NANO: EMT: Scalable DNA Molecular Computation

。根据NSF EMT提案“可扩展的DNA分子计算”，加州大学河滨分校的一组科学家和学生将研究基于DNA分子相互作用制造大规模神经网络计算机的可能性。研究团队包括主要研究员、物理系的Allen Mills教授、化学与环境工程系的Jerome Schultz教授以及几名研究生和本科生。利用一组分子作为计算元素进行实际计算的可能性是很有趣的，因为人们可能能够制造出比目前已知的任何电子计算机都强大得多的分子计算机。这是因为所需尺寸的DNA分子的体积比现代计算机芯片上最小的特征要小数十亿倍，因此，相当于一个功能强大的微处理器的DNA原则上可以装进一个微小的液滴中。我们可以很容易地预测，一台笔记本电脑大小的DNA计算机器可能包含数百万个这样的液滴，其综合计算能力将超过人类的大脑。第一次表明可以用DNA分子进行计算是在1994年，当时在加州大学洛杉矶分校工作的Len Adleman教授戏剧性地证明，溶液中的各种DNA分子找到、识别并结合互补DNA分子的某些反应可以用于用分子进行计算。令人鼓舞的是，Adleman和他的同事最近使用了一个相关的方案来解决一个涉及100万个不同DNA分子的问题。通往DNA巨大计算潜力的道路上的一个绊脚石是由于不完美的分子反应而导致的计算错误的预期积累。NSF的新提案寻求通过将DNA分子计算机配置为类似大脑的神经网络来实现“可扩展的DNA分子计算”，这种计算架构使用冗余来实现其固有的容错性，因此可以扩展到几乎任何规模。第一个实用的神经网络架构，也是测试DNA神经网络的理想架构，是由加州理工学院和贝尔实验室的John Hopfield教授在20世纪80年代初发明的内容寻址记忆神经网络。在提案的第一阶段，学生们将在导师的指导下学习如何利用活细胞中自然发生的生化反应来制造DNA神经网络所需的组件。在第二阶段的工作中，将制造一台小型分子计算机来演示DNA神经网络的操作原理。计划将这个网络扩大一个数量级，以寻找这种类型的分子计算的大小限制。制造接近人脑大小的神经网络的前景非常吸引人，并提出了一些有趣的问题：如何编程或“训练”一个非常大的网络，它如何交流，它的操作是否可以自主，它是否可以表现出自组织和紧急行为。拟议的工作将证明DNA神经网络是否可以在某些利基领域与其他技术竞争。最重要的应用之一可能是使用经过训练的DNA神经网络来识别某些癌细胞类型的DNA特征，并进行廉价的基因分析测试，以确定人们对一系列最常见疾病的易感性，例如各种癌症、哮喘、糖尿病和血色素沉着症，其中可能涉及数十种不同组合的遗传缺陷。因此，即使庞大的网络被证明是不可行的，相对较小的网络的最终受益者可能是病人，他们将获得基于廉价和快速诊断工具的改进治疗，而普通民众将能够根据遗传知识调整他们的生活方式。目前，大型DNA神经网络的前景还不确定，但原则上，它们可能成为全新计算范式的基础。无论如何，拟议的研究将为两名博士生和几名本科生在新兴的生物物理学领域提供优秀的培训，这所大学以其大量少数民族背景的学生而闻名。

. Under NSF EMT Proposal "Scalable DNA molecular computation" a team of scientists and students at the University of California, Riverside will investigate the possibility of making a large-scale neural network computer based on the interactions of DNA molecules. The research team includes the principle investigator, Professor Allen Mills from the Department of Physics, Professor Jerome Schultz from the Department of Chemical and Environmental Engineering and several graduate and undergraduate students. The possibility of doing practical computations using groups of a few molecules as computing elements is intriguing because one might be able to make molecular computers much more powerful than any electronic computer known today. This is because DNA molecules of the required size are billions of times smaller in volume than the smallest feature on a modern computer chip, so that the DNA equivalent of a powerful microprocessor could in principle fit into a tiny droplet of liquid. One can easily project that a DNA computing machine with the physical size of a laptop could contain millions of such droplets that would have a composite computing power exceeding that of a human brain. The first indication that computing could be done with DNA molecules came in 1994 when Prof. Len Adleman, working at UCLA, dramatically demonstrated that certain reactions, in which various DNA molecules in solution find, recognize and bind to complementary DNA molecules, may be utilized for the purpose of doing computations with molecules. It is encouraging that Adleman and his coworkers have recently used a related scheme to solve a problem involving one million different DNA molecules.One stumbling block on the path to the enormous computing potential of DNA is the anticipated accumulation of errors in a computation due to imperfect molecular reactions. The new NSF proposal seeks to achieve "Scalable DNA molecular computation" by configuring a DNA molecular computer as a brain-like neural network, a computing architecture that uses redundancy to make it inherently fault tolerant and therefore possibly scalable to virtually any size. The first practical neural network architecture, and one that is perhaps ideal for testing a DNA neural network, is the content-addressable memory neural network invented by Prof. John Hopfield at Cal Tech and Bell Labs in the early 1980's. In the first phase of the proposal, the students under the guidance of their mentors will learn how to use biochemical reactions naturally occurring in living cells to make the necessary components for a DNA neural network. In the second phase of the work, a small-scale molecular computer will be made to demonstrate the principle of operation of a DNA neural network. Plans will be drawn up for scaling this network up by an order of magnitude to search for size limitations in this type of molecular computing. The prospect of making neural networks with a size approaching that of a human brain is very intriguing and raises interesting questions about how one would program or "train" a very large network, how it would communicate, whether its operation could be autonomous and whether it could exhibit self-organization and emergent behavior.The proposed work will demonstrate whether DNA neural networks could be competitive with other technologies in certain niches. One of the most important applications could be the use of DNA neural networks trained to recognize the DNA signatures of certain cancer cell types and for performing inexpensive genetic profiling tests to determine the susceptibility of people to a spectrum of the most common diseases such as various cancers, asthma, diabetes and hemochromatosis where dozens of genetic defects in different combinations may be involved. Thus even if giant networks prove to be infeasible, the eventual beneficiaries of relatively small networks may be patients that would receive improved therapy based on the availability of an inexpensive and rapid diagnostic tool, and the general populace who will be able to adjust their lifestyles based on genetic knowledge. The prospects for large DNA neural networks are uncertain at this time, but in principle they could be the basis for an entirely new computing paradigm. In any case, the proposed research will provide excellent training for two PhD students and several undergraduate students in the emerging field of biophysics at a University noted for its large component of students from minority backgrounds.