A physics-based approach to artificial intelligence for understanding of biophysical search spaces

近年来，药物研发行业缺乏创新。这主要是由于已知的搜索空间——传统的搜索空间——已经耗尽了。此外，传统的化学方法在寻找新的搜索空间方面的效果正在下降。我的跨学科发现资助研究项目的总体目标是将物理原理和数据驱动的人工智能技术结合起来，对多肽搜索空间的控制原理和穿越它们的方法进行基础科学研究。在我攻读物理学博士期间，我有机会将理论和计算技能运用到生物物理系统中，以理解和设计用于治疗和生物电子应用的短蛋白。在我担任博士后期间，我利用统计学、人工智能技术和合成生物学开发了一种算法，该算法可以识别一组化学组成部分，赋予药物穿透革兰氏阴性细菌外膜的能力。展望未来，我设想在生物物理学和人工智能的界面上工作。我将超越我以前的工作，将基于计算物理的方法的可解释性与深度学习的力量结合起来，以访问和理解新的和不寻常的搜索空间。通过确定重要的基于物理的设计规则，我将阐明基础科学和理论，这些基础科学和理论是开发新型候选治疗方法的基础。深度学习已经成为识别新候选药物的重要工具，但与许多人工智能技术一样，它受到缺乏可解释性的阻碍，这使得从识别到更通用的设计规则变得困难。我将把统计物理学的力量与深度学习的力量结合起来，以理解和识别新的搜索空间，这将为以后的研究提供理论支撑，有助于解决未来制药行业生产力下降的严重问题。我未来五年的具体研究目标是：-目标1：研究构建多肽搜索空间所需的约束条件-目标2：多尺度模型和迭代主动学习，以了解健康状态下抗菌肽作用方法的基础科学-目标3：通过富二硫化物肽及其相应的自由能表面的理论表征开发深度学习模型该计划将为加拿大的生物物理社区带来新技术，并将为科学家和公司提供数据科学与计算物理相结合的新工具。在我的团队中工作的高素质人员将获得计算生物物理学技术的全面基础，重点是分子动力学和蒙特卡罗模拟，以及广泛适用的数据科学技术，重点是数据驱动的机器学习。

The drug discovery industry has been lacking innovation in recent years. This is mostly due to the exhaustion of known places to look--traditional search spaces. In addition, traditional chemistry methods show declining efficacy in finding novel search spaces. The overall goal of my interdisciplinary discovery grant research program is to bring together physical principles and data-driven artificial intelligence techniques to perform basic science research into the governing principles of multi-peptide search spaces and ways to traverse them. During my doctoral work in physics, I had the opportunity to employ skills in theory and computation to biophysical systems to understand and design short proteins for therapeutic and bioelectronic applications. During my postdoctoral appointment, I used statistics, artificial intelligence techniques and synthetic biology to develop an algorithm that identifies a set of chemical building blocks conferring on drugs the ability to permeate the outer membranes of Gram-negative bacteria. Going forward, I envision working at the interface of biophysics and artificial intelligence.  I will go beyond my previous work by marrying the interpretability of computational physics-based approaches with the power of deep learning to access and understand new and unusual search spaces. By identifying important physics-based design rules, I will clarify the basic science and theory that underlies the development of novel therapeutic candidates. Deep learning has emerged as an important tool for identifying new drug candidates, but like many artificial intelligence techniques, it is hampered by its lack of interpretability, which makes it difficult to go from the identification to more general design rules. I will bring together the power of statistical physics with the power of deep learning to understand as well as identify new search spaces, which will help combat the serious problem of dwindling productivity of the drug industry in the future by providing a theoretical scaffold for later research. My specific research aims for the next five years are: - Aim 1: Studying the constraints necessary to construct a multi-peptide search space - Aim 2: Multi-scale model and iterative active learning to understand the basic science of antimicrobial peptide methods of action in a healthy state - Aim 3: Development of deep learning models through theoretical characterization of disulfide-rich peptides and their corresponding free energy surfaces This program will bring new techniques to the biophysical community in Canada and will provide new tools for the integration of data science with computational physics for scientists and companies alike. Highly qualified personnel working in my group will receive a thorough grounding in computational biophysics techniques, with a focus on molecular dynamics and Monte Carlo simulations, and in broadly-applicable data science techniques, with a focus on data-driven machine learning.