Department of Bioengineering, Stanford University, 443 Via Ortega Road, Stanford, CA 94305, United States; Allen Discovery Center for Systems Modeling of Infection, 443 Via Ortega Road, Stanford, CA 94305, United States.
Department of Bioengineering, Stanford University, 443 Via Ortega Road, Stanford, CA 94305, United States; Allen Discovery Center for Systems Modeling of Infection, 443 Via Ortega Road, Stanford, CA 94305, United States.
J Theor Biol. 2019 Jan 14;461:145-156. doi: 10.1016/j.jtbi.2018.10.041. Epub 2018 Oct 24.
The technology for building functionally complete or 'whole-cell' biological simulations is rapidly developing. However, the predictive capabilities of these simulations are hindered by the availability of parameter values, which are often difficult or even impossible to obtain experimentally and must therefore be estimated. Using E. coli's glycolytic network as a model system, we describe and apply a new method which can estimate the values of all the system's 102 parameters - fit to observations from studies of proteomics, metabolomics, enzyme kinetics and chemical energetics - and find that the resulting metabolic models are not only well-fit, but also dynamically stable. An analysis of how well parameter values in the network were determined by the training data revealed that over 80% of the parameter values were not well-specified. Moreover, the distribution of well-determined values was biased to a specific part of the network and against certain types of experimental data. Our results also suggest that perturbing the functional, energetic space of parameters (rather than traditional metabolic parameters) is a superior strategy for exploring the space of biological dynamics. The estimated parameter values matched both training data and previously withheld validation data within an order of magnitude for over 85% of the data points; notably, the area of greatest frustration in the network was also the most fully determined. Finally, our estimation method showed that fidelity to physiological observations such as network response time is enforced at the cost of fit to molecular parameter values. In summary, our reformulation enables estimation of accurate, biologically relevant parameters, generates insight into the biology of the simulated network, and appears generalizable to any biochemical network - potentially including whole-cell models.
构建功能完整或“全细胞”生物模拟的技术正在迅速发展。然而，这些模拟的预测能力受到参数值可用性的限制，这些参数值通常难以甚至不可能通过实验获得，因此必须进行估计。我们使用大肠杆菌的糖酵解网络作为模型系统，描述并应用了一种新方法，可以估计该系统 102 个参数中的所有值——与蛋白质组学、代谢组学、酶动力学和化学能量学研究的观察结果相拟合——并发现由此产生的代谢模型不仅拟合良好，而且动态稳定。对网络中参数值如何通过训练数据确定的分析表明，超过 80%的参数值没有得到很好的规定。此外，参数值的良好确定值的分布偏向网络的特定部分，并反对某些类型的实验数据。我们的结果还表明，扰动参数的功能、能量空间（而不是传统的代谢参数）是探索生物动力学空间的优越策略。估计的参数值与训练数据和之前保留的验证数据在数量级上匹配，超过 85%的数据点；值得注意的是，网络中最令人沮丧的区域也是最完全确定的区域。最后，我们的估计方法表明，对生理观察（例如网络响应时间）的保真度是以牺牲对分子参数值的拟合为代价的。总之，我们的重新表述使得能够估计准确的、与生物学相关的参数，深入了解模拟网络的生物学，并似乎适用于任何生化网络——可能包括全细胞模型。
