Synthyra/PDB-Chain-Complex-Benchmark-Rigor
收藏资源简介:
PDB-Chain-Complex-Benchmark-Rigor是一个基于蛋白质数据银行(PDB)的链和复合物基准数据集,用于机器学习和生物信息学应用。该数据集是对原始数据集Synthyra/PDB-Chain-Complex-Benchmark的严格重新构建,重新计算了拆分分配以确保数据完整性。它包含两种配置:chains(单个蛋白质链)和complexes(蛋白质复合物),数据经过过滤,排除了序列长度小于10或包含X残基的链。拆分策略基于多个组件(如序列、序列哈希、30%序列聚类、结构聚类等)进行连接,目标是训练集占90%、验证集和测试集各占5%的组件计数分布,并严格防止训练、验证和测试集之间的数据泄漏。数据集支持集群平衡的PyTorch采样,适用于训练模型时避免过拟合。构建元数据显示,链数据包含约103.7万行,复合物数据包含约34.5万行,并提供了详细的统计信息和质量控制检查。
PDB-Chain-Complex-Benchmark-Rigor is a chain and complex benchmark dataset derived from the Protein Data Bank (PDB), intended for machine learning and bioinformatics applications. This dataset is a rigorous reconstruction of the original Synthyra/PDB-Chain-Complex-Benchmark dataset, with split assignments recalculated to ensure data integrity. It features two configurations: chains (single protein chains) and complexes (protein complexes). The dataset has been filtered to exclude chains with sequence length less than 10 or containing X residues. The split strategy is formulated by combining multiple components (e.g., sequence, sequence hash, 30% sequence identity clustering, structure clustering, etc.), targeting a component count distribution where the training set accounts for 90% of the total, while the validation and test sets each account for 5%, with strict prevention of data leakage between the training, validation, and test sets. The dataset supports cluster-balanced PyTorch sampling, which is ideal for mitigating overfitting during model training. The metadata generated during dataset construction shows that the chain subset contains approximately 1,037,000 rows, and the complex subset contains approximately 345,000 rows, with detailed statistical information and quality control checks provided.




