scMultiNODE: Recovering Developmental Dynamics in Weakly Resolved Modalities through Multi-Modal Integration

Abstract

Measuring multiple genomic profiles (i.e., modalities) over time enhances our understanding of cell development. However, these modalities are not equally analyzable. Sparse and poorly separated measurements, such as chromatin accessibility, are difficult to interpret on their own, and the limited availability of temporal single-cell co-assay data leaves such modalities without the cross-modal cell correspondence needed to borrow structure from richer ones. As a result, critical developmental analyses and cross-modal comparisons remain underperformed for these weak modalities. Additionally, existing methods cannot integrate separately sequenced measurements while preserving cellular dynamics. We present scMultiNODE, an integration model that combines gene expression and chromatin accessibility measurements from multiple discrete timepoints for single cells, without relying on any prior cell-to-cell correspondence across modalities or time. scMultiNODE employs a scalable Quantized Gromov-Wasserstein optimal transport method to align cells across measurements, and neural ordinary differential equations with a dynamic regularization term to model cell development in a shared latent space, capturing cellular dynamics while preserving cell-type variations. Experiments across four developmental single-cell datasets demonstrate that scMultiNODE integrates separately sequenced measurements over time more effectively than existing methods that overlook cellular dynamics. Crucially, by anchoring chromatin accessibility to the shared dynamical latent space, scMultiNODE recovers developmental trajectories and cell annotations from scATAC-seq that this modality cannot support in isolation. The resulting joint latent space further supports important multi-modal downstream analyses, including complex trajectory investigation and cross-modal label transfer for cell annotation. The data, code, and supplementary notes are publicly available at https://github.com/rsinghlab/scMultiNODE.