Theoretical Chemical Design with Machine Learning: Model Development and Applications

化学家通过研究分子和材料的结构和性质以及它们所经历的变化来确定适合特定任务的物质(例如，与目标蛋白质结合的药物分子)。这是具有挑战性的，因为化合物的空间是巨大的(即，有无数潜在的化学物质)，并且化学性质的实验测量耗时、资源密集，有时甚至是不道德的。传统的理论化学发展物理模型并使用计算资源来计算化学性质。基于精确量子力学计算的传统方法往往不适用，因为它们太昂贵(即计算成本随着分子大小呈指数增长)，而基于快速分子力学模拟的传统方法往往太不准确，无法可靠地预测新的现象。 这些技术和实践的限制激发了我对计算机辅助分子设计的兴趣，特别是开发1)基于最先进的机器学习(ML)算法的新数学模型和2)定性和定量预测化学现象并最终设计具有理想性质的分子的新工具。就像人类化学家从过去的经验中学习预测新分子的性质一样，在ML中，数学模型被训练成利用以前的实验测量或计算研究的结果来预测新分子的性质。所提出的模型适用于化学中的许多问题，旨在以分子力学为代价实现可靠的量子化学方法的准确性。因此，它们可以用于系统、快速和稳健地筛选大分子数据库，从而指导后续的实验和理论研究。值得注意的是，ML方法即使在实验测量(例如，极端环境中的化学、天体化学)和计算模拟(例如，毒性和致癌性等生理反应)不可能的情况下也适用。我们通过包括ChemTools在内的软件包传播我们的模型，ChemTools是一个发现和探索化学概念的免费开源平台，到目前为止已经吸引了来自大约10个国际研究小组的用户。 这对我的研究也有教育上的影响。例如，到目前为止，我们已经组织了3次ChemTools研讨会(智利2017年、中国2018年和法国2019年)，教育学生和博士后关于蟒蛇和概念量子化学。化学工具用于教授研究生课程，以帮助学生掌握理论概念，并使学生熟悉编程，这是最有市场的技术技能之一。更重要的是，通过我的研究，我将培养一支由5名博士、4名硕士组成的多元化团队。和5名本科生研究人员，从而为下一代领导者赋权。

Chemists identify the right substance for a given task (e.g., a drug molecule that binds to a target protein) by studying the structure and properties of molecules and materials, as well as the changes they undergo. This is challenging because chemical compound space is vast (i.e., there are innumerable potential chemical substances) and experimental measurement of chemical properties is time-consuming, resource intensive, and sometimes even unethical. Traditional theoretical chemistry develops physical models and uses computational resources to calculate chemical properties. Conventional methods based on accurate quantum mechanics calculations are often inapplicable because they are too expensive (i.e. their computational cost grows exponentially with the size of the molecule), while conventional methods based on fast molecular mechanics simulations are often too inaccurate to reliably predict new phenomena. These technical and practical limitations motivate my interest in computer-aided molecular design and, specifically, in developing 1) new mathematical models based on state-of-the-art machine learning (ML) algorithms and 2) new tools to qualitatively and quantitatively predict chemical phenomena and ultimately design molecules with desirable properties. Just as human chemists learn from past experiences to make predictions about the properties of new molecules, in ML a mathematical model is trained to leverage the results of previous experimental measurements or computational studies to predict the properties of new molecules. The proposed models are applicable to many problems in chemistry and aim to achieve the accuracy of reliable quantum chemistry methods at the cost of molecular mechanics. Thus, they can be used for the systematic, rapid, and robust screening of large molecular databases, thereby guiding subsequent experimental and theoretical studies. It is important to note that ML methods are applicable even where experimental measurements (e.g., chemistry in extreme environments, astrochemistry) and computational simulations (e.g., physiological responses like toxicity and carcinogenicity) are impossible. We disseminate our models through software packages including ChemTools, a free and open-source platform for discovering and exploring chemical concepts, which so far has attracted users from ~10 international research groups. There is also an educational impact to my research. For example, we have organized 3 ChemTools workshops so far (Chile 2017, China 2018 & France 2019) educating students and postdocs on Python and conceptual quantum chemistry. ChemTools is used in teaching (under)graduate courses to facilitate grasping theoretical concepts and to familiarize students with programming, which is among the most marketable technical skills. More importantly, through my research, I will train a diverse group of 5 Ph.D., 4 M.Sc. and 5 undergraduate researchers, thereby empowering the next generation of leaders.