💻 bioinformatics

Move BeTween modAlities (MBTA) employs flow matching to predict single cell data modalities

The paper introduces Move BeTween modAlities (MBTA), a novel framework that utilizes flow matching to connect modality-specific latent spaces and address structural mismatch in single-cell data, thereby enabling accurate cross-modal translation while preserving the unique structural integrity of each molecular modality.

Original authors: Xu, B., Zhang, Y., Michor, F.

Published 2026-08-09

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Xu, B., Zhang, Y., Michor, F.

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 ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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

Imagine trying to understand a person by looking at three different snapshots: a photo of their face, a recording of their voice, and a scan of their fingerprints. Each picture tells a unique story, but they don't always line up perfectly. The person might look friendly in the photo (a "neighbor" in the visual world), but their voice might sound grumpy to someone who knows them well (a "neighbor" in the audio world). In the world of biology, scientists are trying to do the same thing with cells. They want to take a "snapshot" of a single cell using different molecular cameras: one that reads the cell's instructions (RNA), one that checks how tightly packed its DNA is (chromatin), and one that counts its surface proteins. The goal is to stitch these snapshots together to see the whole cell. But here's the tricky part: the cell doesn't look the same in every snapshot. A cell that is a close neighbor to another in the RNA world might be a stranger in the protein world. This mismatch is the big puzzle scientists are trying to solve, because if you force these different views to look identical, you might erase the very details that make the cell interesting.

Enter MBTA (Move BeTween modAlities), a new computer program designed by researchers to solve this puzzle. Think of MBTA as a super-smart subway map for cells. Instead of trying to force the RNA, protein, and DNA snapshots into one giant, messy room where everything looks the same, MBTA builds separate, perfect rooms for each type of data. Then, it constructs a high-speed, magical train track (called "flow matching") that connects these rooms. If you drop a cell into the RNA room, MBTA can instantly calculate exactly where that same cell would be in the protein room, without squishing or stretching the data to make it fit. The researchers tested this on real data from human breast cancer and mouse embryos, and found that MBTA was much better at translating between these different molecular languages than older methods. It didn't just guess; it learned the specific rules of how a cell's RNA changes its protein shape, even when those changes were complex and non-linear. By keeping the unique "neighborhoods" of each data type intact while still connecting them, MBTA helps scientists see the full, un-distorted picture of how cells grow, change, and sometimes turn into cancer.

The Problem: When Neighbors Don't Agree

In the universe of single-cell biology, scientists have become amazing at taking pictures of individual cells. They can read the cell's RNA (the instructions), check its DNA accessibility (how open the books are), and count its surface proteins (the ID badges). For a long time, the standard way to combine these pictures was to mash them all into one "shared space," like putting all the photos of a person into a single collage. The idea was that if you blend them enough, you get the "true" version of the cell.

But the authors of this paper discovered a hidden flaw in this approach, which they call structural mismatch. Imagine you are at a party. In the "music room," you are standing next to your best friend because you both love jazz. But in the "food room," you are standing next to a stranger because you both love spicy tacos. If you try to force the music room and the food room to be the same room, you have to move your best friend away from you, or move the taco-lover closer, just to make the room fit. You lose the truth of who your neighbors actually are in each specific context.

The researchers found that this happens all the time in cells. A cell's nearest neighbors in the RNA world are often not its nearest neighbors in the protein world. This isn't a mistake; it's a feature! It means each type of data is capturing a different, unique aspect of the cell's life. When older computer programs tried to force these different views into a single shared space, they accidentally erased these differences, smoothing out the unique details that make biology so interesting.

The Solution: A Subway System for Cells

To fix this, the team built MBTA. Instead of forcing all the data into one room, MBTA builds a separate, custom room for each type of data (RNA, proteins, DNA, etc.). It uses a special tool called a Variational Autoencoder (VAE) to shrink each complex dataset down into a neat, manageable map.

Then comes the magic part: Flow Matching. Imagine these separate rooms are train stations. MBTA doesn't just guess how to get from the RNA station to the Protein station; it learns the exact "flow" or current of the river connecting them. It trains a neural network to understand the path a cell takes as it moves from one state to another. This allows the computer to translate a cell from its RNA map to its Protein map with incredible precision, without ever forcing the two maps to look the same.

The beauty of MBTA is its modularity. It's like a subway system where you can add a new line (a new type of data) without having to rebuild the whole city. You can train the "RNA to Protein" line and the "RNA to DNA" line separately, and they work together perfectly. This makes it incredibly fast and efficient, even when dealing with dozens of different molecular measurements at once.

What They Found: Better Maps, Hidden Secrets

The researchers put MBTA to the test against other popular methods (like multiVI, multigrate, and GLUE) using a variety of real-world datasets, including human immune cells, brain cells, and breast cancer samples.

It's More Accurate: In almost every test, MBTA was better at reconstructing the original data and translating it into other forms. While other methods often created "blurry" versions of the cell or lost important details, MBTA kept the sharp edges. It was especially good at handling cases where the structural mismatch was high—exactly the situations where other methods failed.
It Keeps the Truth: The study showed that older methods often forced cells that were different to look the same, or split cells that were the same into different groups just to make the math work. MBTA respected the natural structure of the data. For example, in a dataset of blood stem cells, other methods created fake subgroups that didn't exist in reality, while MBTA correctly identified the true cell types.
It Can Predict the Unseen: The team showed that MBTA could learn the rules of how cells pair up, even if it only saw a few examples. If they hid the pairing information for certain cell types, MBTA could still guess the correct connections for the unseen cells, proving it learned the underlying biology rather than just memorizing the data.
It Reveals Hidden Cancer Patterns: When applied to breast cancer data, MBTA did something remarkable. It could translate RNA data into copy number profiles (which show if parts of the genome are missing or duplicated) more accurately than existing tools. This helped them spot hidden lineage relationships in tumors that other methods missed. They found that certain genes were highly sensitive to changes in their DNA copy number, giving new clues about how these tumors evolve.
It Can Model Time: Finally, the researchers used MBTA to track the development of a mouse embryo. They took gene expression data, evolved it forward in time using a separate tool, and then used MBTA to translate that future gene expression into seven different epigenetic marks (chemical tags on DNA). The result was a complete, moving portrait of a developing embryo, showing how different molecular layers change together over time.

Why It Matters

This paper suggests that the old way of thinking—forcing all data into one shared box—might be holding us back. By acknowledging that different molecular views of a cell have different "neighborhoods," MBTA offers a way to keep the unique character of each view while still connecting them. It's not just a better calculator; it's a new way of seeing cells that respects their complexity. Whether it's figuring out how a tumor grows or how an embryo develops, MBTA provides a clearer, more faithful map of the cellular world, helping scientists ask better questions and find answers that were previously hidden in the noise.

Technical Summary: Move BeTween modAlities (MBTA)

Problem Statement
Integrating diverse molecular modalities (e.g., transcriptomics, epigenomics, proteomics) to construct a comprehensive view of cellular identity remains a significant challenge in single-cell biology. The authors identify a fundamental, underappreciated obstacle termed structural mismatch. This phenomenon occurs when the neighborhood structure of a cell—defined by its nearest neighbors in a low-dimensional space—differs depending on which molecular modality is used to define it.

Existing multimodal integration methods (e.g., multiVI, GLUE, multigrate) typically rely on embedding all modalities into a shared latent space. The authors argue that this approach implicitly assumes a unified geometry that captures the true underlying cell state across all modalities. However, when structural mismatch exists (i.e., when modalities capture distinct, non-equivalent aspects of cell state), forcing them into a shared representation actively erases the structural differences that make multimodal measurements scientifically valuable. This leads to the distortion of modality-specific signals, misrepresentation of biological variation, and the creation of artificial subpopulations in the latent space that do not exist in the raw data.

Methodology: MBTA Framework
To address structural mismatch, the authors introduce Move BeTween modAlities (MBTA), a deep learning framework that avoids forcing modalities into a shared representation. Instead, MBTA maintains modality-specific latent spaces and connects them via flow matching.

Modality-Specific Latent Spaces: Each modality is first projected into its own low-dimensional latent space using a Variational Autoencoder (VAE). This allows each latent space to scale appropriately with its input dimensionality (e.g., accommodating the vast difference between surface protein counts and chromatin accessibility peaks) without under-representing high-dimensional data.
Flow Matching for Translation: Rather than adversarial training or distribution alignment that forces convergence, MBTA learns a nonlinear mapping between the latent spaces of two modalities using flow matching. The system is represented as a neural ordinary differential equation (ODE), where the flow field $f_{z_1 \to z_2}(x, y, t)$ is learned directly from data. Solving the ODE $dy/dt = f(x, y, t)$ yields the corresponding latent coordinate in the target modality ( $z_2$ ) for a cell with a known coordinate in the source modality ( $z_1$ ).
Handling Missing Data and Pairing:
- Co-profiled Data: When paired data exists, the flow matching component is trained directly on the paired latent coordinates.
- Unpaired/Structured Missingness: MBTA can generalize relationships learned from co-profiled cell types to cell types lacking paired measurements, provided cell type annotations are available. It can also operate in settings where only a subset of cells is co-profiled.
Modular and Parallel Architecture: The framework is fully modular. VAEs are fitted to their respective modalities independently, and flow matching models are trained on modality pairs without modifying the underlying VAEs. This allows for parallel training, making the computational cost of integrating $N$ modalities comparable to integrating two, rather than scaling quadratically.

Key Contributions

Identification of Structural Mismatch: The paper formally defines and quantifies structural mismatch as a prevalent property of multi-omic datasets, demonstrating that it is a biological signal rather than a technical artifact.
Novel Architecture: MBTA is presented as the first framework explicitly designed to preserve modality-specific neighborhood structures while enabling cross-modal translation, bypassing the limitations of shared latent space assumptions.
Scalability: The modular design allows for the integration of an arbitrary number of modalities and the handling of datasets with varying degrees of co-profiling.

Results and Benchmarks
The authors evaluated MBTA against state-of-the-art methods (multiVI, multigrate, GLUE, scVI) across diverse datasets, including mouse brain cells (RNA + ATAC), peripheral blood mononuclear cells (RNA + protein), hematopoietic stem cells (RNA + protein), and breast cancer samples (RNA + DNA).

Reconstruction and Translation Accuracy: MBTA consistently outperformed existing methods in data translation tasks. While methods like multigrate performed well in reconstruction, they struggled with translation. MBTA achieved the lowest mean squared error (MSE) and Cramér–von Mises (CvM) statistics in translation, indicating a better match to observed data distributions.
Preservation of Structure: Quantitative analysis of KNN graphs showed that MBTA preserved the distinct neighborhood structures of individual modalities. In contrast, shared-space methods (multiVI, multigrate) produced reconstructed modalities that were artificially similar to each other, exceeding the similarity observed in the original data.
Cell State Classification: In datasets with pronounced structural mismatch (e.g., HSPC data with tree-like RNA structure vs. protein structure), shared-space methods failed to separate distinct cell types, often splitting them into artificial clusters. MBTA successfully separated cell types by concatenating modality-specific latent coordinates.
Downstream Applications:
- Tumor Lineage Inference: Applied to ER+ breast cancer samples, MBTA successfully inferred copy number alterations (CNAs) from scRNA-seq data, outperforming standard tools like inferCNV and copyKat. It identified transcriptomic lineage relationships corroborated by genomic variation and pinpointed dosage-sensitive genes (e.g., MDK, PRLR).
- Temporal Modeling: In mouse embryonic development, MBTA integrated seven epigenetic modalities with scRNA-seq. It enabled an "evolve and translate" strategy: evolving a tractable modality (gene expression) over time using a dynamical model (MIOFlow) and then translating the result to difficult-to-model modalities (e.g., H2A.Z), achieving higher accuracy than direct evolution of the difficult modality.
- Coupling Quantification: MBTA was used to quantify the strength of coupling between modalities (e.g., gene expression vs. protein abundance), revealing cell-type-specific decoupling (e.g., post-translational regulation in T cells vs. stability in erythroid cells).

Significance and Claims
The authors claim that MBTA provides a computational foundation for assembling multi-layered portraits of cells without erasing their individual character. By explicitly addressing structural mismatch, MBTA avoids the "ill-posed problem" of forcing incompatible geometries into a single coordinate system.

The paper positions MBTA not merely as an imputation tool but as a paradigm for integrating the full spectrum of cellular measurements—from transcriptomics and epigenomics to proteomics—into a unified, biologically faithful model. The authors emphasize that MBTA enables the construction of "virtual cells" capable of autonomous temporal evolution and allows researchers to perform in silico perturbation experiments to infer causal relationships between modalities (e.g., how copy number changes drive gene expression). The framework is presented as robust to annotation noise and capable of scaling to arbitrary numbers of modalities through its parallelizable, modular architecture.

The Problem: When Neighbors Don't Agree

The Solution: A Subway System for Cells

What They Found: Better Maps, Hidden Secrets

Why It Matters

Technical Summary: Move BeTween modAlities (MBTA)

More like this