Horizyn-1 Development Dataset: Dual-encoder contrastive learning accelerates enzyme discovery

⚠️ IMPORTANT NOTICE: SUPERSEDED BY VERSION 2 ⚠️ Please use the newest version of this repository. The development model weights (horizyn-v1.ckpt) included in this specific version (v1) are possibly corrupted. A fixed version of the development model (horizyn_v1_0_dev.ckpt), along with the newly released, production-ready inference model weights (horizyn_v1_0_inf.ckpt), are available in Version 2. Overview This repository contains the development dataset used to train and evaluate the Horizyn-1 development model, as described in the paper "Dual-encoder contrastive learning accelerates enzyme discovery." The dataset includes the full set of reaction SMILES, protein-reaction pairs, and pre-computed ProtT5 protein embeddings. The accompanying code for training and evaluation is available at: https://github.com/dayhofflabs/horizyn. A model checkpoint produced using the data and code is also included. Methodology & Splits The training and test sets were created by splitting on reactions to prevent data leakage. The test set was strictly filtered to exclude any reactions with high similarity to any reactions in the training set (see manuscript). For evaluation, the test setup involves identifying the correct enzyme for each test reaction from a total screening pool of 216,132 proteins contained in this dataset. Dataset Statistics Reactions: 10,785 (Train) / 1,012 (Test) Enzymes: 192,769 (Train) / 32,100 (Test) Reaction-Enzyme Pairs: 257,733 (Train) / 33,996 (Test) File Manifest The archive contains the following standardized files: train_rxns.csv & test_rxns.csv Contains reaction SMILES strings. Columns: rs_id, reaction_id, reaction_smiles train_pairs.csv & test_pairs.csv Defines the positive training and testing pairs. Columns: pr_id, reaction_id, protein_id prots_t5.h5 HDF5 file containing pre-computed protein embeddings (ProtT5-XL). Structure: /ids: Dataset of protein IDs (strings) /vectors: Dataset of embeddings (float32, shape: [N, 1024]) prots.fasta (Reference) FASTA file containing protein sequences corresponding to the embeddings in the HDF5 file. horizyn-v1.ckpt Model checkpoint created using this dataset and the official implementation. Code repository: https://github.com/dayhofflabs/horizyn Citation If you use this dataset, please cite the associated preprint: Rocks, J. W., Truong, D. P., Rappoport, D., Maddrell-Mander, S., Martin-Alarcon, D. A., Lee, T., Crossan, S., & Goldford, J. E. (2025). Dual-encoder contrastive learning accelerates enzyme discovery. bioRxiv. DOI: 10.1101/2025.08.21.671639⚠️ 重要提示：本版本已被V2版本替代 ⚠️ 请使用本仓库的最新版本。本特定版本（v1）中包含的开发模型权重（horizyn-v1.ckpt）可能已损坏。 修复版开发模型（horizyn_v1_0_dev.ckpt）以及新发布的可投入生产环境的推理模型权重（horizyn_v1_0_inf.ckpt），均可在V2版本中获取。 ## 数据集概述 本仓库包含用于训练与评估Horizyn-1开发模型的开发数据集，相关细节已在论文《双编码器对比学习加速酶发现》（Dual-encoder contrastive learning accelerates enzyme discovery）中进行了阐述。该数据集包含完整的反应SMILES（reaction SMILES）、蛋白质-反应对以及预计算的ProtT5蛋白质嵌入向量（ProtT5 protein embeddings）。配套的训练与评估代码可通过以下仓库获取：https://github.com/dayhofflabs/horizyn。本仓库同时附带了使用本数据集与官方代码训练得到的模型检查点。 ## 方法与数据集拆分 训练集与测试集通过按反应维度拆分构建，以避免数据泄露问题。测试集经过严格过滤，排除了所有与训练集中反应存在高度相似性的反应（详见论文手稿）。在评估环节中，测试流程需要从本数据集包含的总计216,132个蛋白质的筛选库中，为每个测试反应匹配出正确的酶。 ## 数据集统计 反应：10,785（训练集） / 1,012（测试集） 酶：192,769（训练集） / 32,100（测试集） 反应-酶对：257,733（训练集） / 33,996（测试集） ## 文件清单 本归档文件包含以下标准化文件： 1. "train_rxns.csv" 与 "test_rxns.csv"：存储反应SMILES字符串，列名包括：rs_id、reaction_id、reaction_smiles。 2. "train_pairs.csv" 与 "test_pairs.csv"：定义正样本训练与测试对，列名包括：pr_id、reaction_id、protein_id。 3. "prots_t5.h5"：存储预计算蛋白质嵌入向量的HDF5文件，其内部结构为： - /ids：蛋白质ID数据集（字符串类型） - /vectors：嵌入向量数据集（float32类型，形状为 [N, 1024]） 4. "prots.fasta"（参考文件）：包含与HDF5文件中嵌入向量对应的蛋白质序列的FASTA文件。 5. "horizyn-v1.ckpt"：使用本数据集与官方实现训练得到的模型检查点。 代码仓库地址：https://github.com/dayhofflabs/horizyn ## 引用说明 若您使用本数据集，请引用以下预印本： Rocks, J. W., Truong, D. P., Rappoport, D., Maddrell-Mander, S., Martin-Alarcon, D. A., Lee, T., Crossan, S., & Goldford, J. E. (2025). Dual-encoder contrastive learning accelerates enzyme discovery. bioRxiv. DOI: 10.1101/2025.08.21.671639