DSIMB/PATHOS-PLM-EMBEDDINGS

PATHOS PLM嵌入数据集是一个预计算的蛋白质语言模型嵌入集合，用于20,416个人类SwissProt蛋白质的错义替换和野生型残基。该数据集旨在通过PATHOS工具预测错义突变的致病性。包含三种蛋白质语言模型的嵌入：ESM-C 600M、ESM-2 650M和Ankh2 Large，每种模型都提供突变嵌入（针对生成的错义替换）和野生型嵌入（针对原始蛋白质序列）两种配置。突变嵌入包含每个生成变体的一行数据，记录蛋白质ID、位置、野生型氨基酸、突变氨基酸、变异格式、位置特异性嵌入和平均池化序列嵌入。野生型嵌入包含源WT嵌入矩阵中每个残基的一行数据，记录蛋白质ID、位置、位置特异性野生型残基嵌入和平均池化野生型序列嵌入。总数据集规模约6.5 TiB，包含超过6.8亿行数据。数据集支持流式访问和特定蛋白质查询，适用于生物信息学研究和蛋白质功能预测。

The PATHOS PLM Embeddings dataset is a collection of precomputed protein language model embeddings for missense substitutions and wild-type residues in 20,416 human SwissProt proteins. This dataset is used by PATHOS to predict the pathogenicity of missense mutations. It includes embeddings from three protein language models: ESM-C 600M, ESM-2 650M, and Ankh2 Large, with each model providing both mutation embeddings (for generated missense substitutions) and wild-type embeddings (for the original protein sequences). Mutation embeddings contain one row per generated variant, recording protein ID, position, wild-type amino acid, mutant amino acid, variation format, position-specific embedding, and mean-pooled sequence embedding. Wild-type embeddings contain one row per residue in the source WT embedding matrix, recording protein ID, position, position-specific wild-type residue embedding, and mean-pooled wild-type sequence embedding. The total dataset size is approximately 6.5 TiB, containing over 682 million rows. The dataset supports streaming access and specific protein queries, making it suitable for bioinformatics research and protein function prediction.