PlantAG: Plant Tissue Annotation via Graph Topology

PlantAG is an automated cell-type annotation tool for plant spatial transcriptomics data. It requires no reference atlas, overcoming the fundamental bottleneck that has blocked automated annotation in plant spatial genomics. How it works: For each cell, PlantAG computes Vietoris-Rips persistent homology (H0+H1) on a local k-NN neighbourhood in PCA space, converts the resulting persistence diagrams into vectorised persistence images and scalar TDA features, then trains a 3-layer Graph Convolutional Network (GCNConv, hidden=256) on cosine-similarity cell graphs using marker-gene pseudo-labels derived from Leiden clustering. The penultimate-layer embeddings serve as a biologically meaningful latent space for downstream trajectory analysis (diffusion pseudotime, RNA velocity).Performance: ~90% annotation accuracy across 10 tomato leaf cell types (1,740 cells, 8 Visium sections). Training time ~8 min on GPU, ~45 min on CPU. Introduced in: Luna*, Aberbache* et al. (2026) — TYLCV Cell Atlas. See companion dataset at https://github.com/MorillaLab/tylcv-tomato-visium. This Zenodo deposit contains: 1. Full Python package source code (planttag/). 2. Test suite (tests/). 3. Installation files (setup.py, requirements.txt, environment.yml). 4. Worked example notebooks (examples/). 5. Documentation (docs/), Pre-trained model weights (planttag/weights/planttag_tomato.pt) — NOTE: repository to be realised upon manuscript acceptance.