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

MOSAIC is a spectral framework that integrates population-scale single-cell multi-omics data into a high-resolution feature-by-sample joint embedding to enable systems-level phenotypic characterization through differential connectivity analysis, unsupervised subgroup detection, and improved clinical outcome prediction.

The Gist

Imagine you are trying to understand a massive, bustling city (the human body) by looking at its individual citizens (cells). For a long time, scientists have been counting how many people are in each neighborhood or how loud they are shouting. This is like looking at abundance: "How much of this protein is there?" or "How active is this gene?"

But there's a problem. Two people might be shouting at the exact same volume, but one is shouting to a friend to plan a party, while the other is shouting to a stranger to start a fight. The volume is the same, but the context and the relationships are completely different.

Existing computer tools for analyzing single-cell data mostly just count the volume. They miss the relationships.

Enter MOSAIC. Think of MOSAIC not as a census taker, but as a social network analyst for your cells.

The Core Idea: The "Social Graph" of a Cell

Instead of just asking "How much of Gene X is there?", MOSAIC asks: "Who is Gene X talking to right now?"

In the world of cells, genes and proteins don't work in isolation; they form complex networks. MOSAIC maps these networks. It creates a "social graph" for every single patient in a study, showing who is connected to whom.

Here is how it works, broken down into three simple steps:

1. The Snapshot (The Coupling Matrix): For every single person in the study, MOSAIC takes a snapshot of their cells and draws a map of all the relationships between their genes and proteins. It's like taking a photo of a crowded room and drawing lines between everyone who is talking to each other.
2. The Translation (Spectral Integration): Since every person's room is slightly different, the maps look different. MOSAIC uses a special mathematical trick (called "spectral decomposition") to translate all these different maps into a universal language. It finds the common patterns that exist across all people, creating a shared "playground" where every gene has a specific spot based on who its friends are.
3. The Insight (The Embedding): Now, instead of just seeing a list of gene counts, we see a map of relationships. We can see that in Person A, Gene X is best friends with Gene Y, but in Person B, Gene X has broken up with Gene Y and is now hanging out with Gene Z.

Why Does This Matter? Three Superpowers

The paper shows MOSAIC doing three amazing things that old methods couldn't do:

1. Catching the "Silent Saboteurs" (Differential Connectivity)

Imagine a factory machine (a gene) that is running at the exact same speed as always. A standard inspector would say, "Everything is fine!" But MOSAIC looks closer and sees that the machine has suddenly stopped talking to its safety sensors and started talking to the fire alarm instead.

• Real-world example: The researchers looked at T-cells (immune cells) after a vaccination. They found a gene called STAT5B. Its "volume" (expression) didn't change at all. But MOSAIC saw that it had completely rewired its connections. It stopped talking to "basal" regulators and started talking to "proliferation" and "DNA repair" teams. The cell was preparing to multiply, even though the gene's own volume hadn't changed. Old methods missed this entirely; MOSAIC caught the switch.

2. Finding Hidden Clusters (Unsupervised Subgroup Detection)

Imagine you have a group of people labeled "HIV Positive." A standard tool might say, "They are all the same." But MOSAIC looks at the relationships between their genes and says, "Wait, these 10 people are all stressed out and starving at a cellular level, while these 8 people are fine."

• Real-world example: In a group of HIV+ patients, MOSAIC found a hidden subgroup of neurons that were under extreme metabolic stress (like a city running on emergency power). These patients looked the same on a standard test, but MOSAIC revealed they were biologically different. This could help doctors treat them differently.

3. Predicting the Future (Clinical Outcome)

Imagine trying to predict if a storm will be a "Moderate Breeze" or a "Hurricane."

• Old way: You measure the wind speed (gene abundance).
• MOSAIC way: You measure the wind speed and how the wind is swirling and interacting with the trees.
• Result: When the researchers used MOSAIC to predict how severe a COVID-19 infection would be, it was better than just looking at gene counts. Even better, when they combined the "wind speed" (abundance) with the "swirling patterns" (connectivity), their prediction accuracy skyrocketed. They found patients who looked "safe" based on gene counts but were actually in danger because their cellular networks were falling apart.

The Big Picture

Think of the human body as a giant orchestra.

• Old methods just count how loud each instrument is playing.
• MOSAIC listens to the conductor and the sheet music. It understands that even if the violin is playing at the same volume, it might be playing a different song entirely because the conductor (the disease) has changed the arrangement.

By shifting the focus from "how much" to "how they connect," MOSAIC gives us a new way to see the hidden logic of disease, helping us find new treatments and better predict who is at risk. It turns a static list of parts into a dynamic movie of how life actually works.

Technical Summary

1. Problem Statement

The field of single-cell biology is evolving toward population-scale multi-omics, where datasets capture transcriptomic, epigenomic, and proteomic layers across dozens or hundreds of individuals. However, existing computational frameworks face two critical limitations:

• Cell-Centric Approaches: Methods like MOFA+ or Seurat prioritize learning a unified cell embedding to correct batch effects. While effective for cell type annotation, they treat features (genes/proteins) as passive inputs, failing to model how feature relationships (connectivity) vary across individuals.
• Feature-Centric Approaches: Methods that model cross-modal interactions often generate a global, static feature embedding shared by all samples. This "one-size-fits-all" representation masks patient-to-patient heterogeneity, obscuring biological differences essential for clinical insight (e.g., network rewiring or specific disease subtypes).

The Gap: There is a lack of a framework that is simultaneously feature-centric (modeling regulatory context) and sample-aware (capturing individual-specific variations) to bridge network-level discovery with clinical prediction.

2. Methodology: The MOSAIC Framework

MOSAIC (Multi-Omic Sample-wise Analysis of Inter-feature Connectivity) is a spectral framework designed to learn a high-resolution feature $\times$ sample joint embedding. It operates in three stages:

A. Construction of Sample-Specific Coupling Matrices

For each individual $i$ in a cohort of $S$ samples:

1. Data Integration: Multi-modal data (e.g., RNA, ATAC, Protein) are concatenated into a unified cell-by-feature matrix $X_i$.
2. Coupling Matrix ($U_i$): A symmetric, positive semi-definite matrix of size $F \times F$ (where $F$ is the total number of features) is constructed. Each entry $u_{jl}$ represents the cosine similarity between feature vectors $j$ and $l$ across all cells in sample $i$.
   • This captures the complete intra- and cross-modality network topology specific to that individual.
   • Cosine similarity normalizes for magnitude differences across modalities.

B. Spectral Integration (Aggregation-then-Decompose)

Instead of embedding each sample independently and aligning them post-hoc (which suffers from rotation/sign ambiguity), MOSAIC uses a global spectral approach:

1. Denoising: Each $U_i$ is eigendecomposed. A low-rank approximation $\tilde{U}_i$ is retained using an adaptive "elbow" detection (Kneedle algorithm) to filter noise specific to that sample.
2. Aggregation: The denoised matrices are summed: $P_{agg} = \sum \tilde{U}_i$.
3. Latent Basis Extraction: $P_{agg}$ is eigendecomposed to extract a shared latent basis $W$ (the top $d$ eigenvectors). This defines a common coordinate system for the entire population.

C. Joint Embedding Generation

Each sample's original coupling matrix $U_i$ is projected onto the shared basis $W$:
$$E_i = U_i W$$
The result is a tensor $E \in \mathbb{R}^{F \times d \times S}$, where the slice $e^{(i)}_f$ represents the embedding vector of feature $f$ in sample $i$. Crucially, the same feature has different embedding vectors across different samples, reflecting its unique functional context in each individual.

3. Key Contributions

1. Differential Connectivity (DC) Analysis: MOSAIC detects "rewiring" events where a feature's regulatory partners change between conditions, even if the feature's abundance (expression level) remains constant. This is achieved by testing for statistically significant shifts in a feature's embedding position using PERMANOVA.
2. Modular Patient Stratification: Instead of global clustering (which dilutes signals from small pathways), MOSAIC clusters features into coherent modules based on their stratification profiles. It then identifies patient subgroups driven by specific modules, revealing hidden endotypes.
3. Complementary Clinical Prediction: The framework demonstrates that connectivity profiles (topology) provide predictive information orthogonal to abundance profiles (expression levels). Integrating both improves clinical outcome prediction accuracy.
4. Robustness: By operating on correlation-based coupling matrices rather than raw counts, MOSAIC is inherently robust to batch effects and technical noise that typically plague abundance-based methods.

4. Key Results & Validation

A. Technical Benchmarking

• Embedding Quality: On simulated data with known ground-truth modules, MOSAIC recovered feature groupings as accurately as PCA and MOFA+ under low noise, and outperformed them under high noise.
• Biological Necessity: Analysis of a prefrontal cortex cohort (31 donors) confirmed that feature-feature correlation matrices vary significantly between individuals (Frobenius distance > null distribution), validating the need for sample-specific embeddings.
• Cross-Sample Comparability: In pseudo-replicate tests, MOSAIC achieved higher Silhouette Scores for donor identity recovery compared to independent embedding + Procrustes alignment (PCA, MOFA+, SIMBA), demonstrating superior preservation of biological identity without post-hoc alignment.

B. Biological Discoveries

1. T-Cell Activation (Vaccination Cohort):
   • MOSAIC identified 393 DC features in activated T cells that were missed by standard differential expression analysis.
   • Case Study (STAT5B): The transcription factor STAT5B showed no change in expression between Day 0 and Day 7. However, MOSAIC detected a complete rewiring of its network: it shifted from interacting with basal regulators (e.g., SMARCA2) to proliferation machinery (e.g., CDK13, XPC) and DNA repair factors. This revealed a coordinated proliferation program invisible to abundance-based methods.
2. HIV+ Neuronal Subtypes (PFC Cohort):
   • Applying unsupervised subgroup detection to HIV+ patients, MOSAIC identified a "cryptic" subgroup (HIV-Group1) distinct from the rest of the cohort.
   • This subgroup was characterized by a stress-driven transcriptional program (Integrated Stress Response, metabolic exhaustion) that was not detectable via global clustering or standard differential accessibility analysis.

C. Clinical Outcome Prediction (COVID-19 Severity)

• Applied to 151 COVID-19 patients, MOSAIC identified monocytes as the best predictor of severity.
• Synergy: An integrated model combining abundance (pseudobulk) and connectivity (MOSAIC) features achieved an AUC of 0.879, significantly outperforming abundance-only (0.844) or connectivity-only (0.851) models.
• Mechanistic Insight: The models identified non-overlapping predictive features. Abundance predictors were linked to oxidative stress, while connectivity predictors linked to specific immunoregulatory machinery (e.g., NLRP3 inflammasome activation via KCNK6).

5. Significance

MOSAIC represents a paradigm shift from abundance-centric to connectivity-centric analysis in single-cell multi-omics.

• Conceptual: It decouples a feature's "state" (abundance) from its "function" (connectivity), revealing regulatory rewiring that drives disease mechanisms.
• Methodological: It provides a mathematically rigorous, spectral framework for population-scale integration that avoids the pitfalls of post-hoc alignment and global averaging.
• Clinical: It offers a general-purpose toolkit for systems-level phenotypic characterization, enabling the discovery of hidden patient subtypes and improving the accuracy of clinical risk stratification by leveraging network topology as a complementary biomarker.

The framework is computationally scalable to large cohorts (linear in sample size) and is implemented as an open-source tool available on GitHub.