Discovery & Synthesis Chemputer: An intelligent universal system for automated chemical synthesis and discovery across different hardware and scales

项目摘要 在这个补充的合作项目之间发起的数字化学组在大学 格拉斯哥和NCATS ASPIRE实验室，我们将深化χ DL化学编程的整合 语言与开放反应数据库（ORD）以及集成大型语言模型（LLM）的基础上， AI方法直接从目标化合物的逆合成分析生成χ DL程序。 这项工作将在原赠款期限内完成。提出了两个具体目标：1。发展 一个xDL到ORD桥，可以在基于化学计算机的物理合成平台上实例化。（Coley Lab） 协作）; 2.在化学描述语言（χ DL）中集成大型语言模型（LLM） 为闭环主动学习基础设施（Chopra实验室）生成开发和接口χ DL的框架 协作）。这些目标将在供资期间以高度综合和协作的方式制定。 工作模式对于具体的目标一，我们将开发一套转换器桥接的三个阶段， 实验生命周期：计划、执行和报告。这是通过整合规划和 报告阶段，可以完全由ORD的结构化数据模式表示，中央 执行阶段，可以用xDL完全表示。这些转换器将包括某种程度的推理， 通过物流或其他方式，填写原始计划中可能没有明确定义的程序细节。 他们还可以根据硬件兼容性验证计划是否可以在特定实验室中执行。我们将 将这些转换器实现为开源软件工具，并在计算机硬件平台上测试这些工具 一组基准反应。对于具体的目标二，我们将开发一个扩展到我们的自然语言 通过使用生成的数据集来训练LLM AI系统， 能够直接从所需分子结构的逆合成分析中生成xDL指令文件。 这将通过构建一组定制的LLM代理来实现，这些代理旨在利用χ DL NLP模型来 基于用户输入解释和写入有效的χ DL代码。通过将这些与χ DL蓝图相结合， 作为NCATS ASPIRE合作的一部分，正在为基准反应开发这些χ DL指令， 然后可以从给定分子或一类分子的自动化逆合成分析中产生， 如果所建议的反应在化学文献中还不存在的话。我们将生产规格为进一步 基于LLM的AI系统解释自动分析生成的数据，并建议新的后续 基于预定义的适应度函数优化（例如产物的产率或纯度）的实验， 可以在自动化系统中通过实验来定义。

Project Summary In this supplement to the collaborative project initiated between the Digital Chemistry Group at the University of Glasgow and The NCATS ASPIRE laboratory we will deepen the integration of the χDL chemical programming language with the Open Reaction Database (ORD) as well as integrating Large Language Model (LLM)-based AI approaches into the generation of χDL procedures directly from retrosynthetic analyses of target compounds. This work will be accomplished during the term of the original grant. Two specific aims are proposed: 1. Develop a χDL to ORD bridge which can be instantiated on a chemputer-based physical synthesis platform. (Coley Lab collaboration); 2. Integrate large language models (LLMs) within the Chemical Description Language (χDL) framework to generate develop and interface χDLs for closed-loop active learning infrastructures (Chopra Lab collaboration). These aims will be developed over the term of the funding in a highly integrated and collaborative working modus operandi. For specific aim one we will develop a set of converters bridging the three stages of the experimental life cycle: planning, execution, and reporting. This is achieved by integrating the planning and reporting stages, which can be fully represented by the structured data schema of the ORD, with the central stage of execution, which is fully expressible in χDL. These converters will include some level of inference, through heuristics or otherwise, to fill in procedural details that might not be explicitly defined in the original plan. They can also validate if a plan can be executed in a particular lab in terms of hardware compatibility. We will realize such converters as open-source software tools and test these tools on a chemputer hardware platform for a set of benchmark reactions. For specific aim two we will develop an extension to our Natural Language Processing (NLP) approach to χDL procedure generation by using generated data sets to train a LLM AI system to be able to produce χDL instruction files directly from retrosynthetic analysis of a desired molecular structure. This will be accomplished by building a custom set of LLM agents designed to utilize the χDL NLP model to interpret and write valid χDL code based on user input. By integrating these with the χDL blueprints which are being developed for benchmark reactions as part of the NCATS ASPIRE collaboration, these χDL instructions can then be generated from automated retrosynthetic analysis of a given molecule, or class of molecules even if the suggested reactions do not yet exist in the chemical literature. We will produce specifications for a further LLM based AI system to interpret the data generated by automatic analysis and to suggest new subsequent experiments based on a pre-defined fitness function optimization (for example yield or purity of products) which can be defined experimentally in the automated system.