🧬 biology

Multi-Layer Single-Cell Integration with Deep Embedded Clustering Reveals Metabolic Heterogeneity in Peripheral Blood Mononuclear Cells

This study introduces a multi-layer deep embedded clustering framework that integrates transcriptomic, proteomic, and metabolic data to reveal previously undetected metabolic heterogeneity and mitochondrial cell clusters in peripheral blood mononuclear cells, significantly outperforming standard two-layer clustering methods.

Original authors: alireza ebadi

Published 2026-09-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: alireza ebadi

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 ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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

To understand a living cell, scientists have long relied on reading its instruction manual, the genome, and the messages it sends out, the transcriptome. For years, this was the primary way to tell one type of cell from another. However, a cell is more than just its instructions; it is also defined by the physical tools it carries on its surface and the chemical energy it burns to stay alive. In recent years, technology has advanced to the point where researchers can measure these different layers—what genes are active, what proteins are present, and what metabolic activities are occurring—all within a single cell. This shift allows scientists to see the full picture of cellular life rather than just a single snapshot. The question now is whether combining all these layers together reveals a deeper truth about how cells function, or if the extra data simply adds noise to the signal.

A researcher set out to answer this by building a new way to group human immune cells based on three distinct layers of information: their genetic activity, their surface proteins, and their metabolic state. They focused on peripheral blood mononuclear cells, a diverse group of white blood cells that circulate in the body and fight infection. While previous methods had successfully combined genetic data with surface proteins, they typically left out metabolism, the set of chemical reactions that power the cell. The researcher argued that metabolism is not just a background process but a direct measure of what a cell is actually doing at any given moment. To test this, they analyzed a massive dataset containing nearly 44,000 individual cells, measuring over 20,000 genes and 204 surface proteins for each one. They then calculated the activity of six key metabolic pathways, such as how the cells generate energy, to create a third layer of data for every single cell.

The researcher used a sophisticated computer method to weave these three layers together into a single, unified view of each cell. Instead of just looking at the raw numbers, they trained a digital model to learn the hidden patterns that connect genes, proteins, and metabolism. This model then grouped the cells into clusters based on their combined profile. When they compared this new approach to standard methods that only used genes and proteins, the results were striking. The three-layer method created much clearer and more distinct groups of cells. In fact, the new approach improved the quality of the grouping by more than 50 percent compared to the older methods. The most significant finding, however, was not just the improvement in numbers, but the discovery of a specific group of cells that had been invisible before.

When the researcher looked at the clusters formed by the full three-layer model, they found a distinct group of cells that appeared only when the metabolic data was included. This group was characterized by a high level of activity in the cell's power plants, known as mitochondria, and a low level of activity in the genes that usually define specific immune cell types. In the older methods, which ignored metabolism, these cells were scattered among other groups or hidden within them. The new analysis showed that these cells formed a tight, separate community, suggesting they were in a unique state of activity driven by their energy usage rather than their lineage. This discovery confirmed that metabolism provides a unique dimension of information that genes and proteins alone cannot capture.

The study also tested whether the computer method used to find these groups was better than the traditional tools scientists usually rely on. They compared their new deep learning approach against a standard graph-based method on the exact same data. The new method consistently outperformed the standard tool across every measure of accuracy, creating groups that were more tightly packed and better separated. This suggests that the way the new model learns to connect the different layers of data is more effective at finding the true biological structure of the cells. The researcher verified that these results were stable, meaning they did not change just because the computer started with a different random arrangement, and they held up even when the data was slightly noisy.

Ultimately, this work demonstrates that to truly understand the diversity of immune cells, scientists must look beyond the genetic code and the surface markers to include the cell's metabolic state. The ability to detect a specific mitochondrial cluster that was previously missed shows that ignoring metabolism leaves a blind spot in our understanding of cellular behavior. By successfully integrating these three layers, the researcher has provided a more complete map of cellular identity, one that reveals hidden states of activity and offers a more accurate way to study how cells function in health and disease. This approach sets a new standard for how scientists can combine different types of biological data to uncover the full complexity of life at the single-cell level.

Technical Summary: Multi-Layer Single-Cell Integration with Deep Embedded Clustering

Problem Statement
Current single-cell multi-omics integration methods are predominantly limited to combining two data layers, typically transcriptomics (RNA) paired with either chromatin accessibility or surface proteomics. While these approaches have improved cell-type resolution, they often overlook the metabolic state of cells, which directly dictates cellular function and behavior. The author argues that relying solely on RNA and protein data misses a critical dimension of cellular heterogeneity, particularly in contexts like immune activation where metabolic reprogramming is central. Furthermore, existing integration pipelines often rely on graph-based clustering (e.g., Louvain) on high-dimensional data, which may not fully capture the complex interactions between different molecular layers.

Methodology
The study proposes a multi-layer integration framework applied to 43,791 human Peripheral Blood Mononuclear Cells (PBMCs) from a CITE-seq dataset. The methodology consists of the following stages:

Data Preprocessing and Layer Construction:
- Transcriptome: 20,364 genes per cell, normalized via L2 normalization and log-transformation.
- Proteome: 204 surface epitopes per cell, normalized via L1 normalization and log-transformation.
- Metabolome (Derived): Instead of direct metabolite measurement, the author constructed a metabolic layer by computing pathway activity scores. Six metabolic pathways (Glycolysis, OXPHOS, Glutaminolysis, Fatty Acid Synthesis, Pentose Phosphate, and Serine Synthesis) were defined using curated marker gene sets. For each cell, the mean expression of marker genes for each pathway was calculated to create a six-dimensional metabolic signature.
- Concatenation: The three layers (RNA, Protein, Metabolic) were concatenated to form a multi-layer signature matrix ( $Z$ ) of size $43,791 \times 20,574$ .
Dimensionality Reduction:
- Truncated Singular Value Decomposition (SVD) was applied to $Z$ to reduce dimensionality to 50 components, preserving at least 80% of the cumulative explained variance while removing technical noise.
Deep Embedded Clustering (DEC):
- Architecture: A symmetric autoencoder was employed. The encoder compresses the 50-dimensional input into a 10-dimensional latent space ( $z \in \mathbb{R}^{10}$ ) through layers of sizes 500, 500, and 2000. The decoder mirrors this structure to reconstruct the input.
- Training: The autoencoder was trained using Adam optimization with a Mean Squared Error (MSE) reconstruction loss for 50 epochs (batch size 256).
- Clustering: K-means clustering ( $K=28$ ) was applied to the learned 10-dimensional latent representation.
Baselines and Evaluation:
- Baselines: The author compared their approach against:
  - Graph-based clustering (k-NN graph with $k=5$ and Louvain community detection) on the same latent space.
  - Ablation studies using different input configurations: RNA only, Protein only, RNA + Protein, and the full three-layer model.
- Metrics: Clustering quality was evaluated using Silhouette score, Calinski-Harabasz (CH) index, and Davies-Bouldin (DB) index. Stability was assessed via Adjusted Rand Index (ARI) across random seeds and robustness to Gaussian noise.

Key Results

Performance Improvement: The full three-layer model using DEC significantly outperformed both the RNA-only baseline and the graph-based Louvain clustering.
- Compared to Louvain on the same latent space, DEC achieved a 53.2% improvement in Silhouette score (0.1540 vs. 0.1005), a 62.0% increase in the Calinski-Harabasz index (5,911.96 vs. 3,648.56), and a 24.0% reduction in the Davies-Bouldin index (1.6428 vs. 2.1603).
- The full three-layer DEC model showed a 56% improvement in Silhouette score compared to the RNA-only configuration.
Discovery of Metabolic Heterogeneity: The most significant biological finding was the identification of a distinct mitochondrial cluster that was undetectable using RNA or protein data alone.
- This cluster was characterized by high expression of mitochondrial genes (e.g., MT-CO2, MT-ND4, MT-CYB, MT-ATP6) and low expression of standard immune markers.
- The inclusion of the metabolic layer resulted in an Adjusted Rand Index (ARI) of 0.8154 between RNA-only and full-model clusterings, indicating that approximately 18.5% of cells changed their cluster assignment when metabolism was included.
Layer Contribution: Ablation analysis confirmed that while adding protein provided a marginal gain, the addition of the metabolic layer was crucial for improving cluster compactness (lowest DB index) and revealing the mitochondrial subpopulation.
Stability: The clustering results were robust, with ARI values consistently above 0.95 across five different random seeds and graceful degradation in performance under increasing Gaussian noise.

Significance and Claims
The paper claims to establish a robust strategy for single-cell integration that moves beyond the standard two-layer (RNA + Protein/Chromatin) paradigm. The author asserts that:

Metabolism is Non-Redundant: Metabolic signatures provide a functional dimension of cellular state that is not captured by transcriptomic or proteomic data alone. This is evidenced by the discovery of the mitochondrial cluster, which represents a distinct metabolic state rather than a distinct lineage.
DEC Superiority: Deep Embedded Clustering is superior to conventional graph-based methods (Louvain) for multi-layer single-cell data. The autoencoder's ability to learn a joint latent representation that captures interactions between layers outperforms pairwise distance-based graph construction.
General Applicability: The framework is presented as a generalizable approach applicable to various tissues and disease settings, particularly those where metabolic reprogramming (e.g., in cancer or immune activation) plays a central role.

The author maintains a modest tone regarding limitations, acknowledging that their metabolic layer is derived from RNA (a proxy for pathway activity) rather than direct metabolite measurement, and that the study is currently limited to PBMCs. They suggest that future work should validate these findings with direct single-cell metabolomics and diverse tissue types.

Technical Summary: Multi-Layer Single-Cell Integration with Deep Embedded Clustering

More like this