- 學位論文
生物分子計算系統設計：架構、合成、與模擬
Biomolecular Computing System Design: Architecture, Synthesis, and Simulation
摘要
近年來，系統與合成生物學經歷重要進展，不僅擴展了可實現的生物分子計算系統種類，能夠有效掌握、設計的系統複雜度也不斷提升。從數位邏輯、信號處理、類比運算、程式流程控制系統等，以至上述系統的組合，都已成功在活體內 (in-vivo) 與試管內 (in-vitro) 實現。其中，由多種不同功能的子系統建構而成的異質系統，應用範圍尤其廣泛：例如將針對多種生物標記 (biomarker) 的量測系統與信號處理系統結合，便能形成診斷系統的基本架構。 然而，直至今日，生物分子計算系統設計仍普遍採取針對個別應用獨立擬定方法的策略，缺乏一套具普適性的系統化設計流程，導致較長的開發週期與缺乏調整彈性的系統功能：當實際部署環境與設計時的假設條件出現差異，往往無法藉由有系統的調整原有設計縮短開發週期，需要依照新的環境條件重新設計。考量生物系統本質上顯著的變異性，如何建立一套完整且能應用（或是易於擴充）於各類系統的設計流程更顯重要。 本研究從電子設計自動化於設計複雜電路上的成功獲得啟發。在分析電子電路與生化反應系統的相似與相異處後，許多在電子設計自動化被驗證有效的方法經過調整修改，便能應用於自動化生物分子計算系統的設計。奠基於此，本研究提出一套適於自動化的完整流程，以突破現今設計面臨的障礙。一旦給定要求的系統行為，提出的流程整合了模型建立、組合能滿足所要求行為的化學反應、模擬驗證、對應至實際存在的分子與化學反應等步驟，有系統地產生滿足條件的系統設計。 本研究所提出的設計流程，除了基本的正確性外，亦著重所設計系統的可重組性 (reconfigurability)，產生的系統架構類似現場可規劃邏輯閘陣列（Field Programmable Gate Array; FPGA），由重複的可調功能模組組成，各模組的功能與模組間的連結都可由設計過程中指定的的分子濃度控制。 此外，本文探討了分子計算系統可重組性、與神經形態運算 (neuromorphic computing) 系統經由學習 (learning) 所展現出的自我修正 (self-adaptiveness) 之間的對應關係。經由模擬驗證了在所提出的系統架構下，若將模組用於實踐選定的神經元模型，搭配與神經網路可塑性 (plasticity) 相應的可調參數，所設計出的系統的確能執行希望的神經形態運算。
並列摘要
The advancements in systems and synthetic biology have been broadening the range of realizable systems with increasing complexity both in vitro and in vivo. Systems for digital logic operations, signal processing, analog computation, program flow control, as well as those composed of different functions — for example an on-site diagnostic system based on multiple biomarker measurements and signal processing – have been realized successfully. However, the efforts to date tend to tackle each design problem separately, relying on ad hoc strategies rather than providing more general solutions based on a unified and extensible architecture, resulting in long development cycle and rigid systems that require redesign even for small specification changes. Inspired by well-tested techniques adopted in electronics design automation (EDA), this work aims to remedy current design methodology by establishing a standardized, complete flow for realizing biomolecular systems. Given a behavior specification, the flow streamlines all the steps from modeling, synthesis, simulation, to final technology mapping onto implementing chassis. The resulted biomolecular systems of our design flow are all built on top of an FPGA-like reconfigurable architecture with recurring modules. Each module is designed the function of each module depends on the concentrations of assigned auxiliary species acting as the “tuning knobs.” Reconfigurability not only simplifies redesign for altered specification or post-simulation correction, but also makes post-manufacture fine-tuning — even after system deployment — possible. This flexibility is especially important in synthetic biology due to the unavoidable variations in both the deployed biological environment and the biomolecular reactions forming the designed system. In fact, by combining the system’s reconfigurability and neural network’s self-adaptiveness through learning, we further demonstrate the high compatibility of neuromorphic computation to our proposed architecture — the first work to consciously embed neuromorphic computing capability into molecular computing paradigm. With each module implementing a neuron of selected model (ex. spike-based, threshold-gate-like, etc.), accompanied by an appropriate choice of reconfigurable properties (ex. threshold value, synaptic weight, etc.), simulation results verified that the systems built from our proposed flow can indeed perform desired neuromorphic functions.