Machine Learning in Chemistry and Biology

描述（由申请人提供）：机器学习在化学和生物学领域具有广泛的适用性。这项研究工作的重点是在预测蛋白质和配体之间分子相互作用方面有用的功能的经验推导。从机器学习的角度来看，这个问题的特点提供了独特的挑战，其中的关键是分子相互作用的配置通常是未知的。在小分子蛋白质相互作用的情况下，可以将分子表示为3D对象，这表现在蛋白质和配体的相对构象和排列中的隐藏变量方面。大多数机器学习任务不会以这种方式嵌入隐藏变量，但这个问题并非不可克服。我们已经实现了一些方法，这些方法证明了隐变量问题是可处理的，无论是在模型归纳和评分函数优化的方法上，还是从搜索计算复杂性的角度来看。在这项工作中，我们将在3个问题领域开发新方法并改进现有方法：1)开发与已知蛋白质结构的小分子蛋白质相互作用的评分函数（对接问题）；2)建立针对未知结构蛋白的小分子活性定量模型（3D QSAR问题）；3)开发搜索和优化方法，以提高模型和评分函数的诱导以及对大型小分子文库的高通量应用。目标是以一种可量化的方式解决预测问题，这将允许方法应用的实际改进，并且还将提供对潜在物理分子相互作用的机制方面的见解。

DESCRIPTION (provided by applicant): Machine learning has broad applicability in the fields of chemistry and biology. This research effort is focused on empirical derivation of functions that are useful in the context of predicting aspects of molecular interaction between proteins and ligands. The characteristics of this problem offer unique challenges when approached from the perspective of machine learning, key among them being that the configuration in which molecules interact is not generally known. In the case of small molecule protein interactions, where it is possible to represent molecules as 3D objects, this is manifested in terms of hidden variables in the relative conformation and alignment of protein and ligand. Most machine learning tasks do not embed hidden variables in this fashion, but the problem is not insurmountable. We have implemented a number of methods which demonstrate that the problem of hidden variables is tractable, both methodologically in model induction and scoring function optimization as well as from the perspective of computational complexity in search. In this work, we will develop novel methods and refine existing methods in 3 problem areas: 1) Developing scoring functions for small molecule protein interactions with a known protein structure (the docking problem); 2) Developing quantitative models of small molecule activity against proteins with no known structure (the 3D QSAR problem); and 3) Developing methods for search and optimization that improve both model and scoring function induction and high-throughput application to large libraries of small molecules. The goal is to address the problem of prediction in a quantifiable way, which will allow both practical improvements in applications of the methods, and will also provide insight into the mechanistic aspects of the underlying physical molecular interactions. All methods and data will be made widely available to both academic and industrial investigators.