AlphaFoldDB
收藏资源简介:
AlphaFoldDB Prediction Index是一个基于AlphaFold DB构建的开放数据集,提供大规模预测的蛋白质三维结构及其置信度分数,显著扩展了已知蛋白质序列的结构覆盖范围。数据集以Parquet格式存储,包含约2.47亿行数据,分为训练集(约2.22亿行)和测试集(约0.25亿行),分割方式基于UniProt登录号的哈希值取模10的确定性分割(模0为测试集,模1-9为训练集)。数据统计显示蛋白质序列长度最小为5,最大为4,186,中位数约278,平均长度328.55。最新版本以版本6为主(约2.41亿行)。数据集包含8个关键字段:UniProt登录号(uniprot_accession)、AlphaFold DB标识符(alphafold_id)、最新模型版本(latest_version)、起始和结束残基索引(first_residue_index、last_residue_index)、序列长度(sequence_length)、片段编号(fragment_number,可为空)、是否为片段化预测(is_fragmented_prediction)以及确定性分割桶(split_bucket)。数据来源于AlphaFold DB的镜像,包括48个批量归档文件(16个蛋白质组归档、30个全球健康归档和2个Swiss-Prot归档)。适用于蛋白质结构预测、生物信息学分析、机器学习模型训练与评估等任务,可通过HuggingFace datasets库加载,支持流式处理以提高大规模数据处理效率。
AlphaFoldDB Prediction Index is an open dataset built on AlphaFold DB, providing large-scale predicted protein 3D structures and confidence scores, significantly expanding the structural coverage of known protein sequences. The dataset is stored in Parquet format, containing approximately 247 million rows, divided into a training set (about 222 million rows) and a test set (about 25 million rows), with a deterministic split based on modulo 10 of the hash value of UniProt accession numbers (modulo 0 for test set, modulo 1-9 for training set). Data statistics show a minimum protein sequence length of 5, maximum of 4,186, median of approximately 278, and average length of 328.55. The latest version is predominantly version 6 (about 241 million rows). The dataset includes 8 key fields: UniProt accession (uniprot_accession), AlphaFold DB identifier (alphafold_id), latest model version (latest_version), start and end residue indices (first_residue_index, last_residue_index), sequence length (sequence_length), fragment number (fragment_number, nullable), whether it is a fragmented prediction (is_fragmented_prediction), and deterministic split bucket (split_bucket). Data is sourced from a mirror of AlphaFold DB, comprising 48 batch archive files (16 proteome archives, 30 global health archives, and 2 Swiss-Prot archives). It is suitable for tasks such as protein structure prediction, bioinformatics analysis, machine learning model training and evaluation, and can be conveniently loaded via the HuggingFace datasets library, supporting streaming for efficient large-scale data processing.
数据集概述
AlphaFoldDB Prediction Index 是一个开放的预测蛋白质3D结构索引数据集,包含蛋白质序列的结构预测与置信度评分。数据集以 Parquet 格式存储,总共有 246,689,516 行 数据,涵盖 14 个 Parquet 文件。
数据集划分
| 划分 | 行数 | Parquet 文件数 |
|---|---|---|
| 训练集(train) | 222,017,452 | 12 |
| 测试集(test) | 24,672,064 | 2 |
| 总计 | 246,689,516 | 14 |
- 划分规则:基于
hash(uniprot_accession) % 10的结果,结果为0的行归入测试集,结果1到9的行归入训练集。
数据集统计
| 指标 | 数值 |
|---|---|
| 总行数 | 246,689,516 |
| 最小序列长度 | 5 |
| 近似中位数序列长度 | 278 |
| 平均序列长度 | 328.55 |
| 最大序列长度 | 4,186 |
| 缺少解析片段编号的行数 | 5,619,027 |
最新版本分布:
| 最新版本 | 行数 |
|---|---|
| 1 | 5,271,725 |
| 2 | 347,302 |
| 6 | 241,070,489 |
镜像文件 download_metadata.json 描述了 48 个批量归档文件:16 个蛋白质组归档、30 个全球健康归档和 2 个 Swiss-Prot 归档。
数据集列说明
| 列名 | 描述 |
|---|---|
uniprot_accession |
来自 accession_ids.csv 的 UniProt 登录号。 |
alphafold_id |
AlphaFold DB 标识符,例如 AF-Q5VSL9-F1。 |
latest_version |
条目可用的最新 AlphaFold DB 模型版本。 |
first_residue_index |
UniProt 编号中的第一个残基索引。 |
last_residue_index |
UniProt 编号中的最后一个残基索引。 |
sequence_length |
推导值:last_residue_index - first_residue_index + 1。 |
fragment_number |
从 alphafold_id 解析的 F<数字> 后缀,当后缀缺失或非标准时可为空。 |
is_fragmented_prediction |
指示 fragment_number 是否大于 1。 |
split_bucket |
由 hash(uniprot_accession) % 10 确定的桶,桶 0 为测试集。 |
许可与引用
- 许可协议:Creative Commons Attribution 4.0 International (CC-BY-4.0)
- 引用:
@article{varadi2022alphafolddb, title = {{AlphaFold} Protein Structure Database: massively expanding the structural coverage of protein-sequence space with high-accuracy models}, author = {Varadi, Mihaly and Anyango, Stephen and Deshpande, Mandar and others}, journal = {Nucleic Acids Research}, volume = {50}, number = {D1}, pages = {D439--D444}, year = {2022}, doi = {10.1093/nar/gkab1061} }




