🧬 biology

Mosaic Cell Relation Factorization for Single-Cell Multi-omics Integration Across Simulated Populations and Dataset Sizes

This paper introduces Mosaic Cell Relation Factorization (MCRF), a method for integrating single-cell RNA and chromatin accessibility data that demonstrates favorable performance and stability on smaller simulated datasets but shows limitations compared to existing tools like UINMF and MIDAS on larger datasets, highlighting the need for further evaluation on experimental data.

Original authors: Sandeep chowdary Vejandla, Hemant Tiwari

Published 2026-10-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sandeep chowdary Vejandla, Hemant Tiwari

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

In the microscopic world of living cells, scientists are increasingly trying to read the full story of life by looking at multiple chapters at once. For decades, researchers could only examine one type of molecular instruction at a time, such as the genes that are actively being read to make proteins, or the chemical switches that turn those genes on and off. Today, technology allows scientists to measure both of these layers inside the same tiny cell. However, a practical problem often arises: in many experiments, some cells have both sets of measurements, while others have only one. It is like trying to assemble a complete puzzle when some pieces show the sky and others show the grass, but no single piece shows the whole picture. To make sense of this incomplete mosaic, researchers need a way to connect the dots, figuring out which cells belong together based on the fragments they have, so they can understand how different cell types function and change during health and disease.

A team of researchers at the University of Alabama at Birmingham has developed a new method to solve this specific puzzle, which they call Mosaic Cell Relation Factorization. Instead of trying to guess the missing molecular measurements for cells that only have partial data, this approach focuses on how similar cells are to one another. Imagine looking at a crowd of people where some are holding maps of the city and others are holding photos of the buildings. Rather than trying to draw the missing maps for the people with photos, the new method looks at how the people with maps resemble each other, how the people with photos resemble each other, and then finds the common thread that links the two groups together. By analyzing these patterns of similarity, the method creates a unified map of the entire cell population, allowing scientists to group cells into their correct biological families even when the data is incomplete.

To test if this new approach worked better than existing tools, the researchers created a massive set of simulated experiments. They generated 246 different datasets, ranging from small collections of 500 cells to large ones with 10,000 cells. In these simulations, the scientists knew the true identity of every cell beforehand, which allowed them to act as strict judges. They compared their new method against three other established techniques that are currently used to merge these different types of biological data. The goal was to see which method could most accurately sort the cells into their correct groups, and which one produced the most consistent results when the experiment was run multiple times with slight variations.

The results revealed a nuanced picture of performance that depended heavily on the size of the data and the type of biological structure being studied. In the smaller datasets, which contained fewer cells and features, the new method performed exceptionally well at capturing the overall information shared between the different groups. It was particularly good at preserving the subtle relationships that define how cells are related to one another, often outperforming the older methods in these specific conditions. Furthermore, when the researchers ran the simulations multiple times, the new method produced very stable results, showing little variation from one run to the next. This consistency suggests that the method is robust and reliable when working with smaller, more manageable amounts of data.

However, the story changed when the researchers tested the method on the larger datasets containing thousands of cells. In these more complex scenarios, while the new method still improved its own scores, two of the older methods proved to be better at sorting the cells into the exact groups the scientists expected. The new method remained consistent, but it did not reach the same level of accuracy in identifying the specific cell pairs that belonged together as the other tools did. This finding highlights a limitation: while the new approach is excellent at finding general patterns and relationships, it may struggle to match the precision of established methods when dealing with very large and complex biological populations.

The researchers also discovered that the structure of the cell population mattered greatly. The new method worked best when the cells formed distinct, separate groups, but it faced more challenges when the cells existed along a smooth, continuous path of change, such as during the gradual process of a cell maturing. Additionally, when the biological differences between cells were very subtle, the method's ability to distinguish them decreased, a challenge that affected all the tools they tested. These findings suggest that while the new method offers a powerful and unique way to look at cell relationships, it is not a one-size-fits-all solution. It excels in specific situations, particularly where data is limited or where understanding the broad relationships between cells is the priority, but it may need further refinement to handle the most massive and complex datasets.

Ultimately, this work provides a clearer understanding of how to handle the messy, incomplete data that is common in modern biology. The new method offers a fresh perspective by focusing on the connections between cells rather than just the raw measurements, proving that there are multiple valid ways to reconstruct the biological story from fragmented clues. While it is not yet the perfect tool for every situation, its ability to provide stable and informative results in many conditions makes it a valuable addition to the scientific toolkit. The next step for the researchers is to test these methods on real-world biological samples to see if the patterns observed in the simulations hold true in the complex reality of living tissues.

Technical Summary: Mosaic Cell Relation Factorization (MCRF) for Single-Cell Multi-omics Integration

Problem Statement
Single-cell multi-omics integration faces a "mosaic" challenge: datasets often contain cells with paired modalities (e.g., RNA and ATAC from the same cell) alongside cells with unpaired modalities (RNA-only or ATAC-only). Existing integration methods must reconcile these incomplete observations to create a unified representation that preserves biological differences between cell types and states while correcting for technical batch effects. Current approaches, such as Unshared Integrative Nonnegative Matrix Factorization (UINMF), COBOLT, and MIDAS, utilize diverse strategies including matrix factorization of features, variational autoencoders, and probabilistic frameworks. However, a comprehensive benchmark across varying population structures, dataset sizes, and simulation conditions is required to evaluate their relative strengths and limitations, particularly for methods that do not rely on reconstructing unobserved molecular measurements.

Methodology
The authors introduce Mosaic Cell Relation Factorization (MCRF), a novel integration method that shifts the focus from factorizing raw molecular measurements to factorizing observed cell-cell similarities.

Core Concept: Instead of directly factorizing a cell-by-feature matrix ( $X \approx UV$ ), MCRF constructs a cell similarity matrix $K = X^T X$ (using cosine similarities of normalized, log-transformed counts) and factorizes this relation matrix. The model posits that observed similarities can be explained by a low-rank structure: $K \approx Z Q Z^T + \beta \mathbf{1}\mathbf{1}^T$ , where $Z$ represents cell factors, $Q$ is a modality-specific metric matrix, and $\beta$ is an intercept accounting for constant residual offsets.
Mosaic Design: The method handles three batches: Batch 1 (paired RNA/ATAC), Batch 2 (ATAC-only), and Batch 3 (RNA-only).
- Paired cells in Batch 1 share a single factor representation ( $Z_1$ ) across both modalities, linking the RNA and ATAC relation blocks.
- Unpaired cells in Batches 2 and 3 have modality-specific relations constrained by the shared factors of the paired cells.
- Separate modality matrices ( $Q_R$ and $Q_A$ ) allow the shared factors to contribute differently to RNA and ATAC similarities.
Optimization: The objective minimizes a normalized reconstruction loss (analogous to $1 - R^2$ ) across all six observed relation blocks (within and between batches for both modalities). The model uses Adam optimization with specific initialization strategies and selects the best solution based on the unsupervised objective, without using population labels during training.
Embedding: The final integrated embedding is derived by applying the square root of the averaged, trace-normalized modality matrices to the cell factors ( $E = Z Q^{*1/2}$ ), ensuring that distances in the embedding space reflect the learned modality-specific weighting.

Benchmarking Framework
The study evaluates MCRF against UINMF, COBOLT, and MIDAS using two collections of simulated data generated by scMultiSim:

Original Collection: 144 datasets with 500–800 cells and 110–500 genes.
Larger Collection: 102 datasets with 2,000–10,000 cells and 1,000–2,000 genes.

Variables: Simulations vary by population structure (linear, branching, discrete), Cell Identity Factor (CIF) variability ( $\sigma_{CIF} = 0.1, 0.5$ ), and simulation replicates.
Metrics: Clustering performance is assessed using the Adjusted Rand Index (ARI) and Normalized Mutual Information (NMI) against known ground-truth population labels. Resolution selection for clustering is optimized using NMI on the reference labels.

Key Results

Performance in Original Collection: MCRF demonstrated favorable NMI performance, outperforming UINMF in 22 of 36 configurations and MIDAS in 28 of 36. It showed particular strength in continuous and branching populations. However, its ARI scores were generally lower than competitors, especially in linear configurations.
Consistency: MCRF frequently exhibited lower variability (smaller interquartile ranges) across simulation replicates compared to UINMF and MIDAS, suggesting greater stability in its clustering assignments under fixed simulation conditions.
Performance in Larger Collection: In the larger datasets, MCRF's absolute scores generally improved, but its relative performance declined. UINMF and MIDAS achieved higher median ARI in all 24 configurations with four replicates. MCRF's NMI advantage was minimal or non-existent in these larger settings.
Comparison with COBOLT: MCRF consistently outperformed COBOLT in median NMI across all 60 replicated configurations, though COBOLT's performance in other benchmarks (e.g., batch mixing) suggests its lower scores here may be specific to this evaluation setup.
Impact of Variability: Higher CIF variability generally reduced MCRF's scores, particularly for NMI, mirroring trends seen in other methods.

Significance and Claims
The paper positions MCRF as a promising approach to mosaic integration that leverages cell relation factorization rather than direct feature reconstruction.

Strengths: The method effectively retains information about reference population membership (high NMI) and produces consistent results across replicates, particularly in smaller datasets and complex population structures (branching/continuous). It achieves integration without converting ATAC regions into gene activity features or imputing unobserved modalities.
Limitations: The authors acknowledge that MCRF's clustering agreement (ARI) lags behind established methods like UINMF and MIDAS in larger datasets. The paper explicitly states that the current evaluation relies on simulated ground truth and does not independently assess batch removal, regulatory relationship recovery, or biological validity beyond clustering agreement.
Conclusion: The study concludes that while cell relation factorization is a viable strategy for mosaic integration, its application to larger datasets requires further optimization. The authors emphasize that evaluation on experimental data is the necessary next step to establish the method's value in biological research.

Technical Summary: Mosaic Cell Relation Factorization (MCRF) for Single-Cell Multi-omics Integration

More like this