MANET: Maximum Entropy Neural Networks for Mechanistic Modeling of Single Cell Behavior

项目摘要/摘要 尽管最近在单细胞技术方面取得了实验进展，并同时开发了 统计方法，我们预测单细胞动力学和识别生化的能力 决定细胞间可变性的过程仍然很初级。我们已经确定了关键 实现对单细胞行为的机械化理解的障碍：我们没有 将单细胞数据与机械信令网络模型相结合的计算方法。 在我们以前工作的基础上，利用神经网络的尖端发展， 我们提出了一个全面的研究计划来弥合这一差距。 单细胞数据与机械模型集成的中心问题是，即使是大的- 缩放数据仅对模型进行部分约束，从而生成一系列同样适合数据的模型 井。那么，我们如何从模型中进行选择呢？我们的策略是使用最大熵(Max 一种推断最不复杂模型的方法：一种不反对任何结果的方法 除非有数据和机械约束的保证。在过去的十年里，我们已经 率先使用Max Ent对生物网络的动力学进行建模。在接下来的五年里 几年来，我们计划有两个主要的研究目标：(1)建立和验证计算 将单元格数据与模型集成所需的架构，以及(2)与 实验者，使用开发的框架研究变异性在两个重要的 丝裂原活化蛋白激酶(MAPK)通路和机械转导。我们 设想在探索细胞到细胞的机制起源时，这个框架将是不可或缺的 在广泛的信令网络中的小区可变性。值得注意的是，欠约束模型是 在数量生物学的许多领域中无处不在，包括实验室的另外两项研究 焦点：新陈代谢和微生物组动力学。这里提出的计划将直接受益 大规模数据与机械模型的集成及其原则性探讨 隐藏的假设。

Project Summary/Abstract Despite recent experimental advances in single cell techniques and a concurrent development in statistical methods, our ability to predict single cell dynamics and identify the biochemical processes that dictate cell-to-cell variability remains rudimentary. We have identified the key roadblock in achieving mechanistic understanding of single cell behavior: we do not have computational methods to integrate single cell data with mechanistic signaling network models. Building upon our previous work and leveraging cutting-edge developments in neural networks, we propose a comprehensive research program to bridge this gap. The central problem in integration of single cell data with mechanistic models is that even large- scale data only partially constrain the models, leading to a family of models that fit the data equally well. How do we then choose from the models? Our strategy is to use the Maximum Entropy (Max Ent) approach which infers the least complex model: one that does not disfavor any outcome unless warranted by the data and the mechanistic constraints. Over the past decade, we have pioneered the novel use of Max Ent to model dynamics of biological networks. In the next five years, we plan to have two main research goals; (1) to build and validate the computational architecture required to integrate single cell data with models and (2) in close collaboration with experimentalists, use the developed framework to study the variability in two important pathways; the mitogen activated protein kinase (MAPK) pathway and mechanotransduction. We envision that this framework will be indispensable in exploring the mechanistic origins of cell-to- cell variability across a broad range of signaling networks. Notably, under-constrained models are ubiquitous in many areas of quantitative biology, including two of the laboratory’s other research foci: metabolism and microbiome dynamics. The program proposed here will directly benefit integration of large-scale data with mechanistic models and a principled exploration of otherwise hidden hypotheses.