FET: Small: DNA-based Neural Networks That Learn From Their Environment

FET：小型：基于 DNA 的神经网络，可从环境中学习

• 批准号: 1908643
• 负责人: Lulu Qian
• 金额: $ 50万
• 依托单位: California Institute of Technology
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2019
• 资助国家: 美国
• 起止时间: 2019-10-01 至 2022-09-30
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1908643&HistoricalAwards=false
• 关键词: FET; Small; DNA; based; Neural

The fundamental advantage of DNA circuits, in comparison to electronic circuits, is their capability to detect and act on information in a molecular environment. A significant challenge in engineered molecular systems, DNA circuits included, is embedded learning and adaptive behavior, which is proven to be powerful and pervasive in biology and is promised to open up many doors in molecular technologies. Currently, once built, a DNA circuit always has a fixed function for how to respond to the environment, which means the same input will always trigger the same output. Some DNA circuits are reconfigurable, for example DNA neural networks, but only in the sense that a human user can choose to mix different molecules with desired concentrations for making circuits that perform different tasks. Much effort has been devoted to the design of DNA circuits with embedded learning capabilities. However, successful experimental demonstration has so far been lacking. In this project, the team of researchers will establish new circuit architectures for self-reconfigurable DNA neural networks, and show that a molecular circuit can improve how well it performs a task in a test tube without human intervention. This kind of self-improvement through learning, both supervised and unsupervised, will allow synthetic molecular circuits to gain the adaptive power previously only seen in biology, laying out the foundation for future applications in smart medicine and materials. Moreover, the results will provide experimental evidence for a pure chemical system to spontaneously undergo self-improvement, which will support the hypothesis for learning to guide and effectively accelerate evolution at the origins of life. The scientific understanding will be incorporated into open online software tools, making it accessible to general public and promoting the applications of information-processing molecular circuits. Design principles and wet-lab constructions of DNA neural networks will be introduced into the classroom, and course materials will be shared outside of the researchers' home institution. Students and postdocs will be engaged in interdisciplinary research, with an emphasis to involve more women in science. Lab tours will be provided to local college students, including underrepresented groups. Communications with general public will be facilitated by public talks, news stories, and artistic illustrations and animations of the research. The function of winner-take-all DNA neural networks depends on the concentrations of the weight molecules, which encode the memories that an input pattern is compared with for classifying the pattern. In this project, the weight molecules are designed to be initially inactive, and an appropriate collection of them will become activated when a training pattern and a label strand (indicating which class the pattern is) are simultaneously present in supervised learning. The weight-activation process will be implemented using allosteric toehold strand displacement reactions. Over the course of learning, different sets of input strands representing different training patterns will be sequentially added to the test tube that contains a DNA neural network. Each set of input strands will trigger a response of the neural network to adjust its weights and thus improve its capability for recognizing similar patterns. In unsupervised learning, the DNA neural networks are capable of restoring desired concentrations of circuit components after each round of computation and adjusting the concentrations of active weight molecules based on the circuit output rather than a given class label. The DNA neural networks will be trained and tested using a well-defined and understood task, handwritten digit recognition, to evaluate the complexity and diversity of molecular patterns that a DNA neural network is capable of learning and processing. A software tool will be developed for the design and analysis of DNA neural networks with learning capabilities.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.

与电子电路相比，DNA电路的基本优势是它们能够在分子环境中检测信息并对其进行处理。在包括DNA电路在内的工程分子系统中，一个重大挑战是嵌入学习和适应行为，这在生物学中被证明是强大和普遍的，并有望在分子技术中打开许多大门。目前，一旦建成，DNA电路总是具有如何对环境做出反应的固定功能，这意味着相同的输入将总是触发相同的输出。一些DNA电路是可重新配置的，例如DNA神经网络，但只有在这样的意义上，人类用户才能选择将不同的分子与所需的浓度混合，以制造执行不同任务的电路。人们一直致力于设计具有嵌入式学习能力的DNA电路。然而，到目前为止，还没有成功的实验演示。在这个项目中，研究团队将为可自我重组的DNA神经网络建立新的电路体系结构，并展示分子电路可以在没有人类干预的情况下提高其在试管中执行任务的能力。这种通过学习实现的自我完善，无论是有监督的还是无监督的，都将使合成分子电路获得以前只有在生物学中才能看到的自适应能力，为未来在智能医学和材料方面的应用奠定基础。此外，这些结果将为纯化学系统自发进行自我改进提供实验证据，这将支持学习指导和有效加速生命起源进化的假说。这一科学认识将被纳入开放的在线软件工具，使公众能够接触到它，并促进信息处理分子电路的应用。DNA神经网络的设计原理和湿实验室建设将被引入课堂，课程材料将在研究人员的家庭机构之外共享。学生和博士后将从事跨学科研究，重点是让更多的女性参与科学研究。将向当地大学生提供实验室参观，包括代表人数较少的群体。与公众的交流将通过公开演讲、新闻故事以及研究的艺术插图和动画来促进。赢家通吃的DNA神经网络的功能取决于权重分子的浓度，权重分子编码了输入模式与之进行比较以对模式进行分类的记忆。在这个项目中，权重分子被设计为最初不活跃，当训练模式和标签链(指示模式是哪个类别)同时存在于监督学习中时，适当的权重分子集合将被激活。重量活化过程将使用变构顶端链置换反应来实现。在学习过程中，代表不同训练模式的不同输入链将被顺序添加到包含DNA神经网络的试管中。每一组输入链将触发神经网络的响应，以调整其权重，从而提高其识别相似模式的能力。在无监督学习中，DNA神经网络能够在每一轮计算后恢复所需的电路元件浓度，并基于电路输出而不是给定的类别标签来调整活性重量分子的浓度。DNA神经网络将通过一项定义明确、易于理解的任务--手写数字识别--进行训练和测试，以评估DNA神经网络能够学习和处理的分子模式的复杂性和多样性。将开发一个软件工具，用于设计和分析具有学习能力的DNA神经网络。该奖项反映了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。

项目成果

Simplifying Chemical Reaction Network Implementations with Two-Stranded DNA Building Blocks

使用双链 DNA 构建模块简化化学反应网络实施

• DOI: 10.4230/lipics.dna.2020.2
• 发表时间: 2020
• 期刊: Leibniz international proceedings in informatics
• 作者: Johnson, Robert F;Qian, Lulu
• 通讯作者: Qian, Lulu