SAFE:Synthetically Accessible Fragment Space Extensions by Machine Learning-Based Approaches

该项目的总体目标是建立预测模型，以便扩大合成可获得的化学碎片空间。今天在药物发现，材料设计和合成化学中出现的一个关键问题是如何通过合理的努力精确地绘制可合成的有机化合物的空间。大型制药公司以及几家大型化合物供应商通过定义可合成的化学片段空间来解决这个问题。由于碎片空间是反应模式驱动的，因此可拓问题可以分解为预测与某个反应相容并且在反应中彼此相容的构件。在拟议的项目中，将开发一个化学信息学框架，用于从碎片空间提取反应方案和反应物的数据。此外，针对具体问题的分子描述符将被设计用于反应预测。基于这些数据和表示，耐受的反应物的识别将被期望为选定的反应方案。为此，将使用来自片段空间和文献的数据评估和训练最先进的机器学习模型。从随机森林，支持向量机，梯度提升到神经网络的各种技术将被测试其从训练数据中概括的能力。为了有针对性地提高预测性能，主动学习策略将与筛选技术相结合，以生成大量的实验数据。为了实现这一目标，将开发和评估新的卷积筛选技术。该模型的预测能力将不断通过实验验证，以确保空间扩展的可综合性。将开发用于将确定为生产性的反应转移到碎片空间的软件工具，以及用于将具有最大相容性的反应物分组的优化方案。在项目的最后阶段，将进行工作流程部分自动化的试验。

The overall aim of the project is the creation of predictive models enabling the extension of synthetically accessible chemical fragment spaces. A key question arising today in drug discovery, materials design and also synthetic chemistry is how to precisely map the space of synthetically accessible organic compounds with reasonable efforts. Large pharmaceutical companies as well as several large compound vendors addressed this question with the definition of synthetically accessible chemical fragment spaces. Since fragment spaces are reaction-pattern driven, the extension problem can be broken down to the prediction of building blocks compatible with a certain reaction and compatible with each other in a reaction. In the proposed project a chemoinformatics framework will be developed for the extraction of data on reaction schemes and reactants from fragment spaces. In addition, problem-specific molecular descriptors will be designed for the application in reaction predictions. Based on these data and representations, the identification of tolerated reactants will be aspired for selected reaction schemes. For this purpose, state-of-the-art machine learning models will be evaluated and trained with data from the fragment spaces and the literature. A large variety of techniques from random forests, support vector machines, gradient boosting, to neural networks will be tested for their ability to generalize from the training data. For the targeted improvement of the prediction performance, an active learning strategy will be combined with screening techniques to generate large amounts of experimental data. To achieve this goal, new convolutional screening techniques will be developed and evaluated. The predictive power of the model will be continuously validated experimentally to ensure the synthesizability of the spatial extensions. Software tools for transferring reactions identified as productive into fragment space, and an optimization scheme for grouping reactants with maximum compatibility will be developed. In the final phase of the project, trials for the partial automation of the workflow will be performed.