Cross-species cell-type assignment from single-cell RNA-seq data by a heterogeneous graph neural network

Abstract

Cross-species comparative analyses of single-cell RNA sequencing (scRNA-seq) data allow us to explore, at single-cell resolution, the origins of cellular diversity and evolutionary mechanisms that shape cellular form and function. Cell-type assignment is a crucial step to achieve that. However, the poorly annotated genome and limited known biomarkers hinder us from assigning cell identities for nonmodel species. Here, we design a heterogeneous graph neural network model CAME to learn aligned and interpretable cell and gene embeddings for cross-species cell type assignment and gene module extraction from scRNA-seq data. CAME achieves significant improvements in cell-type characterization across distant species due to the utilization of non-one-to-one homologous gene mapping ignored by early methods. Large-scale benchmarking study showed that CAME significantly outperformed four classical methods in terms of cell-type assignment and model robustness to insufficiency and inconsistency of sequencing depths. CAME could transfer the major cell types and interneuron subtypes of human brains to mouse ones and discover shared cell type-specific functions in homologous gene modules. CAME could well align the trajectories of human and macaque spermatogenesis and reveal their conservative expression dynamics. In short, CAME can make accurate cross-species cell type assignments even for nonmodel species, and uncover shared and divergent characteristics between two species from scRNA-seq data.

GeneCompass: Deciphering Universal Gene Regulatory Mechanisms with Knowledge-Informed Cross-Species Foundation Model

“…To explore the capability of GeneCompass on cross-species downstream tasks, we integrated GeneCompass with the SOTA method CAME 28 for cross-species cell type annotation. Gene embeddings generated by GeneCompass are utilized as initial gene-nodes features within CAME.…”

“…We aim to perform integration and cell-type assignment while preserving biological variability by utilizing the universal gene embeddings from our generative pre-trained model. We apply GeneCompass to CAME 28 which is a heterogeneous graph neural network called GeneCompass-CAME, where cells and genes are modeled as heterogeneous nodes. Also, like CAME, we create the heterogeneous graph with six heterogeneous types: ‘cell to gene’, ‘gene to cell’, ‘cell to cell’, ‘gene to gene’, ‘cell self-loop’, ‘gene self-loop’, where we denote the corresponding weights (shared across species) as W cg , W gc , W cc , W gg , W c and W g , respectively.…”

Unraveling Neuronal Identities Using SIMS: A Deep Learning Label Transfer Tool for Single-Cell RNA Sequencing Analysis

“…However, some neuronal subtypes, such as MGEderived INs, are transcriptomically more conserved across evolution than other primary neurons, including cortical PNs [13,44]. In the future, these IN subtypes could be used as a way to validate SIMS to perform trans-species predictions [96]. Additional modifications, such as gene module extraction could provide increased accuracy for label transfer, as meta-modules could prove to be more conserved between evolutionary distant species than gene orthologs [92,97,98].…”

Representing and extracting knowledge from single-cell data

“…There are cases when two orthogonal graphs are being estimated. One such case is the comparison of cells between species (Liu et al 2023 ); the traditional Euclidean distance between cells is problematic because it is not clear which genes in species A should be compared to which genes in species B. However, assuming that the cells are lined up correctly, and with some knowledge of homology (based on gene sequences), it is possible to find which genes correlate and thus correspond.…”