CAREER: Development of Novel Domain-Tailored Machine Learning Tools for Organic Reaction Development and Discovery

在化学系化学理论、模型和计算方法计划的支持下，麻省理工学院的康纳·W·科利将为合成有机反应的数据驱动建模建立新的算法、计算方法和教育工具，以促进它们的发展和发现。有机合成能够创造出在药物化学、化学探针技术、聚合物科学、农用化学科学、催化剂开发和有机电子学等广泛领域产生影响的功能分子。科利博士和他们的团队将构建新的计算机辅助合成化学工具，使新分子的合成更具预测性和健壮性，增强和增强合成化学专业从业者的直觉。这一目标将通过实验数据、理论化学知识和机器学习建模的紧密结合来实现。通过这项研究产生的新能力将使合成新化学实体的成本和速度明显受益，并应用于人类健康(例如药物)和可持续性(例如催化剂、有机电子设备)。以开放源码、在线课程和培训以及会议专题讨论会等形式提供的教育预计将有助于在数据科学和化学的界面上培训跨学科研究人员。化学的机器学习模型已经开始在计算机辅助合成规划和反应条件优化等孤立的任务中获得牵引力，但在它们帮助解决的挑战范围方面仍然有限。众所周知，最先进的统计模型渴望数据，不能很好地推广，没有建立在成熟的化学理论之上，而且可以说还没有在新的合成方法方面取得真正的发现或产生新的见解。Coley博士将努力通过与以下基本假设相关的四个互补目标来解决这些限制：(1)通过使用文献训练的表示作为先验，可以使反应条件优化算法更有效；(2)基于物理有机化学原理的反应预测模型将显示出比领域无关模型更强的泛化能力；(3)有机合成中的重大发现可以通过新颖性检测算法来表征并用于偏向从头生成；以及(4)可以通过查询驱动的主动学习和贝叶斯最优实验设计来系统地解决知识差距，如底物兼容性的不确定性。每个目标都将涉及重大的技术开发工作，以改善机器学习技术的学习和推广方式，突出进一步算法研究的机会，并最终加速新的合成有机反应的发现和开发。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

WIth support from the Chemical Theory, Models and Computational Methods program in the Division of Chemistry, Connor W. Coley of the Massachusetts Institute of Technology will establish new algorithms, computational approaches, and educational tools for the data-driven modeling of synthetic organic reactions to advance their development and discovery. Organic synthesis enables the creation of functional molecules that have impact in wide range of fields, including medicinal chemistry, chemical probe technology, polymer science, agrochemical science, catalyst development, and organic electronics. Dr. Coley and their team will build novel computer-aided synthetic chemistry tools to make the synthesis of new molecules more predictable and robust, augmenting and enhancing the intuition of expert practitioners of synthetic chemistry. This goal will be pursued through the close integration of experimental data, theoretical chemistry knowledge, and machine learning modelling. The new capabilities generated through this research will tangibly benefit the cost and speed with which new chemical entities can be synthesized, with applications to human health (e.g., medicines) and sustainability (e.g., catalysts, organic electronics). Educational offerings in the form of open-source code, online courses and trainings, and conference symposia are expected to contribute to the training of trans-disciplinary researchers at the interface of data science and chemistry. Machine learning models for chemistry have started to gain traction for isolated tasks such as computer-aided synthesis planning and reaction condition optimization, but remain limited in terms of the scope of challenges they help address. State-of-the-art statistical models are notoriously data-hungry, do not generalize well, do not build on well-established chemistry theory, and arguably have not yet made genuine discoveries or produced novel insights into new synthetic methods. Dr. Coley will endeavor to address these limitations through four complementary aims associated with the following underlying hypotheses: (1) that reaction condition optimization algorithms can be made more efficient by using literature-trained representations as priors; (2) that models for reaction prediction grounded in physical organic chemistry principles will exhibit greater generalization power than domain-agnostic models; (3) that significant discoveries in organic synthesis can be characterized through novelty detection algorithms and used to bias de novo generation; and (4) that knowledge gaps such as uncertainty about substrate compatibility can be systematically resolved via query-driven active learning and Bayesian optimal experimental design. Each aim will involve a significant technology development effort to improve how machine learning techniques learn and generalize, highlight opportunities for further algorithmic research, and ultimately accelerate the discovery and development of new synthetic organic reactions.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.