Plants cannot escape drought, heat, poor soil nutrients or disease, so improving growth and stress tolerance is urgent. High-throughput phenotyping (HTP) uses automated imaging to measure thousands of plants over time, generating massive datasets. Traditional supervised artificial intelligence (AI) depends on labelled images and known patterns, while Euclidean methods often flatten hierarchical and temporal relationships into linear representations. These limits obscure subtle treatment effects and dose responses. Manual feature engineering and supervised labels also slow discovery in multifactorial experiments. Given these challenges, deeper research is needed into geometry-aware, unsupervised analytical frameworks that can resolve hierarchical and dynamic plant phenotypes.
Researchers at the Czech Advanced Technology and Research Institute (CATRIN), Palacký University Olomouc, and VSB–Technical University of Ostrava published (DOI: 10.1016/j.plaph.2026.100186) the study in Plant Phenomics (2026). The team combined topological data analysis (TDA), hyperbolic geometry and contrastive learning to analyse Arabidopsis thaliana seedlings under two nutrient regimes, five cytokinin oxidase/dehydrogenase (CKX) inhibitor compounds and four concentrations over seven days. The work introduces Hyperbolic Topological Data Analysis Mapper (HTDA-Mapper) for both plant-derived descriptors and raw red-green-blue (RGB) images. It provides a geometry-aware alternative to conventional Euclidean Mapper pipelines.
The study analysed 27,761 Arabidopsis seedlings imaged twice daily for seven days. Plants grew under full or low Murashige & Skoog (MS) nutrients and were primed with five CKX inhibitor compounds at 0.01, 0.1, 1 or 10 µM. Seven descriptors were extracted: projected rosette area (AREAPX), perimeter (PERIMETER), Green Leaf Index (GLI), Normalised Green-Red Difference Index (NGRDI), Visible Atmospherically Resistant Index (VARI), absolute growth rate (AGR) and relative growth rate (RGR).
Euclidean UMAP (Uniform Manifold Approximation and Projection) with HDBSCAN (Hierarchical Density-Based Spatial Clustering of Applications with Noise) often represented time linearly. HTDA-Mapper instead used hyperbolic UMAP, a Riemannian K-means cover and hyperbolic DBSCAN (Density-Based Spatial Clustering of Applications with Noise), revealing branching, nested trajectories. It separated low-nutrient plants treated with low versus high compound concentrations, grouped 10 µM treatments in central clusters, and linked related compounds with similar phenotypes, suggesting shared growth-promoting or stress-alleviating modes. Raw-image models SimCLR (Simple Framework for Contrastive Learning of Visual Representations) and BYOL (Bootstrap Your Own Latent) recovered comparable structure without manual labelling. The authors reported that UMAP achieved the highest trustworthiness for descriptor data, supporting its use as the Euclidean baseline. These patterns were less visible in Euclidean maps, where temporal progression frequently collapsed into a near-linear sequence.
The authors said the advance is less about replacing plant biologists than giving them a map that preserves the shape of development. They said hyperbolic geometry reflects the coarse-to-fine hierarchy of nutrient regime, compound identity and concentration, allowing the data to branch naturally rather than collapse into a line. They added that the pipeline can move from handcrafted descriptors to raw images, helping researchers spot compound-specific effects and promising doses without manual labelling. The team said HTDA-Mapper is designed as a practical discovery tool for high-throughput phenotyping and other structured omics data.
Beyond Arabidopsis, HTDA-Mapper could support crop improvement by revealing effective compounds, robust genotypes and adaptive growth strategies that enhance resilience. Its ability to handle quantitative traits and images may extend to transcriptomics, metabolomics and other omics fields where hierarchical, high-dimensional data are common. By combining topological data analysis, hyperbolic geometry and unsupervised contrastive learning, the framework offers an interpretable route from big data to actionable biological hypotheses.
###
References
DOI
10.1016/j.plaphe.2026.100186
Original Source URL
https://doi.org/10.1016/j.plaphe.2026.100186
Funding information
This work was supported by the project Interdisciplinary approaches for the development and application of new materials in industrial, agricultural and medical practice, reg. no. CZ.02.01.01/00/23_021/0008909 of the ERDF Programme Johannes Amos Comenius.
About Plant Phenomics
Plant Phenomics publishes breakthrough research that integrates genomics, genetics, physiology, molecular biology, bioinformatics, statistics, mathematics and computer science to advance plant phenotyping—from the cellular level to entire populations. The journal tackles key scientific challenges in phenomics by showcasing cutting‑edge technologies for data acquisition, management, interpretation and modelling. Its scope spans practical applications across all plant sciences, aiming to translate novel phenotyping tools into real‑world solutions for breeding, crop management and fundamental plant biology. By fostering interdisciplinary innovation, Plant Phenomics serves as a hub for researchers developing high‑throughput, accurate and scalable methods to decipher complex plant traits, ultimately accelerating progress in agriculture and plant science.