Supplementary materials for "MemConverter: An Iterative Pipeline for Reprogramming Protein Localization in Membrane or Aqueous Solution"

Description This dataset contains 85,051 isolated transmembrane domains derived from the tmAFDB and TED. It was specifically curated to fine-tune ProteinMPNN for membrane protein design (MemProtMPNN). We focused on isolated domains to ensure the model learns specific membrane topological constraints rather than soluble features. Data Construction Domain Extraction: Primarily based on domain annotations from The Encyclopedia of Domains (TED). For proteins without TED annotations, Merizo was used to identify domains. Filtering: Validated by TMbed. We excluded proteins with <20% transmembrane residues or domains shorter than 32 residues. Splitting: Clustered using MMseqs2 (30% sequence identity) to strictly separate Training, Validation, and Test sets. Statistics Set Count Percentage Train 72,293 85% Validation 7,655 9% Test 5,103 6% Files Description membrane_domain_dataset.zip: The main dataset archive. data_files.zip: Contains the processed dataset partitions (Train, Validation, Test). test_list: A curated list of representative cases from the test set (one representative sequence per cluster) used for performance evaluation.数据集概述 本数据集包含源自tmAFDB与TED的85051个独立跨膜结构域（transmembrane domains），系为膜蛋白设计场景下的ProteinMPNN微调任务专门整理构建，该微调模型命名为MemProtMPNN。 我们选取独立跨膜结构域作为研究对象，旨在确保模型能够学习膜蛋白特有的拓扑约束特性，而非可溶性蛋白的相关特征。 数据构建流程 结构域提取：主要基于结构域百科（The Encyclopedia of Domains，简称TED）的结构域注释信息进行提取；对于未获取TED注释的蛋白质，则通过Merizo工具识别其结构域。 筛选流程：所有数据均通过TMbed工具进行验证，我们剔除了跨膜残基占比低于20%的蛋白质，以及长度不足32个氨基酸残基的结构域。 数据集划分：采用MMseqs2工具，以30%序列同一性阈值完成聚类，严格划分训练集、验证集与测试集。 数据集统计 | 数据集划分 | 样本数量 | 占比 | | :--------- | :------- | :--- | | 训练集 | 72293 | 85% | | 验证集 | 7655 | 9% | | 测试集 | 5103 | 6% | 文件说明 1. membrane_domain_dataset.zip：主数据集归档文件 2. data_files.zip：包含已完成划分的训练集、验证集、测试集子集数据 3. test_list：从测试集中精选出的代表性样本列表（每个聚类仅保留一条代表性序列），用于模型性能评估