PriorCCI: Interpretable Deep Learning Framework for Identifying Key Ligand-Receptor Interactions Between Specific Cell Types from Single-Cell Transcriptomes.

Understanding the interactions between specific cell types within tissue environments is essential for elucidating key biological processes, such as immune responses, cancer progression, inflammation, and development, in both physiological and pathological studies. The predominant methods for analyzing cell-cell interactions (CCI) rely primarily on statistical inference using mapping or network-based techniques. However, these approaches often struggle to prioritize meaningful interactions owing to the high sparsity and heterogeneity inherent in single-cell RNA sequencing (scRNA-seq) data, where small but biologically important differences can be easily overlooked. To overcome these limitations, we developed PriorCCI, a deep-learning framework that leverages a convolutional neural network (CNN) alongside Grad-CAM++, an explainable artificial intelligence algorithm. This study aims to provide a scalable, interpretable, and biologically meaningful framework for systematically identifying and prioritizing key ligand-receptor interactions between defined cell-type pairs from single-cell RNA-seq data, particularly in complex environments such as tumors. PriorCCI effectively prioritizes interactions between cancer and other cell types within the tumor microenvironment and accurately identifies biologically significant interactions related to angiogenesis. By providing a visual interpretation of gene-pair contributions, our approach enables robust inference of gene-gene interactions across distinct cell types from scRNA-seq data.