lamm-mit/structural-protein-families

该数据集包含来自多个结构/生物材料蛋白质家族（蜘蛛丝蛋白、蚕丝蛋白、胶原蛋白、弹性蛋白、节肢弹性蛋白、角蛋白）以及球状对照蛋白（溶菌酶、肌红蛋白、细胞色素c）的氨基酸序列窗口。序列从UniProt数据库获取，并切分为长度不超过200个氨基酸的重叠窗口（步长为150，每个蛋白质最多4个窗口）；每行数据带有family类别标签。数据集专为迁移学习演示而构建，用于在冻结的ESMC嵌入上训练轻量级分类头部，以识别蛋白质家族。

Amino-acid sequence windows from several structural / biomaterials protein families (spider spidroins, silkworm fibroin, collagen, elastin, resilin, keratin) plus globular controls (lysozyme, myoglobin, cytochrome c). Sequences are fetched from UniProt and sliced into overlapping windows; each row carries a family label. Built for the transfer-learning demo to train a lightweight head on frozen ESMC embeddings to classify family.