Imaging-anchored multiomics in cardiovascular disease: integrating cardiac imaging, bulk, single-cell, and spatial transcriptomics

arXiv:2601.07871 · q-bio.QM, cs.AI, cs.CV, cs.LG · Submitted 2026-01-10 · Read on arXiv

Listen

Transcript

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

Ines: Today's paper: "Imaging-anchored multiomics in cardiovascular disease".

Marcus: The gist: Imaging-anchored multiomics learns joint representations linking cardiac imaging phenotypes to bulk, sc/snRNA-seq, and spatial transcriptomics.

Ines: First, who's behind it and why it matters.

Title and authors: Ines: So we're looking at this paper right now, "Imaging-anchored multiomics in cardiovascular disease: integrating cardiac imaging, bulk, single-cell, and spatial transcriptomics," and the authors are Minh H. N. Le et al. I want to talk about how this whole approach links what we see on an echo or an MRI with the actual molecules inside the tissue.

Marcus: Right. The title tells us they're taking cardiac imaging—that's our spatial phenotype—and anchoring it to bulk RNA sequencing, single-cell data, and spatial transcriptomics to learn joint representations. It sounds like they are trying to find a way to connect what the heart looks like in a picture with what’s happening at the gene level.

Yuki: From a population genetics perspective, I wonder how much of this spatial structure actually reflects broader genetic backgrounds versus just local environmental influences on those molecular states. The paper suggests that imaging captures multiple processes—genetics, haemodynamics, inflammation and tissue remodelling—so we need to be careful about what the models are actually picking up as fundamental biological reality.

Ines: Exactly. The core idea is that cardiac imaging provides this spatial phenotype, which acts like a natural coordinate system for linking it to molecular measurements like bulk RNA-seq or single-cell data. It’s trying to answer why a region looks fibrotic on an MRI by looking at the specific cell types and gene activity in that exact spot.

Marcus: And they are using deep learning, specifically convolutional neural networks and vision transformers for the imaging side, and then deep generative models for the transcriptomic parts to compress all that high-dimensional data into these compact embeddings. It’s a big dimensionality reduction problem.

Yuki: But when we talk about spatial transcriptomics, it brings in another dimension. They use spot-level expression data which gives them molecular annotations of tissue architecture that they can co-register with histology and even in vivo imaging using anatomical landmarks and intermediate registration steps. That’s a lot of layers to stack.

Ines: Right, the paper mentions that cardiac imaging modalities like echocardiography, cardiac MRI, and cardiac CT define this spatial phenotype which can be linked to bulk, single-cell and spatial transcriptomic measurements through their joint representations. This is what they are building into a framework.

Marcus: And they show how these joint representations support several downstream tasks like multimodal fusion, radiogenomics, spatial alignment and image-based prediction of gene-expression patterns. They are not just looking at one thing; they’re aiming for a whole pipeline to do things like subtyping disease or finding biomarkers.

Yuki: That sounds powerful for discovery, but I always wonder about the practical limits of that linkage when we try to apply it across different patient cohorts or species, given how complex the underlying biology is.

Title and authors: Ines: That’s where the paper gets into improvements. They discuss moving beyond just simple concatenation of features—early fusion—to methods like intermediate fusion where they learn modality-specific embeddings and then combine them in a shared latent space.

Marcus: And they explore graph-based methods, which is interesting because it lets you naturally incorporate spatial and biological structure by making nodes represent imaging regions, genes or pathways, and edges encode known relationships like adjacency in the heart. That should help aggregate molecular signals with structural context.

Yuki: If we use graph structures to connect things like LGE-positive segments to fibrosis-related pathways, that could provide a much richer biological prior than just looking at the data in isolation. But does the model always capture those known relationships accurately?

Ines: The paper also talks about contrastive multimodal learning, where they align region-level imaging embeddings with spot-level molecular embeddings so that nearby points in latent space correspond to tissue regions that are both visually and molecularly similar. That’s trying to build a common coordinate system across different data types.

Marcus: That alignment helps when you have missing data or batch effects, which is a huge problem in real-world cohort studies where the samples just don't match up perfectly, so this contrastive learning should help pull those representations closer together.

Yuki: It’s helpful for finding consistent patterns across different studies, but I’m concerned about whether this alignment is truly capturing underlying biological mechanisms or just statistical artifacts from the training data.

Ines: The paper also touches on a specific application called virtual transcriptomics, where they develop models that can predict molecular profiles directly from images, often termed image-based gene-expression imputation.

Marcus: They suggest you can predict myocardial or plaque gene-expression modules using imaging features like LGE patterns or T1 maps. That lets you do something akin to a non-invasive molecular biopsy of the plaque when you only have imaging available.

Yuki: Predicting molecular profiles from images is ambitious, but if it works, it could allow us to study disease progression in ways that are currently impossible with purely clinical data. But what happens if the imaging features don't actually capture the necessary information for that prediction?

Ines: The authors acknowledge this challenge by discussing the need for robust data curation and batch correction across all modalities before you even start training these joint models. They stress that small sample sizes, even with paired data, limit statistical power and encourage overfitting.

Marcus: So they are saying we need really high quality input data just to make this kind of integrated analysis statistically meaningful when dealing with the real-world cohorts we have. The caveats there are pretty sobering for scaling this up right now.

Title and authors: Yuki: It sounds like the immediate implication is that for these complex diseases, we need extremely careful data harmonization before any of these advanced models can yield reliable biological insights about population health or disease mechanisms.

Ines: So to wrap up on "Imaging-anchored multiomics in cardiovascular disease: integrating cardiac imaging, bulk, single-cell, and spatial transcriptomics," the paper shows a way to build joint representations linking cardiac imaging phenotypes to those three molecular layers. It moves us toward using imaging as an anchor for understanding the molecular states underneath.

Marcus: And the authors show that by using graph-based fusion and contrastive learning, they can create methods that align these different data types in a way that addresses missingness and batch effects better than simpler early fusion approaches.

Yuki: It suggests that we can start to move toward inferring coarse molecular phenotypes from imaging alone with appropriate confidence intervals, which is a big step for translating this into clinical practice.

Ines: The paper points toward a future where we might see cross-modal generative models, like diffusion models conditioned on imaging, simulating plausible molecular profiles for diseases. That’s the next frontier they are pointing toward.

Marcus: And agentic orchestration with LLMs is another direction they mention, where an AI agent could orchestrate the whole pipeline—extracting phenotypes from images and projecting patients into a multimodal latent space to draft summaries of findings.

Yuki: It means these systems aren't just about finding correlations anymore; they’re moving toward creating systems that can reason through complex biological questions across multiple data types. That’s a significant shift in how we approach cardiovascular biology.

Ines: So, the paper, "Imaging-anchored multiomics in cardiovascular disease: integrating cardiac imaging, bulk, single-cell, and spatial transcriptomics," is mapping out a path where spatial structure from imaging informs molecular understanding across multiple scales. It’s a lot of heavy lifting with data quality and model design.

Marcus: We need to keep an eye on how they handle those small cohort sizes and the batch correction methods because that’s where the real hurdle for moving these tools from research papers into clinical tools lies.

Yuki: For me, it’s about ensuring that when we start using these image-based predictions, we are rigorously testing them against independent biological validation to make sure we haven't just learned a spurious correlation in the training set.

Ines: That’s the necessary skepticism when dealing with any of this kind of integrated modeling. It’s about moving from simply describing what we see to actually understanding the underlying causal processes driving those phenotypes.

Marcus: So, that’s where we are with this paper on imaging-anchored multiomics in cardiovascular disease: integrating cardiac imaging, bulk, single-cell, and spatial transcriptomics. We'll be looking at how these ideas evolve next time.

The paper's summary: Ines: So, this paper is basically about taking the spatial picture of the heart—that echo or MRI view—and using that as an anchor to connect it to all these different molecular layers we have, like bulk RNA-seq and single-cell data.

Marcus: Yeah, and what they’re doing is building these joint representations. Think of it like trying to find a common language between the visual data of the heart and the gene expression data in the cells.

Yuki: That sounds ambitious because imaging captures structure and function across space and time, while omics capture dynamic molecular states that reflect genetics and environment. So they are trying to bridge those two very different types of information.

Ines: Exactly. They use deep learning models, like vision transformers for the images, and generative models for the transcriptomics, all working together to create these compact summaries that link a specific heart phenotype to what’s happening at the molecular level.

Marcus: The methodology they focus on is multimodal fusion, but they move beyond just sticking everything together in one big vector. They look at early fusion, intermediate fusion where you learn modality-specific embeddings and then combine them in a shared space, and graph-based methods where nodes represent tissue regions or genes with edges showing their known relationships.

Yuki: I’m interested in the graph part because that allows for incorporating biological priors. If the model can use those known structural connections—like how different heart chambers are physically arranged—that might give it a better understanding of what the molecular signals actually mean spatially.

Ines: That's right, and they show how this setup helps with things like radiogenomics, which links image features to molecular measurements directly, and spatial alignment pipelines that put all these different data types into one common coordinate system.

Marcus: They are also using contrastive learning across modalities to align the embeddings. This means they train the system so that if two spots in the latent space look similar visually on an MRI, they should also have similar molecular embeddings from a single-cell sequence, even if they aren't paired data.

Yuki: That sounds like a way to overcome some of those real-world issues with missing data or batch effects in small cohorts because it forces the model to learn a representation that is consistent across different types of data inputs.

Ines: And on the practical side, they discuss how hard it is to do this. They point out that even when you have paired data, the sample sizes for deeply phenotyped patients are still small, which limits how statistically powerful these models can be and increases the risk of overfitting.

Marcus: So we have these really sophisticated integration methods, but they also admit that getting clean input data and handling those small sample sizes is a major hurdle before you can even start trusting the results for real patient populations.

Yuki: It means right now, it’s more about building these conceptual frameworks to see what’s possible than having a ready-to-use tool that works perfectly on any dataset immediately.

Ines: And looking ahead, they suggest three main paths for this work—moving towards richer spatial models with linked imaging and molecular fields, developing foundation models that scale this analysis to more people, and finally integrating these capabilities into digital twins or agentic systems.

Marcus: That moves the goalposts from just describing a condition to actually creating systems that can infer molecular profiles directly from images, which is what they call virtual transcriptomics.

Yuki: I think if they can predict gene expression modules right from an image, it opens up non-invasive ways to study disease progression that we just couldn't do before.

Ines: So the paper lays out a roadmap that goes from learning joint representations now to building AI systems that can orchestrate complex analyses and even simulate molecular profiles.

Marcus: That’s the big picture—moving from correlation to a system that can reason about cardiovascular disease across all these scales.

The paper's improvements: Tom: So, we’re talking about how this paper suggests ways to make these multiomics pipelines better and more robust, moving past just simple feature concatenation.

Ines: Right, they suggest shifting from early fusion—which is just dumping all those data features into one single vector—to intermediate fusion where the system learns modality-specific embeddings first and then merges them in a shared latent space.

Marcus: That makes sense because if you try to merge everything at once, you get messy noise scales and sensitivity issues, especially when modalities have different error profiles.

Yuki: I see that as trying to find a common ground where the biological meaning is preserved even if the data formats are different initially. It’s like finding a universal translation layer before you try to synthesize anything new.

Ines: And they also look at graph-based methods, which use graph neural networks or GNNs where nodes represent imaging regions, genes, or pathways with edges showing their known relationships in the heart structure.

Marcus: That’s interesting because it allows the model to incorporate known biological structures into its learning process. It’s not just looking at data points; it's looking at how those points are physically connected in the heart tissue itself.

Yuki: If you can connect a region showing late gadolinium enhancement, for example, directly to the pathways known to cause fibrosis, that gives the molecular signal a physical context. That should support the findings more than if it were just an isolated data point.

Ines: And they also look at contrastive cross-modal autoencoders. This is where they train them on paired data so that when you see two points in the latent space that are visually close on an image, they *must* also be molecularly close from a single-cell perspective.

Marcus: That’s a way to build that common coordinate system we talked about earlier, making the representations consistent across different types of biological evidence. It handles some of those real-world inconsistencies in the data flow.

Yuki: But the caveat there is still how much paired data they have for these comparisons—if you don't have good paired sets, this alignment might just be learning statistical artifacts instead of true biology.

Ines: Exactly, and that’s why the authors are very careful about rigorous data curation and batch correction before applying these improvements to larger cohorts. The method itself is powerful, but its success depends on the input quality.

Marcus: So the implication for me is that if we want to use this kind of joint representation learning in a large cohort study, we can’t just throw raw data at it; we need very disciplined preprocessing and careful statistical checks to ensure those learned embeddings are actually reflecting biology.

Yuki: And for us as population geneticists, the real question is whether these improved representations can actually generalize across different patient populations or if they only capture the specific environment of the training set.

Ines: That leads right into their future work, where they are talking about cross-modal generative models, like diffusion models conditioned on imaging that could simulate plausible molecular profiles for diseases.

Marcus: That’s moving from analysis to creation—using AI to generate what a disease might look like at the gene expression level, which is a big step beyond just classifying existing data.

Conclusion: Tom: So we’re wrapping up on "Imaging-anchored multiomics in cardiovascular disease: integrating cardiac imaging, bulk, single-cell, and spatial transcriptomics." Basically, this paper shows how to link those spatial heart images to the molecular data from bulk RNA-seq and spatial transcriptomics.

Ines: Yeah, it’s about using that visual information as a reference point to pull out what’s happening at the gene level in specific tissue regions. It changes how we interpret why a heart looks diseased on an MRI by giving us molecular context for those specific spots.

Marcus: The big implication is moving toward more precise subtyping of cardiovascular diseases. If you can link a visual phenotype to a molecular signature, you get much richer data than just looking at one modality in isolation.

Yuki: From my perspective, it suggests we might be able to look at patient populations and identify underlying genetic or environmental drivers that manifest as specific spatial patterns in the heart tissue. That’s something that could help us understand the history of these conditions better.

Ines: And they point toward future work using AI to create generative models that can predict molecular profiles directly from images, which is essentially virtual transcriptomics for things like plaque analysis.

Marcus: It’s a big technical leap, moving from classification to direct prediction of expression modules based purely on imaging features like T1 maps or LGE patterns.

Yuki: That would be powerful if those predictions hold up when you test them in different populations, but the paper itself is focused on the integration framework rather than massive validation across species.

Ines: True, they flag that robust data curation is still a huge hurdle before we can rely on these predictions for clinical decision-making. You need very high quality input data to make this whole system actually work reliably.

Marcus: So yeah, the takeaway is that while the methodology for integrating these different data types is quite sophisticated—using contrastive learning and graph methods—it’s still hitting a wall with sample size limitations in real-world cohorts.

Yuki: It means we need to be patient and build up these foundational models slowly before we can expect them to support broad population genetics or widespread clinical use.

Ines: So, the paper on "Imaging-anchored multiomics in cardiovascular disease: integrating cardiac imaging, bulk, single-cell, and spatial transcriptomics" is a solid framework for connecting structure to molecular function.

Marcus: It’s a clear roadmap for how to build better integrative tools that handle the complexity of combining imaging with different omics layers.

Yuki: It gives us a new lens through which to view cardiovascular pathology, linking physical appearance directly to the underlying molecular history of the heart tissue.

Taipei Medical University · University of Oxford · Carnegie Mellon University · Emory University

q-bio.QM, cs.AI, cs.CV, cs.LG

Submitted: 2026-01-10

Updated: 2026-10-15

Comments: The authors are withdrawing this version of the preprint to incorporate extensive updates, refine the multi-modal fusion pipelines, and expand the evaluation benchmarks for cardiac imaging and spatial transcriptomics integration. A revised and updated version will be resubmitted shortly

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 89/100

The gist: The gist: Imaging-anchored multiomics learns joint representations linking cardiac imaging phenotypes to bulk, sc/snRNA-seq, and spatial transcriptomics.

Key concepts

Spatial Phenotype: This refers to the physical description of the heart derived from imaging techniques like echocardiography, cardiac MRI, and CT. It defines where structures are located within the heart's anatomy in 3D space and time, providing a spatial context for molecular data.
Representation Learning: This is the process of using machine learning models (like CNNs or ViTs) to compress complex, high-dimensional data from imaging and transcriptomics into compact numerical codes. The goal is to extract the most clinically and biologically meaningful patterns while discarding noise.
Multimodal Fusion: This involves combining information from different data types—imaging, bulk RNA-seq, and spatial transcriptomics—into a single predictive model. Techniques range from simple feature concatenation to complex graph-based methods that learn shared latent spaces across all modalities.
Spatial Transcriptomics: This technique maps the location of gene expression directly onto tissue sections using barcoded spots. It acts as a bridge, connecting the macroscopic view from imaging with the microscopic molecular details of specific cell types within the heart tissue.

Terminology

Foundation Model — 2× in this paper · explained in 3 episodes
Foundation Models — 2× in this paper · explained in 7 episodes
Alignment — 1× in this paper
Failure Modes — 1× in this paper · explained in 3 episodes
Multimodal Fusion — 1× in this paper · explained in 4 episodes

Summary

The gist: Imaging-anchored multiomics learns joint representations linking cardiac imaging phenotypes to bulk, sc/snRNA-seq, and spatial transcriptomics.

Foundations: Modalities and Biological Context

Cardiac imaging provides a high-dimensional, spatially and temporally resolved description of heart structure and function Cardiac imaging as a spatial phenotype defines the heart's anatomy and function in space and time. EchoNet family (EchoNet-Dynamic, EchoNet-LVH) has catalysed the use of deep learning for view classification, segmentation and disease detection in echocardiographic videos. Cardiac MRI (CMR) delivers volumetric cine imaging for ventricular function, parametric mapping (T1/T2) for tissue characterisation and late gadolinium enhancement (LGE) for focal fibrosis. These modalities define a spatial phenotype of the heart.

Omic layers capture dynamic molecular states that reflect genetic background and environmental exposures. Bulk RNA-seq of myocardial tissue has been used to profile pathways active in hypertrophic and dilated cardiomyopathies, myocarditis, and heart failure. Single-cell and single-nucleus RNA sequencing (sc/snRNAseq) provide cell-level resolution. Human heart atlases combining sc/snRNA-seq across chambers and regions have catalogued cardiomyocytes, fibroblasts, endothelial and smooth muscle cells, pericytes, immune cells and other stromal populations.

Spatial transcriptomics as a bridge

Spatial transcriptomics extends single-cell approaches by preserving information regarding where transcripts are located in tissue. Array-based platforms such as 10x Genomics Visium place tissue sections on slides with grids of barcoded capture spots. Spatial transcriptomic data provide a molecular annotation of tissue architecture that can be co-registered with histology and, via intermediate registration steps and anatomical landmarks, with in vivo imaging.

Representation Learning Across Modalities

Representation learning aims to compress high-dimensional cardiac imaging data into compact embeddings that preserve clinically and biologically relevant information. Convolutional neural networks (CNNs) and their 3D or spatiotemporal variants remain the workhorses for echocardiography and CMR segmentation, view classification and disease prediction. Vision Transformers (ViTs) and masked autoencoders (MAEs) have been adapted to medical imaging, including cardiac MRI and CT. For bulk transcriptomic data, deep generative models such as autoencoders and variational autoencoders are used for dimensionality reduction.

Multimodal Fusion: Methods, Pipelines and Evaluation

The review considers how imaging, transcriptomics and spatial transcriptomics embeddings can be combined into joint models. A useful taxonomy distinguishes early, intermediate and late fusion. Early fusion concatenates features from all modalities into a single vector per sample However, it suffers from high dimensionality, differing noise scales across modalities and sensitivity to missing data. Intermediate fusion first learns modality-specific embeddings, then combines them in a shared latent space. Graph-based methods provide a natural way to incorporate spatial and biological structure Contrastive multimodal learning aligns representations by pulling together paired samples and pushing apart unpaired ones.

Integrative Pipelines

Integrative pipelines that align imaging with transcriptomic and spatial transcriptomic data are discussed, illustrating how they support disease subtyping, mechanistic biology, and biomarker discovery. Radiogenomics links imaging features with molecular measurements Graph-based radiogenomics introduces biological priors. Spatial molecular alignment pipelines register in vivo imaging, ex vivo imaging, histology and spatial transcriptomics into a common coordinate system.

Practical Considerations

Robust imaging–omics integration begins with rigorous data curation. Batch correction and harmonisation are critical across modalities Sample sizes of deeply phenotyped patients with paired imaging, bulk, single-cell and spatial data remain small, which limits statistical power and encourages overfitting.

Future Directions

Future work is likely to treat the heart as a continuous threedimensional object with linked imaging and molecular fields. Cross-modal generative models are another frontier, where generative adversarial networks and diffusion models conditioned on imaging could simulate plausible molecular profiles. Imaging-informed molecular phenotyping may become feasible to infer coarse molecular phenotypes from imaging alone with appropriate confidence intervals.

Conclusion

Imaging-anchored multiomics is likely to evolve along three main trajectories: towards richer spatial models that integrate imaging and molecular fields, towards foundation models and generative approaches that scale molecularly informed imaging analysis to broader populations and towards incorporation into digital twins and agentic systems that embed these capabilities into clinical workflows.

Key Points

• Imaging-anchored multiomics learns joint representations linking cardiac imaging phenotypes to bulk, sc/snRNA-seq, and spatial transcriptomics.

• Cardiac imaging provides a spatial phenotype that supports voxel/region alignment with molecular states and cell-type programmes.

• Spatial transcriptomics bridges imaging to tissue microdomains via spot-level expression, deconvolution, and registration pipelines.

• Fusion methods (MOFA/scVI, DCCA, multimodal VAEs, GNNs, contrastive learning, transformers) address missingness, batch effects, and small paired cohorts.

• Applications include radiogenomics, virtual transcriptomics, disease subtyping, mechanism discovery and biomarker prioritisation for CVD translation.

Data Availability Statement

No new data were generated or analysed in this review.

Funding

This work was supported by the National Science and Technology Council, Taiwan [grant number NSTC114-2221-E-038-015].

Conflict of Interest

The authors declare no competing interests.

Author Biographies

Minh H. N. Le is an MD–PhD physician–scientist, Harvard alumnus, and incoming postdoctoral researcher at Yale, focusing on artificial intelligence, large language models, machine learning, deep learning, and multimodal cardiovascular imaging applications.

Tuan Vinh is an NIH Oxford–Cambridge Scholar and Emory University graduate whose research spans genomics, multimodal machine learning, causal modeling and AI-driven molecular property prediction at the intersection of chemistry and biomedical data science.

Thanh-Huy Nguyen is a researcher in computational biology at Carnegie Mellon University, specializing in deep learning for biomedical imaging, semi-supervised segmentation, domain adaptation and multimodal data integration across imaging and molecular modalities.

Tao Li is a Ph.D. student in computer science at Emory University, working on causal and interpretable machine learning, graph-based representation learning and multimodal molecular modeling for biomedical and chemical applications.

Bao Quang Gia Le is a Ph.D. student in chemistry at Emory University, focusing on chemical biology, organic synthesis and bioconjugation strategies for developing functional biomolecules and targeted molecular systems.

Han H. Huynh is a translational bioinformatician with expertise in single-cell RNA sequencing, circulating tumor cell analysis and AI-driven studies in cancer biology and cardiovascular medicine.

Monika Raj is a Professor of Chemistry at Emory University whose research centers on site-selective peptide and protein modification, bioconjugation chemistry and chemical biology with applications in biomedical research.

Improvements for AI systems

Bold header: Modality-Specific Representation Learning

This improvement involves training modality-specific encoders (CNNs for imaging, deep generative models for omics) and then combining them in a shared latent space to learn joint representations linking cardiac imaging phenotypes to bulk, sc/snRNA-seq, and spatial transcriptomics. This allows the system to infer which molecular states underlie a given imaging phenotype by leveraging modality-specific strengths.

Bold header: Graph-Based Fusion for Spatial Context

Implement graph neural networks (GNNs) where nodes represent imaging regions, genes or pathways, and edges encode known relationships, such as adjacency in the heart or membership in the same pathway. This enables the system to aggregate molecular signals and molecular nodes to incorporate structural context, which is crucial for identifying structures like stressed border-zone myocardium by connecting LGE-positive segments to fibrosis-related pathways.

Bold header: Contrastive Learning for Cross-Modal Alignment

Utilize contrastive cross-modal autoencoders trained on paired data where they align region-level imaging embeddings with spot-level molecular embeddings, such that nearby points in latent space correspond to tissue regions that are both visually and molecularly similar. This allows the system to learn a common coordinate system for downstream tasks like subtyping by ensuring that nearby points in latent space correspond to tissue regions that are both visually and molecularly similar.

Bold header: Agentic Orchestration with LLMs

Deploy an agentic layer where Large Language Models (LLMs) orchestrate complex analysis pipelines by calling specialised tools such as imaging encoders, fusion models, and statistical tests. This system can perform tasks like extract[ing] regional imaging phenotypes using an imaging model, project[ing] the patient into a multimodal latent space using a fusion model, and then draft[ing] a textual summary integrating these findings with relevant literature.

Bold header: Virtual Transcriptomics for Non-Invasive Biopsies

Develop deep learning models that can predict molecular profiles directly from images, often termed virtual transcriptomics or image-based gene-expression imputation. This allows the system to predict myocardial or plaque geneexpression modules from imaging features such as LGE patterns, T1 maps or strain fields, effectively performing a non-invasive molecular biopsy of the plaque when only imaging is available.

Abstract

Cardiovascular disease arises from interactions between inherited risk, molecular programmes, and tissue-scale remodelling that are observed clinically through imaging. Health systems now routinely generate large volumes of cardiac MRI, CT and echocardiography together with bulk, single-cell and spatial transcriptomics, yet these data are still analysed in separate pipelines. This review examines joint representations that link cardiac imaging phenotypes to transcriptomic and spatially resolved molecular states. An imaging-anchored perspective is adopted in which echocardiography, cardiac MRI and CT define a spatial phenotype of the heart, and bulk, single-cell and spatial transcriptomics provide cell-type- and location-specific molecular context. The biological and technical characteristics of these modalities are first summarised, and representation-learning strategies for each are outlined. Multimodal fusion approaches are reviewed, with emphasis on handling missing data, limited sample size, and batch effects. Finally, integrative pipelines for radiogenomics, spatial molecular alignment, and image-based prediction of gene expression are discussed, together with common failure modes, practical considerations, and open challenges. Spatial multiomics of human myocardium and atherosclerotic plaque, single-cell and spatial foundation models, and multimodal medical foundation models are collectively bringing imaging-anchored multiomics closer to large-scale cardiovascular translation.