SONATA: Disambiguated manifold alignment of single-cell data

Abstract

AbstractRapid advancements in single-cell sequencing technologies have empowered biologists to collect high-resolution omics profiles at the cellular level. Integrating such heterogeneous single-cell data, essential for gaining deeper insights into biological processes, is a formidable challenge. The difficulty arises from the lack of shared cells or features across distinct measurements, compounded by biological and technical variations. To address this, manifold alignment methods have emerged as a crucial tool to align heterogeneous single-cell measurements in a shared space. However, current methods recklessly report an arbitrary alignment solution without considering whether the data can be aligned. In this paper, we introduce SONATA, a novel computational method that utilizes a novel cell-cell ambiguity measurement to gauge the ambiguity of cross-modality correspondences. SONATA identifies mutually ambiguous cell groups, characterized by similar geometric structures, and exhaustively enumerates alignment solutions between datasets. This approach distinguishes between cases with unique solutions or multiple alternatives, ensuring robust and interpretable integration of single-cell data. Finally, we validate the effectiveness of SONATA through extensive experiments on simulated and real datasets.