CHACAM: a cell-cell interaction-guided hierarchical attention model for high-precision cell identity annotation of scRNA-seq data in early C. elegans embryogenesis

Abstract

Accurate cell identity annotation is essential for reconstructing cell lineages and gene-regulatory networks in embryogenesis. However, in Caenorhabditis elegans (C. elegans) single-cell RNA-sequencing (scRNA-seq) atlases, closely related sub-lineage cells often have near-identical transcriptomes, leading to merged labels and limiting downstream analyses. We present CHACAM (Cell-Cell Interaction-guided Hierarchical Attention-based Cell Allocation Model), a supervised machine learning framework that integrates gene expression, a curated C. elegans ligand-receptor (L-R) interaction database, and a high-resolution cell contact map to refine cell annotations. CHACAM learns interpretable "interaction signatures" that distinguish contacting vs. non-contacting neighbors and applies a triangulation strategy using reference cells to resolve ambiguous sister identities. Applied to the Tintori and Cole-Yanai embryo atlases, CHACAM achieves near-perfect annotation accuracy, resolving longstanding ambiguities of the AB lineage and other lineages. The model also identifies new discriminative marker genes and provides mechanistic hypotheses linking contact-specific signaling to fate specification.