🧬 biology

DMT-Dens: Density-preserving manifold visualization for biological data

DMT-Dens is a novel parametric manifold visualization method based on a latent-token Transformer encoder that effectively preserves the sampling density of high-dimensional biological data while maintaining competitive label separability, thereby improving the interpretation of rare and continuous cell-state populations.

Original authors: Ruizhe Wang, Yixuan Dong, Bolin Yang, Bingo Wing-Kuen Ling, Fuji Yang, Zelin Zang

Published 2026-08-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ruizhe Wang, Yixuan Dong, Bolin Yang, Bingo Wing-Kuen Ling, Fuji Yang, Zelin Zang

Original paper licensed under CC BY 4.0 (http://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 ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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

Scientists studying the inner workings of life often face a problem of scale. They collect massive amounts of data from individual cells, measuring thousands of genetic signals at once to understand how a body grows, heals, or fights disease. To make sense of this complexity, researchers use computer programs to shrink these massive, multi-dimensional datasets down to simple two-dimensional maps. These maps act like a visual guide, allowing scientists to see how different cells relate to one another, group together, or transition from one state to another. However, a persistent issue has plagued these maps: the way they arrange the dots often distorts the truth. A cluster of cells that is actually sparse and scattered might be squashed into a tight, dense circle, while a large, abundant group might be stretched out thin. This visual deception makes it difficult to tell if a rare cell type is truly rare or just hidden by the map's own geometry, complicating the search for new biological insights.

A team of researchers has developed a new method called DMT-Dens to fix this specific problem. Their goal was to create a visualization tool that keeps the relative density of the data intact, ensuring that crowded areas on the map stay crowded and sparse areas stay sparse, while still keeping different cell types clearly separated. To test their approach, they applied it to a wide variety of biological data, including detailed maps of human cells from different tissues and developmental stages, as well as standard datasets used to test computer vision. The results showed that their new method was significantly better at preserving the true density of the data than existing popular tools, without sacrificing the ability to distinguish between different cell groups.

The researchers built their solution using a type of artificial intelligence known as a Transformer, which is designed to learn complex patterns. Instead of just trying to keep nearby cells close together on the map, their system adds a specific rule to its learning process: it must also match the local density of the original data. Imagine the data as a landscape where some hills are packed with people and others are empty fields. The new method learns to draw a flat map of this landscape where the crowded hills still look crowded and the empty fields still look empty. It does this by constantly checking the distance between a cell and its nearest neighbors in the original high-dimensional data and comparing it to the distance on the new two-dimensional map. If the map makes a sparse group look too dense, the system adjusts itself to correct the error.

In their experiments, the team compared their new method against several established tools that scientists currently use to visualize cell data. They ran tests on nine different datasets, ranging from synthetic branching structures to real biological samples like human gut cells and embryonic development records. The results were clear: the new method consistently produced maps where the density of points matched the original data much more closely than the other tools. On four different biological datasets, it achieved the highest score for density preservation, and on six out of the nine total datasets, it ranked first. Crucially, it did not lose its ability to separate different cell types; in fact, it ranked among the top two methods for keeping distinct groups of cells apart in seven of the datasets. This balance is vital because a map that preserves density perfectly but mixes up different cell types would be useless for biological discovery.

The researchers also looked deeper into how their system worked by removing specific parts of the design to see what each piece contributed. They found that the specific rule for preserving density was the main driver of the improvement. When they removed this rule, the maps lost their ability to reflect the true density of the data, even though they still managed to separate cell types reasonably well. Conversely, when they changed how the system handled the relationships between all pairs of points, the density improved further, but the ability to separate cell types dropped. This confirmed that the new method successfully balances two competing goals: keeping the crowd sizes accurate and keeping the groups distinct. The team also tested the method on a detailed dataset tracking the development of a roundworm embryo over time. The resulting map showed that the new approach could faithfully represent both the changing density of cell populations and the continuous path of development, offering a clearer view of how cells change as an organism grows.

While the new method offers a significant improvement in visual accuracy, the authors are careful to note what the maps actually represent. The density shown on the map reflects the density of the samples that were collected and processed, not necessarily the absolute number of cells in the body. Factors like how the tissue was prepared or how the cells were sampled can influence the final picture. Therefore, the maps are best understood as a faithful representation of the observed data, helping scientists see the structure of their specific experiment without the distortion introduced by older visualization techniques. By providing a tool that respects both the arrangement and the density of biological data, this work offers a more reliable foundation for exploring the complex, heterogeneous world of living cells.

Technical Summary: DMT-Dens

Problem Definition
Low-dimensional visualization is essential for exploring cell-state heterogeneity in single-cell RNA sequencing (scRNA-seq) and other high-dimensional biological data. While methods like t-SNE and UMAP excel at preserving local neighborhood structures and separating clusters, they often distort the apparent sampling density of the input data. This distortion alters the visual contrast between dense and sparse regions, complicating the interpretation of rare, transitional, or continuous cell-state populations. Existing density-preserving methods (den-SNE, densMAP) address this by adding density terms to non-parametric objectives but remain constrained by their parent methods' neighborhood-affinity formulations. There is a methodological gap for an explicit parametric model that jointly learns neighborhood-rank structure and the relative density profile of processed observations.

Methodology
DMT-Dens is a parametric manifold-visualization method built on a latent-token Transformer encoder. It maps high-dimensional input data $X \in \mathbb{R}^{N \times D}$ to a 40-dimensional latent space and subsequently to a 2D embedding $Y \in \mathbb{R}^{N \times 2}$ . The architecture employs a fixed set of latent tokens per observation, processed via self-attention, independent of the input feature dimension.

The model optimizes a joint objective function $L(\theta) = L_{manifold}(\theta) + \lambda_d L_{dens}(\theta)$ , balancing two primary goals:

Manifold Preservation ( $L_{manifold}$ ): This component utilizes rank-based manifold alignment. It constructs two paired views of the data: the original batch and a neighborhood-augmented view (created by interpolating between an observation and its nearest neighbors in a PCA-reduced space). The loss enforces that the relative ordering of distances in the input space is preserved in the embedding. It employs bidirectional rank affinities and a Student t-kernel similarity, optimized via binary cross-entropy on "hard pairs" (those with the largest mismatch between input and embedding affinities).
Density Preservation ( $L_{dens}$ ): This is the core innovation. The model estimates local density using $k$ -nearest-neighbor (kNN) log-radius profiles. It calculates the Pearson correlation between the log-density vectors of the input space and the 2D embedding space across anchor cells at two scales ( $s = \lfloor k/2 \rfloor$ and $s = k$ ). The loss minimizes the disagreement between these profiles, ensuring that compact and diffuse regions in the embedding reflect the relative concentration and dispersion of the processed input representation.

Key Contributions

Explicit Density Loss: The introduction of a parametric loss function based on the Pearson correlation of kNN log-radius estimates, which explicitly aligns density profiles between high-dimensional input and low-dimensional embedding spaces.
Parametric Framework: A unified deep learning model (Transformer-based encoder) that jointly optimizes neighborhood rank structure and density preservation, distinguishing it from non-parametric baselines.
Rank-Based Alignment with Hard-Pair Aggregation: A manifold alignment strategy that uses bidirectional rank affinities and focuses optimization on the most difficult-to-align pairs to improve structural fidelity.

Results
The authors evaluated DMT-Dens against six baselines (t-SNE, UMAP, den-SNE, densMAP, PaCMAP, PHATE) across nine datasets, including four biological scRNA-seq datasets (HCL, MCA, GAST10K, EPI), a synthetic branching tree, and non-biological datasets (MNIST, EMNIST, NG20, ACT).

Density Preservation: DMT-Dens achieved the highest density correlation on all four biological datasets and six of the nine datasets overall. It significantly outperformed standard t-SNE and UMAP, which showed poor density correlation on biological data.
Label Separability: Despite the focus on density, DMT-Dens maintained competitive label separability, ranking in the top two methods for linear Support Vector Classifier (SVC) accuracy on seven of the nine datasets.
Ablation Studies: Removing the density regularization resulted in the lowest density correlation but higher SVC accuracy, confirming the trade-off. Conversely, using all-pair averaging increased density correlation but degraded kNN preservation and SVC accuracy. The full model represented the best balance, lying on the Pareto frontier of density correlation versus SVC accuracy.
Case Studies: In a C. elegans developmental time-course study, DMT-Dens demonstrated high density correlation (0.909) and SVC accuracy (0.721), outperforming baselines in density metrics while maintaining trajectory continuity.

Significance and Claims
The paper claims that DMT-Dens provides a controlled basis for examining heterogeneous and transitional regions by preserving the relative density of the processed input representation. The authors emphasize that while the method faithfully visualizes sample density, this density reflects experimental sampling and preprocessing (including potential artifacts like tissue dissociation effects) rather than a calibrated measure of absolute biological abundance.

The significance of DMT-Dens lies in its ability to integrate density fidelity with neighborhood structure and label separability within a single parametric framework. This allows researchers to distinguish projection-induced distortions from genuine variation in the data. The authors note that while density fidelity is improved, it does not necessarily align with trajectory topology recovery (e.g., branch identification), suggesting that density-aware projections should be evaluated alongside topology-specific criteria depending on the analytical goal. The parametric nature of the model also supports the projection of new samples, a feature not inherent to non-parametric methods.

More like this