Development of machine learning approaches to population pharmacokinetic model selection and evaluation of application to model-based bioequivalence analysis.

项目总结 拟议的项目是对深度学习/强化的评估 群体药代动力学模型选择的学习方法。我们建议开发一种 使用Python编程语言的命令行应用程序。Python是当前的行业 开发或人工智能应用程序的标准，以及 深度学习/强化学习是现成的(例如，Pytorch和TensorFlow)。这个 申请者之前开发了一种相关的机器学习方法，使用基因 算法。为了比较的目的并使所得到的应用程序普遍可用， 现有的遗传算法解决方案将被移植到Python上。两个应用程序(深度 学习/强化学习和遗传算法)将使用NONMEM®作为参数 被检查的群体药代动力学模型的估计。一个常见的解决方案将 模型选择算法(深度学习/强化学习和遗传算法) NONMEM将用于两者，目前正在开发中，有一个早期版本 在githorb.com上可以买到。此通用解决方案将促进使用其他算法的未来工作 用于模型选择，例如粒子群优化或模拟退火法。这项工作将是 在项目的第一年完成。所有最终代码将放置在公有领域中 Githorb.com网站。 项目的第二年将包括对解决方案的评估(遗传算法 和深度学习/强化学习)。这项评估将包括一系列评估 模型“好”的度量(“遗传算法中的适合度”和“奖励信号” 深度学习/强化学习)。这些衡量模型“好坏”的标准可能包括 目标函数值、简约惩罚、成功协方差步骤的重要性。在 对于这个项目的范围，这些措施将是客观的和数字的。未来的项目可能 在模型选择中加入对模型“善”的主观评价 算法。 机密第1页，共1页

PROJECT SUMMARY The proposed project is for development an evaluation of a deep learning/reinforcement learning approach to population pharmacokinetic model selections. We proposed to develop a command line application using the Python programing language. Python is the current industry standard for development or artificial intelligence applications, and the required packages for deep learning/reinforcement learning are readily available (e.g., Pytorch and Tensorflow). The applicants have previously developed a related machine learning approach using Genetic Algorithm. For purposes of comparison and to make the resulting application generally available, the existing Genetic Algorithm solution will be ported to Python. Both applications (Deep learning/reinforcement learning and Genetic Algorithm) will use NONMEM ® for parameter estimation for the population pharmacokinetic models examined. A common solution linking the model selection algorithm (Deep Learning/Reinforcement Learning and Genetic Algorithm) to NONMEM will be used for both, and is currently under development, with an early version available on github.com. This common solution will facilitate future work using other algorithms for model selection, e.g. particle swarm optimization or simulated annealing. This work will be completed in the first year of the project. All final code will be place in the public domain in github.com. The second year of the project will consist of evaluation of the solutions (Genetic algorithm and Deep Learning/Reinforcement Learning). This evaluation will include assessment of a range of measures of the “goodness” of the model (“fitness in Genetic Algorithm and “reward signal” in Deep Learning/Reinforcement Learning). These measure of model “goodness” may include objective function value, parsimony penalties, importance of successful covariance step. Within the scope of this project, these measures will be objective and numerical. Future projects may include the addition of subjective evaluations of model “goodness” in the model selection algorithm. CONFIDENTIAL Page 1 of 1