CAREER: Machine Learning Approaches to Understanding Molecular Mechanisms Underlying Convergent Evolution of Vocal Learning Behavior

执行各种复杂行为的能力，如人类语言，编码在组成有机体基因组的数十亿个核苷酸中。尽管语言本身是人类独有的，但在多种哺乳动物和鸟类中，包括鸣禽、鹦鹉、蜂鸟、蝙蝠和鲸鱼在内的多种哺乳动物和鸟类的发声学习--根据经验修改声音输出的能力--都是独立进化的，在类人猿中也是独一无二的。在这些物种的进化过程中，数百万年来的基因组序列突变导致了它们大脑中细胞类型的分子特性的差异，从而使人们能够学习它们的发声。这个项目利用物种间的多样性，采用比较基因组学的方法来理解发声学习是如何进化的：相对于没有这种能力的物种，发声学习物种的基因组有哪些共同的特征？为了回答这个问题，这项研究将开发人工智能方法，寻找数十种脑细胞类型的共同基因活动模式，以及数百种哺乳动物的共同基因组序列模式。除了将发声学习行为与特定细胞类型和基因组序列突变联系起来外，将开发的人工智能方法还有可能被应用于研究额外行为和其他特征的进化，这些行为和其他特征在拥有可用基因组的物种中有所不同。为了帮助促进采用这些方法以及其他新的人工智能方法，该项目寻求培养下一代跨学科科学家，他们是人工智能、进化生物学和神经科学方面的专家。神经科学和生物学专业的本科生将有机会通过将计算技术应用于其他行为来参与这项研究。更广泛地说，解释这项研究以及如何进行这项研究的视频和指导教程将被公开提供。词汇学习在多个水平上展示了几个谱系之间惊人的相似之处，包括行为本身、神经回路特征，甚至共同的基因表达模式。尽管有如此丰富的信息，但基因型和表型之间仍然没有牢固的联系。这个项目试图通过将发声学习相关的基因表达模式与形成发声产生的神经回路的特定神经细胞类型联系起来，并通过将它们与候选调控元件的基因组序列的变异联系起来。为了揭示与发声学习行为相关的细胞类型特定基因表达特征，将应用嵌套树概率图形模型--一种同时对细胞类型和物种的层次结构进行建模的机器学习方法。然后，为了追踪与这些基因表达模式相关的调控元件的进化，将应用卷积神经网络模型，该模型根据基因组序列的差异预测开放染色质的差异。作为这些分析的结果，该项目将提出假说，说明有声学习者和非有声学习者基因组序列的差异如何导致细胞类型特定基因表达和开放染色质的差异。除了对发声学习的研究做出贡献外，这里开发的工具还满足了对方法的日益增长的需求，以研究快速增长的基因组中细胞类型和调节元件的进化。该项目的结果将在http://www.pfenninglab.org/project/vocal-2/.This奖项中找到，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。

The ability to perform a variety of complex behaviors, like human speech, is encoded in the billions of nucleotides that make up the genome of an organism. Although speech itself is uniquely human, vocal learning, the ability to modify vocal output as a result of experience, has evolved independently in multiple mammals and birds, including songbirds, parrots, hummingbirds, bats, and whales, as well as humans, uniquely among great apes. During the evolution of each of these species, genome sequence mutations over millions of years have led to differences in the molecular properties of cell types within their brains, allowing for their vocalizations to be learned. This project leverages that diversity across species to take a comparative genomic approach to understanding how vocal learning evolved: what features do the genomes of vocal learning species have in common relative to species without this ability? To answer that question, this research will develop artificial intelligence methods to look for common patterns of gene activity across dozens of brain cell types and common genome sequence patterns across hundreds of mammals. In addition to linking vocal learning behavior to specific cell types and genome sequence mutations, the artificial intelligence methods that will be developed have the potential to be applied to study the evolution of additional behaviors and other traits that vary across species with available genomes. To help facilitate the adoption of these methods, as well as other new artificial intelligence approaches, this project seeks to train the next generation of interdisciplinary scientists, who are experts in artificial intelligence, evolutionary biology, and neuroscience. Undergraduate neuroscience and biology majors will get the opportunity to participate in the research by applying the computational techniques to other behaviors. More broadly, a video and guided tutorials that explain the research and how to conduct it will be made publicly available.Vocal learning demonstrates striking similarities across several of the lineages at multiple levels including the behavior itself, neural circuit features, and even shared gene expression patterns. Despite this wealth of information, there is still no solid connection between genotype and phenotype. This project seeks to make that connection by linking vocal learning-associated gene expression patterns to specific neural cell types that form the neural circuits for the production of vocalization and by linking them to variation in genome sequence at candidate regulatory elements. To uncover the cell type-specific gene expression features associated with vocal learning behavior, a nested tree probabilistic graphical model-- a machine learning approach that simultaneously models hierarchies of cell types and species -- will be applied. Then, to trace the evolution of regulatory elements associated with those gene expression patterns, convolutional neural network models that predict differences in open chromatin from differences in genome sequence will be applied. As a result of these analyses, this project will produce hypotheses for how differences in genome sequence between vocal learners and non-learners lead to differences in cell type-specific gene expression and open chromatin. Beyond the contributions to the study of vocal learning, the tools developed here fill a growing need for methods to study the evolution of cell types and regulatory elements across a rapidly increasing set of genomes. The results of the project will be found at http://www.pfenninglab.org/project/vocal-2/.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.