to technical artifacts and poor factor interpretability. We have developed Spectra, an algorithm that identifies user-provided gene programs, modifies them to dataset context as needed, and detects novel programs that together best explain expression covariation. Spectra overcomes the dominance of cell-type signals by modeling cell-type-specific programs, and can characterize interpretable cell states along a continuum. We show that it outperforms existing approaches in challenging tumor immune contexts; Spectra finds factors that change under immune checkpoint therapy, disentangles the highly correlated features of CD8+ T-cell tumor reactivity and exhaustion, finds a novel program that explains continuous macrophage state changes under therapy, and identifies cell-type-specific immune metabolic programs.

Scalable joint non-negative matrix factorisation for paired single cell gene expression and chromatin accessibility data

“…Inference schemes using EM (Expectation Maximisation) or VI (Variational Inference) for updates could be implemented, but they are usually impractical for large datasets. For example, a recent scRNA-seq NMF based method implements an EM algorithm but favours the use of ADAM (a gradient descent based method) due long run times [40]. A potential extension for intNMF, given its speed, is iterative learning - e .…”

“…intNMF is implemented with an l2 loss, this is a practical decision (similar to MOFA+'s recommendation to use a Gaussian loss for large datasets) due to the efficient update schemes available given an l2 loss. Other cost functions such as the Kullback-Leibler divergence or l1/Poisson could also be applied, which are likely to attend less to highly expressed genes ( [40,41]). Additionally there are flavours of NMF with probabilistic interpretations (primarily by adding sum to one constraints for each cell and each feature) [42], which may further enhance the interpretability of the models.…”

Identification of cell types, states and programs by learning gene set representations

“…Cell-intrinsic gene programs can be identified by scoring single-cell data with gene sets that represent molecular pathways or transcriptional signatures. Scoring methods apply gene signature knowledge-bases using rank-based or factorization-based approaches 11,12,13 to obtain per-cell scores that facilitate the interpretation of cellular phenotypes across a dataset. For example, Spectra 13 is a Bayesian approach that can discern cell type-specific phenotypes from global programs while refining the definition of a gene set depending on the context.…”

“…Scoring methods apply gene signature knowledge-bases using rank-based or factorization-based approaches 11,12,13 to obtain per-cell scores that facilitate the interpretation of cellular phenotypes across a dataset. For example, Spectra 13 is a Bayesian approach that can discern cell type-specific phenotypes from global programs while refining the definition of a gene set depending on the context. Alternatively, to obtain gene programs that span across cell types, recent methods like DIALOGUE 14 , aim to uncover multicellular processes in single-cell RNAseq data.…”

Multi-omics integration of scRNA-seq time series data predicts new intervention points for Parkinson’s disease

“…NMF-based approaches have also shown promising results in dealing with sparse SC samples (Welch et al, 2019;Argelaguet et al, 2020;Jung et al, 2020;Huizing et al, 2023). Furthermore, NMF-based methods have been used to jointly integrate SC data with molecular networks to identify types of SCs (Elyanow et al, 2020), to discover interpretable gene programs (Kunes et al, 2023) and to generate protein representations within various cellular contexts to identify therapeutic targets and nominate cell type contexts for rheumatoid arthritis and inflammatory bowel diseases (Li et al, 2023). Using matrix factorization approaches to integrate SC data with molecular networks (i.e., prior knowledge) allows us to benefit from the biologically relevant information in molecular networks and simultaneously minimize the inherent noisiness of SC data.…”