High-dimensional data is a ubiquitous challenge across modern scientific disciplines. In single-cell genomics, researchers track the expression levels of tens of thousands of genes per cell (Daniel Reichart et al. in Pathogenic variants damage cell composition and single cell transcription in cardiomyopathies). In deep learning, interpretability methods analyze massive matrices of attention patterns across billions of tokens (Jonathan Katzy et al. in Automated Attention Pattern Discovery at Scale in Large Language Models). Even in supercomputing, production logs track dozens of hardware, network, and I/O counters simultaneously (Zhengchun Liu et al. in Characterization and identification of HPC applications at leadership computing facility).
In these hyper-dimensional spaces, human intuition fails, and traditional algorithms falter due to the curse of dimensionality: as the number of dimensions grows, the volume of the space increases exponentially, causing data points to appear uniformly distant from one another. To make sense of this data, we must project it into a lower-dimensional space (typically 2D or 3D) without destroying the meaningful relationships hidden within.
The Manifold Hypothesis
The core reason we can compress high-dimensional data without losing its vital essence is the Manifold Hypothesis. This principle posits that real-world high-dimensional data does not actually fill its embedding space uniformly. Instead, it concentrates around a lower-dimensional, non-linear surface—a manifold—embedded within the higher-dimensional space.
Think of a sheet of paper. It is a two-dimensional object. If you crumple it into a ball, it lives in three-dimensional space, but every point on that paper can still be located using just two coordinates along its surface.
As Yuen Ler Chow et al. demonstrate in Predicting drug polypharmacology from cell morphology readouts using variational autoencoder latent space arithmetic, projecting high-dimensional cellular morphology and gene expression profiles onto a low-dimensional manifold filters out experimental noise and background clutter while retaining the underlying logic of how cells respond to external stimuli. The goal of manifold learning is to "unroll" this crumpled sheet of data.
Linear vs. Non-Linear Projections
The simplest way to reduce dimensions is through linear projection methods like Principal Component Analysis (PCA). PCA rotates the high-dimensional coordinate system to find the axes of maximum variance. Helen Smith utilizes PCA in Systems based analysis of human embryos and gene networks involved in cell lineage allocation to project transcriptomic profiles of early human embryos against reference signatures.
However, linear methods assume that the underlying structure is flat. If the data wraps around itself in a complex curve—like a Swiss roll—a linear projection will squash distinct layers together, completely destroying the true topology of the data. Non-linear dimensionality reduction techniques are required to preserve these intricate geometric structures.
Preserving the Local vs. Global Structure
When mapping a high-dimensional manifold to a 2D plane, a fundamental trade-off arises: do you preserve local structure (keeping close neighbors close) or global structure (keeping distant clusters at the correct relative distance)?
Local Preservation: t-SNE
Techniques like t-Distributed Stochastic Neighbor Embedding (t-SNE) prioritize local structure. t-SNE converts the Euclidean distances between data points into conditional probabilities that represent similarities. It then tries to match these probabilities in a low-dimensional space.
This approach is highly effective at separating distinct cell types or clinical states into clean, isolated clusters. For example, Ye Wang et al. apply t-SNE in Patients with obstructive sleep apnea have suppressed levels of soluble cytokine receptors involved in neurodegenerative disease, but normal levels with airways therapy to map high-dimensional biomarker profiles into distinct clinical clusters. Similarly, Zuqiang Liu et al. use it in A single-cell transcriptional landscape of immune cells shows disease-specific changes of T cell and macrophage populations in human achalasia to visualize distinct immune cell clusters.
The downside of t-SNE is that it often tears apart the global topology. The distance between two distant clusters on a t-SNE plot is largely arbitrary, meaning you cannot reliably infer how biologically or physically related two separate clusters are.
Balancing Both: UMAP
Uniform Manifold Approximation and Projection (UMAP) addresses this limitation by leveraging Riemannian geometry. UMAP operates on three mathematical assumptions:
- The data manifold is locally connected.
- The Riemannian metric is locally constant (or approximately so).
- The data is uniformly distributed on the manifold.
Instead of using simple probabilities, UMAP models the manifold using fuzzy simplicial sets—essentially creating a weighted graph where edge weights represent the likelihood that two points are connected on the manifold. It then optimizes the low-dimensional layout by minimizing the cross-entropy between the high-dimensional and low-dimensional representations.
Because cross-entropy penalizes both moving close points far apart and moving far points close together, UMAP preserves both local neighborhoods and global relationships. This capability is vital when exploring continuous transitions, such as the global structure of physical configuration spaces (Hanlin Sun et al. in Network science: Ising states of matter) or the continuous evolution of tissue states (the authors of Emergent properties of collective gene-expression patterns in multicellular systems).
High-Dimensional Data
│
▼
[Manifold Approximation] ──► Constructs a fuzzy graph of local/global connections
│
▼
[Low-Dimensional Projection] ──► Optimizes layout via cross-entropy minimization
│
▼
Preserved Local Clusters + Meaningful Global Distances
Downstream Synergy: Preprocessing for Clustering
Manifold learning is not merely a tool for human visualization; it fundamentally changes how downstream machine learning algorithms perform.
Standard clustering algorithms like K-means rely on Euclidean distances, which break down in high dimensions. By using a non-linear technique like UMAP as a preprocessing step, the data is compressed onto its true underlying manifold, making distances meaningful again.
Guo-Wei Wei 0001 demonstrates this synergy in UMAP-assisted K-means clustering of large-scale SARS-CoV-2 mutation datasets, showing that UMAP preprocessing significantly improves the accuracy and interpretability of K-means when handling massive genomic datasets.
Furthermore, because UMAP preserves density variations well, it pairs exceptionally with density-based clustering algorithms. Anna E. Bauer et al. leverage this combination in Identification and validation of postpartum depression subtypes: a population-based cohort study, combining UMAP with DBSCAN (Density-Based Spatial Clustering of Applications with Noise). This data-driven pipeline allows for the discovery of novel clinical subtypes without forcing the data into rigid, predefined spherical structures.
By translating hyper-dimensional geometries into intuitive, lower-dimensional spaces, manifold learning bridges the gap between raw computational data and human insight, revealing the structured simplicity hidden within complex systems.
Key Foundational Work
Hematopoietic differentiation: a coordinated dynamical process towards attractor stable states (2010) — contributed 1 concept elements to this direction.
Nonlinear dimensionality reduction methods for synthetic biology biobricks' visualization (2017) — contributed 1 concept elements to this direction.