MOSAIC: A Spectral Framework for Integrative Phenotypic Characterization Using Population-Level Single-Cell Multi-Omics

---

MOSAIC (Multi-Omic Sample-wise Analysis of Inter-feature Connectivity) is a spectral framework designed to learn high-resolution feature $\times$ sample joint embeddings from population-scale single-cell multi-omics data.

Unlike traditional methods that focus on cell embeddings, MOSAIC explicitly models how feature relationships (e.g., Gene-Peak, Protein-Gene) vary across individuals. This enables the detection of regulatory network rewiring and cryptic patient subgroups.

Key Capabilities

• Joint Feature $\times$ Sample Embedding: Projects features and samples into a shared latent space.
• Differential Connectivity (DC) Analysis: Identifies features (genes, proteins) that change their regulatory context between conditions, even without changes in abundance.
• Unsupervised Subgroup Detection: Discovers patient subtypes driven by specific, coherent multi-modal feature modules.
• Scalable & Robust: Linear complexity with respect to sample size (O(S)) and quadratic with respect to features (O(F^2)), utilizing truncated eigendecomposition. Robust to batch effects without explicit correction.

---

Example tutorial

• The MOSAIC tutorial, available in MOSAIC_demo.Rmd and MOSAIC_demo.html.
• The functions defined MOSAIC and downstream analysis are available in MOSAIC_function.R.
• Dataset used in the tutorial can be downloaded through this link.