🤖 machine learning

IRIS: time-structured manifold projections

The paper introduces IRIS, a novel manifold learning algorithm designed to visualize high-dimensional temporal biomedical data by simultaneously preserving chronological ordering and manifold topology, thereby overcoming the limitations of existing methods like t-SNE and UMAP in capturing dynamic biological processes.

Original authors: Brian Ondov, Chia-Hsuan Chang, Weipeng Zhou, Xingjian Zhang, Xueqing Peng, Yutong Xie, Huan He, Qiaozhu Mei, Hua Xu

Published 2026-06-01

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Brian Ondov, Chia-Hsuan Chang, Weipeng Zhou, Xingjian Zhang, Xueqing Peng, Yutong Xie, Huan He, Qiaozhu Mei, Hua Xu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a massive, chaotic library of high-dimensional data. In the world of biology and science, this could be millions of cells, thousands of bacteria samples, or hundreds of thousands of research papers. To make sense of this mess, scientists usually use a tool called "Manifold Learning" (like t-SNE or UMAP). Think of these tools as a magic mapmaker. They take a complex, multi-dimensional object and flatten it onto a 2D piece of paper so humans can see patterns.

However, there's a catch: these standard mapmakers are "time-blind." If you have data that changes over time (like a cell growing from a baby to an adult, or a research paper published in 2010 vs. 2024), the standard map just scrambles the time order. It's like taking a photo album of a child's growth, shuffling the photos randomly, and then trying to arrange them on a wall based only on how the child looks in the picture. You might group all the "baby" photos together, but you lose the story of when they happened.

Enter IRIS.

The authors of this paper created a new algorithm called IRIS (which stands for "time-structured manifold projections"). Think of IRIS as a spiral time machine for data visualization.

Here is how it works, using simple analogies:

1. The "Spiral Galaxy" Layout

Instead of flattening data onto a square grid (like a standard map), IRIS organizes the data into a circular, spiral pattern.

The Center: Represents the earliest time (the "beginning").
The Outer Rings: Represent later times (the "future").
The Angle: Represents the relationships between the data points (who is similar to whom).

Imagine a tree ring or a galaxy. The center is the birth of the galaxy, and the stars spiral outward as time passes. In IRIS, if you look at the map, you can instantly tell "how old" a data point is just by how far out it is from the center.

2. Solving the "Crowding" Problem

Standard maps often have a problem called "crowding." If you have a lot of data from one specific time period, it gets squished together, making it hard to see details.

IRIS's Fix: The algorithm acts like a smart elastic band. It stretches or compresses the "time" axis mathematically before drawing the map. If you have a huge amount of data from one year and very little from another, IRIS stretches the ring for that year so the points have room to breathe, while keeping the overall shape circular. This ensures the map looks balanced and isn't squashed in one spot.

3. Two Steps to the Solution

The paper describes IRIS as working in two main phases, like baking a cake in two steps:

Step 1: Setting the Time (The Radius): First, IRIS figures out the perfect mathematical rule to convert "time" into "distance from the center." It calculates the best way to spread the rings out so they aren't too crowded or too empty.
Step 2: Arranging the Neighbors (The Angle): Once the rings are set, IRIS spins the data points around the center. It tries to keep similar things (like the same type of cell or topic) close to each other, but it forces them to stay on their correct "time ring." It's like organizing a party where everyone must stand on the correct birthday ring, but they can still hug their friends who are on that same ring.

What Did They Test?

The authors tested IRIS on three very different types of "messy" data to see if it worked better than the standard tools (UMAP):

Mouse Embryos (scRNA-seq): Tracking how a mouse grows from a tiny embryo to a baby. IRIS showed the clear progression of development, whereas the standard map scrambled the timeline.
Gut Bacteria (Metagenomics): Looking at bacteria in people of different ages. IRIS clearly showed how the bacterial communities change as people get older.
Scientific Literature: Analyzing thousands of research papers on Alzheimer's and immune systems. IRIS showed how the topics in these papers evolved over the decades, creating a clear timeline of scientific discovery.

The Bottom Line

The paper claims that IRIS is a new way to visualize complex data that respects the timeline.

Standard Maps (UMAP/t-SNE): Good at showing who is similar to whom, but bad at showing when things happened.
IRIS: Good at showing who is similar to whom AND clearly showing when things happened by arranging them in a time-based spiral.

The authors conclude that IRIS is open-source, runs fast, and allows scientists to see the "story" of their data (the dynamics of change) without needing to guess or look at multiple confusing charts. It turns a static snapshot into a clear, chronological movie.

Technical Summary: IRIS – Time-Structured Manifold Projections

Problem Statement
High-dimensional biomedical data, such as single-cell RNA sequencing (scRNA-seq), comparative metagenomics, and scientific literature embeddings, are increasingly generated with temporal dimensions (e.g., developmental stages, subject ages, publication dates). While Manifold Learning algorithms like t-SNE and UMAP are indispensable for visualizing high-dimensional structures by grouping similar items, they are inherently unsupervised and do not incorporate time-ordering into their layout structures. Consequently, the dynamics of cell types or other classes over time are often obfuscated.

Current workarounds are insufficient:

Side-by-side views: Coloring layouts by both class and time in separate plots requires tedious scanning, making it difficult to estimate time ranges for specific classes or identify temporal extremes.
Aligned UMAP: This method produces multiple layouts while penalizing movement across them. However, it requires a small number of discrete time slices and assumes the same items appear in multiple slices, which is rarely the case for scRNA-seq or other temporal datasets. Furthermore, it exacerbates the issue of multiple views, often necessitating animations or unwieldy 3D plots to show trajectories.

Methodology
The authors propose IRIS (Incorporating Radial time-structured Inverse Structure), a new Manifold Learning algorithm that structures low-dimensional layouts both chronologically and by manifold topology. IRIS operates by creating circular layouts where time radiates outward in concentric rings, leaving the angular dimension to encode high-dimensional relationships.

The algorithm proceeds in two main phases:

Phase 1: Optimizing the Radial Exponent (Temporal Mapping)
- Goal: To map timestamps to annular radii such that the density of points is radially uniform, addressing the "crowding problem" and ensuring time is estimable from position.
- Mechanism: Time $t$ is mapped to radius $r$ via a power function: $r_i = \zeta + (1 - \zeta)t_i^\rho$ , where $\zeta$ is a hyperparameter controlling the inner radius ratio, and $\rho$ is a global exponent to be optimized.
- Optimization: The algorithm defines an ideal probability density function (PDF) $P(X)$ that represents a uniform distribution of density across the annulus. It then uses Kullback-Leibler (KL) divergence to find the optimal $\rho^*$ that minimizes the divergence between the ideal PDF and the actual distribution of radii $Q$ derived from the data. This phase accounts for temporal skewness.
- Resampling: To prevent banding artifacts caused by sparse or coarse time labels, IRIS offers optional probabilistic temporal resampling (homogenous or heterogeneous) to create a more continuous time distribution.
Phase 2: Optimizing Angles (Manifold Structure)
- Goal: To optimize the angle $\theta_i$ of each point to preserve high-dimensional relationships while constraining points to the radii determined in Phase 1.
- Mechanism: The algorithm reparameterizes the Euclidean cost function from Cartesian to polar coordinates. It utilizes a probabilistic framework similar to LargeVis (a forerunner of UMAP), defining edge probabilities based on Approximate Nearest Neighbors (ANN).
- Cost Function: The objective function $C$ balances Euclidean distance and angular distance. For positive pairs (neighbors in high-dimensional space), the cost minimizes the distance in both radius and angle. For negative pairs, it maximizes separation. The cost includes a hyperparameter $\beta$ to balance the contribution of Euclidean distance versus angular distance.
- Optimization: Stochastic gradient descent is used to update the angles $\Theta$ to minimize the divergence of distance distributions between high-dimensional and low-dimensional spaces.

Key Contributions

Algorithmic Innovation: IRIS introduces a novel inductive bias that creates circular layouts with time radiating outward, allowing temporal relationships to be visible directly from the layout geometry rather than relying on color channels.
Reparameterization: The paper details a mathematical reparameterization of Euclidean cost functions into polar coordinates, enabling the simultaneous optimization of global time mapping (radius) and local manifold structure (angle).
Temporal Resampling: A proposed strategy to handle sparse or coarse time labels to avoid visual artifacts in the layout.
Implementation: IRIS is implemented in Python and C++, leveraging the LargeVis framework, and is available as an open-source package.

Results
The authors evaluated IRIS on five datasets spanning three biomedical domains:

scRNA-seq (Embryo): Mouse prenatal development (28,604 cells).
scRNA-seq (Neuro): Neurodegenerative disease atlas (30,556 samples).
Metagenomes (Gut): Gut microbiome data across 29 studies (4,479 samples).
Literature (ADRD): Alzheimer's Disease and Related Dementias corpus (17,518 articles).
Literature (PD-1): Immune checkpoint pathway corpus (27,886 articles).

Evaluation Metrics:

Class-structuring (KNN): Measured by the accuracy of a K-Nearest Neighbors classifier predicting class labels from layout coordinates.
Time-structuring (SVR): Measured by the $R^2$ goodness of fit of a Support Vector Regression model predicting timestamps from layout coordinates.
TMPS: A Time-structured Manifold Projection Score defined as the harmonic mean of KNN and SVR.

Findings:

Visual Clarity: IRIS layouts instantly elucidate dynamic phenomena (e.g., branching of cell types, distinct origins of neuronal precursors) that are not readily apparent in paired UMAP plots.
Quantitative Performance: Quantitative metrics confirm that IRIS effectively structures layouts by time (where UMAP performs poorly) while retaining class structuring comparable to UMAP.
Scalability: IRIS runtime scales linearly with the number of samples.

Significance and Claims
The paper claims that IRIS provides a solution to the limitation of existing Manifold Learning algorithms in handling temporal data. By structuring layouts chronologically and by topology simultaneously, IRIS makes temporal relationships clearly visible from the layout itself, freeing the color channel for encoding class information. The authors position IRIS as a tool capable of visualizing a wide range of dynamic biomedical data, offering a more intuitive and information-dense alternative to side-by-side comparisons or multi-view animations. The work is presented as an open-source contribution to the field of data visualization and machine learning in biomedicine.