Multimodal atlas of human atherosclerosis links granular vascular cell states to coronary artery disease risk

MetaPlaq: A Multimodal Atlas Mapping Cellular Drivers of Human Atherosclerosis

If you have ever tried to map disease risk using only bulk tissue samples, you know the frustration of the resolution mismatch. You see the signal, but you cannot find the source. This research addresses that exact pain point.

The goal is to connect inherited genetic risk to specific cellular malfunctions. Genome-wide association studies (GWAS) have identified hundreds of loci (specific genomic locations associated with disease). However, we still struggle to identify which cells those mutations actually affect. We know the "where" in the genome, but we lack the "who" in the tissue.

The Problem

Current approaches to studying atherosclerosis suffer from a resolution gap. Transcriptomic studies (measuring RNA to see which genes are active) provide a high-level view. However, they fail to capture the underlying cis-regulatory architecture (the DNA instructions that control gene expression). Most disease-associated genetic variants reside in non-coding regions. These variants act as dimmers or switches for genes rather than changing the proteins themselves.

Furthermore, traditional single-cell assays lose the "neighborhood" context. You might identify a rare, diseased cell state. But you won't know if it sits in a stable part of the artery or at the volatile edge of a plaque [Figure 2g]. Previous attempts to bridge this gap lacked the scale to resolve rare transitional states. One example is endothelial cells undergoing mesenchymal transition (EndoMT). This is a process where cells change their fundamental identity.

How It Works

The authors built MetaPlaq. This multimodal atlas integrates three distinct data layers to reconstruct the regulatory logic of the artery. Their architecture relies on three primary pillars:

1. Multi-modal Harmonization: They did not treat RNA and DNA accessibility (snATAC-seq) as separate silos. Instead, they used a variational inference model (a probabilistic framework) called scVI to create a unified latent space (a compressed mathematical representation of the data). They utilized the scArches reference mapping approach. This allows them to fine-tune a pre-existing model with new datasets. This effectively performs "transfer learning" to bring disparate studies into one coordinate system [, Figure 2a].
2. Spatial Grounding: They layered high-resolution Visium HD spatial transcriptomics onto the single-cell map. This moves from a "soup" of dissociated cells to a structured anatomical map. It assigns cell-level identities to specific coordinates within the coronary artery [Figure 2g, Figure 3d].
3. Regulatory Network Inference: To link the genome to the phenotype, they employed SCENIC+. This infers enhancer-driven gene regulatory networks (eGRNs). This goes beyond mere correlation. It predicts how specific transcription factors (master controllers of cell identity) interact with enhancers (DNA switches) to drive disease-relevant genes .

Numbers

The scale of MetaPlaq is significant for building a high-fidelity biological reference. The authors report profiling ~1 million cells and nuclei across 254 samples .

On the epigenomic side, they identified 515,498 open chromatin regions (OCRs). These are parts of the genome that are "unzipped" and ready for activity. This is the largest chromatin map reported for these tissues to date .

The paper demonstrates the utility of this granularity. They resolved 9 distinct smooth muscle cell (SMC) states and 6 endothelial cell (EC) states [Figure 3a, Figure 3h]. Crucially, they show that SMC states undergo a predictable phenotypic transition. Cells move from contractile states in the healthy media to osteochondrogenic-like (bone-forming) cells in the advanced plaque [Figure 3d, Figure 6k]. This granularity allowed them to pinpoint that the FMC (fibromyocyte) state carries the heaviest load of coronary artery disease (CAD) and myocardial infarction (MI) heritability [Figure 6p].

What's Missing

Despite the scale, there are clear edges to this system. First, the authors admit that disease-stage definitions were not uniform across all public datasets. Inconsistent labeling is a risk. If "intermediate" means something different in one dataset than another, the resulting disease "gradient" might be an artifact.

Second, the spatial resolution is subject to the limitations of the Visium HD platform. The authors note that necrotic (dead) or fibrotic (scarred) regions in advanced plaques limit transcript detection. This creates a blind spot exactly where the most interesting pathology occurs.

Third, the in-silico (computer-simulated) perturbations were limited to SMCs. While the model can predict how a cell might react to a genetic "knockout," we haven't seen these predictions extended to the complex interplay between immune cells and the vasculature.

Should You Prototype This

Yes, if you are working on target discovery for chronic diseases.

The methodology here is a blueprint for moving from descriptive biology to predictive biology. Specifically, the integration of scArches for harmonization and SCENIC+ for regulatory network inference is valuable. If you want to identify why a specific mutation causes disease, you cannot rely on RNA alone. You need the enhancer-gene link maps the authors have demonstrated.

Code is reportedly available; see the paper for the canonical link. This includes the scRNAutils pipeline. If you have the compute to handle a million-cell integration task, this workflow is a massive leap forward. Don't try to build the atlas yourself. Instead, adopt the multi-modal integration workflow for your own specialized tissue studies.

Figures from the paper