LORE-examples
收藏资源简介：
LORE Examples是一个为MIMIC多模态生物模型设计的小型示例数据集，包含从LORE中提取的匹配多模态样本。该数据集旨在让用户无需准备自有数据即可尝试在DNA、RNA和蛋白质三种模态上进行推理、嵌入和生成任务。每个样本对应一个生物实体（一个转录本和/或其蛋白质），包含多个共观测的模态。样本来自MIMIC训练数据的保留（验证）分割，因此属于分布内数据，且长度受限于模型的上下文窗口。数据集提供两种并行视图配置：raw配置包含原始模态值（如核苷酸/氨基酸字符串、逐位置轨迹等），tokenized配置包含预处理的token ID列表。涵盖的模态分为三类：核酸模态（如RNA/DNA序列、剪接区域、进化保守性评分等）、蛋白质模态（如氨基酸序列、3D结构、溶剂可及表面积等）和文本模态（如功能描述、生物医学文献、生物分类等）。部分测定轨迹（如ATAC-seq、CAGE、RASP2、蛋白质丰度）是上下文条件性的，需要搭配自由文本的context字段使用。每个样本还包含kind（rna/protein/both）、uniprot_id和genome_feature_id等锚点标识。数据集适用于多模态生物数据特征提取、模型测试和验证等场景。
LORE Examples is a small example dataset designed for the MIMIC multimodal biological model, containing matched multimodal samples extracted from LORE. This dataset aims to allow users to try inference, embedding, and generation tasks across three modalities—DNA, RNA, and protein—without preparing their own data. Each sample corresponds to a biological entity (a transcript and/or its protein) and includes multiple co-observed modalities. The samples are sourced from the reserved (validation) split of MIMIC training data, making them in-distribution, with lengths limited by the models context window. The dataset offers two parallel view configurations: the raw configuration contains raw modality values (such as nucleotide/amino acid strings, per-position traces, etc.), and the tokenized configuration contains preprocessed token ID lists. The covered modalities are divided into three categories: nucleic acid modalities (e.g., RNA/DNA sequences, splicing regions, evolutionary conservation scores, etc.), protein modalities (e.g., amino acid sequences, 3D structures, solvent accessible surface area, etc.), and text modalities (e.g., functional descriptions, biomedical literature, biological classifications, etc.). Some assay traces (e.g., ATAC-seq, CAGE, RASP2, protein abundance) are context-dependent and require the use of the free-text context field. Each sample also includes anchor identifiers such as kind (rna/protein/both), uniprot_id, and genome_feature_id. The dataset is suitable for scenarios such as multimodal biological data feature extraction, model testing, and validation.
