VNS spatial transcriptomics results
Description
Overview This dataset bundles spatially referenced observations from spatial transcriptomics and related analyses, packaged for research aligned with a hybrid quantum–classical pipeline for spatial omics: classical dimensionality reduction (e.g., autoencoder on expression-like or derived features), optional quantum kernel similarity, and variational quantum circuits (VQCs) for classification or segmentation-style tasks. In the broader project, inputs are conceptualized as spots (or cells) × features together with 2D coordinates and reference labels (e.g., cortical layers, cell types, or domains), enabling clustering metrics (ARI, AMI), classification (F1, AUC), and spatial coherence (e.g., agreement of predictions with spatial neighbors). The files here are not raw UMI count matrices; they are tabular summaries per spot (or per spatial unit) with coordinates, ground-truth or reference annotations, and integer cluster or assignment IDs from baseline spatial methods and from VNS (Variational Neighbor Search / your spatial method naming in the repo). They are ideal for: Benchmarking spatial domain or cell-type–aware clustering vs. reference labels. Treating method outputs as multi-channel numeric features for downstream ML (including quantum-inspired or hybrid models). Visualizing spatial maps of labels and comparing methods side by side. Teaching spatial transcriptomics evaluation without loading full H5AD or raw 10x folders. Scientific context (project reference) Spatial transcriptomics links gene expression to tissue architecture (e.g., 10x Visium–style spots with (x, y)). The associated work studies high-dimensional, noisy spot-level data and proposes hybrid pipelines: classical encoding of spot-level signals into a latent space, then quantum components for similarity (kernel) or trainable classification (VQC). Real benchmarks with known spatial structure (layers, anatomical domains, or curated cell types) are essential to report ARI/AMI against ground truth and to assess whether predictions respect spatial context. This dataset provides such benchmarks in CSV form: each row is a spatial entity with position, reference label, and multiple algorithmic assignments (baselines + VNS), which can be used as labels for supervised learning or as pseudo-features for unsupervised or hybrid models.
Files
Steps to reproduce
Human DLPFC (dorsolateral prefrontal cortex) — Visium-style slices Files: DLPFC_07_10_VNS_results.csv DLPFC_69_72_VNS_results.csv DLPFC_73_76_VNS_results.csv Content: Rows correspond to spatial spots (10x-style barcodes in the index, e.g. …-1_151507). Columns include: X, Y: Spot coordinates in the tissue array / image space. Ground Truth: Reference spatial domain (e.g., cortical layers such as Layer1–Layer6, WM). Baseline method outputs: Multiple clustering/community-detection results from STAligner (e.g., Leiden, Louvain, KMeans, MClust) as integer cluster IDs. VNS_STAligner: Integer assignment from the VNS pipeline paired with STAligner-style features. batch_name: Slice / batch identifier (e.g., 151507), so you can train per-slice, merge slices, or study batch effects. Use in the project: These slices are standard spatial domain identification benchmarks; they support evaluation of spatial clustering, classification (e.g., layer prediction from engineered features), and spatial coherence of predictions—consistent with the paper’s emphasis on ARI/AMI and spatially aware ML. Forebrain (developmental / brain spatial atlas context) Files: Forebrain_VNS_Ground_truth.csv Forebrain_VNS_vs_other_methods.csv Content: Per spatial location: X, Y: Coordinates. Ground Truth: Fine-grained cell type or developmental labels (e.g., radial glia, glutamatergic/GABAergic lineages, Cajal–Retzius, erythrocytes, etc.). Method columns (in the comparison file): Integer IDs from VNS variants and classical spatial baselines such as CCST, STAGATE, GraphST, PCA-based clustering—allowing direct comparison of spatial clustering quality vs. the same reference annotation. Use in the project: Multi-class spatial structure with rich biology; suitable for multi-class extensions of hybrid pipelines, confusion analysis, and per-class spatial visualization, beyond binary cancer-microenvironment examples in the paper narrative. Dataset “E2” (second spatial benchmark) Files: E2_VNS_Ground_Truth.csv E2_VNS_vs_other_methods.csv Content: Same schema style as Forebrain: coordinates, ground-truth labels, and VNS vs. other spatial methods as integer cluster/assignment columns. Use in the project: Additional held-out-style spatial benchmark for generalization of models or for cross-dataset comparison (E2 vs. DLPFC vs. Forebrain). pciSeq-related auxiliary tables Files: pciSeq_centroids.csv incell_mrna_list_pciseq.csv pciseq_CM.csv Content: pciSeq_centroids.csv: Two columns of numeric coordinates (one row per spatial unit)—likely cell or spot centroids for pciSeq-style in situ resolution. incell_mrna_list_pciseq.csv: A single column of integer indices (e.g., gene or feature indices used in pciSeq-style in-cell mRNA lists). pciseq_CM.csv: Dense numeric matrix (rows of many floating-point entries)—consistent with a confusion-style, co-occurrence, or count matrix across many categories (dimensions align with a larger label or gene space).