III:Small: Interpretable Deep Generative Models for Drug Development

药物发现既耗时又昂贵：完全开发一种新药大约需要10-15年时间，耗资5亿至20亿美元。分子优化是药物发现的关键步骤，通过化学修饰提高候选药物的预期性能。例如，在铅(对给定目标既显示活性又显示选择性的分子)优化中，可以改变铅分子的化学结构以提高它们的选择性和特异性。传统上，这一过程是基于药物化学家的知识、直觉和经验来促进的，并通过基于片段的筛选或合成来完成。这样的方法是不可扩展的。该项目的目标是开发一类新的人工智能(AI)方法和工具来进行硅分子的生成。具体地说，本项目将集中在基于人工智能的硅分子优化中的以下重要方面：1)主要骨架保留，2)分子多样性，3)分子可合成性，4)多性质优化，以及5)可解释性。这项研究的中心假设是，越来越多的公开可用分子数据，包括分子性质、合成途径和药物相似性，包含了丰富的信息，如果适当地分析和利用，可以为揭示、表征和自动化计算分子生成和优化过程提供关键的见解。开发一类新的用于电子药物分子优化的人工智能方法，将需要开发用于电子药物分子优化的新的人工智能模型和方法。研究基于新的深生成模型、深图卷积网络、条件采样方法和强化学习方法的设计是必要的，这些方法从分子图对中学习，从而产生具有改进的生化和生物物理性质的新的分子图。拟议的研究还将提供一个整体框架，以探索彼此之间有足够差异的潜在分子；并将研究分子图搜索方法和贝叶斯优化方法，以指导在潜在嵌入(表示)空间中的搜索。对于多属性优化，本文的研究将提供一种流水线结构和新的强化学习方法。为了理解和促进可解释的生成模型，本研究将开发一套新的方法，包括网络解剖、基于扰动的归因方法、自我解释方法和解缠方法。该项目将产生重大的社会和教育影响，并将通过教育和研究传播加强STEM的多样性。该项目更广泛的科学贡献将是开发创新的人工智能方法和工具，以帮助药物开发。这些技术创新不仅将解决分子生成模型中的关键计算挑战，而且还可能推广到其他问题(如化学信息学、材料设计)，在这些问题中，非常需要生成结构数据，而对这种生成过程的解释是至关重要的。这项拟议的研究可能会降低药物发现过程中的投资成本，显著提高其成功率，并最终有助于改善美国的医疗保健质量。这一奖项反映了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。

Drug discovery is time-consuming and costly: it takes approximately 10-15 years and between $500 million to $2 billion to fully develop a new drug. Molecule optimization is a critical step in drug discovery to improve desired properties of drug candidates through chemical modification. For example, in lead (molecules showing both activity and selectivity towards a given target) optimization, the chemical structures of the lead molecules can be altered to improve their selectivity and specificity. Conventionally, this process is facilitated based on knowledge, intuition and experience of medicinal chemists, and is done via fragment-based screening or synthesis. Such an approach is not scalable. The objective of this project is to develop a new class of Artificial Intelligence (AI) methods and tools to conduct in silico molecule generation. Specifically, this project will focus on the following important aspects in AI-based in silico molecule optimization: 1) major scaffold retention, 2) molecule diversity, 3) molecule synthesizability; 4) multi-property optimization; and 5) interpretability. The central hypothesis underlying the proposed research is that the increasing amount of publicly available molecule data, including molecule properties, synthesis pathways and drug-likeness, contains a wealth of information that, if properly analyzed and utilized, can provide key insights in revealing, characterizing and automating the computational molecule generation and optimization process.Developing a new class of AI methods for in silico drug molecule optimization will require the development of novel AI models and methods for in silico molecule optimization. Examining designs based on new deep generative models, deep graph convolutional networks, conditional sampling approaches and reinforcement learning methods that learn from pairs of molecular graphs, and accordingly generate new molecular graphs with improved biochemical and biophysical properties, is necessary. The proposed research will also provide a holistic framework to explore prospective molecules that are sufficiently different from one another; and will investigate molecular graph search approaches and Bayesian optimization methods to guide search in the latent embedding (representation) space. For multi-property optimization, the proposed research will provide a pipeline structure and new reinforcement learning approaches. To understand and facilitate interpretable generative models, the proposed research will develop a set of novel methods including network dissection, perturbation-based attribution methods, self-explaining methods and disentanglement. This project will have substantial societal and educational impacts, and will enhance diversity in STEM through education and research dissemination. The broader scientific contributions of the will be the development of innovative AI methodologies and tools that will aid drug development. These technical innovations will not only address the key computational challenges in generative models for molecules, but also potentially generalize to other problems (e.g. cheminformatics, materials design) in which generation of structural data is highly needed and interpretation of such generation process is critical. The proposed research can potentially reduce the investment costs during drug discovery, increase its successful rate significantly, and ultimately aid in the improvement of the US health care quality.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.