CAREER: Evaluation of machine learning algorithms for understanding and predicting adaptation to multivariate environments with a Model Validation Program (MVP)

环境变化可能是迅速的，并且涉及环境的多个方面同时发生变化，例如变暖和疾病压力增加。快速的环境变化威胁着人类赖以生存的水产养殖和农作物的生产力。然而，预测生物体对快速和多因素环境变化的脆弱性是一项重大的科学挑战。解决这一挑战的障碍来自生物体通过DNA序列水平的变化同时适应多种环境压力的复杂和非直觉的方式。因此，需要新的方法来理解和预测在多变量环境中的适应。为了满足这一需求，该项目将研究和教育与模型验证程序（MVP）相结合。该研究正在开发和评估机器学习算法（MLAs），以从生物体的DNA序列中理解和预测生物体对多变量环境的适应。为了评估mla，本研究结合了东部牡蛎的数据模拟和实地经验测试，东部牡蛎提供重要的生态系统服务，并支持数百万美元的产业。对于牡蛎来说，这项研究正在研究温度、疾病压力和盐度如何与进化历史相互作用，以决定其在野外的适应性。本研究通过以一种新的方式整合海洋、进化和统计科学前沿的概念，推进了预测复杂环境中适应性的主要科学挑战。机器学习和模型验证传统上不是在海洋和环境科学中教授的，但与这些领域的关系越来越密切。作为更广泛的教育计划的一部分，该研究正在为高中生和本科生开发MVP学习模块，帮助学生建立批判性评估和应用模型所需的基础知识。这些模块正在向大波士顿地区的数百名学生分发，并在网上提供，供广泛使用。MVP指导计划是在海洋进化生态学、统计基因组学和机器学习方面培训研究生、本科生和高中生。这项研究解决了一个迫切的社会需求，即更多地将基因型与恢复、农业和辅助基因流工作的环境相匹配。调查结果正在向牡蛎行业的利益相关者传播。本研究的目的是评估可以模拟非线性的mla是否可以用于理解和预测在各种情况下对多变量环境的适应。在目标1中，首席研究者（PI）正在创建具有不同现实主义方面的模拟数据集，并使用它们来评估和完善mla。这套新颖的模拟是在复杂、多元的环境中研究高基因流下的基因组进化。在目标2中，PI正在利用他们在东牡蛎方面的专业知识，在实地环境中评估mla。PI首先开发了一个全面的海景基因组数据集，并使用它来训练mla，以基于单个核苷酸多态性基因型预测个体的多变量环境。然后，PI正在测试MLA预测是否可以预测不同基因型的适合度，当在普通的花园田间条件下饲养时。在目标3中，PI利用obs . 1和obs . 2获得的数据，将研究和教育结合起来，开发了一系列原创的“MVP学习模块”，并为不同理解水平的人提供了交互式web应用程序，使用牡蛎恢复项目的相关示例。本研究通过生成数据集为未来的研究奠定了基础，这些数据集可能成为开发和基准化创新建模方法的经典示例。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Environmental change can be rapid and involve multiple aspects of the environment changing at the same time, such as warming and increased disease pressure. Rapid environmental change threatens the productivity of aquaculture and crops on which humans depend. Predicting organisms' vulnerabilities to rapid and multifactor environmental change, however, is a major scientific challenge. A hurdle to addressing this challenge arises from the complex and non-intuitive ways that organisms adapt, through changes at the level of the DNA sequence, to many environmental stresses at the same time. Thus, there is a need for new approaches to understand and predict adaptation in multivariate environments. To address this need, this project integrates research and education with a Model Validation Program (MVP). The research is developing and evaluating Machine Learning Algorithms (MLAs) for understanding and predicting adaptation of organisms to multivariate environments from their DNA sequences. To evaluate MLAs, this research combines both data simulation and an empirical test in the field with the Eastern Oyster, which provide important ecosystem services and support a multi-million dollar industry. For oysters, this research is studying how temperature, disease pressure, and salinity interact with evolutionary history to determine fitness in the field. This research advances efforts toward addressing the major scientific challenge of predicting adaptation in complex environments by integrating concepts across the frontiers of marine, evolutionary, and statistical sciences in a new way. Machine learning and model validation are not traditionally taught in the marine and environmental sciences, but are becoming increasingly relevant to these fields. As part of a broader education program, this research is developing MVP Learning Modules for high school students and undergraduates, which help students build the foundational knowledge they need to critically evaluate and apply models. Modules are being disseminated to hundreds of students in the greater Boston area and are being made available online for widespread use. The MVP mentoring program is training graduate students, undergraduates, and high school students in marine evolutionary ecology, statistical genomics, and machine learning. This research addresses a pressing societal need to more informatively match genotypes to environments for restoration, farming, and assisted gene flow efforts. Results are being disseminated to stakeholders in the oyster industry.The goal of this research is to evaluate if MLAs, which can model non-linearities, can be used to understand and predict adaptation to multivariate environments under a wide range of scenarios. In Objective 1, the Principal Investigator (PI) is creating simulated datasets with different aspects of realism, and using them to evaluate and refine the MLAs. This novel set of simulations is studying genome evolution under high gene flow in complex, multivariate environments. In Objective 2, the PI is building on their expertise with the Eastern oyster to evaluate the MLAs in a field setting. The PI is first developing a comprehensive seascape genomic dataset and using it to train MLAs to predict an individual's multivariate environment based on a single nucleotide polymorphism genotype. Then, the PI is testing if the MLA prediction can predict the fitness of different genotypes from across the species range when raised in common garden field conditions. In Objective 3, the PI is integrating research and education by using the data obtained from Objs. 1 and 2 to develop a series of original "MVP Learning Modules" with interactive web apps for persons at different levels of understanding, using the relatable example of an oyster restoration project. This research lays the foundation for future studies by producing datasets that could become classical examples for developing and benchmarking innovative modeling approaches.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.