protein-ligand-design
收藏资源简介:
Protein-Ligand Design Gym(团队JAMMY)是一个专为工具使用强化学习环境设计的数据集,旨在训练大型语言模型像计算化学家或蛋白质工程师一样进行推理,通过实际测量而非直觉猜测来回答蛋白质-配体相关问题。该数据集聚焦于药物发现和蛋白质工程领域,其中蛋白质是由氨基酸字母串构成的分子机器,配体则是小分子(多数为药物),通过与蛋白质结合来调控其功能。数据集中的问题由Claude Opus 4.8生成,并经过严格的工具验证(使用RDKit和Biopython),确保每个问题都有唯一且可计算验证的答案,无法伪造。数据以parquet格式组织,包含训练集和测试集分割。每个数据样本包含以下字段:问题提示(question)、候选分子或蛋白质实体(以SMILES字符串或序列表示,字段为candidates/entities)、参数(params)、经过工具认证的标准答案(answer)、模板(template)、所需工具(tools_required)以及推理过程(reasoning)。代理模型可调用一系列化学和生物学工具来解决问题,包括计算配体分子描述符(如分子量、LogP、氢键供受体等)、药物相似性定量估计(QED)、Lipinski规则违反次数、Veber口服生物利用度规则、Tanimoto相似性、Murcko骨架、子结构匹配(用于检测共价弹头、盐桥、金属螯合剂等)、蛋白质性质(如等电点、亲水性、二级结构比例)、特定pH下的净电荷、糖基化位点计数、BLOSUM62突变评分、序列同一性以及蛋白质基序搜索。该数据集适用于问答任务,特别是需要模型与专业计算工具交互的强化学习场景,以提升在计算化学和生物信息学领域的可靠推理能力。
Protein-Ligand Design Gym (team JAMMY) is a dataset specifically designed for tool-use reinforcement learning environments, aiming to train large language models to reason like computational chemists or protein engineers by answering protein-ligand related questions through actual measurements rather than intuitive guesses. The dataset focuses on the fields of drug discovery and protein engineering, where proteins are molecular machines composed of amino acid strings, and ligands are small molecules (mostly drugs) that regulate protein function through binding. The questions in the dataset are generated by Claude Opus 4.8 and undergo rigorous tool validation (using RDKit and Biopython), ensuring each question has a unique and computationally verifiable answer that cannot be fabricated. The data is organized in parquet format, including training and test set splits. Each data sample contains the following fields: question prompt, candidate molecules or protein entities (represented as SMILES strings or sequences, with fields candidates/entities), parameters, tool-certified standard answers, templates, required tools, and reasoning processes. The agent model can invoke a series of chemical and biological tools to solve problems, including calculating ligand molecular descriptors (such as molecular weight, LogP, hydrogen bond donors/acceptors, etc.), quantitative estimation of drug-likeness (QED), Lipinski rule violations, Veber oral bioavailability rules, Tanimoto similarity, Murcko scaffolds, substructure matching (for detecting covalent warheads, salt bridges, metal chelators, etc.), protein properties (such as isoelectric point, hydrophilicity, secondary structure proportions), net charge at specific pH, glycosylation site counts, BLOSUM62 mutation scores, sequence identity, and protein motif searches. The dataset is suitable for question-answering tasks, particularly in reinforcement learning scenarios that require interaction with professional computational tools to enhance reliable reasoning capabilities in computational chemistry and bioinformatics.
数据集概述:Protein-Ligand Design Gym (Team JAMMY)
该数据集是一个面向药物发现领域的工具使用强化学习环境,旨在训练大语言模型像计算化学家/蛋白质工程师一样进行推理。
核心目标
- 向模型提供分子或蛋白质以及一个关于它的问题,模型必须调用化学/生物学工具(RDKit + Biopython)来测量答案,然后提交最终答案。
- 奖励机制仅基于答案的正确性,每个真实答案均由相同工具计算得出,评分精确且无法伪造。
- 问题由Claude Opus 4.8生成,并通过工具验证,确保每个问题都有一个可通过工具验证的单一答案。
模型可用的工具
| 工具 | 领域 | 计算内容 |
|---|---|---|
mol_descriptors |
配体 (RDKit) | 分子量、LogP、TPSA、氢键供体/受体、可旋转键、芳香环、sp³碳比例、形式电荷 |
qed |
配体 (RDKit) | 定量类药性估计 (0–1) |
lipinski_violations |
配体 (RDKit) | 类药五规则违反次数 |
veber_pass |
配体 (RDKit) | 口服生物利用度规则(可旋转键≤10,TPSA≤140) |
tanimoto_similarity |
配体 (RDKit) | 两个分子之间的ECFP4 (Morgan r=2, 2048-bit) 相似度 |
murcko_scaffold |
配体 (RDKit) | Bemis–Murcko骨架SMILES |
substructure_match |
相互作用 (RDKit) | 统计SMARTS模式——共价弹头、盐桥基团、金属螯合剂、卤键供体、药效团 |
protein_properties |
蛋白质 (Biopython) | 分子量、等电点、GRAVY亲水性、不稳定指数、芳香性、螺旋/转角/折叠比例 |
net_charge_at_ph |
蛋白质 (Biopython) | 在给定pH下的净电荷 |
count_sequon |
蛋白质 (Biopython) | N-连接糖基化序列 (N-X-[S/T], X≠P) |
blosum62_mutation_score |
蛋白质 (Biopython) | 一组点突变的BLOSUM62得分总和 |
sequence_identity |
蛋白质 (Biopython) | 两条序列之间的全局比对百分比一致性 |
motif_search |
相互作用 (蛋白质) | 统计正则表达式模式——催化/结合基序(如锌结合HE..H、糖基化序列) |
submit_answer |
控制 | 提交最终答案 |
数据内容
data/目录:包含问题数据,分为train和test两个Parquet格式的分片。environment/目录:完整的verifiers gym环境,包括protein_ligand_design.py、捆绑的问题文件及pyproject.toml。
数据字段
question:提示问题candidates/entities:以SMILES或序列形式表示的分子/蛋白质params:参数answer:经工具验证的真实答案template:模板tools_required:所需工具reasoning:推理过程
许可与标签
- 许可协议:Apache-2.0
- 语言:英语
- 标签:chemistry, biology, protein, drug-discovery, tool-use, reinforcement-learning, verifiers, prime-intellect
- 任务类别:question-answering
使用方式
通过Hugging Face datasets库加载:
python from datasets import load_dataset ds = load_dataset("poolside-laguna-hackathon/protein-ligand-design")





