scShape-Benchmark/scShapeBench

scShapeBench是一个精心策划的数据集集合，用于在单细胞形状分析中基准测试计算方法。它分为四种配置：scRNAseq（真实世界的单细胞基因表达数据集，包含2,547,517个细胞、102个数据集，以AnnData格式存储，带有预计算的PCA嵌入和Leiden聚类）、synthetic（为受控基准测试生成的合成单细胞数据，以NumPy压缩数组格式存储）、annotations（来自9个独立注释者的每个数据集的形状标签，每个注释者为他们审查的每个数据集分配一个或多个形状类别，如archetypal、multi_branch、simple_traj、clusters，以Parquet格式存储）和labels（使用论文中描述的三种策略从9个注释者标签派生的聚合形状标签，包括majority、soft、confidence_weighted、union，以Parquet格式存储，值为每个形状类别的浮点数在[0, 1]范围内）。数据集还包括元数据文件，如cell_metadata.csv、gene_metadata.csv和dataset_index.csv，以及符合Croissant 1.1标准的croissant.json文件。scShapeBench is a curated collection of datasets assembled for benchmarking computational methods in single-cell shape analysis. It is organized into four configurations: scRNAseq (real-world single-cell gene expression datasets, with 2,547,517 total cells, 102 datasets, stored as AnnData files with precomputed PCA embeddings and Leiden clustering), synthetic (synthetically generated single-cell data for controlled benchmarking, stored as NumPy compressed arrays), annotations (per-dataset shape labels from 9 independent annotators, with shape categories such as archetypal, multi_branch, simple_traj, clusters, stored in Parquet format), and labels (aggregated shape labels derived from the 9 annotator labels using three strategies: majority, soft, confidence_weighted, union, stored in Parquet format with float values in [0, 1] per shape category). The dataset also includes metadata files like cell_metadata.csv, gene_metadata.csv, dataset_index.csv, and a croissant.json file conforming to the Croissant 1.1 standard.