RATAN-PBind: ESM-2 embeddings and baseline models for de novo protein binder nomination

Large companion artefacts for RATAN-PBind (Retrieval-Augmented, Target-Aware Nomination for Protein Binding), a sequence-based pre-screen for de novo protein binders. Code and the headline model are at https://github.com/kartic03/RATAN-PBind. This record holds the files excluded from the repository for size: ESM-2 (esm2_t33_650M_UR50D) mean-pooled embeddings required to reproduce the prototype-similarity features, and the Random Forest, SVM, Extra Trees, and fine-tuned ESM-2 baseline models. This work used Proteinbase by Adaptyv Bio under the ODC-BY licence.本数据集为RATAN-PBind（检索增强型靶标感知蛋白质结合候选物筛选，Retrieval-Augmented, Target-Aware Nomination for Protein Binding）的配套辅助数据集，该工具是面向从头设计蛋白质结合剂的基于序列的预筛选模型。其核心代码与旗舰模型已开源至 https://github.com/kartic03/RATAN-PBind。本数据集记录包含因体积限制未纳入代码仓库的文件：用于复现原型相似性特征的ESM-2（esm2_t33_650M_UR50D）平均池化嵌入向量，以及随机森林（Random Forest）、支持向量机（SVM）、极端随机树（Extra Trees）与微调版ESM-2基线模型。本研究使用了Adaptyv Bio公司开发的Proteinbase数据库，其遵循ODC-BY开源许可协议。