IRIS: time-structured manifold projections

2026-05-28

AI Summary

Existing manifold learning methods, such as t-SNE and UMAP, struggle to preserve temporal ordering in dimensionality reduction visualizations, thereby hindering the analysis of dynamic biological processes like cellular state transitions. To address this limitation, this work proposes IRIS, an algorithm that explicitly incorporates temporal structure into the manifold learning process for the first time. By embedding temporal continuity while preserving nonlinear topological relationships, IRIS enables more faithful representation of time-evolving patterns in high-dimensional data. The method is broadly applicable to diverse time-series datasets, including single-cell RNA sequencing, metagenomics, and scholarly literature evolution. Evaluated across multiple biomedical datasets, IRIS successfully uncovers clear trajectories and dynamic patterns of cellular states over time, offering enhanced interpretability of biological progression.

Abstract

High-dimensional biomedical data, such as cell-by-gene matrices, are increasingly generated temporally. However, Manifold Learning algorithms, like t-SNE and UMAP, cannot incorporate time-ordering in their layouts, obfuscating the dynamics of cell types or other classes. As a solution, we present IRIS, a new Manifold Learning algorithm that structures layouts both chronologically and by manifold topology. IRIS can visualize a wide range of dynamic biomedical data, including scRNA-seq, comparative metagenomics, and literature.

Problem

• manifold learning
• time-structured data
• temporal dynamics
• biomedical data
• dimensionality reduction

Innovation

• Manifold Learning
• time-structured visualization
• IRIS
• single-cell RNA-seq
• temporal dynamics