LiteFold/ESMAtlas

ESMAtlas是ESM宏基因组图谱的Hugging Face镜像，这是一个大规模预测宏基因组蛋白质结构资源，使用Meta AI的ESMFold蛋白质语言模型生成。上游图谱包含来自MGnify蛋白质序列的超过7亿个预测结构；v2023_02版本对应MGnify 2023_02，并使用esm.pretrained.esmfold_v1()模型。该数据集适用于大规模蛋白质结构检索、基于结构的搜索、置信度过滤子集、表示学习，以及研究宏基因组蛋白质的结构多样性，这些蛋白质通常缺乏实验解析结构的覆盖。数据集以WebDataset tar分片形式存储压缩的PDB文件，按预测TM分数（ptm）和平均预测LDDT（plddt）分箱，方便用户选择性下载高置信度区域。包含的元数据文件支持过滤和查找，包括每个蛋白质的置信度分数、序列长度、片段标志、序列校验和、ESMFold版本、图谱版本和源序列数据库成员信息。

ESMAtlas is a Hugging Face mirror of the ESM Metagenomic Atlas, a large-scale resource of predicted metagenomic protein structures generated with Meta AIs ESMFold protein language model. The upstream Atlas contains more than 700 million predicted structures from MGnify protein sequences; the v2023_02 release corresponds to MGnify 2023_02 and uses esm.pretrained.esmfold_v1(). This dataset is useful for large-scale protein structure retrieval, structure-based search, confidence-filtered subsets, representation learning, and studying the structural diversity of metagenomic proteins that are often poorly covered by experimentally solved structures. The repository stores predicted structures as compressed PDB files inside WebDataset tar shards. Shards are grouped by predicted TM score (ptm) and average predicted LDDT (plddt) bins, so users can selectively download higher-confidence regions of the Atlas rather than the full collection. Metadata files are included separately for filtering and lookup, including per-protein confidence scores, sequence lengths, fragment flags, sequence checksums, ESMFold version, Atlas version, and source sequence database membership.