Liu Chuanbo, Wang Jin
State Key Laboratory of Electroanalytical Chemistry, Changchun Institute of Applied Chemistry, Chinese Academy of Sciences, Changchun, Jilin 130022, People's Republic of China.
Center for Theoretical Interdisciplinary Sciences, Wenzhou Institute, University of Chinese Academy of Sciences, Wenzhou, Zhejiang 325001, People's Republic of China.
Proc Natl Acad Sci U S A. 2024 Apr 2;121(14):e2317422121. doi: 10.1073/pnas.2317422121. Epub 2024 Mar 26.
Stochastic reaction networks are widely used in the modeling of stochastic systems across diverse domains such as biology, chemistry, physics, and ecology. However, the comprehension of the dynamic behaviors inherent in stochastic reaction networks is a formidable undertaking, primarily due to the exponential growth in the number of possible states or trajectories as the state space dimension increases. In this study, we introduce a knowledge distillation method based on reinforcement learning principles, aimed at compressing the dynamical knowledge encoded in stochastic reaction networks into a singular neural network construct. The trained neural network possesses the capability to accurately predict the state conditional joint probability distribution that corresponds to the given query contexts, when prompted with rate parameters, initial conditions, and time values. This obviates the need to track the dynamical process, enabling the direct estimation of normalized state and trajectory probabilities, without necessitating the integration over the complete state space. By applying our method to representative examples, we have observed a high degree of accuracy in both multimodal and high-dimensional systems. Additionally, the trained neural network can serve as a foundational model for developing efficient algorithms for parameter inference and trajectory ensemble generation. These results collectively underscore the efficacy of our approach as a universal means of distilling knowledge from stochastic reaction networks. Importantly, our methodology also spotlights the potential utility in harnessing a singular, pretrained, large-scale model to encapsulate the solution space underpinning a wide spectrum of stochastic dynamical systems.
随机反应网络广泛应用于生物学、化学、物理学和生态学等不同领域的随机系统建模。然而，理解随机反应网络中固有的动态行为是一项艰巨的任务，主要原因是随着状态空间维度的增加，可能状态或轨迹的数量呈指数增长。在本研究中，我们引入了一种基于强化学习原理的知识蒸馏方法，旨在将随机反应网络中编码的动态知识压缩到一个单一的神经网络结构中。经过训练的神经网络能够在接收到速率参数、初始条件和时间值时，准确预测与给定查询上下文相对应的状态条件联合概率分布。这消除了跟踪动态过程的需要，能够直接估计归一化状态和轨迹概率，而无需在整个状态空间上进行积分。通过将我们的方法应用于代表性示例，我们在多模态和高维系统中都观察到了高度的准确性。此外，经过训练的神经网络可以作为开发参数推断和轨迹集合生成高效算法的基础模型。这些结果共同强调了我们的方法作为从随机反应网络中提取知识的通用手段的有效性。重要的是，我们的方法还突出了利用单个预训练的大规模模型来封装支撑广泛随机动态系统的解空间的潜在效用。
