AIRI-Institute/genatator-gene-segmentation-dataset
收藏资源简介:
genatator-gene-segmentation-dataset 是一个核苷酸级别的基因分割数据集,专为训练和评估DNA语言模型及相关序列模型在转录本结构预测任务上而设计。该数据集旨在实现转录本架构的生物学详细重建,并支持在从头基因注释背景下的基准测试。每个样本代表一个注释的转录本,并提供核苷酸分辨率的标签,描述转录本的组织结构,包括5非翻译区(5 UTR)、外显子、内含子、3非翻译区(3 UTR)和编码序列(CDS)。数据集包含三个配置:train-human(人类训练数据)、train-multi-specie(多物种训练数据)和val-human(人类验证数据)。所有配置均保留所有基因的所有注释转录本,未应用转录本级别的截断,这支持转录本级别的分割研究,同时允许在多异构体设置中进行基因级别评估。每个样本包含四个字段:dna_sequence(DNA序列)、labels(标签)、metadata(元数据)和status(状态)。数据集适用于训练DNA语言模型、微调预训练的基因组基础模型、基准测试核苷酸级别和基因结构感知的分割方法、评估从仅人类到多物种训练的泛化能力、研究蛋白质编码和长非编码转录本的分割,以及在保留每个基因多个转录本异构体的设置中评估模型。
genatator-gene-segmentation-dataset is a nucleotide-level gene segmentation dataset designed for training and evaluating DNA language models and related sequence models on transcript structure prediction. The dataset targets biologically detailed reconstruction of transcript architecture and supports benchmarking in the context of ab initio gene annotation. Each example represents one annotated transcript and provides nucleotide-resolution labels describing transcript organization, including 5 untranslated region (5 UTR), exon, intron, 3 untranslated region (3 UTR), and coding sequence (CDS). The repository contains three configurations: train-human, train-multi-specie, and val-human. In the current version of the dataset, all configurations retain all annotated transcripts for all genes. No transcript-level truncation is applied. This design supports transcript-level segmentation studies while also enabling gene-level evaluation in multi-isoform settings. Each sample contains exactly four fields: dna_sequence, labels, metadata, and status. It is intended for training DNA language models for transcript segmentation, fine-tuning pretrained genomic foundation models, benchmarking nucleotide-level and gene-structure-aware segmentation methods, evaluating generalization from human-only to multispecies training, studying segmentation of both protein-coding and long non-coding transcripts, and evaluating models in settings where multiple transcript isoforms per gene are retained.




