Scalable querying of human cell atlases via a foundational model reveals commonalities across fibrosis-associated macrophages

Abstract

Single-cell RNA-seq (scRNA-seq) studies have profiled over 100 million human cells across diseases, developmental stages, and perturbations to date. A singular view of this vast and growing expression landscape could help reveal novel associations between cell states and diseases, discover cell states in unexpected tissue contexts, and relatein vivocells toin vitromodels. However, these require a common, scalable representation of cell profiles from across the body, a general measure of their similarity, and an efficient way to query these data. Here, we present SCimilarity, a metric learning framework to learn and search a unified and interpretable representation that annotates cell types and instantaneously queries for a cell state across tens of millions of profiles. We demonstrate SCimilarity on a 22.7 million cell corpus assembled across 399 published scRNA-seq studies, showing accurate integration, annotation and querying. We experimentally validated SCimilarity by querying across tissues for a macrophage subset originally identified in interstitial lung disease, and showing that cells with similar profiles are found in other fibrotic diseases, tissues, and a 3D hydrogel system, which we then repurposed to yield this cell statein vitro. SCimilarity serves as a foundational model for single cell gene expression data and enables researchers to query for similar cellular states across the entire human body, providing a powerful tool for generating novel biological insights from the growing Human Cell Atlas.

Scaling cross-tissue single-cell annotation models

“…We modified the original TabNet implementation in a few crucial ways: scTab’s input data assumption is adapted to the single-cell setting, in particular, the input gene expression is size factor normalized to 10,000 counts per cell and log1p transformed. This common normalization for scRNA-seq data 7,23 cannot be replicated by the simple batch normalization layer used in the original TabNet architecture. We additionally modified the original TabNet architecture to improve computational efficiency, namely by reducing the number of feature and attention blocks (which we found unnecessary after profiling), and training dynamics for faster convergence (Methods).…”

“…Moreover, unlike in the original TabNet model, we normalized the input data before feeding it into the neural network. scRNA-seq data is often normalized to have 10,000 counts per cell and is then log1p transformed afterward 7,12,23 , we applied the same normalization for our scTab model on top of the simple batch normalization layer, which is used in the original TabNet model to normalize the input features, as such a non-linear normalization cannot be achieved by a simple batch normalization layer.…”

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

“…For the ovarian cancer dataset, scDECAF pipeline was run on discovery mode (sparse gene set selection) and gene sets were pruned down to 91 gene sets. To prune gene sets we used a lasso penalty λ = e -2 and UMAP as the input embedding.…”

“…The identification of cell types, states and gene programs allows researchers to decipher the mechanisms of disease, empowering the development of new drugs and therapies. Cell types and states can be identified using various computational methods 2,3,4,5 that transfer annotations from single-cell references 6,7,8 , by means of tissue-specific cell type markers 9,10 or classifiers 3,4 . For example, Seurat V4 6 and SingleR 7 leverage existing reference scRNAseq datasets for cell type annotation, whereas CellAssign 9 and Garnett 10 can deliver the same task with cell type markers.…”

Assessing the limits of zero-shot foundation models in single-cell biology

“…In this study, we assessed the zero-shot performance of two proposed foundation models in single-cell biology: Geneformer [5] and scGPT [6]. We selected these models as representative examples in a rapidly evolving field that includes other approaches like scBERT [4], scFoundation [7], SCimilarity [8], and GeneCompass [9]. Our assessment covers a range of tasks, including the utility of embeddings for cell type clustering, batch effect correction, and the effectiveness of the models’ input reconstruction based on the pretraining objectives (Fig.…”