Molecular Neural Networks for Detection and Control

点击翻译按钮获取中文摘要

The main research question is to discover the necessary components for computation in molecular neural networks and present possible approaches to implement these systems in the context of synthetic biology and DNA nanotechnology. The research will also address the question of learning and autonomy within these systems and determine its importance within implementation of the networks. I will discuss what the basic requirements are for these networks and propose a path towards implementing these networks experimentally through DNA-based computing. Each objective will contribute to the overall aim of this project.The aim of my PhD is to realise proof of principle implementations of scalable chemical molecular neural networks and develop a computational pipeline for the complete development of such networks from abstract specification.Within this, I aim to develop chemical molecular recurrent neural networks with the ability to perform different objectives, such as classification of static input signatures, temporary pattern recognition in a time series or the ability to generate specific pattern outputs from inputs.By training neural networks in-silico and transferring their learned weights in-vitro, molecular networks should be able to perform the same computations. Applications are, for example, in in-vitro cancer diagnostics, where molecular neural networks could be created that could directly classify samples into healthy or cancerous tissues, depending on the presence or absence of specific biomarkers.Objective 1: Design, modelling and simulation of molecular perceptrons and molecular neural networks.Within this objective the basic building blocks of molecular neural networks, perceptrons will be created and analysed. By evaluating different multiplier, threshold, and reporting designs, we can discover which design choices produce the most effective molecular neural network results.Objective 2: Implementation and training of complete feed-forward and recurrent molecular neural networks.With the basic design decisions of molecular perceptrons and neural networks decided, we will design a small library of molecular neural networks: feed-forward networks for classification of static input patterns and recurrent networks for classification and/or generation of time series data. The networks' performance of learning different molecular signatures will be assessed through simulation.Objective 3: Computational pipeline to construct molecular networks as a strategy to shorten nanotechnology development cycles.We will develop a computational pipeline that automatically assembles a functional molecular neural network from the specification provided by the user. Given an abstract network specification, the pipeline will optimize network weights and chemical reaction rates to perform the desired computation, compile a DNA strand displacement system that implements the network, and finally produce DNA sequences for each molecular perceptron.I aim to produce at least 2 research papers, and a review article with the work produced in this project. In addition, I hope to create open-source software that is useful for other researchers who want to create molecular neural networks.