Exploring structured molecular landscape from single-cell multi-omics data by an explainable multimodal model

Abstract

There is an urgent need to understand the molecular landscape beyond the conventional cellular landscape, maximizing the translational use and generalized interpretation of state-of-the-art single-cell genomic techniques in biological studies. We introduced a multimodal explainable artificial intelligence (xAI) model Vec3D to identify a joint definition of cellular states and their distribution in a quantified graphic organization as structured molecular landscape (SML). First, Vec3D substantially improves the accuracy and efficiency of multimodal data analysis. Further, an SML was learned on CITE-seq data of human peripheral blood mononuclear cells (PBMCs), simultaneously revealing the predictive multi-label cell state and corresponding joint cell state markers with complementary effects from genes and proteins. Third, Vec3D demonstrated that the spatial-temporal SML efficiently characterizes molecular dynamics of cell lineages during human lung development. Collectively, Vec3D will be a broadly applicable computational method in the principle of “AI-for-biology”, providing a unified framework for understanding cellular homeostasis and imbalance through SML dynamics.