scMRDR: A scalable and flexible framework for unpaired single-cell multi-omics data integration

summary

The gist

Advances in single-cell sequencing have enabled high-resolution profiling of diverse molecular modalities while integrating unpaired multi-omics single-cell data remains challenging The goal of…

In short

scMRDR is a generative framework for integrating unpaired single-cell multi-omics data into a shared latent space. It uses a $\beta$-VAE with isometric regularization, adversarial training, and masked reconstruction loss to align different omics modalities while preserving unique biological cell types and correcting technical batch variations.

Key concepts

Disentangled Latent Representations
This technique separates the cell's representation into two parts: one shared across all omics (for general biological features) and one specific to each modality. This allows researchers to isolate true biological signals from technical noise and batch effects, making downstream analysis more focused on biology.
Isometric Regularization
This loss function ensures that the distances between cells are preserved across different modalities. It minimizes the difference between distance matrices calculated using the full latent representation versus those calculated using only one modality's data. This is crucial for maintaining intra-omics biological heterogeneity and preserving cell type differences.
Adversarial Training
An adversarial component uses a discriminator to try and tell which omics a latent representation came from. By forcing the model to fool this discriminator, the framework encourages cross-modal alignment, ensuring that embeddings from different data types occupy the same space in the shared latent manifold.

Terminology used across episodes

This episode discusses

The paper

scMRDR: A scalable and flexible framework for unpaired single-cell multi-omics data integration · Read on arXiv

Jianle Sun, Chaoqi Liang, Ran Wei, Peng Zheng, Lei Bai, Wanli Ouyang, Hongliang Yan,  Peng Ye

Shanghai Artificial Intelligence Laboratory

DOI: 10.52202/085713-5170

Transcript

Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.

Ines: We're looking at the paper: "scMRDR: A scalable and flexible framework for unpaired single-cell multi-omics data integration".

Marcus: Here's the gist of the paper. Advances in single-cell sequencing have enabled high-resolution profiling of diverse molecular modalities while integrating unpaired multi-omics single-cell data remains challenging The goal of multi-omics data integration is to map single-cell data in different omics into a shared latent space where representations across modalities are distributionally aligned while preserving biological differences between cell types and correcting for technical variations due to experimental batches Pruning redundant reasoning tokens In this paper, we introduce a scalable and flexible generative framework called single-cell Multi-omics Regularized Disentangled Representations (scMRDR) for unpaired multi-omics integration This method achieves excellent performance on benchmark datasets in terms of batch correction, modality alignment, and biological signal preservation The framework is flexible to completely unpaired data and able to scale to large datasets with more than two modalities naturally How the training signal is built The scMRDR framework utilizes a well-designed β-VAE architecture to disentangle each cell’s latent representations into modality-shared and modality-specific components This disentanglement is augmented with isometric regularization to preserve intra-omics biological heterogeneity Adversarial training is employed to encourage cross-modal alignment Furthermore, a masked reconstruction loss strategy addresses the issue of missing features across modalities The total optimization goal for VAE becomes Ltotal = Lrecon + βLKL + λLalignment + γLpreserve Where the gains come from The authors propose a joint optimization goal incorporating isometric regularization, adversarial training, and masked loss to facilitate modality fusion while preserving biological signals The isometric loss minimizes the discrepancy between pairwise Euclidean distance matrices computed from z and from zu for each modality The adversarial regularity introduces a m-class discriminator D(zu) to distinguish zu of samples from different omics The masked loss strategy prevents gradients from back-propagating through unmeasured features in the reconstruction loss for each modality Pruning redundant reasoning tokens The method demonstrates excellent performance in modality integration, batch correction as well as bio-conservation, surpassing a broad range of existing methods It scales robustly to large datasets and readily accommodates additional omics layers The framework is flexible to completely unpaired data and able to scale to large datasets with more than two modalities naturally Where the gains come from The performance comparisons show that scMRDR outperforms existing methods in modality integration, bio-conservation, and batch correction on real-world unpaired single-cell data The framework maintains stable performance on large-scale datasets despite the strong susceptibility to the choice of preprocessing strategy which itself markedly dependent on data scale The method shows a strong susceptibility to the choice of preprocessing strategy which itself markedly dependent on data scale Where the gains come from The framework is highly scalable for large-level single-cell data integration, enabling the informative linking among different omics and providing more valuable biological insights The method shows a strong susceptibility to the choice of preprocessing strategy which itself markedly dependent on data scale Emerging spatial multi-omics technologies allow us to capture the spatial coordinates of cells, further enriching cellular information The integration of spatial and dynamic information across modalities under different conditions will be an important direction for future exploration and extension of scMRDR Pruning redundant reasoning tokens The authors propose a principled and feasible generative model named scMRDR to integrate unpaired multi-omics single-cell data into a unified latent space The framework is flexible to completely unpaired data and able to scale to large datasets with more than two modalities naturally Where the gains come from The biological significance of integrating single-cell and spatial omics is evaluated by interpolating missing spatial locations in single-cell data by optimal transport between the aligned latent representation zu of samples with spatial locations (merFISH) and samples without spatial information This interpolation performs well where inferred locations of cells align well with the provided cortex layers annotations provided by the original datasets The framework is highly scalable for large-level single-cell data integration, enabling the informative linking among different omics and providing more valuable biological insights Pruning redundant reasoning tokens The limitations of the approach include the trade-off between different optimization goals reflected in the choices of hyperparameters β, λ, and γ The introduction of adversarial loss increases the training difficulty The authors emphasize that isometry is more essential to bio-conservation while adversary more vital to modality alignment Where the gains come from The results show that keeping all other hyperparameters unchanged, simply setting the missing features in the proteomics to zero without applying a loss mask leads to a significant performance drop This highlights the importance of our masked loss strategy in effectively handling integration tasks where some omics layers (e.g.

Title and authors: Ines: So we’re looking at this paper now, "scMRDR: A scalable and flexible framework for unpaired single-cell multi-omics data integration," and it tackles a really tough problem: how do you combine single-cell data from different types of molecules—like genomics, proteomics, whatever—when you don't have matched samples for every cell.

Marcus: Exactly. The title tells us the main idea right off the bat: it’s about making something scalable and flexible for unpaired multi-omics data integration. We're talking about combining data from different sources without needing those perfect paired measurements between each other, which is a huge hurdle in this field one <ref:2510.24987#pg1>.

Ines: Right. The authors are proposing a generative framework called scMRDR. What they’re aiming to do is map all these single-cell data points—from different omics—into one shared latent space where the representations across the different types of molecules are aligned, but crucially, they want to keep the biological differences between cell types intact.

Marcus: And that means correcting for those technical variations you get from different experimental batches too. They’re trying to find a common language for all this data so we can trust the biological signals underneath.

Yuki: From a population genetics view, I’m wondering what kind of biological signals they expect to recover when they force everything into one latent space. Does it capture things that are truly conserved across cell types, or is it just mixing noise?

Ines: That's a fair question, Yuki. The framework uses a specific architecture called a beta-VAE to do the heavy lifting of separating the latent representations into two parts: ones that are shared across modalities and ones that are specific to just one molecule type.

Marcus: So it’s trying to disentangle the data so you can isolate the biological signal from the technical noise inherent in each measurement platform. They augment this with isometric regularization, which is a key part for preserving what they call intra-omics biological heterogeneity one <ref:2510.24987#pg1>.

Yuki: Isometric regularization sounds like it’s enforcing some kind of geometric consistency between the different omics layers. Does that help them separate cell types from just technical drift?

Ines: It does. They define this isometric loss as minimizing the discrepancy between pairwise Euclidean distance matrices computed from the latent representation and those computed only from each individual modality one <ref:2510.24987#pg1>. This constraint is meant to ensure that what 'z' captures—the biological differences between samples, like different cell types—is preserved even when we’re fusing everything together.

Marcus: And they throw in adversarial training too. They use an adversarial objective to encourage cross-modal alignment, which means they have a discriminator trying to tell if a latent variable comes from one omics type or another one <ref:2510.24987#pg1>. This forces the shared latent subspace to be modality invariant across all data types.

Title and authors: Yuki: So, it’s like they’re having two groups of people try to describe the same thing, but only if their descriptions align in a certain geometric way? Is that how they see the biological relationship between these different molecular profiles?

Ines: Not exactly. The adversarial part is about forcing that alignment. But they also have this masked reconstruction loss strategy to handle missing features across modalities one <ref:2510.24987#pg1>. This prevents gradients from flowing through features that aren't actually measured in a specific layer, which helps them utilize all the available information correctly while keeping the data distribution intact.

Marcus: I’m interested in how they scale. They claim this framework is flexible for completely unpaired data and can handle more than two modalities naturally, unlike methods that rely on pre-learned coupling matrices or external prior knowledge one <ref:2510.24987#pg1>. This makes it potentially much more practical for real-world datasets.

Yuki: Scaling to large datasets is important, especially when you’re trying to connect this finding to the wider history of a species. If it can handle big cohorts, then the insights we get from these single cells could support broader population genetic questions.

Ines: The performance comparisons show that scMRDR outperforms existing methods in modality integration and batch correction on real-world unpaired data one <ref:2510.24987#pg1>. It maintains stable performance on large-scale datasets despite the fact that the choice of preprocessing strategy can still influence things depending on the data scale.

Marcus: That dependency on preprocessing is something to keep in mind. The paper shows that keeping all other hyperparameters unchanged, simply setting missing features in proteomics to zero without applying a loss mask leads to a significant performance drop one <ref:2510.24987#pg1>. This really drives home how essential that masked loss strategy is for the integration part of the task.

Yuki: So, if we think about the spatial aspect, which they mention later—spatial multi-omics—does this latent space give us anything meaningful about where these cells are located in a tissue?

Ines: They explore that further in their validation work. They use interpolation by optimal transport between the aligned representation of cells with known spatial locations and those without them, which performs well when the inferred locations match the provided cortex layer annotations one <ref:2510.24987#pg1>. This shows it has potential for spatial position imputation.

Marcus: It’s a strong result if that interpolation works reliably. It suggests that linking different omics layers through this shared latent space provides more valuable biological insights than just looking at each one in isolation.

Title and authors: Yuki: If the method can link genomics to proteomics and still recover meaningful spatial data, then it opens up new ways to connect molecular function directly to tissue architecture across different species.

Ines: The authors propose a principled way, scMRDR, to integrate these unpaired multi-omics single-cell data into one unified latent space one <ref:2510.24987#pg1,unpaired multi-omics single-cell data>. It’s a framework that is flexible enough for completely unpaired data and scales naturally to more than two modalities.

Marcus: Ultimately, the main gain here is providing a robust method for linking different omics layers while preserving the biological signals you care about most, all while handling technical noise across batches.

Ines: The limitations they point out are pretty standard for generative models: there’s a trade-off between the different optimization goals, specifically with the choices of hyperparameters like beta, lambda, and gamma one <ref:2510.24987#pg1>. They note that adversarial loss increases training difficulty but is vital for alignment, while isometry seems more important for bio-conservation.

Yuki: So it’s a balancing act: you can get better alignment by adding complexity to the training objective, but you might lose some of that fine-grained biological detail if you push the regularization too hard.

Marcus: That makes sense. It tells us we need to be careful tuning those parameters based on what we prioritize—whether it's perfect technical alignment or absolute biological fidelity.

Ines: So, to wrap up, scMRDR is a scalable and flexible framework for integrating unpaired single-cell multi-omics data one <ref:2510.24987#pg1,a scalable and flexible framework for>. It uses disentangled representations with isometric regularization, adversarial training for alignment, and masked loss to achieve modality fusion while preserving biological signals.

Yuki: I think the big implication here is that we can start building richer biological models by seeing how different molecular layers interact in a unified space, regardless of whether we have paired data for every single comparison.

Marcus: Yeah, it moves us closer to integrating more complex, real-world multi-omics studies without needing those restrictive paired measurements that used to limit these kinds of analyses one <ref:2510.24987#pg1>. It’s a practical step forward for genomics data scientists working with massive cohorts.

Ines: This paper on scMRDR really shows how we can build generative models that are both powerful enough to handle complexity and flexible enough to adapt to the real challenges in single-cell biology.

Yuki: It’s interesting because it sets a new standard for how we think about data integration when the raw input data is inherently messy and incomplete across different scales.

Marcus: We're looking forward to seeing how this framework gets applied, especially as spatial multi-omics technologies develop, because linking that spatial info with these molecular layers could be really powerful.

Ines: That’s what we’ll be watching next. So that’s our take on scMRDR for now.

The paper's summary: Ines: So, basically, this paper introduces something called scMRDR as a way to take single-cell data from different molecular sources—like genomics and proteomics—and put them into one shared space without needing matched samples for every cell.

Marcus: Right. The core idea is mapping all these different omics data points into a single latent space where the representations across modalities are aligned, but they have to keep the biological differences between cell types intact, you know?

Yuki: So we’re talking about finding this common language that respects the biology while fixing those technical batch effects from different experiments. That sounds like a big hurdle for any integration project.

Ines: Exactly. The framework uses a specific architecture called a beta-VAE to separate the cell's latent representation into parts that are shared across all molecules and parts that are only specific to one molecule type.

Marcus: And they add isometric regularization there, which is key for preserving what they call intra-omics biological heterogeneity, meaning the differences within each omics layer itself.

Yuki: So you’re ensuring that the geometric structure of the data—the distance between cells—remains consistent even after you merge all those different data types together.

Ines: That's right. They minimize a loss that keeps the pairwise Euclidean distance matrices from 'z' and 'zu' consistent across all modalities, which helps capture those biological differences like cell types.

Marcus: Then they also use adversarial training with a discriminator to force cross-modal alignment, pushing the shared latent subspace to be modality invariant across everything.

Yuki: So you’re trying to make sure that the embeddings from genomics and proteomics end up in the same neighborhood in this unified space, regardless of which molecule they came from.

Ines: And they use a masked reconstruction loss strategy to deal with missing features across modalities, preventing gradients from flowing through things that aren't actually measured for every single layer.

Marcus: So it’s trying to get all the information you can use while still keeping the underlying data distribution intact, even when some data is incomplete.

Yuki: It’s a principled generative model, scMRDR, designed to scale and handle more than two omics naturally without needing those pre-learned coupling matrices that limit older methods.

Ines: The performance results show it handles modality integration and batch correction on real unpaired data better than existing methods, even on large datasets.

Marcus: It maintains stable performance on those large-scale datasets, which is important because you usually expect things to break down when you scale up without the right architecture.

Yuki: They also look at spatial multi-omics, using optimal transport to interpolate missing spatial locations in single-cell data based on the aligned representations of cells with known coordinates.

Ines: So the implication is that this unified space allows us to link different omics layers and even impute missing spatial information, which could give us richer biological insights into tissue architecture.

Marcus: But they do admit there’s a trade-off between those optimization goals; you have to balance the adversarial loss for alignment against the isometry loss for preserving biological signals.

Yuki: So it’s a balancing act—you get better alignment by adding complexity, but you might lose some fine-grained biological detail if you tune that regularization too aggressively.

Ines: That's right. The next step is seeing how this framework holds up when we introduce even more complex data structures, like incorporating spatial coordinates alongside the molecular information.

The paper's improvements: Marcus: So, we’re looking at how they actually improved their original scMRDR framework in this paper, and what that means for us down the road.

Ines: Basically, they are tightening up the optimization goals by proposing specific ways to balance those competing needs of alignment and biological fidelity.

Yuki: How do they propose balancing things when you already have two main drivers—the isometric loss for structure and the adversarial loss for alignment?

Marcus: They suggest that isometry is more essential for preserving intra-omics structure, meaning keeping the internal biology of each molecule intact.

Ines: But they show that if you just keep everything else the same, say you set missing features in proteomics to zero without applying a loss mask, the performance drops significantly.

Yuki: So they’re saying that without that masked loss strategy, you lose a lot of information about how the data is distributed.

Marcus: Exactly. The masked reconstruction loss prevents gradients from flowing through unmeasured features, which lets you utilize all the available information while keeping the original distribution intact.

Ines: They are essentially refining how to weigh those different optimization goals by tuning hyperparameters like beta, lambda, and gamma more effectively for real-world data.

Yuki: That sounds like a practical step toward making the model less of a black box and more of a tool you can actually control based on what you need to learn.

Marcus: And they are emphasizing that isometry is the core driver for bio-conservation, while the adversary part is mostly vital for achieving cross-modal alignment between different data types.

Ines: So they’re telling us that if we want perfect biological fidelity, we need to prioritize isometry, but if we want all the molecules to speak a common latent language, the adversarial loss is necessary.

Yuki: It makes sense. We can’t have perfect alignment if it means destroying the underlying biological patterns of one of our omics layers.

Marcus: The implication for cohort studies is that we get a more principled way to choose those hyperparameters based on what we prioritize—fidelity or alignment—instead of just guessing.

Ines: It shows that scMRDR isn't just a static model; it’s a framework where the training signal itself is intelligently designed to handle the challenges of unpaired, multi-omics data.

Yuki: If this works well, it means we can start building more robust models of biological systems across different molecular layers even when our sample sizes or experimental setups are messy.

Marcus: It moves us closer to integrating complex omics studies without needing those restrictive paired measurements that used to limit these kinds of analyses.

Ines: So the next big thing is testing this framework with even more challenging datasets, maybe incorporating dynamic information or spatial coordinates in a more integrated way.

Conclusion: Ines: So, to wrap things up, scMRDR is essentially a scalable framework for taking unpaired single-cell multi-omics data and putting it into one unified latent space while keeping the biology intact.

Marcus: Right. The main takeaway is that you can link different omics layers without needing perfect paired samples between every single cell type, and the method handles batch effects really well across large cohorts.

Yuki: From a population perspective, this suggests we can start building more robust models of species biology using diverse molecular data without having to wait for perfectly matched longitudinal studies.

Ines: It gives us a tool that is flexible enough to scale up and handle more than two omics naturally, which is something we really needed for real-world sequencing data.

Marcus: The caveat is still the trade-off; you have to be careful tuning those loss functions, because isometry and adversarial goals are in competition for what the model prioritizes.

Yuki: I think that tuning part is where the real biological insight comes from, determining whether we’re focusing more on perfect technical alignment or preserving subtle biological signals.

Ines: Exactly. It shows that by designing the training signal this way, we can build models that are both powerful enough to handle complexity and flexible enough to adapt to different data challenges.

Marcus: So moving forward, this framework gives us a solid foundation for linking genomics with proteomics in ways that were previously difficult because of the lack of paired samples.

Yuki: It opens up new avenues for exploring how molecular function relates directly to tissue architecture across different species, which is really exciting.

Ines: We'll definitely be watching how they apply this to spatial multi-omics data next, because linking that spatial info with these molecular layers could be a really powerful direction.

Marcus: Yeah, we’re ready for the next paper whenever it drops.

More episodes

← Home